Keywords: Brain, Functional imaging (e.g. fMRI), Multi- and Hyper-spectral imaging
Abstract: Perinatal hypoxic-ischaemic (HI) brain injury in the term infant remains a significant problem throughout the world. Neonatal encephalopathy (NE) is the clinical manifestation of the ensuing disordered neonatal brain function which can lead to serious consequences including death. The availability of markers of neuronal injury that correlate with disease severity and are predictive of neurodevelopmental disability in childhood would likely facilitate a more targeted therapeutic approach using neuroprotective therapies additionally to hypothermia. To meet the above clinical need we have been developing optical technologies based on broadband Near-Infrared Spectroscopy (or NIRS). Near-infrared (NIR) light (650-950nm) can easily penetrate the skull and reach the brain. By measuring the light attenuation at different wavelengths, one can estimate the concentration of the oxygenated (HbO2) and the deoxygenated-haemoglobin (HHb). These two states of haemoglobin have different absorption spectra, which we can use for our spectroscopic measurements. Another strong NIR absorber is the terminal electron acceptor of the mitochondrial respiratory chain cytochrome-c-oxidase (CCO), which contains a unique Cu-Cu dimer (termed CuA). The NIR absorption spectrum of CCO depends on the redox state of CuA which in turns depends on the availability of oxygen in cells. For several years we have been developing instrumentation and methodology that can non-invasively assess mitochondrial oxygenation through measurement of the changes in oxidation status of cytochrome-c-oxidase ([oxCCO]) [1].

In this talk I will introduce the technology and emphasise several aspects in the instrumentation development. Finally I will be discussing the deployment of the technology from the laboratory, to the preclinical environment and to the neonatal intensive care unit.

References [1] Bale G, Elwell CE, Tachtsidis I. From Jöbsis to the present day: a review of clinical near-infrared spectroscopy measurements of cerebral cytochrome-c-oxidase. J Biomed Opt. 2016 21(9):091307

Keywords: Infrared imaging, Functional imaging (e.g. fMRI)
Abstract: fNIRS (functional near-infrared spectroscopy), an optical neuroimaging method for measuring cortical hemodynamics, offers compact and portable instruments with low running costs, allowing various psychological experiments and measurements to be conducted in relatively unrestricted and natural environments. Owing to such features, fNIRS has been employed in a wide range of psychological applications, which are beyond the reach of conventional neuroimaging devices, including examination of infant cognitive development, visualization of neural bases for neuropsychological evaluation, and diagnostic, neuropharmachological and neuro-feedback assessment of developmental disorders. In short, fNIRS is expanding the horizon of psychological research by uncovering the neural bases of psychological states and processes of the human brain.

Keywords: Image registration, Brain, Other-modality
Abstract: The structural analysis of nerve fibers of the human brain is an important topic in current neuroscience. To obtain information about neural connections with micrometer resolution, polarised light imaging (3D PLI) of histological brain sections is well suited. In our application, both high-resolution (HR, 64m in-plane pixel size) and ultra-high resolution (ultra-HR, 1.3m) 3D PLI data of human brain sections are acquired. However, due to arbitrary translations and rotations caused by the sectioning and mounting process, spatial coherence between sections is lost and image registration is necessary. We introduce a new feature-based approach for registration of ultra-HR 3D PLI data to their corresponding HR images. The approach is based on a novel multi-scale salient feature detection method that is well suited for 3D PLI data. We have successfully evaluated the approach and applied it to 83 sections of a human brain. An experimental comparison with previous state-of-the-art feature detectors demonstrates the superior performance of our approach.

Keywords: Microscopy - Light, Confocal, Fluorescence, Image registration, Cells & molecules
Abstract: Optical flow models can be used for solving non-rigid registration problems of cell nuclei in time-lapse microscopy images. In contrast to local optical flow models employed in existing non-rigid registration methods, we introduce a method based on global optical flow models. We analyze the properties of fluorescence cell microscopy images and find that for regularization of the deformation fields, a convex quadratic regularizer is more suitable than non-convex ones. We further propose an adaptive weighting global method that is derived based on properties of the image noise. Experiments using real live cell microscopy images demonstrate that our proposed method outperforms previous methods as well as general methods based on global optical flow models with non-convex regularizers.

Keywords: Histopathology imaging (e.g. whole slide imaging), Image registration
Abstract: High-throughput serial histology imaging provides a new avenue for the routine study of micro-anatomical structures in a 3D space. However, the emergence of serial whole slide imaging poses a new registration challenge, as the gigapixel image size precludes the direct application of conventional registration techniques. In this paper, we develop a three-stage registration with multi-resolution mapping and propagation method to dynamically produce registered subvolumes from serial whole slide images. We validate our algorithm with gigapixel images of serial brain tumor sections and synthetic image volumes. The qualitative and quantitative assessment results demonstrate the efficacy of our approach and suggest its promise for 3D histology reconstruction analysis.

Keywords: Brain, Microscopy - Light, Confocal, Fluorescence, Image registration
Abstract: The evolutionary success of ants and other social insects is considered to be intrinsically linked to division of labor and emergent collective intelligence. The role of the brains of individual ants in generating these processes, however, is poorly understood. One genus of ant of special interest is Pheidole, which includes more than a thousand species, most of which are dimorphic, i.e. their colonies contain two subcastes of workers: “minors” and “majors”. Using confocal imaging and manual annotations, it has been demonstrated that minor and major workers of different ages of three species of Pheidole have distinct patterns of brain size and subregion scaling. However, these studies require laborious effort to quantify brain region volumes and are subject to potential bias. To address these issues, we propose a group-wise 3D registration approach to build for the first time bias-free brain atlases of intra- and inter-subcaste individuals and automatize the segmentation of new individuals.

Keywords: Microscopy - Light, Confocal, Fluorescence, In-vivo cellular and molecular imaging, Image registration
Abstract: In developmental imaging, 3D+t series of microscopic images allow to follow the organism development at the cell level and have now became the standard way of imaging the development of living organs. Dedicated tools for cell segmentation in 3D images as well as cell lineage calculation from 3D+t sequences have been proposed to analyze these data. For some applications, it may be desirable to interpolate label images at intermediary time-points. However, the known methods do not allow to locally handle the topological changes (ie cell. division). In the present work, we propose an extrapolation method that coherently deformed the label images to be interpolated.

Keywords: Machine learning, Image segmentation, Magnetic resonance imaging (MRI)
Abstract: In this paper, we propose an end-to-end trainable Convolutional Neural Network (CNN) architecture called the M-net, for segmenting deep (human) brain structures from Magnetic Resonance Images (MRI). A novel scheme is used to learn to combine and represent 3D context information of a given slice in a 2D slice. Consequently, the M-net utilizes only 2D convolution though it operates on 3D data, which makes M-net memory efficient. The segmentation method is evaluated on two publicly available datasets and is compared against publicly available model based segmentation algorithms as well as other classification based algorithms such as Random Forest and 2D CNN based approaches. Experiment results show that the M-net outperforms all these methods in terms of dice coefficient and is at least 3 times faster than other methods in segmenting a new volume which is attractive for clinical use.

Keywords: Machine learning, Probabilistic and statistical models & methods, Brain
Abstract: We consider the problem domain shift in analyses of brain MRI data. While many different datasets are publicly available, most algorithms are still trained on a single dataset and often suffer the problem of limited and unbalanced sample sizes. In this work, we propose a surprisingly simple strategy to reduce the impact of domain shift - caused by different data sources and processing pipelines - that typically occurs in cross-dataset analyses. We experimentally evaluate our approach on the problem of using volumetric features to distinguish neurodegenerative diseases and report results using three datasets in two practically relevant scenarios: (1) cross-dataset learning and (2) leveraging pre-trained classifiers across different datasets. We show that our adaptation technique enables both scenarios with performance close to the single-dataset case.

Keywords: Machine learning, Image registration, Brain
Abstract: Cortical thickness estimation performed in vivo via magnetic resonance imaging (MRI) is an important technique for the diagnosis and understanding of the progression of Alzheimer's disease (AD). Directly using raw cortical thickness measures as features with Support Vector Machine (SVM) for clinical group classification only yields modest results since brain areas are not equally atrophied during AD progression. Therefore, feature reduction is generally required to retain only the most relevant features for the final classification. In this paper, a spherical sparse coding and dictionary learning method is proposed and it achieves relatively high classification results on publicly available data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) 2 dataset (N=201) which contains structural MRI data of four clinical groups: cognitive unimpaired (CU), early mild cognitive impairment (EMCI), later MCI (LMCI) and AD. The proposed framework takes the estimated cortical thickness and the spherical parameterization computed by FreeSurfer as inputs and constructs weighted patches in the spherical parameter domain of the cortical surface. Then sparse coding is applied to the resulting surface patch features, followed by max-pooling to extract the final feature sets. Finally, SVM is employed for binary group classifications. The results show the superiority of the proposed method over other cortical morphometry systems and offer a different way to study the early identification and prevention of AD.

Keywords: Machine learning, Connectivity analysis, Diffusion weighted imaging
Abstract: In this work, we study the extent to which structural connectomes and topological derivative measures are unique to individual changes within human brains. To do so, we classify structural connectome pairs from two large longitudinal datasets as either belonging to the same individual or not. Our data is comprised of 227 individuals from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and 226 from the Parkinson's Progression Markers Initiative (PPMI). We achieve 0.99 area under the ROC curve score for features which represent either weights or network structure of the connectomes (node degrees, PageRank and local efficiency). Our approach may be useful for eliminating noisy features as a preprocessing step in brain aging studies and early diagnosis classification problems.

Keywords: Machine learning, Computer-aided detection and diagnosis (CAD), Brain
Abstract: In clinical practice, the magnetic resonance imaging (MRI) is a prevalent neuroimaging technique for Alzheimer’s disease (AD) diagnosis. As a learning using privileged information (LUPI) algorithm, SVM+ has shown its effectiveness on the classification of brain disorders, with single-modal neuroimaging samples for testing but multimodal neuroimaging samples for training. In this work, we propose to apply the multimodal restricted Boltzmann machines (RBM) as an LUPI algorithm for feature learning so as to form an RBM+ algorithm. Furthermore, an ensemble LUPI algorithm is developed, integrating SVM+ and RBM+ by the multiple kernel boosting based strategy. The experimental results demonstrate that the proposed RBM+ works well as an LUPI algorithm for feature learning, and the ensemble LUPI algorithm is superior to the traditional predictive models for the MRI-based AD diagnosis using the positron emission tomography as the privileged information.

Keywords: Magnetic resonance imaging (MRI), Graphical models & methods
Abstract: Autism Spectrum Disorder (ASD) is a complex developmental disorder affecting 1 in 68 children in the United States. While the prevalence may be on the rise, we currently lack a firm understanding of the etiology of the disease, and diagnosis is made purely on behavioral observation and informant report. As one potential method for improving our understanding of ASD, the current study took a network-level approach by assessing the causal interactions among the frontoparietal and default mode networks using volumetric structural covariance of a large Autism dataset. Although preliminary, we report diffuse yet subtle changes throughout these networks when comparing age and sex matched controls to ASD patients.

Keywords: Functional imaging (e.g. fMRI), Brain, fMRI analysis
Abstract: Sub-concussive asymptomatic head impacts during contact sports may develop potential neurological changes and may have accumulative effect through repetitive occurrences in contact sports like American football. The effects of sub-concussive head impacts on the functional connectivity of the brain are still unclear with no conclusive results yet presented. Although various studies have been performed on the topic, they have yielded mixed results with some concluding that sub concussive head impacts do not have any effect on functional connectivity, while others concluding that there are acute to chronic effects. The purpose of this study is to determine whether there is an effect on the functional connectivity of the brain from repetitive sub concussive head impacts. First, we applied a model free group ICA based intrinsic network selection to consider the relationship between all voxels while avoiding an arbitrary choice of seed selection. Second, unlike most other studies, we have utilized the default mode network along with other extracted intrinsic networks for classification. Third, we systematically tested multiple supervised machine learning classification algorithms to predict whether a player was a non-contact sports youth player, a contact sports player with low levels of cumulative biomechanical force impacts, or one with high levels of exposure. The 10-fold cross validation results show robust classification between the groups with accuracy up to 78% establishing the potential of data driven approaches coupled with machine learning to study connectivity changes in youth football players. This work adds to the growing body of evidence that there are detectable changes in brain signature from playing a single season of contact sports.

Keywords: fMRI analysis, Functional imaging (e.g. fMRI), Probabilistic and statistical models & methods
Abstract: Multisubject analysis helps to jointly analyze the medical data from multiple subjects, to make insightful inferences. Multi set canonical correlation analysis (MCCA), which extends the application of canonical correlation analysis to more than two datasets, is one such statistical technique to perform multisubject analysis. MCCA aims to compute optimal data transformations such that overall correlation of transformed datasets is maximized. But, the conventional approach is directly applicable to vector data, which requires the image data to be reshaped into vectors. Vectorization of images disturbs their spatial structure and increases computational complexity. We propose a new two dimensional MCCA approach that operates directly on the image data. Experiments are performed against fMRI data sets acquired through block-paradigm right finger tapping task.

Keywords: Deconvolution, Functional imaging (e.g. fMRI), fMRI analysis
Abstract: Resting-state fMRI provides challenging data that needs to be analyzed without knowledge about timing or duration of neuronal events. The "total activation" framework is one recent approach that combines temporal and spatial regularization to deconvolve the fMRI signals; i.e., undo them from the influence of the hemodynamic response. The temporal regularization is using generalized total variation that promotes piece-wise constant signals of the deconvolved timecourses. In the original formulation, the spatial regularization is expressing 2 -smoothness within regions of a predefined brain atlas. In this work, we replace the latter with 3-D total variation that constrained to the gray matter domain. This allows the recovery of activation clusters with sharp boundaries without any bias from the atlas' partitioning. We propose the corresponding variational formulation and optimization problem, together with results that demonstrate the feasibility of the proposed approach for both simulated and real fMRI data.

Keywords: Functional imaging (e.g. fMRI), image filtering (e.g. mathematical morphology, wavelets,...), Brain
Abstract: Non-local means (NLM) filtering of fMRI can reduce noise while preserving spatial structure. We have developed a vari- ant called temporal-NLM (tNLM) which uses similarity in time-series between voxels as the basis for computing the weights in the filter. Using tNLM, dynamic fMRI data can be denoised while spatial boundaries between functionally dis- tinct areas in the brain tend to be preserved. The degree of smoothing in tNLM is determined by a parameter h. Here we describe a procedure for selection of h to optimize our ability to differentiate functionally discrete brain regions. We demonstrate the method in application to optimized filtering of task fMRI data.

Keywords: Image registration, Computed tomography (CT), Optimization method
Abstract: This paper addresses a dense voxel-wise correspondence of cone-beam computed tomography (CBCT) images towards a non-rigid registration and treatments evaluation in clinical orthodontics. An unsupervised clustering randomized forest is employed to establish voxel-wise correspondence in a reduced subset of the original volume image. A geodesic coordinate is introduced to avoid the structural ambiguities. The geodesic coordinates updated with voxel-wise affinities yield a cascaded geodesic forest. Given a novel volume image, the appearance and cascaded geodesic forests produce a voxel-wise correspondence in the subsets. A regularization scheme is employed to propagate the subset correspondence to the whole images, which results in a dense displacement field of the non-rigid registration between the reference and target volume images. Our technique is based on the unsupervised clustering forests and does not need the predefined atlas for training. Quantitative assessment on practitioner-annotated ground truth demonstrates an improvement to the state-of-the-arts label propagation techniques of CBCT images.

Keywords: Functional imaging (e.g. fMRI), Brain, Image registration
Abstract: Accurate registration plays a critical role in group-wise functional Magnetic Resonance Imaging (fMRI) image analysis, as spatial correspondence among different brain images is a prerequisite for inferring meaningful patterns. However, the problem is challenging and remains open, and more effort should be made to advance the state-of-the-art image registration methods for fMRI images. Inspired by the observation that common functional networks can be reconstructed from fMRI image across individuals, we propose a novel computational framework for simultaneous groupwise fMRI image registration by utilizing those common functional networks as references for spatial alignments. In this framework, firstly, individualized functional networks in each subject are inferred using Independent Component Analysis (ICA); secondly, congealing groupwise registration that takes entropy of stacked independent components (ICs) from all the subjects as objective function is applied to register individual functional maps for maximal matching. The proposed framework is evaluated by and applied to an Alzheimer’s Disease (AD) fMRI dataset and shows reasonably good results.

Keywords: Image registration, Machine learning, Abdomen
Abstract: This paper addresses the estimation of pairwise supervoxel correspondences toward automatic semi-dense medical image registration. Supervoxel matching is performed through random forests (RF) with supervoxel indexes as label entities to predict matching areas in another target image. Ensuring accurate supervoxel boundary adherence requires a fine supervoxel decomposition which highly increases learning complexity. To alleviate this issue, we extend RF based supervoxel matching from single to multi-scale using a recursive hierarchical supervoxel representation. Output RF matching probabilities obtained for the last scale are gathered with ancestor matching probabilities which acts as a coarse-to-fine matching guidance. The effectiveness of our method is highlighted for semi dense abdominal image registration relying on liver label propagation and consistency assessment.

Keywords: Motion compensation and analysis, Computed tomography (CT)
Abstract: Recently, C-arm cone-beam CT systems have been used to acquire knee joints under weight-bearing conditions. For this purpose, the C-arm acquires images on a horizontal trajectory around the standing patient, who shows involuntary motion. The current state-of-the-art reconstruction approach estimates motion based on fiducial markers attached to the knee. A drawback is that this method requires calibration prior to each scan, since the horizontal trajectory is not reproducible. In this work, we propose a novel method, which does not need a calibration scan. For comparison, we extended the state-of-the-art method with an iterative scheme and we further introduce a closed-form solution of the compensated projection matrices. For evaluation, a numerical phantom and clinical data are used. The novel approach and the extended state-of-the-art method achieve a reduction of the reprojection error of 94% for the phantom data. The improvement for the clinical data ranged between 10% and 80%, which is followed by the visual impression. Therefore, the novel approach and the extended state-of-the-art method achieve superior results compared to the state-of-the-art method.

Keywords: X-ray imaging, Lung, Motion compensation and analysis
Abstract: Chest tomosynthesis (CTS) is a newly developed imaging technique which provides pseudo-3D volume anatomical information of thorax from limited angle projections and therefore improves the visibility of anatomy without so much increase on radiation dose compared to the chest radiography (CXR). However, one of the relatively common problems in CTS is the respiratory motion of patient during image acquisition, which negatively impacts the detectability. In this paper, we propose a sin-quadratic model to analyze the respiratory motion during CTS scanning, which is a real time method that generates the respiratory signal by directly extracting the motion of diaphragm during data acquisition. According to the extracted respiratory signal, physicians could re-scan the patient immediately or conduct motion free CTS image reconstruction for patients that could not hold their breath perfectly during the scan time. The effectiveness of the proposed model was demonstrated with both the simulated phantom data and the real patient data.

Keywords: Cells & molecules, Microscopy - Light, Confocal, Fluorescence, Machine learning
Abstract: Morphologic differentiation between abnormal lymphoid cells circulating in peripheral blood (PB) is a difficult task. The goal of this work is to define new features from image analysis that allow quantifying cytological variables in view of a further automatic recognition. We analyzed a total of 12574 digital cell images from 325 patients. PB films were stained with May Grünwald-Giemsa and images were obtained in the CellaVision® DM96. A total of 12 different lymphoid cells were included (Fig.1): mature normal lymphocytes (1085), abnormal lymphoid cells of chronic lymphocytic leukemia (2468), B-prolymphocytic leukemia (292), hairy cell leukemia (611), splenic marginal zone lymphoma (819), mantle cell lymphoma (1438), follicular lymphoma (1119), T-prolymphocytic leukemia (332), T-cell large granular lymphocytic leukemia (894) and Sézary lymphoid cells (756); reactive (1199) and blast lymphoid cells (1561) were also included. We extracted 2676 features (27 geometric and 2649 color and texture) from six color spaces and three regions of interest (nucleus, cytoplasm and the whole cell) for each cell image. Afterwards, by applying information theoretic feature selection, the 20 most relevant and less redundant for the automatic differentiation between the lymphoid cell groups included were analyzed. Statistically significant differences were obtained in all 20 most important features among the median values corresponding to the 12 lymphoid cells groups (p<0.0001). The nucleus/cytoplasm ratio was the best feature to distinguish the different lymphoid cells included in this work. Two additional geometric features were within the first 20: the external profile region (ranked 17th), also called hairiness, and the nuclear circularity (19th). The remaining 15 features that showed relevance for the discrimination among the lymphoid cells subsets were color/texture features (13 statistical and two granulometric). Only three color spaces (CMYK, RGB and HSV) and six color components were involved within them. The contribution of this work is a set of 20 new cytological variables with the following properties: 1) have quantitative formulations, 2) allow qualitative morphological interpretations useful for morphological diagnosis and 3) be efficient to automatically discriminate among a significant number of different lymphoid cell groups through a computerized system.

Keywords: Connectivity analysis, Genes, Animal models and imaging
Abstract: Uncovering genetic influences on brain network organization is critical for explaining the genetic architecture of many psychiatric and neurological disorders, since disturbed brain connectivity of topologically central regions (network hubs) is thought to play a major role in the pathophysiology of a wide array of brain diseases. Hubs have been shown to have distinct transcriptional signatures in humans and mice. Hubs that are more densely interconnected than expected by chance form a rich club, which exists in a range of brain networks. Networks inferred at the macro scale can be imprecise due to the coarse resolution of the imaging techniques used. Caenorhabditis elegans (C. elegans) is the only organism with a nervous system fully mapped at the level of individual neurons thus presents a powerful microscale model for more complex neuronal systems. It provides a way to develop methods that could give insights into the relationship between anatomical connectivity and molecular function by combining gene expression and connectivity data. These methods can later be scaled to the macro level human imaging data. To investigate the gene expression similarity of hub connectivity on a neuronal level, we coupled publicly available synapse-level connectome data of the C. elegans somatic nervous system (279 neurons) with a gene expression dataset containing binary expression profiles for 932 genes (WormBase WS254 release). We confirmed the rich-club organization of the C. elegans connectome with a small number (n=13) of high degree (k = 41-58) nodes. In order to compare different classes of pairwise connections in terms of gene expression similarity we labeled neurons as hubs in the topological rich-club regime (neurons with degree>42) and separated three types of connections accordingly: rich (hub/hub), feeder (hub/nonhub), peripheral (nonhub/nonhub). We computed a measure of transcriptional similarity between each pair of neurons (gene coexpression) as the mean square contingency coefficient between their binary expression profiles. Connected pairs of neurons have more similar expression profiles than unconnected pairs. Gene coexpression for rich links is significantly higher than for any other type of connections. We show the method for linking connectivity to molecular level functioning data which could be applied to a wide range of imaging data.

Keywords: Classification, Microscopy - Light, Confocal, Fluorescence, Cervix
Abstract: In this work, we present preliminary results on the identification of different classes of cells in Pap smear images using a novel image dataset, which consists of images of 4049 annotated cells belonging to five different classes. The classification is based on intensity, texture and shape features which are used as input to a standard neural network.

Keywords: Microscopy - Light, Confocal, Fluorescence, Cells & molecules, Image segmentation
Abstract: Cell division is one of the most fundamental processes of life where the mother cell completely reorganizes its shape to support the partitioning of genetic information and the formation of two identical daughter cells. During this morphological transformation a large number of essential proteins have to carry out their function at specific times and subcellular locations to complete the division successfully. A mechanistic understanding of cell division process requires a comprehensive computational framework that is able to quantitatively analyze the dynamics of cell shape and subcellular protein distributions in an integrated manner is still missing in the state of the art. The proposed research filled this gap by developing a 3D+time (4D) computational model that combines calibrated fluorescence microscopy, image analysis, spatio-temporal registration and computational modelling.

Keywords: Magnetic resonance imaging (MRI), Brain, Classification
Abstract: Patients with major depressive disorder (MDD) who do not achieve full symptomatic recovery after antidepressant treatment have a higher risk of relapse. Compared to pharmacotherapies, electroconvulsive therapy (ECT) more rapidly produces a greater extent of response in severely depressed patients. However, prediction of which candidates are most likely to improve after ECT remains challenging. Using structural MRI data from 42 ECT patients scanned prior to ECT treatment, we developed a random forest classifier based on data-driven shape cluster selection and cortical thickness features to predict remission. Right hemisphere hippocampal shape and right inferior temporal cortical thickness was most predictive of remission, with the predicted probability of recovery decreasing when these regions were thicker prior to treatment. Remission was predicted with an average of 78% accuracy. Classification of MRI data may help identify treatment-responsive patients and aid in clinical decision-making. Our results show promise for the development of personalized treatment strategies.

Keywords: Image segmentation, Brain, Magnetic resonance imaging (MRI)
Abstract: Multi-atlas based label fusion methods have been successfully used for medical image segmentation. In the field of brain region segmentation, multi-atlas based methods propagate labels from multiple atlases to target image by the similarity between patches in target image and atlases. Most of existing multi-atlas based methods usually use intensity feature, which is hard to capture high-order information in brain images. In light of this, in this paper, we endeavor to apply high-order restricted Boltzmann machines to represent brain images and use the learnt feature for brain region of interesting (ROIs) segmentation. Specifically, we firstly capture the covariance and the mean information from patches by three-order restricted Boltzmann Machine and restricted Boltzmann Machine. Then, we propagate the label by the similarity of the learnt high-order features. We validate our feature learning method on two well-known label fusion methods e.g., local-weighted voting (LWV) and non-local mean patch-based method (PBM). Experimental results on the NIREP dataset demonstrate that our method can improve the performance of both LWV and PBM by using the high-order features.

Keywords: Functional imaging (e.g. fMRI), Brain, Graphical models & methods
Abstract: Functional architecture of the brain is organized into functionally specified networks or modules. Many methods have been employed to detect modules in the brain network, for example, Newman’s modularity method and Louvain method for community detection. However, these methods suffer from a resolution limit and the detected number of modules is often inaccurate. In this work, we adopt Eigen Value Decomposition (EVD) on the signless Laplacian of the functional connectivity matrix to detect modules. This method is unaffected by the resolution-limit and could identify the number of clusters more accurately. We tested the EVD method on two datasets. On a cat’s cortex connectome, 5 modules were identified, agreeing with anatomical knowledge while Newman’s and Louvain methods performed unstably. On the 872 fMRI scans in the Human Connectome Project, 9 modules were identified in the functional brain network, which complies well with the field knowledge.

Keywords: fMRI analysis, Connectivity analysis, Brain
Abstract: Functional connectivity of the human brain and the hierarchical modular architecture of functional networks can be investigated using functional MRI (fMRI). Various network models, such as power-law networks and modular networks have been earlier explored to study brain networks. In order to investigate the plausibility of modeling functional brain networks with network models based on distribution of node degree and connection weights, we will compute the goodness-of-fit of several network models on resting-state fMRI scans gathered in the Human Connectome Project. Our experiments suggest that the power-law networks and stochastic block models validly fit on functional connectivity of the subjects and the stochastic block models have the potential to detect functional modules of the brain.

Keywords: Multi-modality fusion, Diffusion weighted imaging, Shape analysis
Abstract: In order to improve the preclinical diagnose of Alzheimer's disease (AD), there is a great deal of interest in analyzing the AD related brain structural changes with magnetic resonance image (MRI) analyses. As the major features, variation of the structural connectivity and the cortical surface morphometry provide different views of structural changes to determine whether AD is present on presymptomatic patients. However, the large scale tensor-valued information and relatively low imaging resolution in diffusion MRI (dMRI) have created huge challenges for analysis. In this paper, we propose a novel framework that improves dMRI analysis power by fusing cortical surface morphometry features from structural MRI (sMRI). We first compute the hyperbolic harmonic maps between cortical surfaces with the landmark constraints thus to precisely evaluate surface tensor-based morphometry. Meanwhile, the graph-based analysis of structural connectivity derived from dMRI is conducted. Next, we fuse these two features via the optimal mass transportation (OMT) and eventually the Wasserstein distance (WD) based single image index is computed as a potential clinical multimodality imaging score. We apply our framework to brain images of 20 AD patients and 20 matched healthy controls, randomly chosen from the Alzheimer's Disease Neuroimaging Initiative (ADNI2) dataset. Our preliminary experimental results of group classification outperformed those of some other single dMRI-based features, such as regional hippocampal volume, mean scores of fractional anisotropy (FA) and mean axial (MD). The novel image fusion pipeline and simple imaging score of structural changes may benefit the preclinical AD and AD prevention research.

Keywords: Brain, Computer-aided detection and diagnosis (CAD), Magnetic resonance imaging (MRI)
Abstract: The geometry of the human brain changes due to age and neurodegeneration. The brain geometry is expected to undergo a similar change in shape with a normal aging, however such change may differ in patients suffering from neurodegenerative disorders. In the novel framework proposed in this work, we model the brain geometry as a 3D point cloud and study the algebraic topology features of this point cloud. Specifically, we compute the persistence timelines of a simplicial complex in a multiscale simplicial homology of the underlying topology space. Further, persistence landscape summary features are obtained from the timelines and studied for their difference between the two groups. The statistical significance obtained in a permutation testing experiments highlights the ability of the persistence landscape features to differentiate between the PD and healthy control brain geometry.

Keywords: Magnetic resonance imaging (MRI), Brain, Pattern recognition and classification
Abstract: The differential diagnosis among ADHD subtypes is an important research area for the neuroimaging community. We pursue this goal by using machine learning techniques in this study. Selective subjects matched by age and handedness information from publicly available ADHD-200 dataset were used in this study. In addition, this work is based only on the resting-state fMRI images. We calculated the global connectivity maps from the fMRI images and used the average of the connectivity measure of each atlas-based cortical parcellation as a feature for the classifier input. For the classification, we used hierarchical extreme learning machine (H-ELM) classifier. By using the proposed feature extraction method, we achieved a 71.11% (p<0.0090) nested cross-validated accuracy and a kappa score of 0.57 in multiclass classification settings.

Keywords: Diffusion weighted imaging, Shape analysis, Brain
Abstract: Cortical folds encode crucial information of brain development, cytoarchitecture and function. It is widely accepted that common anatomy is preserved across individuals within species, while huge variation still hamper establishing fine-grained anatomical correspondences and predicting the locations of a specific anatomical pattern via conventional image registration methods, especially for complex cortical folding pattern, such as gyral 3-hinge. Recently, white matter axonal wiring patterns have been suggested to be strongly correlative to cortical folding patterns. Therefore, in this work, we studied the relation between complex 3-hinge folding patterns and structural connective patterns, and proposed effective methods to predict the locations of 3-hinges by using structural connective features and spatial distribution patterns. The prediction accuracy of our methods outperforms conventional image registration methods.

Keywords: Image reconstruction - analytical & iterative methods, Motion compensation and analysis, Optimization method
Abstract: Three-dimensional (3D) biomedical image sets often have in-plane resolutions that are exceedingly higher than the out-of-plane spacing between images. Image interpolation can be used to reduce the effective out-of-plane spacing. Optical flow and/or other registration-based interpolators have proven useful in interpolating this type of data in the past. When acquired images are comprised of signals that describe the flow velocity of incompressible fluids, additional information is available to guide the interpolation process. In this paper, we present an optical-flow based framework for image interpolation that also minimizes resultant divergence in the interpolated data.

Keywords: Image reconstruction - analytical & iterative methods, Computed tomography (CT), Inverse methods
Abstract: Iterative image reconstruction algorithms have the potential to reduce the radiation dose in computed tomography (CT), but they are computationally intensive and their performance usually depends on careful parameter tuning. In this paper, we propose an iterative CT reconstruction algorithm based on the new class of variance-reduced stochastic gradient descent (VR-SGD) algorithms. Our experiments show that the proposed algorithm has a very fast convergence, while also eliminating the need for step size tuning. This study shows that VR-SGD can be used to devise very efficient iterative CT reconstruction algorithms.

Keywords: Electrical impedance tomography, Brain, Image reconstruction - analytical & iterative methods
Abstract: Electrical Impedance Tomography (EIT) is an alternative way to image brain functions, in the form of conductivity distribution image, by using the boundary voltage information while a small current is injected. In head applications, due to the lack of accurate head models and the high-degree nonlinearity, the image reconstruction tends to fail. Recently, a nonlinear difference imaging approach has been proposed to mitigate modeling error. This approach, however, is based on unconstrained modeling that allows tissue conductivity values to be unrealistically negative. Consequently, substantial image artifacts are possibly conducted. In this work, two methods of constrained modeling were demonstrated they are able to substantially reduce artifacts and improve localization performance. New images of conductivity distribution of the mapped constraint domains, derived from the use of constrained modeling, are also exhibited here. The simulation result shows that the new images achieve better localization performance than those of using unconstrained modeling.

Keywords: Breast, Machine learning, X-ray imaging
Abstract: Detection of Architectural distortion (AD) is important for ruling out possible pre-malignant lesions in breast, but due to its subtlety, it is often missed on the screening mammograms. In this work we suggest a novel AD detection method based on region proposal convolution neural nets (R-CNN). When the data is scarce, as typically the case in medical domain, R-CNN yields poor results. In this study, we suggest a new R-CNN method addressing this shortcoming by using a pretrained network on a candidate region guided by clinical observations. We test our method on the publicly available DDSM data set, with comparison to the latest faster R-CNN and previous works. Our detection accuracy allows binary image classification (normal vs. containing AD) with over 80% sensitivity and specificity, and yields 0.46 false-positives per image at 83% true-positive rate, for localization accuracy. These measures significantly improve the best results in the literature.

Keywords: Computer-aided detection and diagnosis (CAD), Vessels, Computed tomography (CT)
Abstract: Aortic dissection is a condition in which a tear in the inner wall of the aorta allows blood to flow between two layers of the aortic wall. Aortic dissection is associated with severe chest pain and can be deadly. Contrast-enhanced CT is the main modality for detection of aortic dissection. Aortic dissection is one of the target abnormalities during evaluation of a triple rule-out CT in emergency cases. In this paper, we present a method for automatic patient-level detection of aortic dissection. Our algorithm starts by an atlas-based segmentation of the aorta which is used to produce cross-sectional images of the organ. Segmentation refinement, flap detection and shape analysis are employed to detect aortic dissection in these cross-sectional slices. Then, the slice-level results are aggregated to render a patient-level detection result. We tested our algorithm on a data set of 37 contrast-enhanced CT volumes, with 13 cases of aortic dissection. We achieved an accuracy of 83.8%, a sensitivity of 84.6% and a specificity of 83.3%.

Keywords: Computer-aided detection and diagnosis (CAD), Skin, Image segmentation
Abstract: Segmentation of skin lesions is considered as an important step in computer aided diagnosis (CAD) for melanoma diagnosis. There have many attempts to segment skin lesions in a semi- or fully-automated manner. Existing methods, however, have problems with over- or under-segmentation and do not perform well with challenging skin lesions such as when a lesion is partially connected to the background or when image contrast is low. To overcome these limitations, we propose a new semi-automated skin lesion segmentation method that incorporates fully convolutional networks (FCNs) with multi-scale integration. We leverage the use of FCNs to derive high-level semantic information with simple user interaction e.g., a single click to accurately segment skin lesions of various complexity. Our experiments with 379 skin lesion images show that our proposed method achieves better segmentation results when compared to the state-of-the-art skin lesion segmentation methods for challenging skin lesions.

Keywords: Computer-aided detection and diagnosis (CAD), Image enhancement/restoration(noise and artifact reduction), Image segmentation
Abstract: Segmentation of white blood cells (i.e. leukocytes) is a crucial step toward the development of haematological images analysis of peripheral blood smears. This step, however, is complicated by the complex nature of the different types of white blood cells and their large variations in shape, texture, color, and density. This study addresses this issue and presents a single fully automatic segmentation framework for both nuclei and cytoplasm of the five classes of leukocytes in a microscope blood smears. The proposed framework integrates a priori information of enhanced nuclei color with Gram-Schmidt orthogonalization, and multi-scale morphological enhancement to localize the nuclei, whereas clustering-based seed extraction and watershed are utilized to segment the cytoplasm. The experimental results on two different datasets show that the proposed method works successfully for both nuclei and cytoplasm segmentation, and achieves more accurate segmentation results compared to the other methods in the literature.

Keywords: Computer-aided detection and diagnosis (CAD), Lung, Computed tomography (CT)
Abstract: Non-invasive imaging biomarkers identifying has become a powerful tool for predictive diagnosis given increasingly available clinical imaging data. In parallel, molecular profiles has been well documented in non-small cell lung cancers (NSCLCs) over the past decade. However, there has been limited studies on leveraging the two major sources for improving computer-aided diagnosis. In this paper, we investigate the problem of predicting molecular profiles with CT imaging arrays. In particular, we formulate a discriminative convolutional neural network to learn deep features for predicting epidermal growth factor receptor (EGFR) mutation states which are biomarkers associated with cancer cell growth. We evaluate our approach on two independent datasets including Dataset1 (discovery set, 595 patients) and Dataset2 (validation set, 89 patients). Extensive experimental results demonstrate that the learned CNN-based features are effective in predicting EGFR mutation states (AUC=0.828, ACC=76.16%) on Dataset1, and it further demonstrates generalized predictive performance (AUC=0.668, ACC=67.55%) on Dataset2.

Keywords: Optimization method, Tissue, X-ray imaging
Abstract: Osteoarthritis is a common cartilage disease, particularly in societies with aging population. Over 80% of the people over 75 years are affected in the USA. MRI and X-ray can be used to image cartilage, but both approaches suffer from specific drawbacks. X-ray Talbot-Lau interferometers (TLI) have the potential to unite benefits from both modalities. However, TLI setups require to be carefully designed for an imaging task, and the design process itself is not yet well understood.

In this paper, we present an optimization framework for directly visualizing cartilage in the knee with phase-contrast imaging. First, we create simulated phantoms and make a setup-independent choice for an X-ray spectrum that maximizes the theoretically possible contrast to noise ratio over dose. Then, we analytically adapt a Talbot-Lau interferometer to the best spectrum for a knee phantom. It turns out that cartilage can be visualized with an effective dose of 1.16 mSv.

Keywords: Machine learning, Whole-body, Computed tomography (CT)
Abstract: Difficulty on collecting annotated medical images leads to lack of enough supervision and makes discrimination tasks challenging. However, raw data, e.g., spatial context information from 3D CT images, even without annotation, may contain rich useful information. In this paper, we exploit spatial context information as a source of supervision to solve discrimination tasks for fine-grained body part recognition with conventional 3D CT and MR volumes. The proposed pipeline consists of two steps: 1) pre-train a convolutional network for an auxiliary task of 2D slices ordering in a self-supervised manner; 2) transfer and fine-tune the pre-trained network for fine-grained body part recognition. Without any use of human annotation in the first stage, the pre-trained network can still outperform CNN trained from scratch on CT as well as MR data. Moreover, by comparing with pre-trained CNN from ImageNet, we discover that the distance between source and target tasks plays a crucial role in transfer learning. Our experiments demonstrate that our approach can achieve high accuracy with a slice location estimation error of only a few slices on CT and MR data. To the best of our knowledge, our work is the first attempt studying the problem of robust body part recognition at a continuous level.

Keywords: Computer-aided detection and diagnosis (CAD), Kidney, Computed tomography (CT)
Abstract: Detection and segmentation of small renal mass (SRM) in renal CT images are important pre-processing for computer-aided diagnosis of renal cancer. However, the task is known to be challenging due to its variety of size, shape, and location. In this paper, we propose an automated method for detecting and segmenting SRM in contrast-enhanced CT images using texture and context feature classification. First, kidney ROIs are determined by intensity and location thresholding. Second, mass candidates are extracted by intensity and location thresholding. Third, false positive reduction is applied with patch-based texture and context feature classification. Finally, mass segmentation is performed, using the detection results as a seed, with region growing, active contours, and outlier removal with size and shape criteria. In experiments, our method detected SRM with specificity and PPV of 99.63% and 64.2%, respectively, and segmented them with sensitivity, specificity, and DSC of 89.91%, 98.96% and 88.94%, respectively.

Keywords: EEG & MEG, Brain, Classification
Abstract: Surgical resection of portions of the temporal lobe is the standard of care for patients with refractory mesial temporal lobe epilepsy. While this reduces seizures it often results in an inability to form new memories, which leads to difficulties in social situations, learning, and suboptimal quality of life. Learning about the success or failure to form new memory in such patients is critical if we are to generate neuromodulation-based therapies. To this end we tackle the many challenges in analyzing memory formation when their brains are recorded using stereoencephalography (sEEG) in a Free Recall task. Our contributions are threefold. First we compute a rich measure of brain connectivity by computing the phase locking value statistic (synchrony) between pairs of regions, over hundreds of word memorization trials. Second we leverage the rich information (over 400 values per pair of probed brain regions) to form consistent length feature vectors for classifier training. Third we train and evaluate seven different types of classifier models and identify which ones achieve the highest accuracy and which brain features are most important for high accuracy. We assess our approach on data from 37 patients awaiting resection surgery. We achieve up to 73% accuracy distinguishing successful from unsuccessful memory formation in the human brain from just 1.6 sec epochs of sEEG data.

Keywords: Histopathology imaging (e.g. whole slide imaging), Computer-aided detection and diagnosis (CAD), Tissue
Abstract: Tissue segmentation is an important pre-requisite for efficient and accurate diagnostics in digital pathology. However, it is well known that whole-slide scanners can fail in detecting all tissue regions, for example due to the tissue type, or due to weak staining because their tissue detection algorithms are not robust enough. In this paper, we introduce two different convolutional neural network architectures for whole slide image segmentation to accurately identify the tissue sections. We also compare the algorithms to a published traditional method. We collected 54 whole slide images with differing stains and tissue types from three laboratories to validate our algorithms. We show that while the two methods do not differ significantly they outperform their traditional counterpart (Jaccard index of 0.937 and 0.929 vs. 0.870, p < 0.01).

Keywords: Dimensionality reduction, Histopathology imaging (e.g. whole slide imaging), Machine learning
Abstract: One of the main challenges of histological image analysis is the high dimensionality of the images. This can be addressed via summarizing techniques or feature engineering. However, such approaches can limit the performance of subsequent machine learning models, particularly when dealing with highly heterogeneous tissue samples. One possible alternative is to employ unsupervised learning to determine the most relevant features automatically. In this paper, we propose a method of generating representative image signatures that are robust to tissue heterogeneity. At the core of our approach lies a novel deep-learning based mechanism to simultaneously produce representative image features as well as perform dictionary learning to further reduce dimensionality.By integrating this mechanism in a broader framework for disease grading, we show significant improvement in terms of grading accuracy compared to alternative local feature extraction methods.

Keywords: Histopathology imaging (e.g. whole slide imaging), Breast, Pattern recognition and classification
Abstract: Histopathology image classification can provide automated support towards cancer diagnosis. In this paper, we present a transfer learning-based approach for histopathology image classification. We first represent the image feature by Fisher Vector (FV) encoding of local features that are extracted using the Convolutional Neural Network (CNN) model pretrained on ImageNet. Next, to better transfer the pretrained model to the histopathology image dataset, we design a new adaptation layer to further transform the FV descriptors for higher discriminative power and classification accuracy. We used the publicly available BreaKHis image dataset for classifying between benign and malignant breast tumors, and obtained improved performance over the state-of-the-art.

Keywords: Ultrasound, Stereoscopy, Heart
Abstract: An increasing number of minimally invasive cardiac surgeries are guided by real-time 3D ultrasound. We developed a multi-view autostereoscopic system for real-time image guidance with 3D ultrasound. Furthermore, we performed a user study to test the effect of the system on the completion times and a number of errors made during specially designed tasks. The system maintained rendering frame rate at 20 fps and satisfactory added latency of 45-70 ms. In the study, the test subjects controlled a MitraClip device, while being guided by 3D ultrasound presented on either a standard 2D display or a multi-view autostereoscopic display. No significant difference in the completion times between the two visualization systems was measured. However, the test subjects performed 2 errors while using a standard 2D display.

Keywords: Optical coherence tomography, Vessels, Image-guided treatment
Abstract: Automatic identification of side branch and main vascular measurements in IVOCT images take critical roles in pre-interventional decision making for coronary artery disease treatment. But there is very little works have been presented on these tasks. In this paper, we proposed a novel side branch identification algorithm which utilizes a new defined global curvature feature to identify the ostium of side branch. Based on the identification results, the main vascular can be segmented automatically for measurements. In the measurements, the diameter of maximum inscribed circle of main vascular is proposed in the first time, which could be helpful in stent size decision. The qualitative and quantitative validation results demonstrated that the proposed algorithm is effective and accurate.

Keywords: Nuclear imaging (e.g. PET, SPECT), Heart, Quantification and estimation
Abstract: Triple-vessel disease (TVD) is often difficult to detect with single-photon emission computed tomography (SPECT) studies due to the global and potentially uniform reduction of tracer uptake throughout the heart. We aim to study if the stress-to-rest ratio of the signal-to-noise ratio (RSNR) between cardiac rest and stress studies acquired with novel cadmium zinc telluride SPECT scanners could be used to distinguish between patients with and without TVD. Methods: One hundred and two patients with suspected coronary artery disease were retrospectively involved. Each subject underwent a Tl-201 SPECT scan and subsequent coronary angiography within a two-month period and then separated into TVD (n=41) and control (n=61) groups based on the coronary angiography. Automatic segmentation was performed to extract voxel intensities from the left ventricular myocardium. The RSNR was calculated by dividing the stress SNR by the rest SNR. Summed rest, stress, and difference scores were calculated using quantitative perfusion SPECT (QPS) for all subjects. Receiver-operating characteristic analysis was performed using the coronary angiography results as the gold standard for diagnosing TVD. Results: The RSNR in the TVD group was found to be significantly lower than that in the control group (0.83 +/- 0.15 and 1.06 +/- 0.17, respectively; p<0.01). Receiver-operating characteristic analysis showed that the RSNR can detect TVD more accurately than the summed difference score from QPS with higher sensitivity (85% vs. 68%), higher specificity (90% vs. 72%), and higher accuracy (88% vs. 71%).Conclusion: The RSNR may serve as a useful index to assist in the diagnosis of TVD when a fully automatic quantification method is used in cadmium zinc telluride-based SPECT studies.

Keywords: Brain, Electrical impedance tomography, Probabilistic and statistical models & methods
Abstract: This work describes a paradigm shift in brain and central nervous system tract imaging. Low intensity electric field sensing technologies are used passively and actively with advanced spatial-temporal signal processing algorithms to diagnose and non-invasively treat a variety of neurological disorders.

Keywords: Ultrasound, Image acquisition, Tissue
Abstract: This study proposed an angular coherence-based reconstruction scheme in ultrasound plane wave compounding to achieve high image quality at ultrafast frame rates (> 1 kHz). The angular coherence is calculated by measuring the similarity of aperture domain signals among multiple transmissions. Image contrast of the proposed reconstruction scheme was improved by 70.3% (3 angles), 43.2% (21 angles), and 38.7% (63 angles) without compromising the spatial resolution in both the phantom experiments and the in vivo carotid artery.

Keywords: Brain, Angiographic imaging, Infrared imaging
Abstract: A method for visualization of cerebral blood flow based on near infra-red fluorescence imaging is presented. The method can be used intra-operatively to provide low-latency feedback on cerebral perfusion to the surgeon.

Keywords: X-ray imaging
Abstract: Abstract— abnormal lesion detection is one of the most important problems in radiology. We proposed convolutional neural network (CNN) for abnormal multi-lesion detection in chest radiographs

I. MATERIALS AND METHODS We have extended the You Only Look Once (YOLO) model for detecting and classifying normal and 6-classes abnormalities in chest-PA X-ray, which include nodule, consolidation, interstitial opacity, cardiomegaly, pleural effusion, and pneumothorax. For training and testing, 4070 data were enrolled including 252, 80 patients with nodule, 148, 41 patients with consolidation, 247, 82 patients with interstitial opacity, 1210, 757 patients with cardiomegaly, 797, 202 patients with pleural effusion, 197, 57 patients with pneumothorax, respectively, which were randomly split into 80 percent for training and 20 percent for test. All data were from radiology department of Asan Medical Center (AMC). We used labeled and region of interests (ROI) images dataset which is drawn respectively by two thoracic expert radiologists in AMC, which are considered as ground truths. This project is built on top of the YOLO architecture, focusing on fine-tuning the pre-trained model based on the ImageNet dataset. We have adopted the full YOLO model-26 layers including 24 convolution layers and 2 fully connected layers. Our architecture is modified from the last layers after convolution layers. Since there are only 6 classes of abnormal lesion, our last layer requires C = 6. And we randomly crop 448 x 448 size images from the original chest X-ray images.

II. RESULTS AND CONCLUSION Prediction images with rectangular ROI on localized multi abnormalities as well as classification results were displayed. The color scheme of lesions is nodule (pink), consolidation (yellow), interstitial opacity (green), cardiomegaly (emerald), pleural effusion (sky blue) and pneumothorax (blue). The mean average precision (mAP) is about 74.3%. The highest accuracy of lesion detection is 100% of pleural effusion and lowest accuracy is approximately 15% of the consolidation and the bounding box must have an intersection over union (IOU)>0.5. Lastly, the suggested approach could predict other diseases or same diseases without any ground truth mask.

Keywords: Computer-aided detection and diagnosis (CAD), Skin, Eye
Abstract: This special session aims to explore some of the constraints faced by real-world health care providers and clinics and some of the potential ways to resolve or mitigate those constraints. The talks will address one or more of the following aspects of tele health: 1. relevance of telehealth solutions; 2) challenges - in terms of infrastructure, technology, clinical and research facets; 3) feasible computational algorithms; 4) overview of current scenario.

Keywords: Machine learning, Computer-aided detection and diagnosis (CAD), Retinal imaging
Abstract: Diabetic retinopathy (DR) is a major cause of irreversible vision loss, and DR screening relies on retinal clinical signs. Opportunities for computer-aided DR diagnosis have emerged with the development of Ultra-Wide- Field (UWF) digital scanning laser technology. UWF imaging covers 82% greater retinal area (200°), against 45° in conventional cameras, allowing more clinically relevant retinopathy to be detected. UWF images also provide a high resolution of 3078 x 2702 pixels. We are developing methods for automatic detection of DR and its severity, called staging, based on recognition of bright (cotton wool spots and exudates) and dark lesions (microaneurysms and blot, dot and flame haemorrhages) in UWF images. The main contribution of this study is the automated examination of the peripheral area of the retina for the first time, where first or early clinical signs of DR can appear.

Keywords: Magnetic resonance imaging (MRI), Quantification and estimation, Image segmentation
Abstract: The research involves developing image processing techniques for magnetic resonance imaging (MRI) that can be used to improve the clinical picture of the pathophysiological processes preceding the development of osteoarthritis.

Keywords: Breast, Image enhancement/restoration(noise and artifact reduction), Nuclear imaging (e.g. PET, SPECT)
Abstract: Molecular breast imaging (MBI) has been shown to be a highly sensitive imaging modality for detection of breast cancer, especially for patients with dense breasts. Since MBI exams involve the injection of a radioactive tracer, it is important to optimize the tradeoff between radiation dose and image quality. In this paper, we explore the advantages of denoising MBI data. We propose the acquisition of lower-dose images, which are then restored through a denoising algorithm. Denoised and noisy low-dose images are combined through an weighted average to achieve image quality comparable to the full dose noisy image.

Keywords: Brain, Magnetic resonance imaging (MRI), Computational Imaging
Abstract: Few studies have quantified in systematic detail (A) how white matter (WM) connectivity is affected by blood-brain barrier (BBB) disruption, and (B) how micro-hemorrhages resolved via susceptibility-weighted imaging (SWI) lead to changes in the structural organization of the connectome. Here we investigate the extent to which such micro-hemorrhages are associated with WM connectivity alterations and with changes in neurological function in a preliminary sample comprising both old-aged and middle-aged adults with mild traumatic brain injury (mTBI). Our results provide novel insights into how BBB disruption and mechanical shearing of axons during traumatic axonal injury (TAI) may differentially affect neural and cognitive dysfunction in individuals belonging these two age groups, and underline the need to study the heterogeneity of brain responses to mTBI as a function of age at injury.

Keywords: Other-modality, Brain, Functional imaging (e.g. fMRI)
Abstract: Functional near infrared spectroscopy (fNIRS) is an emerging modality in neuroimaging. Major development in fNIRS technology has seen modern fNIRS instrumentation with more than 100 channels and the introduction of wearable/ wireless fNIRS systems. The fNIRS has been applied in the fields of neurology, psychology, education, psychiatry and basic research such as brain computer interface, neuroergonomics and sport sciences research. This special session aims to provide a platform to discuss and share latest advances of fNIRS technology and its applications.

Keywords: Computer-aided detection and diagnosis (CAD), Breast, X-ray imaging
Abstract: To develop deep-learning based computer aided diagnosis (CAD) for chest-PA X-ray images, we proposed class activation map (CAM) with convolutional neural network (CNN) based on manual drawing of the lesions presence in the chestPA X-ray images.

I.MATERIALS AND METHODS Our method extends the concepts of CNN for extracting feature and classifying 6-classes (nodule, consolidation, interstitial opacity, cardiomegaly, pleural effusion, pneumothorax) abnormalities on chest radiographs. To train, validate and test, images were selected from a collection of images of 6072 healthy subjects and 5050 patients with multi-classes including 944 images with nodule, 550 consolidation, 280 interstitial opacity, 2143 cardiomegaly, 1364 pleural effusion, and 331 pneumothorax. Each image were labeled and region of interests (ROI) images were drawn respectively by two thoracic radiologists, which are considered as ground truths. For training the patterns, these data were randomly split into 60% for training, 20% for validation and 20% for test and trained by transfer learning using pre-trained CNN on the ImageNet dataset. Because abnormal patterns could be various and too small on chest radiographs, we preprocessed whole image to patches having various location based on the drawn ROI images. After fine-tuning with patch images, we let the whole radiographic image trained again.

II.RESULTS We developed the system to highlight the abnormal lung area with attention map as well as classification results. The average accuracy of classification was 94.71% for 6-classes based on the weak labeled classes. Because healthy subject was much more than each class, the accuracy was high. The specificity in classifying interstitial opacity was the highest with 87%, whereas the specificity in classifying consolidation was the lowest with 61%.

III.CONCLUSION We successfully let the patterns of abnormalities train and localize on the chest-PA X-ray images with weak labeled data. Most of results of localization showed accurately trained abnormality patterns and localize these regions, which could assist radiologists to diagnose lung diseases.

Keywords: X-ray imaging, Classification, Machine learning
Abstract: Deep learning needs high quality and tremendous number of data set. Therefore, it is necessary to automatically detect outlier in massive data before training. We explore effective data cleansing using convolutional neural network.

Keywords: Deconvolution, Contrast agent quantification, Nuclear imaging (e.g. PET, SPECT)
Abstract: Kinetic analysis of dynamic medical images is useful but not yet supported by open-source software unless Matlab is used. itk::ka2 addresses this gap by extending ITK, a C++ library widely used within medical imaging research. This module has been recently updated and is available from: bitbucket.csiro.au/projects/ITKKA. In addition to a simpler interface for users and programmers, updates include the inclusion of the basis pursuit method, more generalized cost function metrics, faster optimization utilizing analytically defined models and two-stage linear/non-linear modelling. The library is documented and designed to be easily extensible.

Keywords: Quantification and estimation, Magnetic resonance imaging (MRI), Brain
Abstract: Abstract— Multiple sclerosis (MS) is characterised by T2 white matter lesions but may not correlate to a patient’s functional state or impending disease progression. Quantitative susceptibility mapping (QSM) is a biomarker associated with iron concentration and demyelination in white matter. This pilot study examined clinical and MRI parameter relationships to the QSM value over brain regions in mild (EDSS = or <2) relapsing-remitting MS patients.

Keywords: Ultrasound, Muscle, Medical robotics
Abstract: Miniaturization of ultrasound technology has provided an opportunity to utilize imaging in new settings, where imaging has traditionally not been used. Previously, our research group has demonstrated the feasibility of real-time analysis of ultrasound images of dynamic muscle activity for the purpose of controlling an external device such as a multi-articulated prosthetic hand. In this work, we describe the results from investigation of different image analysis algorithms on data acquired from an amputee subject. Our proposed strategy overcomes a number of disadvantages of conventional myoelectric control, which suffers from low signal to noise and limited specificity for deep muscles.

Keywords: Functional imaging (e.g. fMRI), Vessels, Brain
Abstract: We report a novel haemodynamic signal using BOLD-weighted fMRI that is restricted to the internal cerebral veins and identifiable in all subjects assessed to date. This signal is oscillatory with peak power at 0.054 Hz. In early MS patients, venous power was diminished compared to controls and was inversely correlated with T2 lesion volume. These results indicate that neuroinflammation is associated with altered venous haemodynamics in MS.

Keywords: Endoscopy, Gastrointestinal tract, Surgical guidance/navigation
Abstract: Estimating the position of the colonoscope tip with respect to the colon surface would help navigation and is essential for 3D automatic scene reconstruction as we investigated previously by generating visibility maps of the colon internal surface. This paper presents a method to estimate the pose of the colonocope camera with six degrees of freedom (DOF) by training a deep convolutional neural network (CNN). We trained the CNN by the optical flow pattern which was estimated from 30,000 realistic simulated frames. The camera poses for these frames were known from our simulator software. We validated our proposed method to estimate camera pose on simulated video datasets, and on actual colonoscopy videos. Our results showed that the colonoscopy camera pose could be estimated with higher accuracy and speed by leveraging simulated data and the prediction ability of the CNN in comparison to conventional computer vision methods such as classical structure from motion (SFM) pipeline.

Keywords: Perfusion imaging, Quantification and estimation, Lung
Abstract: In this study, we examined the utility of quantitative dynamic contrast-enhanced (DCE) MRI for assessing tumor angiogenesis in patients with non-small-cell lung cancer (NSCLC) by correlating pharmacokinetic parameters estimated from the extended Tofts model (ETM) and the adiabatic approximation to the tissue homogeneity (AATH) model and microvessel density estimated from histopathologic data. We prospectively enrolled 55 patients with NSCLC on the basis of clinical staging, laboratory findings, and enhanced CT. DCE-MRI data were acquired at 1.5-T MR system. The MR signal was converted into concentration of contrast agent using variable flip angle method. For quantification of DCE-MRI data, two different pharmacokinetic models were used: extended Tofts model (ETM), and adiabatic approximation to the tissue homogeneity (AATH) model. The population-averaged arterial input function (AIF) was used for this study. The mean concentration time curve extracted from the manual-drawn tumor areas were non-linearly fitted to ETM and AATH model using the population-averaged AIF. Our results showed that Fp, and vp parameters estimated from AATH model were significantly positive correlation with MVD (r=0.41, 0.26 respectively), and MTT, E, and T0 parameters were significantly negative correlation with MVD (r=-0.37, -0.29, -0.39 respectively) in NSCLC patients. In ETM, Ktrans, ve, and vp parameters showed significantly positive correlation with MVD (r=0.35, 0.30, 0.38 respectively), and T0 parameter showed significantly negative correlation with MVD (r=-0.39). In conclusion, our findings demonstrated that as the microvessel density increases, flow/permeability and plasma volume increases (positive correlation), and tissue enhancement starts earlier (negative correlation), suggesting the utility of DCE-MRI for assessing tumor angiogenesis in NSCLC patients.

Keywords: Skin, Probabilistic and statistical models & methods, Image segmentation
Abstract: Skin lesion segmentation is a challenging task. It is an issue for different automated melanoma diagnosis approaches (ABCD, CASH). Kernel density estimation (KDE) has the ability of differentiating color probability density functions of a lesion and its surrounding healthy tissue in dermoscopic images, which allows accurate lesion segmentation.

Keywords: Ultrasound
Abstract: Theory of and experimental data obtained using field expansion technique that permits immediate visualization of vibrating structures are presented. The approach is based on the angular spectrum method of wave-field analysis where acoustic propagation between parallel planar surfaces is modeled using the two-dimensional Fourier transform. Selected example of the remotely reconstructed surface velocity distribution of a complex acoustic radiator is presented; the 3D plot of the distribution confirms the applicability of this approach in design optimization of sources with arbitrary geometries, including ultrasound (diagnostic and therapeutic) arrays and as a tool in monitoring of structural health.

Keywords: Magnetic resonance imaging (MRI), Image acquisition
Abstract: 23Na-MRI provides a promising tool for the detection of metabolic processes and consequently offers potential as a non-invasive biomarker. Challenging imaging characteristics due to fast bi-exponential signal decay necessitate dedicated acquisition strategies. Zero Echo Time (zTE) sequences have been successfully applied to imaging of solid materials. This work demonstrates initial results of zTE in 23Na-MRI with efficient sampling through twisted trajectories. Phantom measurements over varying saline concentrations in liquid and agar-doped media show good edge characteristics.

Keywords: Population analysis, Angiographic imaging, Probabilistic and statistical models & methods
Abstract: Coronary artery disease severely affects every fourth person across the globe, yet the preferred stent intervention still fails in 2-8% of patients. Local hemodynamics is a major determining factor in adverse outcomes and can be explored through medical imaging with 3D-reconstruction for computational flow analysis. Shape biomarkers were successfully extracted from 20 individual normal cases to reveal predictors of disease probability. In future, these biomarkers may inform clinical intervention.

Keywords: Computer-aided detection and diagnosis (CAD), Machine learning, Vessels
Abstract: Early and more sensitive detection of small aneurysms in 3D cerebral angiograms is required to prevent potentially fatal rupture events. Herein, we propose a novel method that entails structure enhancement filtering to highlight potential aneurysm locations, intra-vascular distance mapping for regional vascular shape encoding and dimensionality reduction and a convolutional neural network to automatically determine optimal features and classification rules for aneurysm detection. Evaluation on 15 3D digital subtraction angiograms showed better performance of the proposed method compared to enhancement filtering and random forest based methods, as it achieved a 100% detection sensitivity at a low number of false positives (2.4 per dataset). The proposed method is also applicable to other angiographic modalities.

Keywords: Machine learning, Heart, Other-modality
Abstract: An automatic, real-time localization of ventricular tachycardia (VT) can improve the efficiency and efficacy of interventional therapies. Because the exit site of VT gives rise to its QRS morphology on electrocardiograms (ECG), it has been shown feasible to predict VT exits from 12-lead ECGs. However, existing work have reported limited resolution and accuracy due to a critical challenge: the significant inter-subject heterogeneity in ECG data. In this paper, we present a method to explicitly separate and represent the factors of variation in data throughout a deep network using denoising autoencoder with contrastive regularization. We demonstrate the performance of this method on an ECG dataset collected from 39 patients and 1012 distinct sites of ventricular origins. An improvement in the accuracy of localizing the origin of activation is obtained in comparison to a traditional approach that uses prescribed QRS features for prediction, as well as the use of a standard autoencoder network without separating the factors of variations in ECG data.

Keywords: Machine learning, Computer-aided detection and diagnosis (CAD), Classification
Abstract: Automatic recognition of surgical workflow is an unresolved problem among the community of computer-assisted interventions. Among all the features used for surgical workflow recognition, one important feature is the presence of the surgical tools. Extracting this feature leads to the surgical tool presence detection problem to detect what tools are used at each time in surgery. This paper proposes a deep learning based multi-label classification method for surgical tool presence detection in laparoscopic videos. The proposed method combines two state-of-the-art deep neural networks and uses ensemble learning to solve the tool presence detection problem as a multi-label classification problem. The performance of the proposed method has been evaluated in the surgical tool presence detection challenge held by Modeling and Monitoring of Computer Assisted Interventions workshop. The proposed method shows superior performance compared to other methods and has won the first prize of the challenge.

Keywords: Ultrasound, Vessels, Image segmentation
Abstract: The segmentation of media-adventitia and lumen-intima boundaries of the Carotid Artery forms an essential part in assessing plaque morphology in Ultrasound Imaging. Manual methods are tedious and prone to variability and thus, developing automated segmentation algorithms is preferable. In this paper, we propose to use deep convolutional networks for automated segmentation of the media-adventitia boundary in transverse and longitudinal sections of carotid ultrasound images. Deep networks have recently been employed with good success on image segmentation tasks, and we thus propose their application on ultrasound data, using an encoder-decoder convolutional structure which allows the network to be trained end-to-end for pixel-wise classification. Concurrently, we evaluate the performance for various configurations, depths and filter sizes within the network. In addition, we further propose a novel fusion of envelope and phase congruency data as an input to the network, as the latter provides an intensity-invariant data source to the network. We show that this data fusion and the proposed network structure yields higher segmentation performance than the state-of-the-art techniques.

Keywords: Machine learning, Histopathology imaging (e.g. whole slide imaging), Vessels
Abstract: We report an automated method for characterization of microvessel morphology in micrographs of brain tissue sections to enable the facile, quantitative analysis of vascular differences across large datasets consisting of hundreds of images with thousands of blood vessel objects. Our objective is to show that virtual 3D parametric models of vasculature are adequately capable of representing the morphology of naturally acquired data in neuropathology. In this work, we focus on three distinct morphologies that are most frequently observed in formalin-fixed, paraffin-embedded (FFPE) human brain tissue samples: single blood vessels showing no (or collapsed) significant lumen (“RoundLumen-”); single blood vessels with distinct lumen (“RoundLumen+”); two blood vessels bundled together in close proximity (“Twins”). The analysis involves extraction of features using pre-trained convolutional neural networks. A hierarchical classification is performed to distinguish single blood vessels (RoundLumen) from Twins; followed by a more granular classification between RoundLumen- and RoundLumen+. A side-by-side comparison of the virtual and natural data models is presented. We observed that classification models built on the virtual data perform well achieving accuracies of 92.8% and 98.3% for the two aforementioned classification tasks respectively.

Keywords: Microscopy - Electron, Image segmentation, Optimization method
Abstract: Biological membranes are one of the most basic structures and regions of interest in cell biology. In the study of membranes, segment extraction is a well-known and difficult problem because of impeding noise, directional and thickness variability, etc. Recent advances in electron microscopy membrane segmentation are able to cope with such difficulties by training convolutional neural networks. However, because of the massive amount of features that have to be extracted while propagating forward, the practical usability diminishes, even with state-of-the-art GPU's. A significant part of these network features typically contains redundancy through correlation and sparsity. In this work, we propose a pruning method for convolutional neural networks that ensures the training loss increase is minimized. We show that the pruned networks, after retraining, are more efficient in terms of time and memory, without significantly affecting the network accuracy. This way, we manage to obtain real-time membrane segmentation performance, for our specific electron microscopy setup.

Keywords: Machine learning, Brain, Magnetic resonance imaging (MRI)
Abstract: 1p/19q co-deletion is an important prognostic factor in low grade gliomas. However, determination of the 1p/19q status currently requires a biopsy. To overcome this, we investigate a radiogenomic classification using support vector machines to non-invasively predict the 1p/19q status from multimodal MRI data. Different approaches of predicting this status were compared, a direct approach which predicts the 1p/19q co- deletion status and an indirect approaches which predicts the mutation status of 1p and 19q individually and combines these predictions to predict the 1p/19q co-deletion status. Using the indirect approach based on both the T1-weighted and T2- weighted images delivered the best result and resulted in a 95% confidence interval for the sensitivity and specificity of [0.44; 0.89] and [0.70; 1.00] respectively.

Keywords: Computer-aided detection and diagnosis (CAD), Magnetic resonance imaging (MRI), Prostate
Abstract: Prostate Cancer (PCa) is highly prevalent and is the second most common cause of cancer-related deaths in men. Multiparametric MRI (mpMRI) is robust in detecting PCa. We developed a weakly supervised computer-aided detection (CAD) system that uses biopsy points to learn to detect PCa on mpMRI. Our CAD system, which is based on a deep convolutional neural network architecture, yielded an area under the curve (AUC) of 0.903+/-0.009 on a receiver operation characteristic (ROC) curve computed on 10 different models in a 10 fold cross-validation experiment. 9 of the 10 ROCs were statistically significantly different from a a competing support vector machine based CAD that yielded a 0.86 AUC when tested on the same dataset (alpha = 0.05). Furthermore, our CAD system proved to be more robust in detecting high-grade transition zone lesions.

Keywords: Microscopy - Light, Confocal, Fluorescence, Cells & molecules, Tracking (time series analysis)
Abstract: Observing the dynamic characteristics of proteins by multi-channel fluorescence microscopy allows studying the interactions between different subcellular structures. We introduce a probabilistic approach for tracking and colocalization analysis of viral proteins in two-channel microscopy image sequences. Our approach is based on particle filters and the Kalman filter， and performs tracking and colocalization analysis jointly. The transition probability matrix is adjusted dynamically. We have applied our approach to two-channel synthetic data and real live-cell fluorescence microscopy images of hepatitis C virus proteins required for the assembly of infectious virus particles. It turned out that the approach accurately determines the colocalization state of the observed particles and robustly tracks the particles.

Keywords: Blind source separation & Dictionary learning, Holography, Cells & molecules
Abstract: We propose a convolutional sparse dictionary learning and coding approach for detecting and counting instances of a repeated object in a holographic lens-free image. The proposed approach exploits the fact that an image containing a single object instance can be approximated as the convolution of a (small) object template with a spike at the location of the object instance. Therefore, an image containing multiple non-overlapping instances of an object can be approximated as the sum of convolutions of templates with spikes. Given one or more images, one can learn a dictionary of templates using a convolutional extension of the K-SVD algorithm for sparse dictionary learning. Given a set of templates, one can efficiently detect object instances in a new image using a convolutional extension of the matching pursuit algorithm for sparse coding. Experiments on red blood cell (RBC) and white blood cell (WBC) detection and counting demonstrate that the proposed method produces promising results without requiring additional post-processing.

Keywords: Probabilistic and statistical models & methods, Classification, Microscopy - Electron
Abstract: Single particle reconstruction (SPR) from cryo-electron microscopy (EM) is a technique in which the 3D structure of a molecule needs to be determined from its contrast transfer function (CTF) affected, noisy 2D projection images taken at unknown viewing directions. One of the main challenges in cryo-EM is the typically low signal to noise ratio (SNR) of the acquired images. 2D classification of images, followed by class averaging, improves the SNR of the resulting averages, and is used for selecting particles from micrographs and for inspecting the particle images. We introduce a new affinity measure, akin to the Mahalanobis distance, to compare cryo-EM images belonging to different defocus groups. The new similarity measure is employed to detect similar images, thereby leading to an improved algorithm for class averaging. We evaluate the performance of the proposed class averaging procedure on synthetic datasets, obtaining state of the art classification.

Keywords: Vessels, Pattern recognition and classification, Infrared imaging
Abstract: Finger vein patterns are unique biometric features, which differentiate from individual to individual, so they are suitable for authentication applications. Systems based on the use of this feature have numerous advantages such as low cost and high accuracy. A new finger vein authentication approach is proposed, which is based on the efficient detection of the non-vein regions, in order to define the main vein patterns. The proposed method is robust in extracting and depicting not only the finger’s vein pattern, but also other important features such as the veins’ width. The authentication algorithm has been evaluated on a finger vein database of 400 images. The false acceptance and false rejection rates achieved are exceptionally small.

Keywords: Ultrasound, Image segmentation, Machine learning
Abstract: Prenatal biometry interpretation from ultrasound image is crucial for fetal growth monitoring. Computerized ultrasound image segmentation methods can greatly improve the efficiency and objectiveness of the interpretation. However, the boundary incompleteness and ambiguity in ultrasound images hinder the automatic solutions severely. In this paper, we propose a cascaded framework for fully automatic prenatal ultrasound image segmentation. Our framework firstly utilizes a customized Fully Convolutional Network (FCN) to exploit feature extractions from multiple visual scales and distinguish the fetal anatomy with a dense prediction map. To enhance the local spatial consistency of the prediction map, we further implant the customized FCN core into an Auto-Context scheme for successive refinement. By modifying the join operator in traditional Auto-Context scheme from parallelization to summation, our framework gains extra considerable improvement. We demonstrate the efficacy of our proposed method on two challenging datasets: fetal head and abdomen ultrasound images. Extensive experimental results show that our method can bridge severe boundary incompleteness and achieves the best segmentation accuracy in various scenarios when compared with other methods.

Keywords: Machine learning, Classification, Other-method
Abstract: As medical imaging datasets grow, we are approaching the era of big data for radiologic decision support systems. This requires renewed efforts in dataset curation and labeling. We propose a methodology for weak labeling of medical images for attributes such as anatomy and disease that relies on image to sentence transformation. The methodology consists of three models, a convolutional neural network that is trained on a coarse classification task and acts as an image feature generator, a language model to map sentences to a fixed length space, and a multi-layer perceptron that acts as a function approximator to map images to the sentence space. The transform model is trained on matched image-sentence pairs on a dataset of echocardiography studies. For a given image, labels are extracted from the closest sentences to the output of the image-sentence transform. We show that the resulting solution has an 78.2% accuracy in labeling Doppler images with aortic stenosis. We also show that the retrieved sentences are consistent with the true sentences in terms of meaning with an average BLEU score of 0.34, matching the current highly performing machine translation solutions.

Keywords: Ultrasound, Fetus, Machine learning
Abstract: Automatic analysis of fetal echocardiography screening images could aid in the identification of congenital heart diseases. The first step towards automatic fetal echocardiography analysis is locating the fetal heart in an image and identifying the viewing (imaging) plane. This is highly challenging since the fetal heart is small with relatively indistinct anatomical structural appearance. This is further compounded by the presence of artefacts in ultrasound images. Herein we provide a state-of-art solution for detecting the fetal heart and classifying each individual frame as belonging to one of the standard viewing planes using fully convolutional neural networks (FCNs). Our FCN model achieves a classification error rate of 23.48% on real-world clinical ultrasound data. We also present comparative performance for analysis of different FCN architectures.

Keywords: Ultrasound, Bone
Abstract: This study aims to investigate the apparent backscatter of cancellous bone from the second ultrasonic harmonics using an imaging system. A cylinder model was used as the bone trabecular model and a positive correlation was found between the power spectrum of the second harmonics and trabecular thickness. The result of the in vitro experiment using ultrasound imaging system shows that the apparent backscatter parameters from the second harmonics (R=0.53- 0.71) correlates better with bone density than the fundamental backscatter signal (R=0.39-0.47).

Keywords: Image enhancement/restoration(noise and artifact reduction), Brain, Ultrasound
Abstract: Hyperechogenicity of the substantia nigra (SN) in the "butterfly shaped" midbrain is a widely recognized diagnostic marker to differentiate between the early stages of Parkinsons Disease (PD) and other diseases which cause parkinsonian symptoms. While clinical differentiation of these diseases can be difficult, hyperechogenicity of the SN is only common in PD patients. Transcranial B-mode Ultrasound Imaging (TCUI) has become a heavily relied upon method to detect echogenicity in the brain. While standard B-mode imaging can show the presence of SN hyperechogenicity, it may not be able to do so with high enough specificity for reliably accurate diagnoses. The cutoff of what is considered a normal echogenic size is 0.2cm^2. Clearly, boundary definition is of the utmost importance to avoid overestimating the size of the echogenic area. Many studies have shown that the harmonic component of ultrasound images have better contrast and dynamic range than standard B-mode images. However a simple bandpass filter across the harmonic frequency would contain interference from the noisy fundamental component due to overlap of the frequency bands. We propose the use of an adaptive Third Order Volterra Filter (TOVF), which is a nonlinear filter that separates a B-mode image into its linear, quadratic, and cubic components regardless of spectral overlap. This paper investigates several variants of the commonly used adaptive Least Mean Squared (LMS) algorithm for determining filter coefficients, and their potential to improve dynamic range and resolution in B-mode images compared to a standard LMS algorithm. We found that several variant algorithms indeed show improvement in terms of Power Spectral Density (PSD) at the harmonics.

Keywords: Retinal imaging, Machine learning, Computer-aided detection and diagnosis (CAD)
Abstract: Retinal Neovascularization (NV) is a critical stage of Diabetic Retinopathy (DR) and its detection is critical to prevent blindness. Existing fully supervised frameworks typically take a patch-based approach and report good results on only limited number of images due to sparsity of annotated data. We propose a patch-based semi-supervised framework which paves way for including unlabeled data in training. In this framework, NV patches are modeled using oriented energy and vesselness based features. These features are fused within a cotraining based semi-supervised framework by ensuring spatial consistency across patches. Rule-based criteria on patch-level neovascularity scores is used to derive the final image-level decision. The proposed approach was evaluated on 3 public and 1 private datasets, both at patch and image level detection on nearly 200,000 patches. An AUC of 0.985 with sensitivity of 96.2% at specificity of 92.6% was obtained for abnormality detection at patch-level, while at the image-level, a sensitivity of 96.76% at specificity of 91.85% were obtained. The achieved performance on a large number of patches indicates the robustness of our approach.

Keywords: Retinal imaging, Eye, Image segmentation
Abstract: Retinal nerve fiber layer defect (RNFLD) is the earliest objective evidence of glaucoma in fundus images. Glaucoma is an optic neuropathy which causes irreversible vision impairment. Early glaucoma detection and its prevention are the only way to prevent further damage to human vision. In this paper, we propose a new automated method for RNFLD detection in fundus images through patch features driven recurrent neural network (RNN). A new dataset of fundus images is created for evaluation purpose which contains several challenging RNFLD boundaries. The true boundary pixels are classified using the RNN trained by novel cumulative zero count local binary pattern (CZC-LBP), directional differential energy (DDE) patch features. The experimental results demonstrate high RNFLD detection rate along with accurate boundary localization.

Keywords: Magnetic resonance imaging (MRI), Compressive sensing & sampling, Optimization method
Abstract: In this study, we deal with the problem of image reconstruction from compressive measurements of multi-contrast magnetic resonance imaging (MRI). We propose a synthesis based approach for image reconstruction to better exploit shared information across contrasts, while retaining unique features of each contrast image. For fast recovery, we propose an augmented Lagrangian based algorithm, using Alternating Direction Method of Multipliers (ADMM). We then compare the proposed algorithm to the state-of-the-art Compressive Sensing-MRI algorithms, and show that the proposed method produces higher-quality images in shorter computation time.

Keywords: Magnetic resonance imaging (MRI), Optimization method, Computational Imaging
Abstract: Estimating the main magnetic field inhomogeneity is important for many magnetic resonance imaging (MRI) techniques. Regularized estimation methods can provide accurate estimates that intrinsically avoid phase wrapping, account for the chemical shift due to fat, and reduce noise. However, these methods require minimizing nonconvex cost functions and existing algorithms are undesirably slow or do not scale to realistic 3D datasets due to memory limitations. This paper proposes a new algorithm that overcomes these limitations. The algorithm adapts the nonlinear conjugate gradient method by incorporating an monotonic line search and efficient iteration-dependent preconditioning. Experiments on multi-echo field map estimation show that our algorithm is competitive with state-of-the-art methods in 2D, and scale successfully to 3D datasets, where current fast methods fail due to memory limitations.

Keywords: Image segmentation, Machine learning, Microscopy - Light, Confocal, Fluorescence
Abstract: Fluorescence microscopy has emerged as a powerful tool for studying cell biology because it enables the acquisition of 3D image volumes deeper into tissue and the imaging of complex subcellular structures. Quantitative analysis of these structures, which is needed to characterize the structure and constitution of tissue volumes, is facilitated by nuclei segmentation. However, manual segmentation is a laborious and intractable process due to the size and complexity of the data. In this paper, we describe a nuclei segmentation method using a deep convolutional neural network, data augmentation to generate training images of different shapes and contrasts, a refinement process combining segmentation results of horizontal, frontal, and sagittal planes in a volume, and a watershed technique to count the number of nuclei. Our results indicate that compared to 3D ground truth data, our method is able to successfully segment and count 3D nuclei.

Keywords: Image segmentation, Microscopy - Light, Confocal, Fluorescence, Microscopy - Multi-photon
Abstract: Image segmentation is an important step in the quantitative analysis of fluorescence microscopy data. Since fluorescence microscopy volumes suffer from intensity inhomogeneity, low image contrast and limited depth resolution, poor edge details, and irregular structure shape, segmentation still remains a challenging problem. This paper describes a nuclei segmentation method for fluorescence microscopy based on the use of three dimensional (3D) active contours with inhomogeneity correction. The correction information utilizes 3D volume information while addressing intensity inhomogeneity across vertical and horizontal directions. Experimental results demonstrate that the proposed method achieves better performance than other reported methods.

Keywords: Validation, Modeling - Anatomical, physiological and pathological, Computational Imaging
Abstract: We introduce a novel model-based generator that produces biologically grounded synthetic volumes of the cerebrovasculature. Our models are synthesized stochastically, according to the biological characteristics of venule arborescence in the human collateral sulcus. Each synthetic volume produced is individually unique, yet representative of this cerebral region. As the locations and characteristics of filaments embedded within our models is known, ground truth data is easily derived. Therefore, our synthetic volumes provide a feasible foundation for the model-based validation of vascular segmentation algorithms.

Keywords: Image segmentation, Shape analysis, Histopathology imaging (e.g. whole slide imaging)
Abstract: In this paper, we propose a novel segmentation method for cells in histopathologic images based on a sparse shape prior guided variational level set framework. We automate the cell contour initialization by detecting seeds and deform contours by minimizing a new energy functional that incorporates a shape term involving sparse shape priors, an adaptive contour occlusion penalty term, and a boundary term encouraging contours to converge to strong edges. As a result, our approach is able to accommodate mutual occlusions and detect contours of multiple intersected cells. We apply our algorithm to a set of whole-slide histopathologic images of brain tumor sections. The proposed method is compared with other popular methods, and demonstrates good accuracy for cell segmentation by quantitative measures, suggesting its promise to support biomedical image-based investigations.

Keywords: Deconvolution, Image enhancement/restoration(noise and artifact reduction)
Abstract: Three-dimensional (3D) deconvolution microscopy is very effective in improving the quality of fluorescence microscopy images. In this work, we present an efficient approach for the deconvolution of 3D fluorescence microscopy images based on the recently developed PURE-LET algorithm. By combining multiple Wiener filtering and wavelet denoising, we parametrize the deconvolution process as a linear combination of elementary functions. Then the Poisson unbiased risk estimate (PURE) is used to obtain the optimal coefficients. The proposed approach is non-iterative and outperforms existing techniques (usually, variants of Richardson-Lucy algorithm) both in terms of computational efficiency and quality. We illustrate its effectiveness on both synthetic and real data.

Keywords: Image reconstruction - analytical & iterative methods, Image enhancement/restoration(noise and artifact reduction), Deconvolution
Abstract: We consider the restoration of blurred images corrupted by Poisson noise using variance-stabilizing transformations (VST). Although VST are an established tool used extensively for denoising, their adoption in deconvolution problems is problematic because VST are necessarily nonlinear operators, and thus break the linear image-formation model typically adopted in deconvolution. We propose a deblurring framework where the image is 1) deconvolved by a linear regularized inverse filter, 2) transformed by VST into an image which can be treated as corrupted by strong spatially correlated noise with constant variance and known power spectrum, 3) denoised by a filter for additive colored Gaussian noise, 4) returned to the original range via inverse VST. We particularly analyze the stabilization of Poisson variates after linear filtering and characterize the noise power spectrum before and after application of VST. We present an efficient implementation of this original deblurring framework using the BM3D denoising filter, demonstrating state-of-the-art results which are especially appealing in low SNR imaging conditions.

Keywords: Image enhancement/restoration(noise and artifact reduction), Probabilistic and statistical models & methods, Endoscopy
Abstract: Laparoscopic images exhibit artifacts resulting from surgical smoke, specular highlights, and noise. These artifacts degrade the results of subsequent processing (e.g., tracking, segmentation, and depth analysis) and compromise surgical quality. We formulate a unified Bayesian inference problem for desmoking, specularity removal, and denoising in laparoscopic images. We propose novel probabilistic graphical models and sparse dictionary models as image priors. For inference, we rely on variational Bayesian expectation maximization. Results on simulated and real-world laparoscopic images, including clinical expert evaluation, show that our joint optimization method outperforms the state of the art.

Keywords: Visualization, Heart, Image enhancement/restoration(noise and artifact reduction)
Abstract: Three dimensional Polarized Light Imaging (3D-PLI) is a contrast-enhancing technique that measures the spatial fiber architecture in the human postmortem brain or heart at a submillimeter resolution. In a voxel, the 3D fiber orientation is defined by the direction angle and the inclination angle whose sign is unknown. To have an accurate explanation of fiber orientation, it is compulsory to clear up this sign ambiguity. A tilting process provides information about the true inclination sign, however the technique is highly sensitive to noise. In this work, a partial differential equations based method is proposed to reduce the noise: the total variation model of Rudin-Osher-Fatemi is extended to 3D orientation vector images to restore the sign. The proposed algorithm is evaluated on synthetic and human heart data and results show that the true sign of the inclination angle can be successfully extracted.

Keywords: Holography, Blind source separation & Dictionary learning, image filtering (e.g. mathematical morphology, wavelets,...)
Abstract: Mitigating the effects of the twin image artifact is one of the key challenges in holographic lens-free microscopy. This artifact arises due to the fact that imaging detectors can only record the magnitude of the hologram wavefront but not the phase. Prior work addresses this problem by attempting to simultaneously estimate the missing phase and reconstruct an image of the object specimen. Here we explore a fundamentally different approach based on post-processing the reconstructed image using sparse dictionary learning and coding techniques originally developed for processing conventional images. First, a dictionary of atoms representing characteristics from either the true image of the specimen or the twin image are learned from a collection of patches of the observed images. Then, by expressing each patch of the observed image as a sparse linear combination of the dictionary atoms, the observed image is decomposed into a component that corresponds to the true image and another one that corresponds to the twin image artifact. Experiments on counting red blood cells demonstrate the effectiveness of the proposed approach.

Keywords: Classification, Machine learning, Magnetic resonance imaging (MRI)
Abstract: The classification of MRI images according to the anatomical field of view is a necessary task to solve when faced with the increasing quantity of medical images. In parallel, advances in deep learning makes it a suitable tool for computer vision problems. Using a common architecture (such as AlexNet) provides quite good results, but not sufficient for clinical use. Improving the model is not an easy task, due to the large number of hyper-parameters governing both the architecture and the training of the network, and to the limited understanding of their relevance. Since an exhaustive search is not tractable, we propose to optimize the network first by random search, and then by an adaptive search based on Gaussian Processes and Probability of Improvement. Applying this method on a large and varied MRI dataset, we show a substantial improvement between the baseline network and the final one (up to 20% for the most difficult classes).

Keywords: Image segmentation, Machine learning, Prostate
Abstract: Automatic MR whole prostate segmentation is a challenging task. Recent approaches have attempted to harness the capabilities of deep learning for MR prostate segmentation to tackle pixel-level labeling tasks. Patch-based and hierarchical features-based deep CNN models were used to delineate the prostate boundary. To further investigate this problem, we introduce a Holistically-Nested Edge Detector (HED) MRI prostate deep learning segmentation and 3D surface reconstruction model that facilitate the registration of multi-parametric MRI with histopathology slides from radical prostatectomy specimens and targeted biopsy specimens. Application of this technique combines deep learning and computer aided design to provide a generalized solution to construct a high-resolution 3D prostate surface from MRI images in three orthogonal views. The performance of the segmentation is evaluated with MRI scans of 100 patients in 4-fold cross-validation. We achieve a mean Dice Similarity of 88.6%.

Keywords: Image quality assessment, Machine learning, Magnetic resonance imaging (MRI)
Abstract: In medical imaging, images are usually evaluated by a human observer (HO) depending on the underlying diagnostic question which can be a time-demanding and cost-intensive process. Model observers (MO) which mimic the human visual system can help to support the HO during this reading process or can provide feedback to the MR scanner and/or HO about the derived image quality. For this purpose MOs are trained on HO-derived image labels with respect to a certain diagnostic task. We propose a non-reference image quality assessment system based on a machine-learning approach with a deep neural network and active learning to keep the amount of needed labeled training data small. A labeling platform is developed as a web application with accounted data security and confidentiality to facilitate the HO labeling procedure. The platform is made publicly available.

Keywords: Magnetic resonance imaging (MRI), Abdomen, Computer-aided detection and diagnosis (CAD)
Abstract: In recent years, magnetic resonance imaging (MRI) has been explored for non-invasive assessment of renal transplant function. This paper proposes a computer-aided diagnostic (CAD) system for the assessment of renal transplant status, which integrates both clinical and MRI-derived biomarkers. The latter are derived from either 3D (2D + time) dynamic contrast-enhanced MRI or 4D (3D + b-value) diffusion-weighted (DW) MRI. In order to extract the MRI-based biomarkers, our framework performs multiple image processing steps, including MRI data alignment to handle the motion effects, kidney segmentation using a geometric deformable model, local motion correction, and estimation of image-based biomarkers. These biomarkers are fused with clinical biomarkers (creatinine clearance and serum plasma creatinine) for the classification of transplant status using a machine learning classifier. Our CAD system has been tested on a cohort of 100 subjects (50 DCE-MRI and 50 DW-MRI) using a “leave-one-subject-out” approach and distinguished rejection from non-rejection transplants with an overall accuracy of 98% for both DCE-MRI and DW-MRI data sets. These preliminary results demonstrate the promise of the proposed CAD system as a reliable non-invasive diagnostic tool for renal transplant assessment.

Keywords: Image segmentation, Prostate, Magnetic resonance imaging (MRI)
Abstract: The automated segmentation of the prostate gland from MR images is increasingly used for clinical diagnosis. Since deep learning demonstrates superior performance in computer vision applications, we propose a coarse-to-fine segmentation strategy using ensemble deep convolutional neural networks (DCNNs) to address prostate segmentation in MR images. First, we use registration-based coarse segmentation on pre-processed prostate MR images to define the potential boundary region. We then train four DCNNs as voxel-based classifiers and classify the voxel in the potential region is a prostate voxel when at least three DCNNs made that decision. Finally, we use boundary refinement to eliminate the outliers and smooth the boundary. We evaluated our approach on the MICCAI PROMIS12 challenge dataset and our experimental results verify the effectiveness of the proposed algorithms.

Keywords: Diffusion weighted imaging, Validation, Machine learning
Abstract: Axon diameter estimation has been a focus of the diffusion MRI community for the past decade. The main argument has been that while diffusion models always overestimate the true axon diameter, their estimation still correlates with changes in true value. Until now, this remains more as a discussion point. The aim of this paper is to clarify this hypothesis using a recently acquired cat spinal cord data set, where the diffusion MRI signal of both a multi-shell and AxCaliber acquisition have been registered with the underlying histology values. We find that the axon diameter as estimated by signal models and AxCaliber does not correlate with their true sizes for axon diameters smaller than 3 mu m. On the other hand, we also train a random forest machine learning algorithm to map signal-based features to histology values of axon diameter and volume fraction. The results show that, in this dataset, this approach leads to a more reliable estimation of physically relevant axon diameters than using sophisticated diffusion models.

Keywords: Diffusion weighted imaging, Computational Imaging, Image reconstruction - analytical & iterative methods
Abstract: This paper proposes a multi-shell sampling scheme and corresponding transforms for the accurate reconstruction of the diffusion signal in diffusion MRI by expansion in the spherical polar Fourier (SPF) basis. The sampling scheme uses an optimal number of samples, equal to the degrees of freedom of the band-limited diffusion signal in the SPF domain, and allows for computationally efficient reconstruction. We use synthetic data sets to demonstrate that the proposed scheme allows for greater reconstruction accuracy of the diffusion signal than the multi-shell sampling scheme obtained using the generalised electrostatic energy minimisation (gEEM) method used in the Human Connectome Project. We also demonstrate that the proposed sampling scheme allows for increased angular discrimination and improved rotational invariance of reconstruction accuracy than the gEEM scheme.

Keywords: Diffusion weighted imaging, Brain, Surgical guidance/navigation
Abstract: This work presents an initial exploration of joint cortical surface and diffusion MRI analysis for neurosurgical patient data. We propose a groupwise cortical modeling strategy that performs an embedding of cortical points from a healthy population and a method for transferring the embedding (with associated information of anatomical label) to patient datasets for cortical parcellation prediction. Our proposed method correlates cortical surfaces based on groupwise white matter connectivity characteristics via a fiber clustering scheme. Unlike other parcellation methods, correspondence of cortical surface vertices is not required. Thus the proposed method can be applied to datasets of patients with brain tumors, using an approximate cortical surface such as a white matter/gray matter boundary derived from diffusion anisotropy. Our initial results on patient data showed good overlap of functional ground truth (subject-specific functional MRI activation areas) with predicted cortical parcels, with 10 of 13 activations overlapping an anatomically corresponding prediction.

Keywords: Diffusion weighted imaging, Machine learning, Brain
Abstract: We propose to use Gaussian process regression to accurately estimate the diffusion MRI signal at arbitrary locations in qspace. By estimating the signal on a grid, we can do synthetic diffusion spectrum imaging: reconstructing the ensemble averaged propagator (EAP) by an inverse Fourier transform. We also propose an alternative reconstruction method guaranteeing a nonnegative EAP that integrates to unity. The reconstruction is validated on data simulated from two Gaussians at various crossing angles. Moreover, we demonstrate on nonuniformly sampled in vivo data that the method is far superior to linear interpolation, and allows a drastic undersampling of the data with only a minor loss of accuracy. We envision the method as a potential replacement for standard diffusion spectrum imaging, in particular when acquistion time is limited.

Keywords: Diffusion weighted imaging, Brain
Abstract: Purpose: To increase the sensitivity in longitudinal analysis of DW-MRI data with the ball-and-sticks model. Methods: Longitudinal DW-MRI data (baseline and two follow-up scans) of 25 middle-aged subjects (47 to 61 years at base line) were acquired. After coregistering all the diffusion-weighted images (DWIs) from the baseline and follow-up scans to a subject-specific intermediate space, an extended ball-and-sticks model was fitted. Stick orientations were constrained such that they did not change over time. The stick fractions were warped and projected onto the TBSS (tract-based spatial statistics) skeleton, and were compared to a reference framework in which all scans were processed independently. Results: Compared to the reference framework, the standard deviation of the apparent noise on the primary stick fractions on the TBSS skeleton was reduced with approximately a factor two. Conclusion: The use of the proposed longitudinal DW-MRI pipeline may significantly increase the precision compared to a default cross-sectional image processing pipeline.

Keywords: Image segmentation, Shape analysis, Magnetic resonance imaging (MRI)
Abstract: In this paper a hierarchical brain segmentation from multiple MRIs is presented for a global-to-local shape analysis. The idea is to group voxels into clusters with high within-cluster and low between-cluster shape relations. Doing so, complementing voxels are analysed together, optimally wheeling the power of multivariate analysis. Therefore, we adapted hierarchical spectral clustering to volumetric image datasets. The outcome is a segmentation of the brain into regions at different levels of detail and anatomical overlap, which are tested in their ability to predict ADAS-cog (a score for cognitive function in Alzheimer’s disease (AD) diagnosis) from brain shape. The results show a benefit for a global-to-local analysis compared to a typical voxel-based and whole brain-based analysis. Additionally, knowledge on brain variations in AD is perfectly confirmed and even elaborated on.

Keywords: Magnetic resonance imaging (MRI), Brain, Quantification and estimation
Abstract: Cortical Thickness (CTh) estimation from Magnetic Resonance Imaging (MRI) data of Multiple Sclerosis (MS) patients is biased at variable extent by the presence of white matter lesions. To overcome this limitation, several methods have been developed. In this study, we evaluate the impact on CTh measurements of different lesion corrections obtained combining three lesion segmentations (manual or automatic) with three intensity filling methods at whole brain and regional scale. Mean relative CTh differences (MRE) after lesion correction with automatic or manually-based methods was used to size the correction effects. Considered the full 3x3 factorial design, an analysis of variance was performed with lesion segmentation and filling as factors. The estimated CTh was remarkably similar between manually-based (gold standard) and automatic corrections with MRE generally well under 2% in all pairwise comparisons and spatial scale. However, a higher MRE was observed using FSL filling. Although the overall CTh agreement, these results suggest that the lesion filling approach provided with FSL library (FMRIB group, Oxford, UK), regardless of the lesion segmentation method used, can deliver an underestimation in the order of 1% of MRE compared to other corrections.

Keywords: Diffusion weighted imaging, Brain, Computational Imaging
Abstract: White matter segmentation is an essential step to study whole-brain structural connectivity via dMRI white matter tractography. One important goal of different segmentation methods is to improve consistency of the white matter segmentations across multiple subjects. In this study, we quantitatively compare two popular white matter segmentation strategies, i.e., a cortical-parcellation-based method and a groupwise fiber clustering method, to investigate their performance on the consistency. Our experimental results indicate that the groupwise fiber clustering tended to generate more consistent segmentations across multiple subjects. This suggests that the fiber clustering strategy could provide a potential alternative to the traditional cortical-parcellation-based brain connectivity modeling methods.

Keywords: Modeling - Anatomical, physiological and pathological, Brain, Magnetic resonance imaging (MRI)
Abstract: We present a method based on deformable meshes for the reconstruction of the cortical surfaces of the developing human brain at the neonatal period. It employs a brain segmentation for the reconstruction of an initial inner cortical surface mesh. Errors in the segmentation resulting from poor tissue contrast in neonatal MRI and partial volume effects are subsequently accounted for by a local edge-based refinement. We show that the obtained surface models define the cortical boundaries more accurately than the segmentation. The surface meshes are further guaranteed to not intersect and subdivide the brain volume into disjoint regions. The proposed method generates topologically correct surfaces which facilitate both a flattening and spherical mapping of the cortex.

Keywords: Image segmentation, Diffusion weighted imaging, Brain
Abstract: Segmentation of the developing cortical plate from MRI data of the post-mortem fetal brain is highly challenging due to partial volume effects, low contrast, and heterogeneous maturation caused by ongoing myelination processes. We present a new atlas-free method that segments the inner and outer boundaries of the cortical plate in fetal brains by exploiting diffusion-weighted imaging cues and using a cortical thickness constraint. The accuracy of the segmentation algorithm is demonstrated by application to fetal sheep brain MRI data, and is shown to produce results comparable to manual segmentation and more accurate than semi-automatic segmentation.

Keywords: Microscopy - Light, Confocal, Fluorescence, Tracking (time series analysis), Modeling - Image formation
Abstract: Fluorescence microscopy has allowed studying dynamical biological processes in vivo with an ever increasing accuracy. Nonetheless, the physically inherent resolution limits impede the study of very dynamical intracellular processes such as microtubule dynamics. One way to overcome this limited resolution is to reconstruct the underlying object dynamics from the image data by using Bayesian statistics. This framework allows combining statistical models about the image formation process and the dynamical process driving the biological function under scrutiny. In this work we show that the accuracy and robustness of tracking microtubule dynamics can be improved by imposing a weak dynamical prior about the hidden geometry of the microtubule and by accounting for the overall photobleaching.

Keywords: Modeling - Anatomical, physiological and pathological, Histopathology imaging (e.g. whole slide imaging), High-content (high-throughput) screening
Abstract: Exploring the spatial interactions between tumor and the inflammatory microenvironment using digital pathology image analysis can contribute to a better understanding of the immune function and tumor heterogeneity. We address this by providing tools able to reveal various metrics describing spatial relationships in the cancer ecosystem. The approach comprises nuclei segmentation and classification using supervised learning algorithm, lymphoid aggregates and tumor patterns detection and spatial distribution quantification using sparse sets' mathematical morphology. Tumor patterns were classified into three groups: surrounded by lymphocytes, close to lymphoid aggregates or distant and might be protected from immune attack. The approach provides statistical assessment and comprehensive visual representation of the inflammatory tumor microenvironment.

Keywords: Modeling - Anatomical, physiological and pathological, Prostate, Radiation therapy, planing and treatment
Abstract: Objectives: Using in silico simulations from histopathological cancer prostate specimen, the objectives were to identify the total dose corresponding to various fractionations necessary to destroy the tumor cells (50% to 99.9%) and to assess the impact of the Gleason score on those doses. Materials and methods: Histopathological specimens were extracted from 7 patients with delineated tumor foci and assigned Gleason scores (GS, indicator of tumor aggressiveness). Slide samples were scanned and used within a simulation model developed in the Netlogo software. The model contained tumor cells, endothelial cells and normal cells. We used the equations of the model simulating the radiation response of hypoxic tumors published by Espinoza et al. [1]. The model parameters were adjusted to biological values from the literature. Three fractionations were tested, at 2, 2.5 and 3 Gy/fraction at 24h interval and using two α/β ratios. Results: A total of 730 simulations were performed. With an α/β of 3.1 Gy, the mean (SD) doses (Gy) to kill 99% of the tumor cells were 72 (±14), 68 (±13) and 65 (±12) for fractionations (Gy) of 2, 2.5 and 3, respectively. With an α/β of 1.5 Gy, the mean (SD) doses (Gy) were 353 (±87), 282 (±66) and 244 (±56) for fractionations (Gy) of 2, 2.5 and 3, respectively. The foci with GS 7: 4+3 needed significantly higher doses than the foci with GS 7: 3+4 to destroy the tumor cells from 50% to 99.9%, at all fractionations. Conclusions: Simulations with α=0.15 Gy-1and β=0.048 Gy-2 show total dose to destroy the tumor closed to clinical experience, depending on the fractionation. Highest Gleason score tissues appear more radio-resistant.

Keywords: Image synthesis, Cells & molecules, Microscopy - Light, Confocal, Fluorescence
Abstract: The existence of benchmark datasets is essential to objectively evaluate various image analysis methods. Nevertheless, manual annotations of fluorescence microscopy image data are very laborious and not often practicable, especially in the case of 3D+t experiments. In this work, we propose a simulation system capable of generating 3D time-lapse sequences of single motile cells with filopodial protrusions, accompanied by inherently generated ground truth. The system consists of three globally synchronized modules, each responsible for a separate task: the evolution of filopodia on a molecular level, linear elastic deformation of the entire cell with filopodia, and generation of realistic, time-coherent cell texture. The capability of our system is demonstrated by generating a synthetic 3D time-lapse sequence of a single lung cancer cell with two growing filopodia, visually resembling its real counterpart acquired using a confocal fluorescence microscope.

Keywords: Ultrasound, Deconvolution
Abstract: Scanning acoustic microscopy (SAM) is a well-accepted imaging modality for forming quantitative, two-dimensional maps of acoustic properties of soft tissues at microscopic scales. The quantitative maps formed using our custom SAM system using a 250-MHz single-element transducer have a nominal resolution of 7 mum, which is insufficient for some investigations. To enhance spatial resolution, a SAM system operating at even higher frequencies could be designed, but associated costs and experimental difficulties are challenging. Therefore, the objective of this study is to evaluate the potential of super-resolution (SR) image processing to enhance the spatial resolution of quantitative maps in SAM. To the best of our knowledge, this is the first attempt at using post-processing, image-enhancement techniques in SAM. Results of realistic simulations and experimental data acquired from a standard resolution test pattern confirm the improved spatial resolution and the potential value of using SR in SAM.

Keywords: Heart, Magnetic resonance imaging (MRI), Image segmentation
Abstract: For electrophysiology procedures, obtaining the information of scar within the left ventricle is very important for diagnosis, therapy planning and patient prognosis. The clinical gold standard to visualize scar is late-gadolinium-enhanced-MRI (LGE-MRI). The viability assessment of the myocardium often requires the prior segmentation of the left ventricle (LV). To overcome this problem, we propose an approach for fully automatic LV segmentation in 2-D LGE-MRI. First, the LV is automatically detected using circular Hough transforms. Second, the blood pool is approximated by applying a morphological active contours approach. The refinement of the endoand epicardial contours is performed in polar space, considering the edge information and scar distribution. The proposed method was evaluated on 26 clinical LGE-MRI data sets. This comparison resulted in a Dice coefficient of 0.85 ± 0.06 for the endocardium and 0.84 ± 0.06 for the epicardium.

Keywords: Machine learning, Magnetic resonance imaging (MRI), Classification
Abstract: In the recent years there have been a number of studies that applied deep learning algorithms to neuroimaging data. Pipelines used in those studies mostly require multiple processing steps for feature extraction, although modern advancements in deep learning for image classification can provide a powerful framework for automatic feature generation and more straightforward analysis. In this paper, we show how similar performance can be achieved skipping these feature extraction steps with the residual and plain 3D convolutional neural network architectures. We demonstrate the performance of the proposed approach for classification of Alzheimer's disease versus mild cognitive impairment and normal controls on the Alzheimer’s Disease National Initiative (ADNI) dataset of 3D structural MRI brain scans.

Keywords: Liver, Computer-aided detection and diagnosis (CAD), Diffusion weighted imaging
Abstract: Automatic non-invasive assessment of hepatocellular carcinoma (HCC) malignancy has the potential to substantially enhance tumor treatment strategies for HCC patients. In this work we present a novel framework to automatically characterize the malignancy of HCC lesions from DWI images.

We predict HCC malignancy in two steps: As a first step we automatically segment HCC tumor lesions using cascaded fully convolutional neural networks (CFCN). A 3D neural network (SurvivalNet) then predicts the HCC lesions' malignancy from the HCC tumor segmentation. We formulate this task as a classification problem with classes being "low risk" and "high risk" represented by longer or shorter survival times than the median survival. We evaluated our method on DWI of 31 HCC patients. Our proposed framework achieves an end-to-end accuracy of 65% with a Dice score for the automatic lesion segmentation of 69% and an accuracy of 68% for tumor malignancy classification based on expert annotations. We compared the SurvivalNet to classical handcrafted features such as Histogram and Haralick and show experimentally that SurvivalNet outperforms the handcrafted features in HCC malignancy classification. End-to-end performance of our proposed framework is with p>0.953 equivalent to tumor malignancy classification based on expert annotations.

Keywords: Computer-aided detection and diagnosis (CAD), Image segmentation, Heart
Abstract: Precise and objective segmentation of atrial scarring (SAS) is a prerequisite for quantitative assessment of atrial fibrillation using non-invasive late gadolinium-enhanced (LGE) MRI. This also requires accurate delineation of the left atrium (LA) and pulmonary veins (PVs) geometry. Most previous studies have relied on manual segmentation of LA wall and PVs, which is a tedious and error-prone procedure with limited reproducibility. There are many attempts on automatic SAS using simple thresholding, histogram analysis, clustering and graph-cut based approaches; however, in general, these methods are considered as unsupervised learning thus subject to limited segmentation accuracy. In this study, we present a fully-automated multi-atlas based whole heart segmentation method to derive the LA and PVs geometry objectively that is followed by a fully automatic deep learning method for SAS. Our deep learning method consists of a feature extraction step via super-pixel over-segmentation and a supervised classification step via stacked sparse auto-encoders. We demonstrate the efficacy of our method on 20 clinical LGE MRI scans acquired from a longstanding persistent atrial fibrillation cohort. Both quantitative and qualitative results show that our fully automatic method obtained accurate segmentation results compared to the manual segmentation based ground truths.

Keywords: Brain, Machine learning, Magnetic resonance imaging (MRI)
Abstract: Estimating human age from brain MR images is useful for early detection of Alzheimer's disease. In this paper we propose a fast and accurate method based on deep learning to predict subject's age. Compared with previous methods, our algorithm achieves comparable accuracy using fewer input images. With our GPU version program, the time needed to make a prediction is 20 ms. We evaluate our methods using mean absolute error (MAE) and our method is able to predict subject's age with MAE of 4.0 years.

Keywords: Image registration, Prostate, Ultrasound
Abstract: Despite the common invisibility of cancerous lesions in trans-rectal ultrasound (TRUS), TRUS-guided random biopsy is considered the gold standard to diagnose prostate cancer. Pre-interventional magnetic resonance imaging (MRI) has been shown to improve the detection of malignancies but fast and accurate MRI/TRUS registration for multi-modal biopsy guidance remains challenging. In this work, we derive a statistical deformation model (SDM) from 50 automatically segmented patient datasets and propose a novel registration scheme based on a lesion-specific, anisotropic preconditioned similarity metric. The approach is validated on a dataset of 10 patients, showing landmark registration errors of 1.41 mm in the vicinity of suspicious areas.

Keywords: Image registration, Machine learning
Abstract: We introduce a deep encoder-decoder architecture for image deformation prediction from multimodal images. Specifically, we design an image-patch-based deep network that jointly (i) learns an image similarity measure and (ii) the relationship between image patches and deformation parameters. While our method can be applied to general image registration formulations, we focus on the Large Deformation Diffeomorphic Metric Mapping (LDDMM) registration model. By predicting the initial momentum of the shooting formulation of LDDMM, we preserve its mathematical properties and drastically reduce the computation time, compared to optimization-based approaches. Furthermore, we create a Bayesian probabilistic version of the network that allows evaluation of registration uncertainty via sampling of the network at test time. We evaluate our method on a 3D brain MRI dataset using both T1- and T2-weighted images. Our experiments show that our method generates accurate predictions and that learning the similarity measure leads to more consistent registrations than relying on generic multimodal image similarity measures, such as mutual information. Our approach is an order of magnitude faster than optimization-based LDDMM.

Keywords: fMRI analysis, Image registration
Abstract: An accurate registration plays a critical role in group-wise fMRI image analysis. Inspired by the observations that common functional networks can be reconstructed from fMRI image across individuals and in different brain states, we propose a novel computational framework for fMRI image registration by using these common function networks as references for correspondence between individuals. This framework innovatively utilizes averaged functional networks from fMRI images instead of structure MRI templates as registration reference and aims to maximize the match between each functional network across individuals. The proposed method is examined on the resting state fMRI images of 184 subjects publicly released by an autism study. Our results show increased registration accuracy and revealed meaningful findings on brain functional network changes in autism patients.

Keywords: Optimization method, Image registration, Shape analysis
Abstract: Diffeomorphic image registration algorithms are widely used in medical imaging, and require optimization of a high- dimensional nonlinear objective function. The function being optimized has many characteristics that are relevant for op- timization but are typically not well understood. Due to that complexity, most authors have used a simple gradient de- scent, but it is not often discussed how step sizes are chosen or if line searches are used. Further, if a system is to be robust to a range of input images, that may differ to varying degrees, the optimization must be adaptable. Here, we present two methods of adaptable gradient descent with line searches, and test how they affect image registration. The optimization schemes are deployed for geodesic shooting in diffeomor- phisms - an approach that is used to quantify anatomical changes, such as atrophy, in longitudinal image pairs. We evaluate the optimization schemes on their convergence char- acteristics and based on how well the resulting atrophy scores correlate with diagnostic group and mini mental state exam (MMSE) scores. We find that the Barzilai-Borwein method with a backtracking line search outperforms other optimiza- tion schemes in convergence time and adaptability by a wide margin. We also find that the variable optimization schemes do not significantly affect the ability to measure atrophy with clinical significance.

Keywords: Other-method, Image registration, Shape analysis
Abstract: Longitudinal registration has been used to map brain atrophy and tissue loss patterns over time, in both healthy and demented subjects. However, we have not seen a thorough application of the geodesic shooting in diffeomorphisms framework for this task. The registration model is complex and several choices must be made that may significantly impact the quality of results. One of these decisions is which image matching functional should drive the registration. We investigate four matching functionals for atrophy quantification using geodesic shooting in diffeomorphisms. We check if the choice of matching functional has an impact on the correlation of atrophy scores with clinical variables. We also check the impact of matching functional choice on estimates of the N80 sample size for hypothetical clinical trials that test for slowing of brain atrophy. We find that the mutual information function, which has primarily been used for linear and multi-modal registration, achieves comparable correlation with clinical variables to other matching functionals while yielding better sample size estimates.