Keywords: Blind source separation & Dictionary learning, Computer-aided detection and diagnosis (CAD), Classification
Abstract: Mouse models are broadly used to study the mechanisms of neuropsychiatric disorders and to test potential treatments. In these models, automation to monitor behavioural differences during social interactions is currently limited. We propose in the present study a new method to conduct automatic behavioural classification, using an original unsupervised machine learning. We applied the proposed method to mice mutated in Shank2, a gene associated with autism spectrum disorders. We validated our results by comparing automatically extracted results to rule-based classifier labelling. We discovered seven behavioural states matching from 80 to 95% previous rule-based classification, and two unsuspected behaviours. Interestingly, we also highlighted genotype-related differences in two behavioural categories, namely locomotion and facing the conspecific.

Keywords: Retinal imaging, Eye, Computer-aided detection and diagnosis (CAD)
Abstract: Retinal vessel keypoint detection and classification is a fundamental step in tracking the physiological changes that occur in the retina which is linked to various retinal and systemic diseases. In this paper, we propose a novel Vessel Keypoint Detector (VKD) which is derived from the projection of log-polar transformed binary patches around vessel points. VKD is used to design a two-stage solution for junction detection and classification. In the first stage, the keypoints detected using VKD are refined using curvature orientation information to extract candidate junctions. True junctions from these candidates are identified in a supervised manner using a Random Forest classifier. In the next stage, a novel combination of local orientation and shape based features is extracted from the junction points and classified using a second Random Forest classifier. Evaluation results on five datasets show that the designed system is robust to changes in resolution and other variations across datasets, with average values of accuracy/sensitivity/specificity for junction detection being 0.78/0.79/0.75 and for junction classification being 0.87/0.85/0.88. Our system outperforms the state of the art method [1] by at least 11%, on the DRIVE and IOSTAR datasets. These results demonstrate the effectiveness of VKD for vessel analysis

Keywords: Retinal imaging, Vessels, Computer-aided detection and diagnosis (CAD)
Abstract: Proliferative diabetic retinopathy is an important public health issue with deteriorating impact on the vision of its patient. In this study, a novel approach is proposed for the characterization of abnormal vessels based on a spatial point pattern method. Points of interest corresponding to vascular junctions are detected by a perceptual organization technique, and then second-order statistical measures are computed. Significant differences (p<0.05) between healthy retinal regions and areas with neovascularizations were obtained, which suggests that the second-order statistics could be used as a relevant feature to discriminate the abnormal from the normal vasculature. The relevance of the new measures was also evaluated with respect to an existing set of features using classification. The inclusion of a new second-order measure increases the characterization sensitivity against the already existing features from 75.76% to 84.85%.

Keywords: Computer-aided detection and diagnosis (CAD), Brain, Magnetic resonance imaging (MRI)
Abstract: This paper presents a novel longitudinal framework for clinical score prediction in Alzheimer’s disease (AD) diagnosis. In contrast to the previous approaches that use the data collected at a single time point only for the clinical score prediction, we propose to exploit the imaging data of multiple time points. Furthermore, a spatial-temporal group sparse method is proposed for robust feature selection through imposing a fused smoothness term and a locality-preserving-projection based term as well as integrating correntropy into the framework, which is able to promote the prediction consistency and reduce the adverse effect of noises and outliers. Ensemble learning of support vector regression (SVR) is exploited to predict the AD scores more accurately with the selected features. The proposed approach is extensively evaluated on the Alzheimer’s disease neuroimaging initiative (ADNI) dataset. The experiments demonstrate that our proposed approach not only achieves promising regression accuracy, but also recognizes disease-related biomarkers successfully.

Keywords: Computer-aided detection and diagnosis (CAD), Computational Imaging, Computed tomography (CT)
Abstract: Adrenal lesions include a wide variety of benign and malignant neoplasms of the adrenal gland, and are seen in up to 5% of computed tomography (CT) examinations of the abdomen. Better identification of these lesions is important for effective management and patient prognosis. Detection on low-contrast CT images, however, even for experienced physicians can be difficult and error-prone, because the lesions are often problematic to be separated from the normal surrounding structures. Existing lesion detection techniques have problems in identifying and differentiating low-contrast tumors, which is related to their use of low-level features rather than high level of semantics. Hence we propose an automated approach using fully convolutional networks (FCNs) and multi-scale integration to detect adrenal lesion on low-contrast CT scans. The architecture of FCNs includes deep, coarse, semantic information and shallow, fine, appearance information in a hierarchical manner and it enables the encoding of image-wide location and semantics, which are desirable characteristics for adrenal lesion detection. We also propose a multi-scale integration with a superpixel based random walk (MI-SRW) approach to refine the lesion boundaries on different scales. The MI-SRW technique enables us to constrain the spatial and appearance consistency and then use complementary information provided on different scales to detect adrenal lesions of various sizes and characteristics. We used 38 adrenal lesions detected on low-contrast CT and compared our approach to existing ‘state-of-the-art’ methods and found that our approach had superior detection performance.

Keywords: Ultrasound, Perfusion imaging, Contrast agent quantification
Abstract: The onset and progression of immune-mediated inflammatory arthritis, such as rheumatoid arthritis and spondyloarthritis, are linked to the IL23-IL17 immune axis, so that many therapeutic strategies aim at modulating this pathway. However, there is so far no possibility of an in vivo direct monitoring, without a biopsy, of the specific T cells involved in this modulation. Synovial perfusion, and thus synovial angiogenesis, has been recognized as a sensitive and early marker of inflammation that can be evaluated via quantitative analysis of contrast-enhanced ultrasound imaging data. We propose a quantitative analysis of contrast enhanced ultrasound data, exploiting both a pixel-wise analysis for characterizing the perfusion patterns and their heterogeneity within a patient’s synovia, and a model that add to the gamma-variate function a term accounting for a possible slow-flow component, whose presence and amplitude is estimated via a variational Bayesian method. We show that this quantification allows to find a relationship between perfusion parameters to CD4+ T helper cells subtypes that are believed to be involved in the IL23-IL17 immune axis modulation: significant correlations are as high as 0.90, suggesting the possibility of estimating T cells concentration from non invasive imaging data.

Keywords: Ultrasound, Image acquisition, Elasticity measures
Abstract: In elastography, Conventional Linear Array (CLA) - based RF data acquisition can only provide good quality displacement measurements in the direction of beam propagation (axial direction). For obtaining high-precision Lateral Displacement Estimation (LDE), one of the popular methods is by interpolating A-lines in between neighboring RF A-lines. However, acquiring and utilizing the actual data from sub-pitch location will yield fundamentally better estimation. In this paper, we explore a novel method of acquiring and augmenting post-beamformed RF A-line in sub-pitch locations by electronically translating the sub-aperture by activating odd and even number of elements alternatively. We compare this approach to another recently described method where sub-pitch translations of beams were accomplished by actuator-assisted translation of the linear array transducer. The performances of the methods were studied through simulation and experiments on phantoms. The results demonstrate that these methods yield better quality LDE compared to those obtained from interpolation of RF A-lines.

Keywords: Ultrasound, Vessels, Compressive sensing & sampling
Abstract: Spectral Doppler ultrasound imaging allows quantification of blood flow by estimating the velocity distribution of the blood within a range gate on a single image line. In applications, it is desirable to have both brightness mode (B-mode) at high frame rate, to allow the medical doctor to visualize and follow vessel movement, as well as Doppler mode with high spectral resolution of the velocity distribution, so that changes in blood flow can be tracked. In this paper, we propose a slow-time sparse emission pattern of the blood Doppler signal which significantly reduces the number of transmissions sent. For the proposed sampling scheme, we derive the minimal number of Doppler emissions allowing reconstruction of the signal’s power spectrum and provide power spectrum recovery techniques that achieve this minimal rate. Using realistic Field II simulations, we show that accurate estimation of blood velocity spectrum can be performed from only 12% of the transmissions required in conventional scanning. Thus, several vessel regions may be investigated while keeping a high frame rate of the B-mode images.

Keywords: Motion compensation and analysis, Image acquisition, Ultrasound
Abstract: In the context of ultrasound (US) deformation imaging it has been shown that lateral motion estimation can be improved by simultaneously combining transverse oscillation (TO) beamforming with dedicated motion estimators and regularisation techniques. This paper provides insights into the relative contributions of beamforming strategies (focused imaging [FOC], plane wave TO beamforming [PWTO], and PWTO after heterodyning [PWTO*]) and regularisation (unregularised RF tracking, combined envelope/RF tracking without regularisation, and combined envelope/RF tracking with explicit regularisation) to the performance of US displacement estimation algorithms. More specifically, a non-rigid image registration technique was used to track lateral tissue motion in an in-silico and in-vitro phantom setup for all beamforming-tracking combinations. It was found that PWTO and PWTO* tracking benefited more from regularisation than FOC tracking, and was even a necessary requirement to warrant the use of TO beamforming over traditional FOC imaging. For example, for a range of lateral displacements (0–1000 micrometer), in-silico errors were 56±56 micrometer and 118±293 micrometer for the unregularized FOC and PWTO* scenario respectively. After regularisation, these errors remained relatively stable for the FOC case but drastically reduced for the PWTO* scenario: 51±45 micrometer versus 14±12 micrometer respectively.

Keywords: Ultrasound, Thermal imaging, Computational Imaging
Abstract: A dual-mode ultrasound array (DMUA) imaging system for monitoring and delivery of transcranial focused ultrasound (tFUS) therapy is described. A DMUA prototype operating in the frequency range 2.4 - 4.2 MHz has been shown to generate tFUS beams at therapeutic levels in rat brains textit{in vivo} based on two-dimensional image guidance. Furthermore, real-time ultrasound thermography was shown to be feasible using beamformed DMUA imaging data. In this paper, we demonstrate the use of 3D DMUA imaging for guidance of tFUS beam to specific locations within the brain defined with reference to the bregma. The results demonstrate the ability of the DMUA prototype to clearly resolve the lambda, bregma and the medial suture lines, which allows the precise placement of the focal spot at the target location in a repeatable manner within a single experiment session and from one session to another. This has significant impact on tFUS applications requiring multiple treatments of the same target region, e.g. temporal lobe epilepsy.

Keywords: Image segmentation, Atlases, Brain
Abstract: Microscopic analysis of histological sections is considered the “gold standard” to verify structural parcellations in the human brain. Its high resolution allows the study of laminar and columnar patterns of cell distributions, which build an important basis for the simulation of cortical areas and networks. However, such cytoarchitectonic mapping is a semiautomatic, time consuming process that does not scale with high throughput imaging. We present an automatic approach for parcellating histological sections at 2um resolution. It is based on a convolutional neural network that combines topological information from probabilistic atlases with the texture features learned from high-resolution cell-body stained images. The model is applied to visual areas and trained on a sparse set of partial annotations. We show how predictions are transferable to new brains and spatially consistent across sections.

Keywords: Breast, Histopathology imaging (e.g. whole slide imaging), Computer-aided detection and diagnosis (CAD)
Abstract: One of the most important prognostic markers to assess proliferation activity of breast tumors is estimating the number of mitotic figures in H&E stained tissue. We propose the use of a recently published convolutional neural architecture, Wide Residual Networks, for mitosis detection in breast histology images. The model is trained to classify each pixel of on an image using as context a patch centered on the pixel. We apply post-processing on the network output in order to filter out noise and select true mitosis. Finally, we combine the output of several networks using majority vote. Our approach ranked 2nd in the MICCAI TUPAC 2016 competition for mitosis detection, outperforming most other contestants by a significant margin.

Keywords: Computer-aided detection and diagnosis (CAD), Histopathology imaging (e.g. whole slide imaging), Breast
Abstract: Diagnosis of breast carcinomas has so far been limited to the morphological interpretation of epithelial cells and the assessment of epithelial tissue architecture. Consequently, most of the automated systems have focused on characterizing the epithelial regions of the breast to detect cancer. In this paper, we propose a system for classification of hematoxylin and eosin (H&E) stained breast specimen based on convolutional neural networks that primarily targets the assessment of tumor-associated stroma to diagnose breast cancer patients. We evaluate the performance of our proposed system using a large cohort containing 646 breast tissue biopsies. Our evaluations show that the proposed system achieves an area under ROC of 0.92, demonstrating the discriminative power of previously neglected tumor-associated stroma as a diagnostic biomarker.

Keywords: Image segmentation, Histopathology imaging (e.g. whole slide imaging), Machine learning
Abstract: Analysis and interpretation of stained tumor sections is one of the main tools in cancer diagnosis and prognosis, which is mainly carried out manually by pathologists. The avent of digital pathology provides us with the challenging opportunity to automatically analyze large amounts of these complex image data in order to draw biological conclusions from them and to study cellular and tissular phenotypes at a large scale. One of the bottlenecks for such approaches is the automatic segmentation of cell nuclei from this type of image data. Here, we present a fully automated workflow to segment nuclei from histopathology image data by using deep neural networks trained from a set of manually annotated images and by processing the posterior probability maps in order to split jointly segmented nuclei. Further, we provide the image data set that has been generated for this study as a benchmark set to the scientific community.

Keywords: Histopathology imaging (e.g. whole slide imaging), Machine learning, Pattern recognition and classification
Abstract: The detection of cells and nuclei is a crucial step for the automatic analysis of digital pathology slides and as such for the quantification of the phenotypic information contained in tissue sections. This task is however challenging because of high variability in size, shape and textural appearance of the objects to be detected and of the high variability of tissue appearance. In this work, we propose an approach to specifically tackle the variability in size. Modeling the detection problem as a local maxima detection problem on a center probabilistic map, we introduce a nuclear surface area map to guide the selection of local maxima while releasing a-priori knowledge on the size or structure of the objects to be detected. The good performance of our approach is quantitatively shown against state-of-the-art nuclei detection methods.

Keywords: Microscopy - Electron
Abstract: Understanding the mechanism of pore formation by perforin-like proteins using Cryo-Electron Microscopy and X-ray Crystallography.

James C. Whisstock, ARC Centre of Excellence in Advanced Molecular Imaging, Monash University.

Members of the perforin-like superfamily of pore forming proteins are found in all classes of life and perform key roles in immunity, in defence, as bacterial virulence factors and in developmental biology. In humans, family members include perforin, Complement component-9 (C9) and Macrophage Expressed Gene-1 (MPEG-1). These proteins are all known to transition from a soluble monomeric form to adopt a substantial, oligomeric membrane inserted pore that is of sufficient size (>150 Å)to permit the passage of fully folded proteinaceous toxins. In this presentation I will describe our current state of knowledge with respect to the structure of perforin-like pores and how these molecules transition from the monomeric to the oligomeric form.

Keywords: Microscopy - Super-resolution, Single cell & molecule detection, Visualization
Abstract: Fluorescence flow cytometry quantifies biomolecules on a cellular basis for population discrimination with high sample throughput, but provides no spatial information about the detected biomolecules. dSTORM super-resolution microscopy allows for imaging at the nanometer scale revealing spatial distribution and interactions of biomolecules in situ, but its time-consuming imaging process inevitably limits sample throughput.

To synergistically join flow cytometry and dSTORM, we developed SortdSTORM, a stand-alone technique with a single flexible and time-efficient experimental workflow. We apply SortdSTORM to delineate and characterize subpopulations of immune cells involved in the signaling of IL-37, a powerful anti-inflammatory cytokine with great therapeutic potential.

Whereas by flow cytometry alone both unstimulated and inflammatory stimulated cells look similar, SortdSTORM reveals functionally crucial differences in molecular distribution, including complex formation under stimulated conditions. We rapidly visualize, characterize and quantify IL-37 and its two receptor proteins IL-18Rα and IL-1R8 on clearly defined, low-abundant cell populations at nanometer-resolution. Quantitative assessment of cluster formation provides critical insights into the potential and readiness of a cell to carry out the anti-inflammatory signaling of IL-37 and therewith the mechanisms and prerequisites for IL-37 to execute its anti-inflammatory program in human immune cells.

Keywords: Microscopy - Electron, Microscopy - Light, Confocal, Fluorescence, Brain
Abstract: Identifying stem cell phenotypes involved in brain repair using novel correlative-light electron microscopy methods Viola Oorschot2, Benjamin W. Lindsey1, Jan Kaslin1, Georg Ramm2

1Australian Regenerative Medicine Institute, Monash University Clayton Campus, Clayton, Victoria, Australia 2Monash Ramaciotti Centre for Structural Cryo EM, Monash University, Melbourne, Victoria, 3800, Australia.

Which stem cells contribute to the reparative process following injury to the adult brain remains a fundamental question. Uncovering the identify of endogenous adult neural stem cell populations has great promise for developing future therapeutic strategies for patients with brain injury or neurodegenerative disease. While glial-scarring following lesion largely prevents neural regeneration in mammals, the adult zebrafish has become a leading model for CNS repair owing to its exquisite ability to regenerate major regions of its brain following injury using stem/progenitor cells. The heterogeniety of forebrain stem/progenitor cells in the zebrafish has been an obstacle however in elucidating the subset of these cells involved in the regenerative process given the lack of cell-specific markers available. To overcome this problem, we have developed novel correlative-light electron microscopy (CLEM) techniques that take advantage of the high degree of fluorescent labelling possible using the Tokuyasu cryo-embedding approach. By combining Tokuyasu sample preparation with scanning electron microscopy (SEM), scanning-transmission electron-microscopy (STEM), and transmission electron microscopy (TEM), we are able to seamlessly observe and quantify the morphological profiles of adult neural stem cells using both immuno-fluorescent antibody markers and ultrastructural details. Our data thus far suggests that proliferative cells of a non-glial phenotype may play a larger role in the regenerative process than initially thought. The strength of this approach will enable us to conclusively identify and compare the population and hierarchy of stem cells in the healthy and injured brain, and will serve as a leading method to reliable distinguish stem cells across organ tissues.

Keywords: Kidney
Abstract:

Keywords: Classification, Other-modality
Abstract: The water quality of lakes, rivers, and streams has long been monitored by the government and other agencies; however, many of them still do not meet the standards of drinkable or swimmable. These problems are in part due to the lack of knowledge of the source of bacterial contamination. Optical imaging exploring the morphological features of bacterial colonies provides a promising solution to discriminate different bacteria species recently. The bacterial scattering image features were extracted by wavelet decomposition after image preprocessing and normalization. The support vector machine was chosen as the pattern classifier to identify the unknown bacterial scattering image by comparing the feature vectors of those in the database. It achieved high accuracy at the subspecies level for source tracking with limited hosts (human, farm animals and water birds). However, the accuracy drops quickly when the number of hosts increases. The deep convolutional neural networks (CNN) have outperformed many other cutting edge techniques in image classification recently. While image classification is a typical task for CNN, successful training depends on a large number of labeled training samples which is a challenge to this research in acquiring and annotating large number of images. Therefore, excessive data augmentation was used in this study with affine and elastic deformation to increase training images, and improve simulation effectiveness and learning invariance. Because laser scattering pattern is centered at the incident point, all deformation operations are referenced to the laser incident location as origin to avoid possible side effects. A five-layer convolutional neural network followed by two fully connected networks was constructed for layered feature extraction. The new classifier has clear performance improvement compared with previous classification results for 5 and more hosts with >90% accuracy at the subspecies level. The accuracy keeps 100% for different bacteria in the species level. Therefore, this new technique can be a potentially useful tool for not only stringent source tracking but also low cost, rapid, and accurate identification of different bacteria.

Keywords: Classification, Machine learning
Abstract: Abnormal lymphoid and blast cells are mononuclear cells (MNC) circulating in peripheral blood (PB) in malignant diseases such as lymphoid neoplasms and acute leukemias. An automated morphological detection of these cells would help as a support for the morphologic diagnosis. We present a system for the automatic recognition of four groups obtained from PB: 1) MNC 2) granulocytes 3) smudge cells and 4) platelets, using deep features from different convolutional neural networks (CNN) and classification with a quadratic support vector machine (SVM). We compiled a dataset of 22371 digital cell images from the routine workload of the Core Laboratory of the Hospital Clinic of Barcelona. PB films were stained automatically with May Grünwald-Giemsa and images were obtained using the CellaVision DM96 system and taken from a total of 404 patients. The dataset was organized in four groups: MNC (10982 images including lymphocytes, monocytes, reactive lymphocytes, abnormal lymphocytes, plasma cells, atypical promyelocytes, blast cells and erythroblasts), Granulocytes (7186 images including basophils, eosinophils, neutrophils, band cells, myelocytes, metamyelocytes, promyelocytes), Platelets (1962 images corresponding to giant platelets and aggregates) and Smudge cells (2241 images of isolated cell nuclei). We extracted a total of 4096 features from the images using six pre-trained CNN’s: Alexnet, vgg-s,m,f , vd-16 and vd-19 . Then, with these features we trained a quadratic SVM using Matlab with a 5-fold cross validation. Total accuracies in all cases were above 94%. In addition, the AUC ROC was close to 1 for all the analysis. The use of pre-trained CNN to extract deep features from PB images allowed obtaining very high accuracy percentages in their automatic classification without requiring previous segmentation of the regions of interest of the images. Results are encouraging towards the development of an automatic recognition system for a broader family of abnormal mononuclear cells as a diagnosis support tool.

Keywords: Optoacoustic/photoacoustic imaging, Brain
Abstract: We have developed a fast and compact OR-PAM system for the clinical applications. Thanks to a waterproof 2-axis MEMS scanner, the size of developed OR-PAM system can be minimized to only 30 × 90 × 30 mm3 along x, y, and z axes, respectively. Using the fast pulsed laser with a repetition rate of 50 kHz, volumetric imaging speed is increased up to 0.625 Hz Finally, we have successfully obtained the in vivo PA images of microvasculatures in mouse brain.

Keywords: Retinal imaging, Connectivity analysis, Dimensionality reduction
Abstract: We propose a novel, robust and fully automatic method for retrieving the connections among blood vessels in retinal images, which might be rotated, curved or interrupted due to several diseases, poor imaging conditions or segmentations. The main core of our solution is a five-dimensional connectivity kernel modelling the contextual connections in the primary visual cortex in lifted space of positions, orientations, curvatures and intensity. This kernel is used for creating an affinity matrix, which is further processed in a spectral clustering step. This method detects the challenging vessel connections perfectly.

Keywords: Magnetic resonance imaging (MRI), Brain, image filtering (e.g. mathematical morphology, wavelets,...)
Abstract: The automated detection of cerebral vessels is of great importance in understanding of the mechanism, diagnosis and treatment of many brain vascular pathologies. However, automatic vessel detection from 3D angiography continues to be an open issue. In this paper we introduce a novel 3D symmetry filter that has excellent performance on enhancing vessels in magnetic resonance angiography (MRA). The proposed filter not only takes into account of local phase features estimated by using a quadrature filter so as to distinguish between lines and the edges, but also uses the weighted geometric mean of the blurred and shifted responses of the quadrature filter, which allows more tolerance in the position of the respective contours. As a result this filter can produce a strong response to the vascular features despite variations in scale, contrast, and bifurcations in images. Our results demonstrate its superior performance to other state-of-the-art methods.

Keywords: Brain, Functional imaging (e.g. fMRI), fMRI analysis
Abstract: Traditional task-based fMRI activation detection methods, such as the general linear model (GLM), assume that the fMRI signals of activated brain regions follow the external stimulus paradigm. Typically, these activated regions are detected independently in a voxel-wise fashion, and the interaction among voxels is nevertheless neglected. Despite the wide use and remarkable success of GLM, the temporal and spatial relationships among activated regions remain unveiled. In response to this challenge, we present a novel method that combines two-stage sparse representation framework and the operator modulations (integral and derivative) to explore the temporal and spatial organizations underlying fMRI-derived activations in the brain. The two-stage sparse representation framework is designed to deal with big data and the functional operator is focused on finding the refined activation areas in the brain under task performances. Experiments demonstrated that diverse temporal and spatial organizations between activated regions exist and different functional operators may lead to different activation areas, thus significantly supplementing to the available principle of GLM that has been widely used in the human brain mapping field.

Keywords: Functional imaging (e.g. fMRI), Brain, fMRI analysis
Abstract: Resting-state functional magnetic resonance imaging (rs-fMRI) has become a powerful technique for analyzing cognitive function and its disruption in mental diseases, including schizophrenia. It has played an important role in analyzing and diagnosing mental diseases. In this paper, we present a novel machine learning approach called group-wise sparse representation of rs-fMRI signals to find differences between schizophrenia patients and healthy controls. Firstly, we extract the fMRI signals from all subjects that are registered into the MNI atlas space in the pre-processing step to build a large input signal matrix. Secondly, we use online dictionary learning and sparse coding methods to derive the coefficient matrix. Thirdly, we use two-sample t-test to analyze the regions of increased and decreased activity in schizophrenia patients compared to healthy controls. Finally, by using the AAL atlas the distributions of the detected interested regions are obtained. We test our approach on the COBRE dataset and the experimental results show that the schizophrenia patients have increased activities mainly at the fronto-parietal network. All other networks excluding the fronto-parietal network show decreased phenomenon. These results provide novel insights into better understanding of schizophrenia.

Keywords: Blind source separation & Dictionary learning, Classification, Functional imaging (e.g. fMRI)
Abstract: Task-based functional magnetic resonance imaging (tfMRI) is widely used to localize brain regions or networks in response to various cognitive tasks. However, given two groups of tfMRI data acquired under distinct task paradigms, it is not clear whether there exist intrinsic inter-group differences in signal composition patterns, and if so, whether these differences could be used for data discrimination. The major challenges originate from the high dimensionality and low signal-to-noise ratio of fMRI data. In this paper, we proposed a novel framework using hybrid temporal and spatial sparse representation to tackle above challenges. We applied the proposed framework to the Human Connectome Project (HCP) tfMRI data. Our experimental results demonstrated that the task types of fMRI data can be successfully classified, achieving a 100% classification accuracy. We also showed that both task-related components and resting state networks (RSNs) can be reliably identified. Our study provides a novel data-driven approach to detecting discriminative inter-group differences in fMRI data based on signal composition patterns, and thus potentially can be used to control fMRI data quality and to infer biomarkers in brain disorders.

Keywords: Connectivity analysis, Atlases, Classification
Abstract: Human brain connectomics is a rapidly evolving area of research, using various methods to define connections or interactions between pairs of regions. Here we evaluate how the choice of (1) regions of interest, (2) definitions of a connection, and (3) normalization of connection weights to total brain connectivity and region size, affect our calculation of the structural connectome. Sex differences in the structural connectome have been established previously. We study how choices in reconstruction of the connectome affect our ability to classify subjects by sex using a support vector machine (SVM) classifier. The use of cluster-based regions led to higher accuracy in sex classification, compared to atlas-based regions. Sex classification was more accurate when based on finer cortical partitions and when using dilations of regions of interest prior to computing brain networks.

Keywords: Diffusion weighted imaging, Brain
Abstract: A variety of parametric models based on the Apparent Diffusion Coefficient (ADC) have been proposed to describe signal decay in the interpretation of diffusion weighted Magnetic Resonance Imaging data. In this work, we investigate the robustness of several of these models, including the exponential model, the bi-exponential model and the recently proposed gamma distribution model, using a Cr'{a}mer-Rao lower bound analysis. It is demonstrated that the parameters of all models are susceptible to significant estimation errors at the current gradient strengths of clinical scanners. To complement the theoretical analysis, the robustness of the models is compared through application to experimental ovine optic nerve and rat brain MRI data. Our results indicate that gamma distribution model fits experimental data better than the commonly employed bi-exponential model.

Keywords: Connectivity analysis, Functional imaging (e.g. fMRI), Brain
Abstract: Functional and structural connectivity of the convey different information about the brain. The integration of this different approach is receiving growing attention from the research community, as it can shed new light on brain functions. In this context, this manuscript proposes a constrained autoregressive model with different lag-order allowing to generate an “effective” connectivity matrix that model the structural connectivity integrating the functional activity. Several order of lag are investigated to observe how different time points relate each other and influence structural connectivity. In practice, an initial structural connectivity representation is altered according to functional data, by minimizing the reconstruction error of an autoregressive model of different orders constrained by the structural prior. The model is further validated in a case-control experiment, which aims at differentiating healthy subject and young patients in the autism spectrum disorder.

Keywords: Blind source separation & Dictionary learning, Functional imaging (e.g. fMRI), Brain
Abstract: Equipped with selective auditory attention (SAA), people are able to rapidly shift their attention to auditory events of interest. Although abstract neuroimaging paradigms are fundamental for exploring the neural basis of SAA, whether those findings are valid in a more naturalistic condition and how the types of auditory stimuli affect SAA are largely unknown. Here we propose a brain decoding study to explore SAA using naturalistic auditory excerpts in three categories (pop music, classical music and speech) as stimuli for functional magnetic resonance imaging (fMRI). We adopted a computational auditory attention model to estimate attentional allocation for the excerpts. We then extracted brain activity features from fMRI data via sparse representation and used them to decode the auditory attention allocation. Our experimental results showed that the primary auditory cortex was commonly involved in the attentional processing of the three categories and the contribution of distinct brain networks to the decoding model in each group. Our study on the one hand provides novel insights into neural SAA in naturalistic experience, on the other hand shows the possibility of leveraging neuroimaging studies by integrating naturalistic stimuli and computational auditory information processing approaches.

Keywords: Computer-aided detection and diagnosis (CAD), Breast, X-ray imaging
Abstract: In computerized detection of clustered microcalcifications (MCs) from mammogram images the occurrence of false positives (FPs) varies greatly from case to case. In this work, we develop a probabilistic modeling approach to estimate the number of individual FPs present in a detected MC lesion. We describe the number of true positives (TPs) by a Poisson-Binomial probability distribution, wherein a logistic regression model is trained to determine the probability for a detected MC to be a TP based on its characteristics in the detector output. We demonstrate the proposed approach on a set of 188 full-field digital mammography images with two existing MC detectors [difference-of-Gaussians (DoG) and SVM detector]. The results show that the error level in the estimated number of FPs can be as low as 2.51 on average when the number of FPs is as high as 11.38 in a detected MC lesion.

Keywords: Computer-aided detection and diagnosis (CAD), Ultrasound, Bone
Abstract: Developmental dysplasia of the hip (DDH) is a congenital deformity occurring in ~3% of infants. If diagnosed early most cases of DDH can be effectively treated using a Pavlik harness. However, current diagnosis of DDH using 2D ultrasound is and can have high inter-operator variability. In this paper we propose a method to automatically segment the acetabulum bone and derive geometric indices of hip dysplasia from this model. In the proposed method, using multi-scale superpixels, we incorporate global and local image features into a Deep Learning framework to obtain a probability map of the bone to be segmented and then use this map in probabilistic graph search to guide the segmentation. Clinically relevant geometric measures of hip dysplasia, including a new index of acetabular rounding, are then automatically calculated from the segmented acetabulum contour. We tested this method on 2D ultrasound of 50 infant hips and the contours generated matched closely with manual segmentations at root mean square error 1.8±0.7 mm and Hausdorff distance 2.1±0.9 mm. In this pilot data, the measured indices of dysplasia give an area under the curve of 86.2% for classifying normal vs dysplastic hips. The proposed approach could be used clinically for accurate and automatic diagnosis of hip dysplasia in infants.

Keywords: Skin, Classification, Machine learning
Abstract: Similarity in appearance between various skin diseases, often makes it challenging for clinicians to identify the type of skin condition, and the accuracy is highly reliant on the level of expertise. There is also a great degree of subjectivity and inter/intra observer variability found in the clinical practices. In this paper, we propose a method for automatic skin diseases recognition that combines two different types of deep convolutional neural network features. We hold the hypothesis that it is equally important to capture global features such as color and lesion shape, as well as local features such as local patterns within the lesion area. The proposed method leverages deep residual network to represent global information, and bilinear pooling technique which allows to extract local features to differentiate between skin conditions with subtle visual differences in local regions. We have evaluated our proposed method on MoleMap dataset with 32,195 and ISBI-2016 challenge dataset with 1,279 skin images. Without any lesion localisation or segmentation, our proposed method has achieved state-of-the-art results on the large-scale MoleMap datasets with 15 various disease categories and multiple imaging modalities, and compares favorably with the best method on ISBI-2016 Melanoma challenge dataset.

Keywords: Classification, Brain, Magnetic resonance imaging (MRI)
Abstract: Neurodegenerative pathologies, such as Alzheimer's disease, are linked with morphological alterations and tissue variations in subcortical structures which can be assessed from medical imaging and biological data. In this work, we present an unsupervised framework for the classification of Alzheimer's disease (AD) patients, stratifying patients into four diagnostic groups, namely: AD, early Mild Cognitive Impairment (MCI), late MCI and normal controls by combining shape and voxel-based features from 12 sub-cortical areas. An automated anatomical labeling using an atlas-based segmentation approach is proposed to extract multiple regions of interest known to be linked with AD progression. We take advantage of gray-matter voxel-based intensity variations and structural alterations extracted with a spherical harmonics framework to learn the discriminative features between multiple diagnostic classes. The proposed method is validated on 600 patients from the ADNI database by training binary SVM classifiers of dimensionality reduced features, using both linear and RBF kernels. Results show near state-of-the-art approaches in classification accuracy (> 88%), especially for the more challenging discrimination tasks: AD vs. LMCI (76.81%), NC vs. EMCI (75.46%) and EMCI vs. LMCI (70.95%). By combining multimodality features, this pipeline demonstrates the potential by exploiting complementary features to improve cognitive assessment.

Keywords: Functional imaging (e.g. fMRI), Brain, Classification
Abstract: This paper introduces a comprehensive computer-aided diagnosis (CAD) system for autism diagnosis that integrates anatomical and functional information of the brain using both structural and functional magnetic resonance (MR) brain images. In order to move towards the idea of personalized medicine, analysis of the brain’s Brodmann areas (BAs) is conducted to reach a diagnosis decision on the local areas. This local analysis will help clinicians allocate the subject on the autism spectrum as well as correlate autism abnormalities to the areas responsible for certain skills, such as cognitive ones. The diagnosis is done through several analysis steps on both structural and functional MRI volumes. First, anatomical features are extracted from the cerebral cortex (Cx) and the cerebral white matter (CWM) of structural MR images. A monetary reward functional MR experiment is also held to get the areas of activation in the brains of the participants in response to the applied task. Next, all the extracted features are fed to a multi-level deep network for both local and global diagnosis. The CAD system has been evaluated using subjects from the NDAR database (23 – 210 months), achieving a global classification accuracy of 94.7% based on fusing both modalities. Moreover, brain maps are shown for different autistic subjects to indicate the strength of association of each BA with autism, which supports the idea of personalized medicine. The proposed CAD system, along with the idea of local feature extraction and diagnosis, holds the promise for being capable to resolve autism endophenotypes.

Keywords: Shape analysis, Probabilistic and statistical models & methods, Classification
Abstract: Statistical analysis of longitudinal data is a significant problem in Biomedical imaging applications. In the recent past, several researchers have developed mathematically rigorous methods based on differential geometry and statistics to tackle the problem of statistical analysis of longitudinal neuroimaging data. In this paper, we present a novel formulation of the longitudinal data analysis problem by identifying the structural changes over time (describing the trajectory of change) to a product Riemannian manifold endowed with a Riemannian metric and a probability measure. We present theoretical results showing that the maximum likelihood estimate of the mean and median of a Gaussian and Laplace distribution respectively on the product manifold yield the Fr'{e}chet mean and median respectively. We then present efficient recursive estimators for these intrinsic parameters and use it to classify MR brain scans (acquired from the publicly available OASIS database) of patients with and without dementia.

Keywords: Image segmentation, Computed tomography (CT), Machine learning
Abstract: Cancer is one of the leading causes of death worldwide. Radiotherapy is a standard treatment for this condition and the first step of the radiotherapy process is to identify the target volumes to be targeted and the healthy organs at risk (OAR) to be protected. Unlike previous methods for automatic segmentation of OAR that typically use local information and individually segment each OAR, in this paper, we propose a deep learning framework for the joint segmentation of OAR in CT images of the thorax, specifically the heart, esophagus, trachea and the aorta. Making use of Fully Convolutional Networks (FCN), we present several extensions that improve the performance, including a new architecture that allows to use low level features with high level information, effectively combining local and global information for improving the localization accuracy. Finally, by using Conditional Random Fields (specifically the CRF as Recurrent Neural Network model), we are able to account for relationships between the organs to further improve the segmentation results. Experiments demonstrate competitive performance on a dataset of 30 CT scans.

Keywords: Lung, Computer-aided detection and diagnosis (CAD), Machine learning
Abstract: Characterization of lung nodules as benign or malignant is one of the most important tasks in lung cancer diagnosis, staging and treatment planning. While the variation in the appearance of the nodules remains large, there is a need for a fast and robust computer aided system. In this work, we propose an end-to-end trainable multi-view deep Convolutional Neural Network (CNN) for nodule characterization. First, we use median intensity projection to obtain a 2D patch corresponding to each dimension. The three images are then concatenated to form a tensor, where the images serve as different channels of the input image. In order to increase the number of training samples, we perform data augmentation by scaling, rotating and adding noise to the input image. The trained network is used to extract features from the input image followed by a Gaussian Process (GP) regression to obtain the malignancy score. We also empirically establish the significance of different high level nodule attributes such as calcification, sphericity and others for malignancy determination. These attributes are found to be complementary to the deep multi-view CNN features and a significant improvement over other methods is obtained.

Keywords: EEG & MEG, Blind source separation & Dictionary learning, Brain
Abstract: Currently, the standard event-related potentials (ERP) technique consists in averaging many on-going electroencephalogram (EEG) trials using the same stimuli. Key questions are how to extract the ERP from on-going EEG with fewer average times and how to further decompose ERP into basic components related to cognitive process. In this paper we introduce a novel Blind Source Separation (BSS) approach based on a weak exclusion principle (WEP) to solve these problems. The superior aspect of this algorithm is that it is based on a deterministic principle, which is more appropriate to analyze non-stationary EEG signals than most other BSS methods based on statistical hypotheses. The results show that by using our BSS algorithm can quickly and effectively extract ERPs using fewer average times than traditional averaging method. We show that, via BSS, we can isolate two main ERP components, which are respectively related to an exogenous process and a cognitive process, and can discriminate between the occipital lobe and the frontal lobe responses from the brain, agreeing with the classical component modeling in ERPs. Single-trial ERP separation results have demonstrated the consistency of these two main ERP components. Thus, BSS based on WEP can provide a window to better understand ERP, not only averaging behavior, but the complexities of moment-to-moment dynamics as well.

Keywords: Histopathology imaging (e.g. whole slide imaging), Classification, Computer-aided detection and diagnosis (CAD)
Abstract: Automatic detection of lymphocytes could contribute to develop objective measures of the infiltration grade of tumors, which can be used by pathologists for improving the decision making and treatment planning processes. In this article, a simple framework to automatically detect lymphocytes on lung cancer images is presented. This approach starts by automatically segmenting nuclei using a watershed-based approach. Nuclei shape, texture, and color features are then used to classify each candidate nucleus as either lymphocyte or non-lymphocyte by a trained SVM classifier. Validation was carried out using a dataset containing 3420 annotated structures (lymphocytes and non-lymphocytes) from 13 1000x1000 pixels patches extracted from lung cancer whole slide images. A Deep Learning model was trained as baseline. Results show an F-score 30% higher with the presented framework than with the Deep Learning approach. The presented strategy is, in addition, more flexible, requires less computational power, and requires much lower training times.

Keywords: Computer-aided detection and diagnosis (CAD), Classification, Machine learning
Abstract: In this paper, we present a novel deep learning model termed Deep Autoencoding-Classification Network (DACN) for HEp-2 cell classification. The DACN consists of an autoencoder and a normal classification convolutional neural network (CNN) with the two sharing the same encoding pipeline. The DACN model is jointly optimized for the classification error and the image reconstruction error based on a multi-task learning procedure. We evaluate the proposed model using the publicly available ICPR2012 benchmark dataset. We will show that this architecture is particularly effective when the training dataset is small which is often the case in medical imaging applications. We will present experimental results to show that the proposed approach outperforms all known state of the art HEp-2 cell classification methods.

Keywords: Machine learning, Histopathology imaging (e.g. whole slide imaging), Breast
Abstract: Histopathology forms the gold standard for confirmed diagnosis of a suspicious hyperplasia being benign or malignant and for its sub-typing. While techniques like whole-slide imaging have enabled computer assisted analysis for exhaustive reporting of the tissue section, it has also given rise to the big-data deluge and the time complexity associated with processing GBs of image data acquired over multiple magnifications. Since preliminary screening of a slide into benign or malignant carried out on the fly during the digitization process can reduce a Pathologist's work load, to devote more time for detailed analysis, slide screening has to be performed on the fly with high sensitivity. We propose a deep convolutional neural network (CNN) based solution, where we analyse images from random number of regions of the tissue section at multiple magnifications without any necessity of view correspondence across magnifications. Further a majority voting based approach is used for slide level diagnosis, i.e., the class posteriori estimate of each views at a particular magnification is obtained from the magnification specific CNN, and subsequently posteriori estimate across random multi-views at multi-magnification are voting filtered to provide a slide level diagnosis. We have experimentally evaluated performance using a patient level 5-folded cross-validation with 58 malignant and 24 benign cases of breast tumors to obtain average accuracy of 94.67%, sensitivity of 96.00%, specificity of 92.00% and F-score of 96.24% while processing each view in approx. 10 ms.

Keywords: Prostate, Multi- and Hyper-spectral imaging, Image segmentation
Abstract: Detection and segmentation of glandular structures are important, since these structures contain clinical relevant information regarding the disease status and Gleason grade of prostate cancer. Manual gland segmentation process is very time consuming and subjective, also existing automated methods are not robust and reliable. We set out to design an automated, fast and objective method. In this paper we present an automated methodology for automated detection of structures of interest in digitized histopathology images of a Tissue Micro Array (TMA).We show a successful method for detection of prostate glandular structures and its nuclei. Our method integrates different techniques: (1) construct hyperspectral transmission images using sixteen light wavelengths, (2) use Principal Component Analysis (PCA) to construct new RGB images, (3) use clustering to segment different structures in an unsupervised fashion, and (4) apply post-processing morphological cleaning as the final step in our pipeline. We detected 80% plus of the glandular structure in 61% of cores , 80% -50% of the glands in 15% of cores and less than 50% of the glands in 24% of cores.

Keywords: Virtual/augmented reality, Angiographic imaging, Vessels
Abstract: Biplane X-ray angiography is currently the gold standard for navigational guidance during percutaneous interventions in vascular structures; but it remains limited to 2D projections. In this study, we propose a novel graph-based voxel coloring method for 3D reconstruction of vascular structures from biplane X-ray angiography sequences. The reconstruction is obtained by using the random walks algorithm on a graph-based representation of a discretized visual hull, to obtain the probability of belonging to the vascular structure. A multi-scale scheme is introduced to reconstruct at a finer level, while being computationally efficient. The proposed method was validated using the XCAT motion simulator and on calibrated clinical data.

Keywords: Radiation therapy, planing and treatment, Machine learning, Shape analysis
Abstract: In radiation therapy, tumor tracking allows to adjust the beam such that it follows the respiration-induced tumor motion. However, most clinical approaches rely on implanted fiducial markers to locate the tumor and, thus, only provide sparse information. Motion models have been investigated to estimate dense internal displacement fields from an external surrogate signal, such as range imaging. With increasing surrogate complexity, we propose a respiratory motion estimation framework based on kernel ridge regression to cope with high-dimensional domains. This approach was validated on five patient datasets, consisting of a planning 4DCT and a follow-up 4DCT for each patient. Mean residual error was at best 2.73 +- 0.25 mm, but varied greatly between patients.

Keywords: Contrast agent quantification, Histopathology imaging (e.g. whole slide imaging), Computed tomography (CT)
Abstract: Biological tissues have a complex 3D structure, and thus 2D measurements like histomorphometry only partially reveal the full 3D structure. As a solution, we present a novel contrast agent for contrast-enhanced microfocus computed tomography (CE-CT) allowing to simultaneously visualize in 3D the bone and bone marrow vascularization and adiposity in murine long bones. We validated the method by detailed comparison with histology. Furthermore, we quantified the 3D structure of the different tissues, as well as their spatial correlation, in different mouse models (i.e. ageing and type 2 diabetes), showing clear differences in tissue structure and composition between the groups, thus highlighting the added value of CE-CT compared to histomorphometry.

Keywords: Surgical guidance/navigation, Medical robotics
Abstract: Registration is important in surgical navigation system (IGS), which aligns the view coordinates of pre-scanned computed tomographic (CT) or magnetic resonance (MR) images to real world coordinates of the patient in operating room. The most widely used landmark registration method aligns two coordinate systems by finding pairs of fiducial markers which may case large errors because only a few points are taken into account. The z-touchTM (BrainLab, Heimstetten, Germany) uses laser beam to capture multiple points for registration while requiring patients exposed to a special tracker during the process. In this study, a hand-hold 3D scanner is employed to increase registration accuracy and comfort of patients, as well as simplify workflow of surgeons. The mean deviation of this method is about 1.032 mm with variance 1.413.

Keywords: Diffusion weighted imaging, Functional imaging (e.g. fMRI), Tractography
Abstract: Diffusion weighted MRI (DWI) fiber tractography is the primary tool for mapping neuronal fiber tracts in vivo. Recently, growing evidence demonstrates that functional MRI (fMRI) signals can be reliable detected in white matter. With this in hand, we introduced the concept of a spatio-temporal correlation tensor (STCT) to describe the functional architecture of white matter [1]. In this work, we proposed a novel algorithm for tractography by encapsulating DWI and fMRI information and the fiber orientation density function (ODF) was estimated using Bayesian method. The proposed method improves accuracy of tractography and could particularly benefits tracking neuronal circuits with specific functional activities.

Keywords: Computed tomography (CT), Image segmentation, Computer-aided detection and diagnosis (CAD)
Abstract: We propose a method to improve lesion segmentation accuracy in breast CT images. We used histogram equalization and exponential function as pre-processing. We validated the effectiveness of the proposed method using simulated breast CT images. Our results show the proposed method has better lesion segmentation performance than conventional method.

Keywords: Ultrasound
Abstract: Ultrasound Nakagami imaging is a useful technique to estimate variations in envelope statistics of fatty liver. Nowadays there are several estimators of Nakagami parameters to be developed. Therefore, this study compared the performances of Nakagami imaging based on different estimators in assessing the severity of fatty liver.

Keywords: Functional imaging (e.g. fMRI), Brain, Surgical guidance/navigation
Abstract: Stimulation of the thalamus is an emerging treatment for patients with refractory epilepsy, including Lennox-Gastaut syndrome (LGS). However, the optimal thalamic targets remain uncertain. Furthermore, functional organization of thalamic nuclei, and their cortical projections, is poorly understood in LGS. Here we used task-free fMRI to compare functional connectivity between thalamic sub-regions and cortical networks in LGS patients and healthy controls. We hypothesized that patients would show abnormal functional connectivity in specific thalamocortical circuits.

Keywords: Nuclear imaging (e.g. PET, SPECT), Motion compensation and analysis, Shape analysis
Abstract: The ability to image freely moving rodents using positron emission tomography presents many exciting possibilities for exploring links between brain function and behaviour. A key requirement for this approach is obtaining accurate estimates of animal pose throughout a scan. The goal of this project is to develop a markerless shape-from-silhouette approach which requires only a limited number of camera views by exploiting an apriori animal model. Our specific aim here was to compare techniques for extracting silhouettes from image streams of a freely moving rat and develop our calibrated multi-camera camera setup for motion estimation. Four segmentation methods were studied for silhouette extraction from a multi-camera setup: thresholding, region-growing, random walk and k-means clustering. Of these methods, thresholding and region-growing were the simplest to implement but performed poorly in uneven lighting. The random walk and k-means clustering approaches produced excellent segmentations and adapted well to changing light. We slightly favoured the random walk approach despite its manual initialisation because of much better efficiency. Our multi-camera setup comprised four monochrome 1280x1024 fast frame rate cameras, spatially calibrated to one another (accuracy 0.15 pixels +/- 0.09 pixels) and synchronised and cross-calibrated to an independent marker-based optical tracking system used for ground truth. Cross-calibration accuracy was 0.9, 1.2 and 0.6 mm in the x, y and z directions, respectively. The system allowed a space of 250 mm x 250 mm for a rat to maneuver while being fully inside the field-of-view of all cameras. Using a taxidermal rat specimen we obtained test motion sequences using our multi-camera setup in conjunction with marker-based tracking as a reference and then generated a space-carved reconstruction of the specimen to validate it. The next stage of this work is to incorporate the apriori model of the rat so that we can extract pose estimates using a 3D-2D registration between the model and our silhouettes. Initially this will be done initially a rigid-bodyvbut is easily extensible to an articulated (non-rigid) head/body model.

Keywords: Quantification and estimation, Spine, Computed tomography (CT)
Abstract: Upon aging, spine can develop different extent of degeneration causing osteoporosis or bone fracture. Traditional evaluation on spine was to evaluate the bone mineral density and wedging effect of vertebral body along different vertebrae by measuring average bone intensity and vertebral body heights in Computed Tomography images respectively. This project aimed to investigate morphology changes of vertebral bodies in thoracic and lumbar region of the normal human population from adult to elderly. Results showed that the bone mineral density and height has decreased along age. Further researches could be done on analysis of diseased spine with healthy ones.

Keywords: Image segmentation
Abstract: Healthy food consumption has become increasingly important around the world as people attempt to stay slim and avoid the dangers associated with obesity. Therefore, a means of measuring the calorie content of a person’s daily food intake represents a significant aid to the process of watching one’s weight. In this paper, we proposed the thermal imaging system to calculate the correct food volume, especially for Thai Food Image as oppose to some recent approaches that calculate based on the table of calorie. We employed the SLIC Super Pixels technique, and DBSCAN clustering algorithm for the food image segmentation. The Localized Multi-texton Histogram (LMTS) and RGB histogram are implemented for shape, texture, and colour feature extraction. Finally, the Weighted Fuzzy C-Means is adopted to classify Thai Food Images.

Keywords: Image reconstruction - analytical & iterative methods, Vessels
Abstract: In this work, we lay the foundation from which multiscale atlases characterizing cerebrovasculature structural variation across the entire brain of small animal models can be constructed from Knife-Edge Scanning Microscopy datasets. Through the geometric reconstruction of the vascular filaments embedded in the volumetric imaging dataset, the capability to distinguish cerebral vessels based on diameter and other morphological properties across the whole-brain is provided. This results in a means to study local variations in small vessel morphometry which has a profound effect on surrounding neuronal composition in different cerebral regions, as well as the robust and fragile aspects of the cerebrovasculature system across the larger vessels.

Keywords: Single cell & molecule detection, Machine learning
Abstract: Collective migration describes the coordinated movement of larger groups of individuals and can usually be observed at high densities in confined systems. Quantitative image analyses are essential to reveal fundamental principles and correlations underlying the coordinated movement and observed pattern formation. In our project, we investigate the temporal motion of cells in 2D/3D to study highly directed interactions and unique motility features on single cell level. To extract unbiased information describing these highly complex systems, we established a custom designed image analysis pipeline, which covers the automated segmentation, tracking and analysis of large time-resolved in-vitro fluorescence microscopy image sequences. The pipeline allows characterizing the highly dynamic behavior of single cells within a very complex system. Image segmentation is conducted using the open source software tool ‘ilastik’ based on semi-supervised machine learning procedures [C. Sommer et. al., ISBI Pro. 2011]. We train the pixel classifier using a random forest by providing sparse data annotations. The generated probability map is transformed into a final segmentation, which we use for the following object tracking. Trajectories of cells are reconstructed finding the most probable configuration of a tracking-by-assignment model as in ‘ilastik’ [M. Schiegg et al., ICCV Pro. 2013]. Afterwards, cells are subsequently analyzed regarding well-defined biophysical features and quantities by applying our custom developed bioinformatics tool. Furthermore, advanced characteristics such as next neighbor behavior is extracted. Moreover, cells are affiliated to characteristic cell accumulations by applying an unsupervised clustering algorithm. Additionally, we address errors during our analysis by using custom designed filter options to correct for bias in the analysis. By changing various parameters of the initial experiments, we are able to reveal unbiased critical parameters necessary for the appearance of specific pattern formation and maintaining the integrity of the collective. This will improve our understanding of the underlying mechanism and dynamics of cell behavior, which will give new insights into e.g. the complex healing process of skin wounds or the mechanism underlying malaria parasite movement for revealing potential weak spots to target in drug development.

Keywords: Magnetic resonance imaging (MRI), Tracking (time series analysis), Image segmentation
Abstract: An active-shape-model and registration based method is used to segment the knee bones and cartilages from six magnetic resonance (MR) images of the knee joint with increasing degrees of flexion. Cartilage contact areas are estimated using proximity algorithms. Validation is performed against manual segmentations performed for 10 image series.

Keywords: X-ray imaging, Image segmentation, In-vivo cellular and molecular imaging
Abstract: Synchrotron radiation (SR) is one of the important technologies in medical imaging. In order to use data from SR X-ray as a medical image, the problem of segmentation must be considered in digital image processing. Segmentation is also an essential step in the generation of accurate information on three-dimensional reconstruction. The purpose of this study was to find an algorithm suitable for segmentation in SR X-ray images.

Keywords: Quantification and estimation, Lung, Computed tomography (CT)
Abstract: To evaluate the effect of image dimension of region of interest (ROI) in developing convolutional neural network (CNN) based automatic classification system of regional patterns on volumetric chest CT with diffuse interstitial lung disease (DILD).

I.MATERIALS AND METHODS From 764 volumetric CT images of DILD, a thoracic radiologist selected cubic 3D region of interest (ROI) and labeled each ROI with following regional patterns; 106 ROIs with normal pattern, 135 ROIs with ground-glass opacity, 160 ROIs with reticular opacity, 123 ROIs with honeycombing, 137 ROIs with emphysema, and 103 ROIs with consolidation. Because each 3D volume has different pixel spacing and size, we resampled various sizes to 30x30x30 images with same size using tri-cubic interpolation. Resampled data were randomly split into 80% for training, 10% for validation, and 10% for test. To augment the data set, five 20x20x20 ROI were generated from each 30x30x30 ROI. Finally, we made axial (2D) images, axial/sagittal/coronal reconstructed (2.5D) images, and three-dimensional (3D) images respectively. We constructed relatively simple CNN architectures having four-convolutional layer, two-pooling layer, and two-dense layer. The basic architectures are totally same but applied with different dimension of convolution filter─totally 3x3 convolution filter for 2D image data, 3x3x3 only first convolution filter for 2.5D image data, and totally 3x3x3 convolution filter for 3D image data. The accuracy in classifying each regional disease patterns were evaluated and compared.

II.RESULTS Each method is compared with test set at the lowest loss point for validation set. The results show that using 3-dimentional ROD data and 2.5D have similar accuracy of 83% and these of two results have higher accuracy than 2D (80%) in classifying regional patterns of DILD. We assume accuracy of 2D was the lowest due to reduced information, on the contrary, accuracy of 3D and 3x2D having much more information of patterns was the highest.

III.CONCLUSION Using 3-dimentional image data in developing automatic system to classify regional disease pattern of DILD CT based on CNN seems to be more desirable than using axial images, if available. Further study to optimize the network using augmentation methods based on 3D and 3x2D is awaited.

Keywords: Magnetic resonance imaging (MRI), Brain, Probabilistic and statistical models & methods
Abstract: We aimed to determine required participant counts for a longitudinal rehabilitative study of children with unilateral cerebral palsy that aims to measure changes in either cortical thickness or tractography-derived diffusion metrics of the more-affected corticospinal tract. Results highlighted the importance of scan preparation and the utilization of anatomically specific measures of change.

Keywords: Nuclear imaging (e.g. PET, SPECT), Motion compensation and analysis, Brain
Abstract: Patient motion during neurological PET corrupts images and in the worst case may require repeat imaging, which is seldom feasible. Such events only occur in a subset of scans, and it may be unclear ahead of time which patients will be at risk. Thus, it is convenient to have a means of retrospectively applying motion correction. A methodology for extracting a signal representing motion in a number of dimensions, applicable to both periodic and aperiodic motion, is introduced. Results of a phantom study demonstrate its utility.

Keywords: Brain, fMRI analysis, Connectivity analysis
Abstract: In this abstract, we hypothesize that dynamic energetic cost of the healthy brain and its dynamic local functional connectivity are tightly related, and therefore neuroimaging of local connectivity could provide a convenient marker of time-varying metabolism in the human brain.

Keywords: fMRI analysis, Image enhancement/restoration(noise and artifact reduction), Functional imaging (e.g. fMRI)
Abstract: We empirically verified on real “near-null” (resting-state) fMRI data that the fully-automated noise classification algorithm SOCK, used as a noise filter, does not artificially inflate Z-scores of block design fMRI studies.

Keywords: Computed tomography (CT), Tooth, Image acquisition
Abstract: We took spectral CT images of a molar tooth with a micro-CT equipped with a CdTe photon counting x-ray detector. The spectral CT images show different contrasts at different energy bands demonstrating the potential of a spectral dental CT for future clinical application.

Keywords: Functional imaging (e.g. fMRI), Brain, Blind source separation & Dictionary learning
Abstract: On the assumption that some a priori information can be used, a multi-objective optimization based temporal ICA with reference (MOPtICA-R) method is proposed in this paper. The experimental results have demonstrated that the proposed MOPtICA-R method not only can greatly improves the stability of temporal ICA, but also can improves the ability of functional connectivity detection compared with the classical ICA methods.

Keywords: Optical coherence tomography, Eye, Modeling - Image formation
Abstract: Choroid neovascularization (CNV) usually causes varying degrees of retinal degradation and visual loss, most of them are permanent. For CNV prediction, a framework basing a noninvasive image-based approach is proposed in this paper.The method consists of four steps: pre-processing, meshing, CNV growth modeling and optimization. The modeling part, which combining biomechanical model and FEM-based model , with mass effect on three patient data set, the recall ratio, error ratio and dice coefficient are: 77.1±3.2%, 16.4±4.3% and 77.2±2.9% respectively.

Keywords: Image segmentation, Other-modality, Skin
Abstract: Conventional image processing techniques applied to diagnose malignant melanoma in dermoscopic images provide unsatisfactory results due to hair occlusion of the lesion region. An automatic algorithm based on Hessian filtering of hair in the lesion is presented to improve automatic hair removal.

Keywords: Optimization method, Computational Imaging, Image registration
Abstract: While segmentation aims to partition a given image into significant constituents and identify structures such as homogeneous regions or edges in order to quantify information, registration, given two images called Template and Reference consists of determining an optimal deformation φ that maps the structures visible in the Template into the corresponding ones in the Reference. This latter technique is encountered in the domain of shape averaging, multi-modality fusion to facilitate diagnosis and treatment planning, progression disease evaluation, shape tracking, etc.. The sought deformation φ is seen as the optimal solution of a specifically designed cost function, the problem being mathematically hard to solve due to its ill-posedness, to the involved non-linearity, to its non-convexity and to its versatile formulation. As structure/salient component/shape/geometrical feature matching and intensity distribution comparison rule registration, it sounds relevant to intertwine the segmentation and registration tasks into a single framework: accurate segmented structures allow to drive the registration process correctly providing then a reliable deformation field between the encoded structures, whilst the difficulty of weak edge segmentation can be overcome thanks to registration. The scope of the proposed work is thus to provide a novel unified variational model for joint segmentation and registration (and its thorough theoretical study) in which the shapes to be matched are viewed as hyperelastic materials, and more precisely, as Saint Venant-Kirchhoff ones. The dissimilarity measure relies on weighted total variation and nonlocal shape descriptor inspired by the piecewise constant Mumford-Shah model for segmentation. Theoretical results among which relaxation, existence of minimizers, analysis of two numerical methods of resolution (a first one based on the dual formulation of weighted total variation, on the introduction of auxiliary variables and on L2/L1 penalizations yielding asymptotic results, and a second one based on the approximation of the weighted total variation by a sequence of integral operators involving a differential quotient and a suitable sequence of radial mollifiers, and on a splitting approach leading again to Γ-convergence results), asymptotic results and Γ-convergence properties are provided.

Keywords: Ultrasound, Computer-aided detection and diagnosis (CAD), Heart
Abstract: Evaluation of regional systolic function based on qualitative or semi-quantitative visual assessment of myocardial thickening is useful in the detection of cardiac diseases. In particular, analysis of wall motion abnormality is an established method for detecting myocardial ischemia. The primary aim of this study was to investigate whether the enhancement of 2D+t echocardiography datasets using a hybrid approach of local phase-based feature asymmetry (FA), oriented feature symmetry (OFS), as well as Eulerian Motion Magnification (EMM) video processing techniques, improved visual assessment of wall motion abnormality. Classification accuracy by respondents (n=48) improved from 70.8% on the non-enhanced B-mode to 73.4% with the enhanced images.

Keywords: Machine learning, Single cell & molecule detection, Histopathology imaging (e.g. whole slide imaging)
Abstract: Automated cell detection is a critical step for a number of computer-assisted pathology related image analysis algorithm. However, automated cell detection is complicated due to the variable cytomorphological and histological factors associated with each cell. In order to efficiently resolve the challenge of automated cell detection, deep learning strategies are widely applied and have recently been shown to be successful in histopathological images. In this paper, we concentrate on bone marrow trephine biopsy images and propose a hybrid deep autoencoder (HDA) network with Curvature Gaussian model for efficient and precise bone marrow hematopoietic stem cell detection via related high-level feature correspondence. The accuracy of our proposed method is up to 94%, outperforming other supervised and unsupervised detection approaches.

Keywords: Computer-aided detection and diagnosis (CAD), Endoscopy, Machine learning
Abstract: This work addresses Transfer Learning via Convolutional Neural Networks (CNN's) for the automated classification of colonic polyps in eight HD-endoscopic image databases acquired using different modalities. For this purpose, we explore if the architecture, the training approach, the number of classes, the number of images as well as the nature of the images in the training phase can influence the results. The experiments show that when the number of classes and the nature of the images are similar to the target database, the results are improved. Also, the better results obtained by the transfer learning compared to the most used features in the literature suggest that features learned by CNN's can be highly relevant for automated classification of colonic polyps.

Keywords: Machine learning, Image segmentation, Heart
Abstract: A significant amount of research has been done on the segmentation of coronary arteries. However, the resulting automated boundary delineation is still not suitable for clinical utilization. The convolutional neural network was driving advances in the medical image processing. We propose the brief convolutional network (BCN) that automatically produces the labeled mask with the luminal and wall boundaries of the coronary artery. We utilized 50 patients of CCTA – intravascular ultrasound matched image data sets. Training and testing were performed on 40 and 10 patient data sets, respectively. The prediction of luminal and wall mask was performed using stacked BCN on the each image view: axial, coronal, and sagittal of straightened curved planar reformation. We defined the vector that includes probability from BCN result on each image view and proposed amplified probability. We used an Adaptive Boost regressor with an extremely randomized tree regressor to determine the label for unknown probability vector.

Keywords: Classification, Machine learning
Abstract: Modality corresponding to medical images is a vital filter in medical image retrieval systems. This article presents the classification of modalities of medical images based on the usage of principles of hyper-dimensional computing and reservoir computing. It is demonstrated that the highest classification accuracy of the proposed method is on a par with the best classical method for the given dataset (83% vs. 84%). The major positive property of the proposed method is that it does not require any optimization routine during the training phase and naturally allows for incremental learning upon the availability of new training data.

Keywords: Endoscopy, Gastrointestinal tract, Classification
Abstract: In wireless capsule endoscopy (WCE), a swallowable miniature optical endoscope is used to transmit color images of the gastrointestinal tract. However, the number of images transmitted is large, taking a significant amount of the medical expert's time to review the scan. In this paper, we propose a technique to automate the abnormality detection in WCE images. We split the image into several patches and extract features pertaining to each block using a convolutional neural network (CNN) to increase their generality while overcoming the drawbacks of manually crafted features. We intend to exploit the importance of color information for the task. Experiments are performed to determine the optimal color space components for feature extraction and classifier design. We obtained an area under receiver-operating-characteristic (ROC) curve of approximately 0.8 on a dataset containing multiple abnormalities.

Keywords: Ultrasound, Heart, Machine learning
Abstract: Medical ultrasound is a ubiquitous, non-invasive, and relatively inexpensive technology used for a wide array of diagnostic tasks. In the case of 2D handheld ultrasound, the positioning of the probe has a direct impact on the diagnostic relevance of the acquired images; shifts of as little as a few millimeters can render the images unusable. We present a method which interpolates the predictions of a deep convolutional neural network classifier to estimate the viewpoint of the imaging probe directly from the visual data. For the discrete version of echocardiogram view classification, our method outperforms recent approaches on real-world data.

Keywords: Image segmentation, Heart, Magnetic resonance imaging (MRI)
Abstract: Detection of infarcted myocardium in the left ventricle is achieved with delayed enhancement magnetic resonance imaging (DE-MRI). However, manual segmentation is tedious and prone to variability. We studied three texture analysis methods (run-length matrix, co-occurrence matrix, and autoregressive model) in combination with histogram features to characterize the infarcted myocardium. We evaluated 10 patients with chronic infarction to select the most discriminative features and to train a support vector machine (SVM) classifier. The classifier model was then used to segment five human hearts from the STACOM DE-MRI challenge at ICCAI 2012. The Dice coefficient was used to compare the segmentation results with the ground truth available in the STACOM dataset. Segmentation using texture features provided good results with an overall Dice coefficient of 0.71 ± 0.12 (mean ± standard deviation).

Keywords: Image segmentation, Magnetic resonance imaging (MRI), Tissue
Abstract: Accurate segmentation of hip joint cartilage from magnetic resonance (MR) images provides a basis for obtaining morphometric data of articular cartilages for investigation of pathoanatomical conditions such as osteoarthritis. In this paper, we present an automated MR-based cartilage segmentation method using an ensemble of neural networks for the individual femoral and acetabular cartilage plates of the hip joint. The segmentation is performed in two stages with different image resolution levels for segmentation of the combined hip cartilage and separation of the individual cartilage plates, respectively. Neural networks used in both stages are trained in an over-complete manner using 20 training MR images with manual labeled images. Compared with expert manual segmentations, the automated method achieved mean Dice’s similarity coefficients of 0.805, 0.766 and 0.712 for segmentation of the combined, femoral and acetabular cartilage volumes in a set of 26 testing MR images.

Keywords: Deconvolution, Brain, Nuclear imaging (e.g. PET, SPECT)
Abstract: The partial volume effect (PVE) due to the low resolution of SPECT in brain SPECT volumes can be modeled as a convolution of a three-dimensional point-spread function (PSF) with the underlying true radioactivity. In this paper, a deconvolution guided by the edge locations in the geometric transfer matrix (GTM) method as a weighted regularization, denoted as RGTM, was proposed to take into account both the discrepancy from the convolution and the regional-homogeneity prior information in the correction of the PVE (PVC). Two steps were conducted: GTM and then a weighted regularization. Twenty digital phantom simulations were made to compare the performance of RGTM with those of Van-Cittert deconvolution (VC), GTM, and the region-based voxel-wise correction (RBV). Clinical data from eighty-four healthy adults with 99mTc-TRODAT-1 SPECT and MRI scans were also tested. Because the proposed RGTM was good in both constant and non-constant ROIs, its robustness is better than other methods if the distribution of the underlying radioactivity is not known exactly.

Keywords: Retinal imaging, Eye, Machine learning
Abstract: Diabetic Retinopathy (DR) is one of the leading causes of blindness worldwide. Detecting DR and grading its severity is essential for disease treatment. Convolutional neural networks (CNNs) have achieved state-of-the-art performance in many different visual classification tasks. In this paper, we propose to combine CNNs with dictionary based approaches, which incorporates pathology specific image representation into the learning framework, for improved DR severity classification. Specifically, we construct discriminative and generative pathology histograms and combine them with feature representations extracted from fully connected CNN layers. Our experimental results indicates that the proposed method shows improvement in quadratic kappa score ( kappa^2 = 0.86) compared to the state-of-the-art CNN based method (kappa^2 = 0.81).

Keywords: Retinal imaging, Eye, Machine learning
Abstract: The fovea is one of the most important anatomical landmarks in the eye and its localization is required in automated analysis of retinal diseases due to its role in sharp central vision. In this paper, we propose a two-stage deep learning framework for accurate segmentation of the fovea in retinal colour fundus images. In the first stage, coarse segmentation is performed to localize the fovea in the fundus image. The location information from the first stage is then used to perform fine-grained segmentation of the fovea region in the second stage. The proposed method performs end-to-end pixelwise segmentation by creating a deep learning model based on fully convolutional neural networks, which does not require the prior knowledge of the location of other retinal structures such as optic disc (OD) and vasculature geometry. We demonstrate the effectiveness of our method on a data-set with 400 retinal images with average localization error of 14 pixels.

Keywords: Probabilistic and statistical models & methods, Tracking (time series analysis), In-vivo cellular and molecular imaging
Abstract: When working with real data, underlying parameters such as the detection or clutter rates are generally unknown and possibly varying over time, however the right parametrisation is crucial to extract proper statistics about the monitored objects. In this article, a single cluster Probability Hypothesis Density (PHD) filter is used to jointly estimate the location and number of a set of objects and the clutter rate. The algorithm is verified on a simulated scenario designed to emulate the challenging nature of Single-Molecule Localisation Microscopy (SMLM) imaging sequences and demonstrated on a similar scenario with real data.

Keywords: Microscopy - Light, Confocal, Fluorescence, Tracking (time series analysis), Cells & molecules
Abstract: In this work we set several mathematical tools to study the role of Dynamin and Endophilin in the Clathrin-mediated endocytosis process. Their different dynamics and co-localizations are observed by using TIRF microscopy. We define in particular a novel tracking method in order to track the Clathrin-coated pits and quantify their co-localization with the proteins involved in the process. Thereafter, we perform a statistical analysis of the lifetimes of Clathrin-coated pits tracks by a suitable mixture model. This allows to define different populations of interest and point out the role of Endophilin and Dynamin in the process.

Keywords: Classification, Thyroid, Ultrasound
Abstract: Ultrasonography is a valuable diagnosis method for thyroid nodules. Automatically discriminating benign and malignant nodules in the ultrasound images can provide aided diagnosis suggestions, or increase the diagnosis accuracy when lack of experts. The core problem in this issue is how to capture appropriate features for this specific task. Here, we propose a feature extraction and selection method for ultrasound images based on the convolution neural networks (CNNs), try to introduce more meaningful and specific features to the classification. A CNN model trained with ImageNet data is transferred to the ultrasound image domain, to generate semantic deep features under small sample condition. Then, we combine those deep features with conventional features such as Histogram of Oriented Gradient (HOG) and Scale Invariant Feature Transform (SIFT) together to form a hybrid feature space. Furthermore, to make the general deep features more pertinent to our problem, a feature subset selection process is employed for the hybrid nodule classification, followed by a detailed discussion on the influence of feature number and feature composition method. Experimental results on 1037 images show that the accuracy of our proposed method is 0.929, which outperforms other relative methods by over 10%.

Keywords: Quantification and estimation, Pattern recognition and classification, Probabilistic and statistical models & methods
Abstract: Estimation of hemoglobin is important to diagnose anaemia which is a grave public health problem in developing and in other less developed countries. Hemoglobin, which normally is present within red blood cells, is the compound responsible for coloring blood red. Therefore redness of blood and consequently of skin, is a measure of hemoglobin concentration in blood. We utilize redness of palm to estimate hemoglobin non-invasively. We propose a machine vision based portable, user-friendly, non-invasive and cost effective approach to measure hemoglobin exploiting the redness measure of skin of palm. A camera captures the video of a palm of a human subject before and after the blood flow is restricted to the palm using a sphygmomanometer cuff in forearm close to the wrist. The video continues till the blood flow is released after sudden and rapid release of pressure in cuff. Measuring the redness of skin color after occlusion and after resumption of blood flow to palm, we propose a regression based classifier to predict the hemoglobin content of the blood. Through a human subject study, we show that our approach can estimate hemoglobin content up to an accuracy of 91%.

Keywords: Functional imaging (e.g. fMRI), Inverse methods, Image reconstruction - analytical & iterative methods
Abstract: Surface-based approaches have proven particularly relevant and reliable to study cortical functional magnetic resonance imaging (fMRI) data. However projecting fMRI volumes onto the cortical surface remains a challenging problem. Very few methods have been proposed to solve it and most of them rely on a simple interpolation. We propose here an original surface-based method based on a model representing the relationship between cortical activity and fMRI images, and a resolution through an inverse problem. This approach shows interesting perspectives for fMRI data processing as it is highly robust to noise and offers a good accuracy in terms of activations localization.

Keywords: Ultrasound, Image acquisition, Image reconstruction - analytical & iterative methods
Abstract: In this paper, a new method inspired by the synthetic aperture approach is proposed that aims at reducing the system0s complexity (only 8 or 16 active elements) without compromising the image quality, and at frame rates comparable to or higher than conventional focused linear array technique. The novel method has been investigated in simulations using Field II software and experiments performed on a wire phantom using an ultrasound scanner. Results show that the proposed method provides better Lateral Resolution (LR) to that obtained when conventional focused linear array technique is used. The estimated LR at the focal point was 1.09 mm and 0.29 mm for conventional and the proposed method, respectively,in simulations. These were estimated to be 1.03 mm and 0.38 mm, respectively, in experiments. The image quality is shown to improve further when sign coherence factor weighting is incorporated during beamforming.

Keywords: X-ray imaging, Kidney, Image reconstruction - analytical & iterative methods
Abstract: Kidney stones is a renal disease with high prevalence and one of the major reasons for emergency room visits. The prevalence of kidney stones is increasing, and the lifetime recurrence rate is estimated as almost 50 %. Thus, treatment of kidney stones becomes an increasingly important topic. However, different types of kidney stones require specific treatments, which creates the need for accurate diagnosis of the stone type prior to the intervention. Imaging techniques that are commonly used for the detection of kidney stones, such as X-ray CT and ultrasound, are insufficient to differentiate the types of kidney stones.

In this paper, we present a proof-of-concept study for differentiating kidney stones using X-ray dark-field tomography. The most important advantage of this method is its ability to image non-homogeneous kidney stones, i.e., to localize and identify the individual components of mixed-material kidney stones. We use a weighted total-variation regularized reconstruction method to compute the ratio of dark-field over absorption signal (DA Ratio) from noisy projections. We evaluate the performance of the proposed approach on two kidney stones of homogeneous composition, and one well-defined numerical phantom with known ground truth for mixed types of stones. We illustrate that the DA Ratio is significantly distinguished for different materials from the experiments. Reconstruction of phantom data recovers voxel-wise material information with high accuracy. We show that X-ray dark-field tomography has a significant potential in selective characterization of kidney stones.

Keywords: Machine learning, Heart, Ultrasound
Abstract: In this paper, we address the problem of automated pose classification and segmentation of the left ventricle (LV) in 2D echocardiographic images. For this purpose, we compare two complementary approaches. The first one is based on engineering ad-hoc features according to the traditional machine learning paradigm. Namely, we extract phase features to build an unsupervised LV pose estimator, as well as a global image descriptor for view type classification. We also apply the Supervised Descent Method (SDM) to iteratively refine the LV contour. The second approach follows the deep learning framework, where a Convolutional Network (ConvNet) learns the visual features automatically. Our experiments on a large database of apical sequences show that the two approaches yield comparable results on view classification, but SDM outperforms ConvNet on LV segmentation at a significantly lower training computational cost.

Keywords: Computer-aided detection and diagnosis (CAD), Heart, Ultrasound
Abstract: This paper describes an artificial neural network (ANN) method that employs a feature-learning algorithm to detect the lumen and MA borders in intravascular ultrasound (IVUS) images. Three types of imaging features including spatial, neighboring, and gradient features were used as the input features to the neural network, and then the different vascular layers were distinguished using two sparse autoencoders and one softmax classifier. To smooth the lumen and MA borders detected by the ANN method, we used the active contour model. The performance of our approach was compared with the manual drawing method and another existing method on 538 IVUS images from six subjects. Results showed that our approach had a high correlation (r = 0.9284 ~ 0.9875 for all measurements) and good agreement (bias = 0.0148 ~ 0.4209 mm) with the manual drawing method, and small detection error (lumen border: 0.0928±0.0935 mm, MA border: 0.1056±0.1088 mm). The average time to process each image was 14±4.6 seconds. The obtained results indicate that our proposed approach can be used to efficiently and accurately detect the lumen and MA borders in IVUS images.

Keywords: Brain, Animal models and imaging, Angiographic imaging
Abstract: The relation between normal and pathological aging and the cerebrovascular component is still unclear. In this context, the common marmoset, which has the advantage of enabling longitudinal studies over a reasonable timeframe, appears as a good pre-clinical model. However, there is still a lack of quantitative information on the macrovascular structure of the marmoset brain. In this paper, we investigate the potentiality of multi-angle TOF MR angiography using a 3T MRI scanner to perform morphometric analysis of the marmoset brain vasculature. Our image processing pipeline greatly relies on the use of multiscale vesselness enhancement filters to help extract the 3D macovasculature and perform subsequent morphometric calculations. Although multi-angle acquisition does not improve morphometric analysis significantly as compared to single-angle acquisition, it improves the network extraction by increasing the robustness of image processing algorithms.

Keywords: Image segmentation, Microscopy - Light, Confocal, Fluorescence
Abstract: We investigate a novel, parallel implementation of active contours for image segmentation combining a multi-agent system with a quad-edge representation of the contour. The control points of the contour evolve independently from one another in a parallel fashion, handling contour deformation, and convergence, while the quad-edge representation simplifies contour manipulation and local re-sampling during its evolution. We illustrate this new approach on biological images, and compare results with a conventional active contour implementation, discussing its benefits and limitations. This preliminary work is made freely available as a plug-in for our open-source Icy platform, where it will be developed with future extensions.

Keywords: Image segmentation, Magnetic resonance imaging (MRI), Vessels
Abstract: The magnetic susceptibility of haemoglobin is modulated by oxygen saturation, providing a mechanism to non-invasively measure oxygen extraction fraction. When combined with perfusion techniques, quantitative susceptibility mapping facilitates regional measurement of cerebral metabolic rate of oxygen consumption. However, accurate measurement requires a complete vein map to measure anatomical variance in the metabolic demands of tissue. In this work we present a novel shape-based Markov Random Field technique to segmentation the cerebral veins that provides accurate and complete vein maps. The shape-based graph underpinning the model controls the spatial relationships between voxels and enforces cylindrical geometry, allowing increased sensitivity with accurate vein boundaries.

Keywords: Histopathology imaging (e.g. whole slide imaging), Quantification and estimation, Image segmentation
Abstract: The translation of genomic sequencing technology to the clinic has greatly advanced personalized medicine. However, the presence of normal cells in tumors is a confounding factor in genome sequence analysis. Tumor purity, or the percentage of cancerous cells in whole tissue section, is a correction factor that can be used to improve the clinical utility of genomic sequencing. Currently, tumor purity is estimated visually by expert pathologists; however, it has been shown that there exist vast inter-observer discrepancies in tumor purity scoring. In this paper, we propose a quantitative image analysis pipeline for tumor purity estimation and provide a systematic comparison between pathologists' score and our image-based tumor purity estimation.

Keywords: Computed tomography (CT), Brain, Atlases
Abstract: In the field of medical imaging, atlases are generally used for computer-aided anatomical and functional parcellation of a brain, and for distinguishing which tissue is normal and which is pathologic. The purpose of this paper is to create a set of human brain atlas probability maps, which would be publicly available for clinical and research community and could be applied to computer tomography (CT) images in clinical studies. By utilizing the state of the art deformable image registration, three publicly available datasets were aligned to an age-specific symmetric multimodal human brain template represented in CT and MR. The validation of the cortical parcellation is based on 5 patients with multimodal acquisitions including non-contrast CT, CT angiography and MR T1. By complementing the multimodal CT-MR template with probability maps for the territory of Middle cerebral artery and its cortical regions, this dataset may be valuable for development of computer aided detection and navigation systems addressing neurovascular diseases.

Keywords: Image registration, Visualization, Liver
Abstract: Fast and accurate registration of image data is a key component of computer-aided medical image analysis. Instead of performing the registration directly on the input images, many algorithms compute the transformation using a sparse representation extracted from the original data. However, in order to apply the resulting transformation onto the original images, a dense deformation field has to be reconstructed using a suitable inter-/extra-polation technique.

In this paper, we employ the radial basis function (RBF) to reconstruct the dense deformation field from a sparse transformation computed by a model-based registration. Various kernels are tested using different scenario. The dense deformation field is used to warp the source image and compare it quantitatively to the target image using two different metrics. Moreover, the influence of the number and distribution of the control points required by the RBF is studied via two different scenarios. Beside the accuracy, the performance of the method accelerated using a GPU is reported.

Keywords: Image registration, Brain, Other-modality
Abstract: We introduce a new approach for the elastic registration of high-resolution 3D polarised light imaging (3D PLI) data of histological sections of the human brain. For accurate registration of different types of 3D PLI modalities, we propose a novel intensity similarity measure that is based on a least-squares formulation of normalized cross-correlation. Moreover, we present a fully automatic registration pipeline for rigid and elastic registration of high-resolution 3D PLI images with a blockface reference including preprocessing such as segmentation. We have successfully evaluated our approach using manually obtained ground truth for five sections of a human brain and experimentally compared it with previous approaches. We also present experimental results for 60 brain sections.

Keywords: Image registration, Diffusion weighted imaging, Liver
Abstract: Registration of diffusion weighted datasets remains a challenging task in the process of quantifying diffusion indexes. Respiratory and cardiac motion, as well as echo-planar characteristic geometric distortions, may greatly limit accuracy on parameter estimation, specially for the liver. This work proposes a methodology for the non-rigid registration of multiparametric abdominal diffusion weighted imaging by using different well-known metrics under the groupwise paradigm. A three-stage validation of the methodology is carried out on a computational diffusion phantom, a watery solution phantom and a set of voluntary patients. Diffusion estimation accuracy has been directly calculated on the computational phantom and indirectly by means of a residual analysis on the real data. On the other hand, effectiveness in distortion correction has been measured on the phantom. Results have shown statistical significant improvements compared to pairwise registration being able to cope with elastic deformations.

Keywords: Whole-body, Visualization, Pattern recognition and classification
Abstract: Action recognition systems have the potential to support clinicians, coaches and physical therapists in identifying important adopted movement patterns which could aid injury detection potential or inform rehabilitation strategies. Currently, motion capture systems, structured light pattern and time-of-flight sensors have utilization limitations that place constraints on their use outside of the laboratory setting. For this reason, we propose a system for human action recognition from video. The method presented in this work has utility with patient populations, such as Parkinson’s disease, Alzheimer’s disease, multiple sclerosis and dementia, outside of laboratory setting to detect the degree of which, and progression of, gait pathology. We developed a novel vision algorithm for template matching—the characterization of the motion in a video sequence. The method, titled Frequency Divergence Image, is a paradigm shift in template matching methods. Template matching methods measure macro-motion, whereas the proposed method detects micro-motion that differs from the flow of the action. We show that micro-cues improve prediction performance of human action on a real-world data set. We demonstrate a 9.15% improvement in classification accuracy over the original Motion History Image formulation when used with a convolutional neural network. Future work will focus on the deployment of the system to identify gait pathology from various patient populations.

Keywords: Image enhancement/restoration(noise and artifact reduction)
Abstract: A wide range of biomedical applications require detection, quantification and modelling of curvilinear structures in 3D images. Here we propose a 3D contrast-independent approach to enhance curvilinear structures based on the 3D Phase Congruency Tensor concept. The results show that the proposed method is insensitive to intensity variations along the 3D curve, and provides successful enhancement within noisy regions. The quality of the 3D Phase Congruency Tensor is evaluated by comparing it with state-of-the-art intensity-based approaches on both synthetic and real biological images.

Keywords: Optical coherence tomography, Image enhancement/restoration(noise and artifact reduction), Quantification and estimation
Abstract: In this paper, we present a real-time instantaneous phase estimation technique for the calibration of swept source optical coherence tomography (SS-OCT) systems. The proposed algorithm is able to accurately estimate both the amplitude and phase content of a swept source signal. The results are robust in the presence of substantial noise. Furthermore, the resulting phase profile is readily unwrapped, allowing for the generation of a k-linear sampling clock by means of a simple comparison step, which enables the SS-OCT system operator to determine the required number of points for accurate image reconstruction without the need for MZI path length modification. Our Simulations exhibit equivalent results with Hilbert transform based calibration techniques and low computational time rendering it suitable for real time applications.

Keywords: Skin, Pattern recognition and classification
Abstract: Skin cancer has a significant economic burden on the health system. Current methods used for the early detection of skin cancers are inefficient and have limited efficacy due to the constraint of the availability of limited experts required to make a reliable diagnosis. Thus, attempts to fund skin cancer screening programmes have been problematic. MoleMap has operated a tele-dermatology programme for the early detection of skin cancer for 20 years. Analysis of the data from this programme shows that the efficacy is very high >95%, but the efficiency is relatively low – 94% of lesions imaged are not suspicious of skin cancer. The time to conduct a screening session is long as the protocols require that many lesions must be imaged for subsequent tele-diagnosis by the expert.

A recent development has been to allow highly skilled skin cancer nurses to select only a few lesions – those that are assessed as being possibly suspicious for skin cancer - to be imaged for subsequent tele-diagnosis by the experts. This has been shown to improve the efficiency whilst not compromising the efficacy. The challenge to implementing this in a public screening environment is the time taken to train a skin cancer nurse – currently this involves 100 hours of formal training and 12 months of operational experience in a skin cancer clinic.

We have identified two pathways to accelerating this new approach to providing a larger base of skin cancer nurses: 1) Develop new training programmes using perceptive and adaptive learning techniques (PALM) that are created from the data held in the MoleMap skin cancer database. 2)Develop a computer assisted imaging system built around IBM Watson’s deep learning algorithms and the MoleMap skin cancer database.

Both approaches are important in allowing the skin cancer nurse to rapidly and accurately determine which lesions may be suspicious of skin cancer that would then be imaged and assessed by the experts using tele-dermatology.

Initial results suggest that the use of PALMs will allow skin cancer nurses training to be reduced to weeks rather than the months currently required. Further work is required to test the results on a broader population of nurses.

The initial work in developing the deep learning algorithms is promising as it suggests that it will be possible to reliably determine if a lesion is potentially a skin cancer with the

Keywords: Skin, Pattern recognition and classification
Abstract: Skin diseases, in general, have a high economic burden on health systems. Skin clinics perform a very important role of screening malignant cases at an early stage when it can make the most difference. However, in practice, these skin clinics, though highly accurate, are inefficient. A research programme using a dataset collected by MoleMap NZ Ltd. has enabled us to bridge these gaps in the existing literature. A DL based algorithm is developed to identify risks at multiple levels, e.g. “Cancer or Benign”, “Melanoma vs Any disease” or “Differential disease diagnosis”. The Deep Convolutional Neural Network (D-CNN) architecture uses temporal information and can combine information from images taken by multiple devices. Early results show a high accuracy for differential diagnosis based on single image modality. Moreover, a combination of these images yields an even higher accuracy. More interestingly, the performance using “clinical” images are very close to that of using “dermatoscopic” images. This holds real potential of using this technology in practice and especially in a tele-dermatology setting where specialised devices may not be available. Successful application of the developed assistive technology in these areas hold the potential of making the skin clinic more efficient. This will lead to better patient care and more meaningful public screening methodologies than currently possible.

Keywords: Computer-aided detection and diagnosis (CAD), Parallel computing
Abstract: Health systems, whether public or private are a significant economic burden and often are very inefficient. Access to high quality care is also dependent on socio-economic factors as well as a patient’s location – regional and third world patients do not have the same access to highly trained and knowledgeable clinicians, or treatment facilities as first world metropolitan patients. The long term goal of tele-health has been to alleviate the access problem as well as to provide economies of scale in disease diagnosis. But tele-health has always been constrained by technology, most notable network bandwidth and computing power.

With the proliferation of the internet and broadband networks, as well as large computing infrastructure sites, cloud computing can now support tele-health at national levels. It is possible to transmit and store patient data, including imaging studies into the cloud and perform near real-time analytics on the data to provide clinical benefits, such as diagnosis assistance, triage and screening results. To support such computing workloads a scalable and flexible platform has been built using many cloud technologies such as containerization (Docker), resource management (Mesos) and resource scheduling (Marathon) and utilizing DevOps and agile processes. In addition, to exploit deep machine learning for image analytics, the platform can employ GPU hardware available from the cloud infrastructure.

Using two example tele-health applications – melanoma detection and diabetic retinopathy screening it is shown how such applications can be easily built using web user interfaces and componentized analytics accessed via a simple RESTful API. The applications can be scaled based on workloads and a flexible “pluggable” architecture allows new or updated analytics to be added at will. These applications can provide primary care physicians or nursing care, with the diagnostic expertise of highly skilled clinical specialists, regardless of where they are located in the world at a fraction of the cost.

Keywords: Computer-aided detection and diagnosis (CAD), Eye, Retinal imaging
Abstract:

Keywords: Retinal imaging, Computer-aided detection and diagnosis (CAD), Machine learning
Abstract: It is estimated that 415 million people had diabetes in 2016 and the number is set to increase to in excess of 640 million people by 2040. Eye disease (diabetic retinopathy) is one of the most common complications of diabetes and it is a leading cause of vision loss and blindness worldwide. Screening for diabetic retinopathy involves the interpretation of retinal photographs and systematic screening programs have been shown to be effective in reducing blindness through timely treatment. At present the bulk of screening is performed by specialist medical practitioners or certified retinal image graders at significant cost and with limited capacity. Machine learning systems have the potential to transform diabetic retinopathy screening. Scalability, as well as the potential for interval change detection and integration with health metadata for individualised risk prediction are potential advantages of such systems. The capacity for deployment at the point-of-care or remotely offers great promise to improve eye health care delivery to remote and rural communities. Challenges in the development of these approaches include variations in image quality related to camera performance, image compression, pupil size, ocular media opacities and variations in retinal anatomy. The availability of reference image sets with large numbers of graded images across the spectrum of pathology in a wide range of ethnicities is also important for the training of machine learning systems. Other potential challenges relate to privacy and the sharing of health data, integration with digital health record systems and adoption by clinicians. This presentation aims to highlight some clinical perspectives about emerging machine learning approaches to retinal image analysis and tele-ophthalmology for diabetic retinopathy screening.

Keywords: Computer-aided detection and diagnosis (CAD), Eye, Retinal imaging
Abstract:

Keywords: Pattern recognition and classification, Eye, Retinal imaging
Abstract: Eye disease has emerged as a significant health concern in the world. It is estimated that the number of patients affected by diabetic retinopathy (DR) is set to increase to in excess of 640 million people by 2040. The number of projected Glaucoma patients may touch 110.8 million by 2040. Screening for eye disease involves the interpretation of retinal photographs and image analysis on a small scale have been shown to be effective in reducing blindness through timely treatment.

With the availability of large number of retinal images it is important to design algorithms that take full advantage of these large datasets. This presentation highlights a generalized workflow for accurate and efficient large scale analysis of retinal images obtained as part of a tele-ophthalmology setting.

The presentation describes different algorithms for image quality assessment (IQA), left/right eye detection and landmark detection that act as first components in the workflow. IQA is extremely important to determine whether a given image is of sufficient quality to be further used for automatic image analysis. The left/right eye detection is also crucial for downstream analysis where knowledge of actual eye type is helpful in devising accurate treatment plans. A landmark detection module that can accurately identify and segment prominent landmarks such as the fovea, optic cup and optic disc, and the vasculature is also described.

The talk will also describe different image analysis components for DR and glaucoma diagnosis. A deep learning (DL) based method for determining DR severity from colour fundus images is described. This method leverages left/right eye detection and pathology specific DL pipelines to augment the main DR severity estimation pipeline. Deep learning methods for glaucoma diagnosis is also discussed. This method leverages optic cup and disc segmentation results to measure the optic cup-to-disc ratio for accurate glaucoma diagnosis.

Keywords: Nuclear imaging (e.g. PET, SPECT), Quantification and estimation, Brain
Abstract: In vivo tau imaging with PET is a promising new modality that offers a unique insight into Alzheimer’s disease (AD) pathology. 18F-AV1451 is a tau tracer currently being evaluated in both AIBL and ADNI. While MR-based quantification remains the gold standard, there is great interest in PET-only quantification techniques for use in patients who cannot undergo MRI. In this study, 3 PET-only methods (single atlas, adaptive atlas, and PCA-based atlas) are evaluated and compared to an MR-based quantification in both AIBL (94 subjects) and ADNI (87 subjects). Results show that all quantifications performed using PET-only normalization approaches did equally well, with an average quantification error around 2% in both cohorts.

Keywords: Image segmentation, Nuclear imaging (e.g. PET, SPECT), Dimensionality reduction
Abstract: While accurate tumor delineation in FDG-PET is a vital task, noisy and blurring imaging system makes it a challenging work. In this paper, we propose to address this issue using the theory of belief functions, a powerful tool for modeling and reasoning with uncertain and/or imprecise information. An automatic segmentation method based on clustering is developed in 3-D, where, different from available methods, PET voxels are described not only by intensities but also complementally by features extracted from patches. Considering there are a large amount of features without consensus regarding the most informative ones, and some of them are even unreliable due to image quality, a specific procedure is adopted to adapt distance metric for properly representing clustering distortions and neighborhood similarities. A specific spatial regularization is also included in the clustering algorithm to effectively quantify local homogeneity. The proposed method has been evaluated by real-patient images, showing good performance.

Keywords: Nuclear imaging (e.g. PET, SPECT), Machine learning, Image reconstruction - analytical & iterative methods
Abstract: The performance of smoothness-enforced Bayesian PET image reconstruction is strongly affected by the weight on regularization. Compromises need to be made between variance and spatial resolution. In this work, we propose to use an artificial neural network (ANN) to fuse the image versions reconstructed from a maximum a posteriori (MAP) algorithm with different regularizing weights for quantitative improvement. Using the BrainWeb phantoms, we simulated PET data at different count levels for different subjects with and without lesions. We designed an ANN and trained it using MAP reconstructions with different regularization parameters for one normal subject at a specific count level. Reconstructed images from the simulations with lesions, of other count levels, and of other subjects were fused using the trained ANN. In all of the testing experiments, the designed ANN fusion keeps the noise level as low as what the MAP algorithm achieves at heavy regularization while significantly reduces the bias or improves the lesion contrast. We conclude that the proposed ANN fusion method removes the need for tuning the regularization of the MAP algorithm.

Keywords: Nuclear imaging (e.g. PET, SPECT), Heart, Image reconstruction - analytical & iterative methods
Abstract: Cardiac gated images in single photon emission computed tomography (SPECT) are known to suffer from increased noise due to low data counts. In this work, we investigate a 4D reconstruction approach based on an adaptive spatiotemporal smoothing prior, which is used to exploit the common signal component among the different cardiac gates in a sequence. In the experiments, we evaluated this approach with both simulated NCAT imaging data and two sets of clinical acquisitions. The results demonstrate that the proposed 4D approach can be more effective for improving the heart wall in terms of both noise levels and spatial resolution than motion-compensated 4D reconstruction. The proposed approach was also found to be robust for noise suppression when the imaging dose was reduced.

Keywords: Image reconstruction - analytical & iterative methods, Nuclear imaging (e.g. PET, SPECT), Optimization method
Abstract: 4D-PET reconstruction has the potential to significantly increase the signal-to-noise ratio in dynamic PET by fitting smooth temporal functions during the reconstruction. However, the optimal choice of temporal function remains an open question. A 4D-PET reconstruction algorithm using adaptive-knot cubic B-splines is proposed. Using realistic Monte-Carlo simulated data from a digital patient phantom representing an [18-F]-FMISO-PET scan of a non-small cell lung cancer patient, this method was compared to a spectral model-based 4D-PET reconstruction and the conventional MLEM and MAP algorithms. Within the entire patient region the proposed algorithm produced the best bias-noise trade-off, while within the tumor region the spline- and spectral model-based reconstructions gave comparable results.

Keywords: Inverse methods, Computed tomography (CT), Bone
Abstract: The analysis of trabecular bone micro structure from textit{in-vivo} CT images is still limited due to insufficient spatial resolution. In a previous work, we have investigated the use of super resolution techniques to improve image quality based on a TV based approach. However, the method is limited to recover the bimodal nature of the image. In this work, we investigate the use of a double well non convex constraint to solve the joint super resolution/segmentation problem. Two different minimization schemes are proposed to obtain a critical point of the non convex functional. The two methods improve the reconstruction results on real data. %X-ray Computed Tomography (CT) techniques are increasingly used to image the micro-structure of the trabecular bone in studying the osteoporosis, a bone fragility disease, still difficult to diagnose.

Keywords: Tooth, Image segmentation, Computed tomography (CT)
Abstract: We propose a new semi-automatic framework for tooth segmentation in Cone-Beam Computed Tomography (CBCT) combining shape priors based on a statistical shape model and graph cut optimization. Poor image quality and similarity between tooth and cortical bone intensities are overcome by strong constraints on the shape and on the targeted area. The segmentation quality was assessed on 64 tooth images for which a reference segmentation was available, with an overall Dice coefficient above 0.95 and a global consistency error less than 0.005.

Keywords: Computed tomography (CT), Lung, Image segmentation
Abstract: Delineation of blurry boundary from medical images is challenging in particular when the target object or region of interest is adjacent to other tissues with similar or overlapping intensity distributions. To address this challenge, we propose a graph model with adaptive global and geodesic constraints to contour the indistinct boundary from CT images. The global factor reflects the appearance affinities and better differentiates background and foreground objects. The geodesic compartment is capable to capture and highlight the thin and weak boundary information. These complementary terms are incorporated in a transductive graph model for segmentation. The model was tested on 20 low contrast CT studies of patients with non-small cell lung cancer. The segmented tumor volume was compared with manual delineations and evaluated with respect to spatial overlap and shape similarity. The experimental results and student’s t-test demonstrated that considering the complementary global and geodesic factors contributed to the improved boundary definition.

Keywords: Computed tomography (CT), Lung, Image segmentation
Abstract: We propose an automated framework for lung nodule segmentation from pulmonary CT scan using graph cut with a deep learned prior. The segmentation problem is formulated as a hybrid cost function minimization task, which combines a domain specific data term with a deep learned probability map. The proposed segmentation framework embodies the robustness of deep learning in object localization, while retaining the hallmark of traditional segmentation models in addressing the morphological intricacies of elaborate objects. The proposed solution offers more than 20% performance improvement over a contemporary data driven model, and also outperforms traditional graph cuts especially in situations where model initialization is slightly inaccurate.

Keywords: Machine learning, Image segmentation, Computed tomography (CT)
Abstract: Mandible bone segmentation from computed tomography (CT) scans is challenging due to mandible’s structural irregularities, complex shape patterns, and lack of contrast in joints. Furthermore, connections of teeth to mandible and mandible to remaining parts of the skull make it extremely difficult to identify mandible boundary automatically. This study addresses these challenges by proposing a novel framework where we define the segmentation as two complementary tasks: recognition and delineation. For recognition, we use random forest regression to localize mandible in 3D. For delineation, we propose to use 3D gradient-based fuzzy connectedness (FC) image segmentation algorithm, operating on the recognized mandible sub-volume. Despite heavy CT artifacts and dental fillings, consisting half of the CT image data in our experiments, we have achieved highly accurate detection and delineation results. Specifically, detection accuracy more than 96% (measured by union of intersection (UoI)), the delineation accuracy of 91% (measured by dice similarity coefficient), and less than 1 mm in shape mismatch (Hausdorff Distance) were found.

Keywords: Magnetic resonance imaging (MRI), Brain, Machine learning
Abstract: Detection of brain metastases in patients with undiagnosed primary cancer is unusual but still an existing phenomenon. In these cases, identifying the cancer site of origin is non-feasible by visual examination of magnetic resonance (MR) images. Recently, radiomics has been proposed to analyze differences among classes of visually imperceptible imaging characteristics. In this study we analyzed 46 T1-weighted MR images of brain metastases from 29 patients: 29 of lung and 17 of breast origin. A total of 43 radiomics texture features were extracted from the metastatic lesions. Support vector machine (SVM) and k-nearest neighbors (k-NN) classifiers were implemented to evaluate the classification performance. The influence of gray-level quantization for computation of texture features was also examined. The best classification (AUC = 0.953 ± 0.061), evaluated with nested cross-validation, was obtained using the SVM classifier with two texture features derived from the 16 gray-level quantization co-occurrence matrix.

Keywords: fMRI analysis, Brain, Classification
Abstract: Several recent studies have shown that dictionary learning and sparse representation can effectively reconstruct hundreds of interacting functional brain networks simultaneously from whole-brain fMRI data. However, accurate classification and recognition of those hundreds of functional networks from an individual or a population of many subjects is still a challenging and open problem due to the intrinsic variability of functional networks and other noise sources. To tackle this problem, this paper presents an effective deep learning framework to train convolutional neural networks from a large dataset of hundreds of thousands of available brain network volume maps, which was then applied on testing samples for network classification and recognition. We effectively applied computer-labeled data as training set so the whole process can be automated. Experimental results showed that the proposed method is quite robust in handling noisy patterns in the dataset, which suggests that our work offers a new computational framework for modeling functional connectomes from fMRI big data in the future.

Keywords: Computed tomography (CT), Heart, Machine learning
Abstract: Accurate volume estimation of left atrial aneurysm plays an essential role in the early diagnosis and therapy planning. However, it is a challenging task due to huge shape variabilities of aneurysms and great appearance variations of images, which tends to be intractable for segmentation methods. In this paper, we propose a novel estimation method for direct estimation of atrial aneurysm volumes without segmentation. To handle the high variabilities and variations, we propose a new multi-view semi-supervised manifold learning (MSML) algorithm, which fuses multiple complementary features to generate compact, informative and discriminative aneurysm image representation by leveraging both labeled and unlabeled data. Based on the obtained image representation by the MSML, we adopt random regression forests to conduct direct and efficient volume estimation. Our method for the first time achieves a fully automatic estimation of left atrial aneurysm volumes. Experiments on a clinical dataset of 67 subjects with a total of 1220 images show that our method achieves a high correlation coefficient of 0.91 with ground truth manually labelled by clinical experts and largely outperforms other methods, which demonstrates the effectiveness for aneurysm volume estimation and indicates its potential use in clinical practise.

Keywords: Brain, Magnetic resonance imaging (MRI), Genes
Abstract: Optimal representations of the genetic structure underlying complex neuroimaging phenotypes lie at the heart of our quest to discover the genetic code of the brain. Here, we suggest a strategy for achieving such a representation by decomposing the genetic covariance matrix of complex phenotypes into maximally heritable and genetically independent components. We show that such a representation can be approximated well with eigenvectors of the genetic covariance based on a large family study. Using 520 twin pairs from the QTIM dataset, we estimate 500 principal genetic components of 54,000 vertex-wise shape features representing fourteen subcortical regions. We show that our features maintain their desired properties in practice. Further, the genetic components are found to be significantly associated with the CLU and PICALM genes in an unrelated Alzheimer’s Disease (AD) dataset. The same genes are not significantly associated with other volume and shape measures in this dataset.

Keywords: Multi-modality fusion, Computer-aided detection and diagnosis (CAD), Magnetic resonance imaging (MRI)
Abstract: In this study, a novel feature selection framework is proposed to simultaneously perform classification and clinical sores prediction of Parkinson’s disease (PD) via multi-modal neuroimaging data. Specifically, a new feature selection model is devised to capture discriminative features to train support vector regression model for clinical scores (e.g., sleep scores and olfactory scores) prediction and support vector classification model for class label identification. Our method is evaluated on a public dataset of 208 subjects including 56 normal controls (NC), 123 PD and 29 scans without evidence of dopamine deficit (SWEDD) via a 10-fold cross-validation method. The experimental results demonstrate that multi-modal data can effectively improve the performance in disease status identification and clinical scores prediction compared to one single modality. Our proposed method also outperforms the related methods.