Keywords: Computational Imaging, Compressive sensing & sampling, Image reconstruction - analytical & iterative methods
Abstract: We propose a structured low rank matrix completion algorithm to recover a time series of images consisting of linear combination of exponential parameters at every pixel, from undersampled Fourier measurements. The spatial smoothness of these parameters is exploited along with the exponential structure of the time series at every pixel, to derive an annihilation relation in the k-t domain. This annihilation relation translates into a structured low rank matrix formed from the k-t samples. We demonstrate the algorithm in the parameter mapping setting and show significant improvement over state of the art methods.

Keywords: Quantification and estimation, Magnetic resonance imaging (MRI)
Abstract: MRI parameter quantification has diverse applications, but likelihood-based methods typically require nonconvex optimization due to nonlinear signal models. To avoid expensive grid searches used in prior works, we propose to learn a nonlinear estimator from simulated training examples and (approximate) kernel ridge regression. As proof of concept, we apply kernel-based estimation to quantify six parameters per voxel describing the steady-state magnetization dynamics of two water compartments from simulated data. In relevant regions of fast-relaxing compartmental fraction estimates, kernel estimation achieves comparable mean-squared error as grid search, with dramatically reduced computation.

Keywords: Image registration, Machine learning, Brain
Abstract: Registration involving one or more images containing pathologies is challenging, as standard image similarity measures and spatial transforms cannot account for common changes due to pathologies. Low-rank/Sparse (LRS) decomposition removes pathologies prior to registration; however, LRS is memory-demanding and slow, which limits its use on larger data sets. Additionally, LRS blurs normal tissue regions, which may degrade registration performance. This paper proposes an efficient alternative to LRS: (1) normal tissue appearance is captured by principal component analysis (PCA) and (2) blurring is avoided by an integrated model for pathology removal and image reconstruction. Results on synthetic and BRATS 2015 data demonstrate its utility.

Keywords: Magnetic resonance imaging (MRI), Machine learning, Compressive sensing & sampling
Abstract: Compressed sensing (CS) enables significant reduction of MR acquisition time with performance guarantee. However, computational complexity of CS is usually expensive. To address this, here we propose a novel deep residual learning algorithm to reconstruct MR images from sparsely sampled k-space data. In particular, based on the observation that coherent aliasing artifacts from downsampled data has topologically simpler structure than the original image data, we formulate a CS problem as a residual regression problem and propose a deep convolutional neural network (CNN) to learn the aliasing artifacts. Experimental results using single channel and multi channel MR data demonstrate that the proposed deep residual learning outperforms the existing CS and parallel imaging algorithms. Moreover, the computational time is faster in several orders of magnitude.

Keywords: Compressive sensing & sampling, Magnetic resonance imaging (MRI), Image reconstruction - analytical & iterative methods
Abstract: High-dimensional signals, including dynamic magnetic resonance (dMR) images, often lie on low dimensional manifold. While many current dynamic magnetic resonance imaging (dMRI) reconstruction methods rely on priors which promote low-rank and sparsity, this paper proposes a novel manifold-based method, we term M-MRI, for dMRI reconstruction from highly undersampled k-space data. Images in dMRI are modeled as points on or close to a smooth manifold, and the underlying manifold geometry is learned through training data, called “navigator” signals. Moreover, low-dimensional embeddings which preserve the learned manifold geometry and effect concise data representations are computed. Capitalizing on the learned manifold geometry, two regularization loss functions are proposed to reconstruct dMR images from highly undersampled k-space data. The advocated framework is validated using extensive numerical tests on phantom and in-vivo data sets.

Keywords: Image reconstruction - analytical & iterative methods, Compressive sensing & sampling, Microscopy - Electron
Abstract: We designed a complete acquisition-reconstruction framework to reduce the radiation dosage in 3D scanning transmission electron microscopy (STEM). Projection measurements are acquired by randomly scanning a subset of pixels at every tilt-view (i.e., random-beam STEM or “RB-STEM”). High-quality images are then recovered from the randomly downsampled measurements through a regularized tomographic reconstruction framework. By fulfilling the compressed sensing requirements, the proposed approach improves the reconstruction of heavily-downsampled RB-STEM measurements over the current state-of-the-art technique. This development opens new perspectives in the search for methods permitting lower-dose 3D STEM imaging of electron-sensitive samples without degrading the quality of the reconstructed volume. A Matlab code implementing the proposed reconstruction algorithm has been made available online.

Keywords: Microscopy - Super-resolution, Image reconstruction - analytical & iterative methods, Inverse methods
Abstract: Single molecule localization microscopy has made great improvements in spatial resolution achieving performance beyond the diffraction limit by sequentially activating and imaging small subsets of molecules. Here, we present an algorithm designed for high-density molecule localization which is of a major importance in order to improve the temporal resolution of such microscopy techniques. We formulate the localization problem as a sparse approximation problem which is then relaxed using the recently proposed CEL0 penalty, allowing an optimization through recent nonsmooth nonconvex algorithms. Finally, performances of the proposed method are compared with one of the best current method for high-density molecules localization on simulated and real data.

Keywords: Compressive sensing & sampling, Microscopy - Super-resolution, Image reconstruction - analytical & iterative methods
Abstract: In this work, we introduce an original strategy to apply the Compressed Sensing (CS) framework to a super-resolution Structured Illumination Microscopy (SIM) technique. We first define a framework for direct domain CS, that exploits the sparsity of fluorescence microscopy images in the Fourier domain. We then propose an application of this method to a fast 4-images SIM technique, which allows to reconstruct super-resolved fluorescence microscopy images using only 25% of the camera pixels for each acquisition.

Keywords: Probabilistic and statistical models & methods, Cells & molecules, Microscopy - Light, Confocal, Fluorescence
Abstract: Microscopic analysis of neuronal cell morphology is required in many studies in neurobiology. The development of computational methods for this purpose is an ongoing challenge and includes solving some of the fundamental computer vision problems such as detecting and grouping sometimes very noisy line-like image structures. Advancements in the field are impeded by the complexity and immense diversity of neuronal cell shapes across species and brain regions, as well as by the high variability in image quality across labs and experimental setups. Here we present a novel method for fully automatic neuron reconstruction based on sequential Monte Carlo estimation. It uses newly designed models for predicting and updating branch node estimates as well as novel initialization and final tree construction strategies. The proposed method was evaluated on 3D fluorescence microscopy images containing single neurons and neuronal networks for which manual annotations were available as gold-standard references. The results indicate that our method performs favorably compared to state-of-the-art alternative methods.

Keywords: Computational Imaging, Microscopy - Light, Confocal, Fluorescence, Inverse methods
Abstract: Traction Force Microscopy (TFM) is a commonly used technique to compute cellular tractions that cells exert to the surrounding substrate. Traction recovery is an ill-posed inverse problem, which needs regularization to stabilize the solution. Due to its simplicity, Tikhonov or L2-regularization is usually used, but recent studies have demonstrated the increase of sensitivity and resolution in the recovered tractions using an L1-regularization scheme. In this manuscript, we present an approximation that makes feasible the traction recovery on full-size microscope images when working in the spatial domain. We perform also a comparison between the two regularization schemes named before (relying in L2-norm for the data fidelity term) and the full L1-regularization (using L1-norm for both the regularization and data fidelity terms). Our proof-of concept using real data reveal that L1-regularizations might give an improved resolution (more accused for full L1-regularization) and a reduction in the background noise with respect to the classical zero-order Tikhonov regularization.

Keywords: Compressive sensing & sampling, Image enhancement/restoration(noise and artifact reduction), Image reconstruction - analytical & iterative methods
Abstract: The design of coded apertures for compressive X-ray tomosynthesis is addressed here, based on the analysis of the coherence of the sensing matrix. It is shown that minimizing the inner products between the columns of the transfer function matrix, it is possible to obtain significant improvements with respect to the results obtained with the use of random codes. The computational cost of optimization is dramatically lower than other approaches introduced recently.

Keywords: Computed tomography (CT), Compressive sensing & sampling, Image enhancement/restoration(noise and artifact reduction)
Abstract: We propose a stable and fast reconstruction technique for parallel-beam (PB) tomographic X-Ray imaging, relying on the discrete pseudo-polar (PP) Radon and PP Fourier transforms. Our main contribution is a resampling method, based on modern sampling theory, that transforms the PB measurements to a PP grid. The resampling process is both fast and accurate, and in addition, simultaneously denoises the measurements, by exploiting geometrical properties of the tomographic scan. The transformed measurements are then reconstructed using an iterative solver with TV regularization. We show that reconstructing from measurements on the PP grid, leads to improved recovery, due to the inherent stability and accuracy of the PP Radon transform, compared with the PB Radon transform. We also demonstrate recovery from a reduced number of PB acquisition angles and low SNR measurements. Our approach is shown to achieve superior results over other state-of-the-art solutions, that operate directly on the PB measurements.

Keywords: Angiographic imaging, Image reconstruction - analytical & iterative methods, Vessels
Abstract: Tomographic reconstruction of cardiovascular structures from rotational angiograms acquired with interventional C-arm devices is challenging due to cardiac motion. Gating strategies are widely used to reduce data inconsistency but come at the cost of angular undersampling. We employ a spatio-temporally regularized 4-D reconstruction model, which is solved using a proximal algorithm, to handle the substantial undersampling associated with a strict gating setup. In a numerical phantom study based on the Cavarev framework, similarity to the ground truth is improved from 82.3% to 87.6% by this approach compared to a state-of-the-art motion compensation algorithm, whereas previous regularized methods evaluated on this phantom achieved results below 80%. We also show first image results for a clinical patient data set.

Keywords: X-ray imaging, Bone, Image reconstruction - analytical & iterative methods
Abstract: X-ray phase computed tomography (CT) offers higher sensitivity than conventional X-ray CT. A new phase-CT instrument producing a nano-focused beam has been developed at the ESRF (European Synchrotron Radiation Facility) for nano-imaging. In order to obtain final images, a suited phase retrieval algorithm is necessary, which is attracting broader interest recently. In this paper, we explicit the 3D phase CT image reconstruction problem, including the stage of phase retrieval prior to 3D CT reconstruction. The phase retrieval problem is solved by extending the single distance Paganin method to multi-distance acquisitions, followed by an iterative non-linear conjugate gradient descent optimization method. The method is evaluated on bone tissue samples imaged at voxel sizes of 120 nm. The results obtained from acquisition at 1 and 4 distances, with and without the iterative refinement are compared. The results show that this method yields improved images compared to other methods.

Keywords: Computed tomography (CT), Image acquisition, Image reconstruction - analytical & iterative methods
Abstract: High radiation dose impedes the development of in-vivo micro-CT. In this paper, we presents a low dose and fast in-vivo micro computed tomography (micro-CT) system based on equally sloped tomography (EST) technique and the monochromatic synchrotron X-ray source. Comparing with regular CT, the projection number required for our imaging system can be reduced by about 75%. In addition, combining with an X-ray shutter, total exposure time of 4 seconds and about 0.67Gy absorption dose for a set of CT data have been achieved. These results demonstrate that micro-CT with monochromatic synchrotron X-rays has great potential in the investigation into the microstructure evolution inside a small animal for biomedical research.

Keywords: Magnetic resonance imaging (MRI), Whole-body, Computer-aided detection and diagnosis (CAD)
Abstract: Deep learning has made rapid advances recently in computer vision, robotics and natural language processing and the novel applications have changed the landscape significantly. However the progress of deep learning in medical imaging has not been as rapid, predominantly due to the inherent nature of the images. In the computer vision community, many popular deep learning methods were trained on natural images. However, natural and medical images differ significantly: natural images are usually three channel colour images(RGB), while medical images tend to be single channel grayscale. For natural images the goal is to infer the context, isolate multiple foreground from background and identify objects of interest. In medical images, however, we are interested in identifying regions of interests that may not have a clear boundary or discernible shape. Further, some characteristics of medical images such as rotational invariance do not apply to natural images. Another significant problem faced by medical imaging is the paucity of data. Typical deep learning networks are trained on a large amount of data, while such an annotated dataset in medical imaging applications is both expensive and time consuming to acquire. Visualization of the trained deep networks has also been a problem as medical professionals require a different set of visualizations to those required for natural images. In fact, we believe that lack of visualization has added to the lack of transparency of deep learning techniques, encouraging a black-box view of deep learning. Finally, training a deep network from scratch on a large dataset takes a considerable amount of time. Training also requires an expert to optimize the network hyper-parameters, as heuristics for hyper-parameter optimization may not be available just yet. Transfer learning may be a solution but comes with its own problems. Most pre-trained networks used for transfer learning are trained on natural images, and adaptation to medical imaging is problematic. Even after adaptation, transfer learning with deep learning techniques performs best when pre-trained on similar datasets. In this talk, we shall discuss the challenges in applying deep learning techniques in medical imaging and propose intuitive solutions to overcome them.

Keywords: Optimization method, Pattern recognition and classification, Computational Imaging
Abstract:

Keywords: Eye, Machine learning
Abstract:

Keywords: Image segmentation, Prostate, Magnetic resonance imaging (MRI)
Abstract: Manual segmentation of organs from medical images is time consuming, and prone to inter- and intraexpert variability. Software algorithms should be able to overcome challenges like (1) variability of organ size and shape between subjects; (2) variable image appearance and intensity ranges from different scanning protocols; and (3) lack of clear boundaries due to similar intensity profiles of surrounding tissues. Due to poor contrast, segmentation methods using specific image features find it hard to distinguish between organ and non-organ regions. Machine learning (ML) methods present a formal way to learn and identify features that are highly discriminative for classification and segmentation purposes. A robust algorithm requires many example cases to learn from a wide range of image appearances. However, obtaining sufficient manual annotations is very expensive, time consuming, and requires personnel with high expertise.

This talk presents a ML-based segmentation method that requires significantly fewer labeled samples, yet achieves higher segmentation accuracy than conventional ML methods. Our algorithm uses (1) an active learning (AL) strategy that incorporates principles from visual saliency to query labels of informative samples; and (2) semisupervised learning (SSL) that uses a few labeled samples and many unlabeled samples to construct a highly accurate classifier. Recent work has shown that expert feedback can greatly improve the accuracy of medical image-segmentation algorithms. This motivated us to explore the efficacy of incorporating expert feedback for improved medical image segmentation. The important contribution of our work is a visual saliency-based approach to select the most informative samples for AL. We show that many of the principles of salient region detection are applicable to query selection in active learning tasks. Hence, selecting the most informative region in an image becomes a problem of salient region detection by defining an appropriate measure of a region’s importance.

Keywords: Image reconstruction - analytical & iterative methods, Angiographic imaging
Abstract: Magnetic particle imaging (MPI) is a relatively new imaging modality that images the spatial distribution of superparamagnetic iron oxide nanoparticles administered to the body. In this study, we use a new method based on Alternating Direction Method of Multipliers (a subset of Augmented Lagrangian Methods, ADMM) with total variation and l1 norm minimization, to reconstruct MPI images. We demonstrate this method on data simulated for a field free line MPI system, and compare its performance against the conventional Algebraic Reconstruction Technique. The ADMM improves image quality as indicated by a higher structural similarity, for low signal-to-noise ratio datasets, and it significantly reduces computation time.

Keywords: Ultrasound, Computer-aided detection and diagnosis (CAD)
Abstract: This paper proposes an accurate ultrasound probe localization method for 3D US image reconstruction. The proposed method consists of (i) feature tracking of a video sequence and (ii) camera pose estimation using structure from motion (SfM). SfM is used to reconstruct 3D point clouds from multiple-view images and simultaneously estimate each camera position. To apply SfM to a video sequence, the accurate method is required to track features between adjacent frames. We employ a sub-pixel image matching technique using Phase-Only Correlation (POC) for feature tracking. POC is a technique of image matching using phase components in the Fourier transforms of images. Through a set of experiments, we demonstrate that the proposed method can estimate the location of the US probe with about 2mm error for the probe travel distance of 150--200mm.

Keywords: Shape analysis, Microscopy - Multi-photon, Computational Imaging
Abstract: Although parameterizing the shape of objects is one of the important research topics in three dimensional data processing, systematic and data-driven evaluation of 3D shape, such as cellular shapes, is difficult. In this presentation, we describe the workflow of feature extraction based on mathematically robust procedures; curvature flow-based conformal transformation and spherical harmonics decomposition. Simulation dataset analysis demonstrated the method clustered triangulated objects appropriately.

Keywords: Microscopy - Light, Confocal, Fluorescence, Machine learning, Cells & molecules
Abstract: Early detection of the presence of leukemic cells in the peripheral blood (PB) and the subsequent possibility of a prompt treatment are essential for the patient survival. The morphologic differentiation among different types of abnormal lymphoid cells and blast cells in PB is a complex work. Moreover, subtle morphologic characteristics are exhibited by some abnormal neoplastic lymphoid cells, which are shared with reactive lymphoid cells. This study aims to automatically classify several types of cells using PB digital images, including normal and reactive lymphocytes, abnormal lymphoid cells and blast cells. We included a total of 9048 PB digital cell images of 10 different cell types with a resolution of 360 x 363 pixels, obtained from 223 patients, stained with May Grünwald-Giemsa and acquired in the CellaVision DM96. Lymphoid cells were segmented by color clustering using spatial kernel fuzzy c-means and Gaussian mixture models, and the watershed transformation. Thus, three regions of interest were obtained: the entire cell, the nucleus and the peripheral zone around the cell. We extracted geometric features about the size and shape of the nucleus, cell and cytoplasm. Four types of color and texture features on several color components were calculated: statistical features, wavelet statistical features, Gabor features and granulometric features. Each of these features was applied to the entire cell, the nucleus and the cytoplasm. Thus, the full set combined 27 geometric features and 2649 color-texture features. Then, a recognition module was built, which included a classifier based on support vector machines (SVM) with a 3 rd order polynomial kernel, whose inputs were the 140 most relevant and less redundant features selected from the full set. The number of selected features and the tuning of the best classifier were performed jointly through an iterative process, involving 10-fold cross validation. We obtained a cell classification accuracy of 86.6 %. In this paper, ten types of lymphoid cells have been automatically recognized showing promising results in the respective classifications. The main contribution of this work is the development of a complete method that could allow to design a practical diagnosis support tool of lymphoid neoplasms in the near future.

Keywords: Image segmentation, Tracking (time series analysis), image filtering (e.g. mathematical morphology, wavelets,...)
Abstract: Imaging of live cell is an important step for understanding and analyzing biological functions via studying cellular dynamics. Cell segmentation and tracking in microscopy images are challenging tasks due to embedded noise. In our paper, adaptive dictionary learning is proposed for filtering microscopy images. To the best of our knowledge, this method has not been applied before for this task, despite it provides good results for filtering typical and astronomical images. We applied our method to two different types of microscopy images. Preliminary results show that the proposed algorithm outperforms the state of the art techniques.

Keywords: Image segmentation, Microscopy - Light, Confocal, Fluorescence, Molecular and cellular screening
Abstract: Computerized image systems are able to recognize normal blood cells, but fail with abnormal lymphoid cells associated to lymphomas. The main challenge lies in the subtle differences existing in morphologic features among these classes, what requires a refined segmentation of three regions of interest (ROIs): nucleus, cytoplasm and peripheral zone around the cell. This paper presents a new efficient segmentation framework for this problem, which have three stages: (1) leukemic cells isolation; (2) cell and nucleus segmentation; and (3) separation of touching cells.

The image is transformed to the YCbCr color space. Through spatial kernel fuzzy c-means clustering of the new color components, three membership images are generated: background, cell and red blood cells (RBC). The background and the RBC memberships are binarized and combined to obtain a modified foreground mask. Then, the watershed transformation (WT) is performed over the distance transform of this mask, obtaining a separation of the cells in a label matrix. From here, the peripheral ROI is generated. Subsequently, the original image is cropped to the bounding box of this ROI and the region outside the peripheral region is replaced by the mean of the background. Thus, a new isolated cell image is obtained.

In the second stage, the new cell image is transformed to the YCbCr space. Then, a clustering using Gaussian mixture models is done to obtain three posterior probability images: background, cytoplasm and nucleus. Finally, a combination of these probability images produces the cell and nucleus ROIs.

Finally in the third stage, the circle variances of the cell and nucleus ROIs are used to decide if there are touching cells. If this is the case, the new isolated cell and the three ROIs are processed using the distance transform, the H-minima transform and the WT with markers to separate all the cells.

12 leukemic cell types with a total of 16408 images from 374 patients were considered. They were segmented to evaluate the efficiency of the framework. The overall segmentation efficiency was 98.9 %.

The proposed segmentation framework has been able to process a wide variety leukemic cell types, which have distinct morphologic characteristics, achieving an excellent overall efficiency.

Keywords: Fluorescence tomography, Breast, image filtering (e.g. mathematical morphology, wavelets,...)
Abstract: This novel imaging platform utilizing high-speed photography and customized image analysis algorithms was developed for predicting efficacy of a novel medical device designed to deliver medicines to breastfeeding infants. It provides a whole new platform for experiments that previously were not possible.

Keywords: Functional imaging (e.g. fMRI), Brain, fMRI analysis
Abstract: Functional network analysis based on matrix decomposition/factorization methods including ICA and dictionary learning models have become a popular approach in fMRI study. Yet it is still a challenging issue in interpreting the result networks because of the inter-subject variability and image noises, thus in many cases, manual inspection on the obtained networks is needed. Aiming to provide a fast and reliable functional network identification tool for both normal and diseased brain fMRI data analysis, in this work, we propose a novel supervised dictionary learning model based on rank-1 matrix decomposition algorithm (S-r1DL) with sparseness constraint. Application on the Autism Brain Imaging Data Exchange (ABIDE) database showed that S-r1DL can fast and accurately identify the functional networks based on the given templates, comparing to unsupervised learning method.

Keywords: Genes, Population analysis, Magnetic resonance imaging (MRI)
Abstract: The influence of genes on cortical structures has been assessed through various phenotypes already. However, the heritability of the sulcal pits, which are the deepest points of the sulci, has not yet been estimated. These pits are assumed to be under tight genetic control and to have close relationship with underlying functional organization. We estimated the heritability of these pits depth using the HCP pedigree data. Our results confirm previous hypothesis, with significant heritability estimates found for the pits of central and superior temporal sulci, and underline the pit as biomarkers of interest for future genome wide association studies.

Keywords: Blind source separation & Dictionary learning, fMRI analysis, Functional imaging (e.g. fMRI)
Abstract: Independent vector analysis (IVA) has exhibited promising applications to complex-valued fMRI data, however model order effects on complex-valued IVA have not yet been studied. As such, we investigate model order effects on IVA using 16 task-based complex-valued fMRI data sets. A noncircular fixed-point complex-valued IVA (non-FIVA) algorithm was utilized. The model orders were varied from 10 to 160. The ICASSO toolbox was modified for selecting the best spatial estimates across all runs to assess the IVA stability. Non-FIVA was compared to a complex-valued independent component analysis (ICA) algorithm as well as to real-valued IVA and ICA algorithms which analyzed magnitude-only fMRI data. The complex-valued analysis detected component splitting at higher model orders, but in a different way from the magnitude-only analysis in that a complete component and its sub-components exist simultaneously. This suggests that the incorporation of phase fMRI data may better preserve the integrity of the larger networks. Good stability was also achieved by non-FIVA with different orders.

Keywords: Functional imaging (e.g. fMRI), Brain, Classification
Abstract: Alzheimer’s disease (AD) and its early stage, mild cognitive impairment (MCI), have been widely analyzed by brain functional connectivity network (BFCN) due to its promising potential in identifying biomarkers for the neurodegenerative disorders and understanding the brain functions. The accurate construction of biologically meaningful brain network plays an essential role in these applications. Sparse learning has been widely applied for the network construction. However, the previous sparse learning studies often fail to consider the smoothness penalty from multiple time points. To address these problems, we propose a new multi-task learning method to integrate the fused penalty with smooth regularization. Specifically, a novel objective function is developed to consider fused learning of multiple time points via smoothness constraint. We evaluate our method for MCI identification via the resting-state functional magnetic resonating imaging (rs-fMRI) from Alzheimer’s Disease Neuroimaging Initiative Phase-2 (ADNI-2) dataset. The extensive experimental results demonstrate that the proposed algorithm is quite effective for human brain connectivity modeling of MCI. In addition, our sparse BFCN modeling outperforms the conventional work and will benefit both basic and clinical neuroscience studies.

Keywords: Animal models and imaging, Connectivity analysis, Atlases
Abstract: Diffusion tensor imaging allows to infer brain connectivity from white matter, which can then be investigated aiming at finding possible biomarkers of disease. The usual initial step in graph construction is to identify the nodes in the brain using a predefined atlas. However, atlases are usually not considering the white matter structure. As a result, atlas-based brain parcellation and, hence, brain graphs are not fully considering the white matter organization. In this work, we are proposing an atlas-free scheme to map the structural brain networks. The idea is to identify the nodes in the brain exploiting the white matter structure inferred from the data. We first retrieve the white matter pathways from DTI, grouping fiber tracts into bundles. We then use these pathways in a clustering pipeline to identify the brain regions to map into the graph nodes, which are used to define the brain connectivity. We empirically tested the goodness of the proposed approach on a known case-control study obtaining results confirming findings in related literature.

Keywords: Brain, Magnetic resonance imaging (MRI), Shape analysis
Abstract: The dynamic development of brain cognition and motor functions during infancy are highly associated with the rapid changes of the convoluted cortical folding. However, little is known about how the cortical folding, which can be characterized on different scales, develops in the first two postnatal years. In this paper, we propose a curvature-based multi-scale method using spherical wavelets to map the complicated longitudinal changes of cortical folding during infancy. Specifically, we first decompose the cortical curvature map, which encodes the cortical folding information, into multiple spatial-frequency scales, and then measure the scale-specific wavelet power at 6 different scales as quantitative indices of cortical folding degree. We apply this method on 219 longitudinal MR images from 73 healthy infants at 0, 1, and 2 years of age. We reveal that the change patterns of cortical folding are scale-specific and region-specific. Specifically, at coarser spatial-frequency levels, the majority of the primary folds flatten out, while at finer spatial-frequency levels, the majority of the minor folds become more convoluted. This study provides valuable insights into the longitudinal changes of infant cortical folding.

Keywords: Magnetic resonance imaging (MRI), Brain, Quantification and estimation
Abstract: Quantitative medical imaging often utilizes intensity normalization based on a signal from a neighboring region. The choice of this region can substantially affect the quantification, and is potentially confounded by the presence of pathology or other limitations such as partial volume effect. In this paper we outline the desirable list of criteria for selecting a normalization region of interest, and utilize this approach for quantitative susceptibility mapping (QSM) MRI in a study of neurodegeneration. The proposed criteria includes (i) association between reference region and demographics such as age, (ii) diagnostic group separation effect in the reference region, (iii) correlation between reference and target regions, (iv) local variance in the reference region, and (v) reduced cross-sectional variance within the diagnostic groups using the reference region. The intensity normalization was then evaluated using 119 subjects with normal cognition, mild cognitive impairment and Alzheimer’s disease. For QSM applications in ageing we found that normalizing by the white matter regions not only satisfies the criteria but it also provides the best separation between clinical groups in the brain nuclei regions.

Keywords: Diffusion weighted imaging, Tractography, Brain
Abstract: Brain development is a protracted and dynamic process. Many studies have charted the trajectory of white matter development, but here we sought to map these effects in greater detail, based on a large set of fiber tracts automatically extracted from HARDI (high angular resolution diffusion imaging) at 4 tesla. We used autoMATE (automated multi-atlas tract extraction) to extract diffusivity measures along 18 of the brain’s major fiber bundles in 667 young adults, aged 18-30. We examined linear and non-linear age effects on diffusivity measures, pointwise along tracts. All diffusivity measures showed both linear and non-linear age effects. Tracts with the most pronounced age effects were those that connected the temporal lobe to the rest of the brain. Nonlinear age effects were picked up strongly in the anterior corpus callosum and right temporo-parietal tracts.

Keywords: Diffusion weighted imaging, Connectivity analysis, Brain
Abstract: Traumatic brain injury (TBI) is the leading cause of death and disability in children, and can lead to long lasting functional impairment. Many factors influence outcome, but imaging studies examining effects of individual variables are limited by sample size. Roughly 20-40% of hospitalized TBI patients experience seizures, but not all of these patients go on to develop a recurrent seizure disorder. Here we examined differences in structural network connectivity in pediatric patients who had sustained a moderate-severe TBI (msTBI). We compared those who experienced early post-traumatic seizures to those who did not; we found network differences months after seizure activity stopped. We also examined correlations between network measures and a common measure of injury severity, the Glasgow Coma Scale (GCS). The global GCS score did not have a detectable relationship to brain integrity, but sub-scores of the GCS (eyes, motor, verbal) were more closely related to imaging measures.

Keywords: Computer-aided detection and diagnosis (CAD), Machine learning, Breast
Abstract: Automatic detection and classification of lesions in medical images is a desirable goal, with numerous clinical applications. In breast imaging, multiple modalities such as X-ray, ultrasound and MRI are often used in the diagnostic work-flow. Training robust classifiers for each modality is challenging due to the typically small size of the available datasets. We propose to use cross-modal transfer learning to improve the robustness of the classifiers. We demonstrate the potential of this approach on a problem of identifying masses in breast MRI images, using a network that was trained on mammography images. Comparison between cross-modal and cross-domain transfer learning showed that the former improved the classification performance, with overall accuracy of 0.93 versus 0.90, while the accuracy of de-novo training was 0.94. Using transfer learning within the medical imaging domain may help to produce standard pretrained shared models, which can be utilized to solve a variety of specific clinical problems.

Keywords: Breast, Computer-aided detection and diagnosis (CAD), X-ray imaging
Abstract: Breast mass detection and segmentation are difficult tasks due to the variation in size and shape of breast masses. Constructing classifiers for this problem is also challenging due to the fact that normal tissue regions overwhelmingly outnumber abnormal regions. In this paper, we propose a novel approach for detecting and segmenting breast masses in mammography based on multi-scale morphological filtering and a self-adaptive cascade of random forests (CasRFs). CasRFs can cope with severe class imbalance by adding layers to the cascade until a minimum number of false-positives (FPs) is reached. The approach achieves an average sensitivity of 0.94 with 1.99 FPs/image on INbreast and a sensitivity of 0.77 with 3.93 FPs/image on DDSM BCRP.

Keywords: Computer-aided detection and diagnosis (CAD), Computational Imaging, Computed tomography (CT)
Abstract: A major functional role of the nasal cavity is air conditioning of the inspired environmental air to nearly alveolar conditions in order to protect the moist and warm alveolar lining where gas exchange takes place. It is well known that the morphological disparities among nasal passages can change airflow patterns to a great extent. However, this morphological variation effect on air conditioning performance remains unclear. This research aims to investigate the nasal air conditioning performance among nasal models with distinct vestibule phenotypes across subjects with normal healthy nasal anatomy and with symptoms of nasal disease. Numerical simulations considering humidified air with room temperature condition was conducted. Airflow pattern, heat and mass transfer between the inhaled airflow and the nasal mucosa were analysed and compared among selected nasal models. Results showed that the nasal air conditioning performance is closely related to nasal structures, and its morphological variations, especially in the nasal vestibule region, can induce up to 25% relative humidity, 3 ℃ temperature changes in anterior nasal region.

Keywords: Skin, Computer-aided detection and diagnosis (CAD)
Abstract: We have designed a content based image retrieval (CBIR) system for dermoscopic images focusing on images with pigment networks. The system locates and matches a query image that has a pigment network with the most similar images containing pigment networks in a database of dermoscopic images. Dermoscopy interest points in the query image are detected and a vector of 128 features is extracted as the descriptor from each keypoint. Then, the descriptors are matched according to our matching algorithm to similar features arising in the database images. This leads to a meaningful matching as we are matching similar dermoscopy structures with each other. The performance of the system has been tested on more than 1000 images. Results show that our system will locate and retrieve similar images with pigment networks, with accuracy > 75.4%. This system can help physicians in diagnosis as they are shown similar looking dermoscopy images with known pathology.

Keywords: Atlases, Diffusion weighted imaging, Brain
Abstract: Spatio-temporal atlas is a useful tool in imaging studies of neurodevelopment, which characterizes the growth of brain, and allows the monitoring of its development. The imaging of preterm and term born infants provides opportunities to develop a series of spatio-temporal atlases that track the changes during the particular period of neurodevelopment between. The aim of this paper is to develop a spatio-temporal atlas of diffusion MRI for neonatal brains between 32 to 42 weeks postmenstrual age (PMA). We subdivided the cohort consisting of preterm- and term-born infants according to their PMA at the MRI scan based on a kernel ridge regression, and generated the atlases based on Fibre Orientation Distribution (FOD) reconstruction of the diffusion data.

Keywords: Machine learning, Computed tomography (CT), Abdomen
Abstract: In this paper, we propose new prognostic methods that predict 5-year mortality in elderly individuals using chest computed tomography (CT). The methods consist of a classifier that per- forms this prediction using a set of features extracted from the CT image and segmentation maps of multiple anatomic structures. We explore two approaches: 1) a unified frame- work based on two state-of-the-art deep learning models ex- tended to 3-D inputs, where features and classifier are automatically learned in a single optimisation process; and 2) a multi-stage framework based on the design and selection and extraction of hand-crafted radiomics features, followed by the classifier learning process. Experimental results, based on a dataset of 48 annotated chest CTs, show that the deep learn- ing models produces a mean 5-year mortality prediction AUC in [68.8%,69.8%] and accuracy in [64.5%,66.5%], while radiomics produces a mean AUC of 64.6% and accuracy of 64.6%. The successful development of the proposed mod- els has the potential to make a profound impact in preventive and personalised healthcare.

Keywords: Nerves, Computer-aided detection and diagnosis (CAD)
Abstract: Unilateral peripheral facial paralysis (UPFP) is a form of facial nerve paralysis and clinically classified according to facial asymmetry. Prompt and precise assessment is crucial to the neural rehabilitation of UPFP. For UPFP assessment, most of the existing assessment systems are subjective and empirical. Therefore, an objective assessment system will help clinical doctors to obtain a prompt and precise assessment. Distinguishing precisely between degrees of asymmetry is hard using pure pattern recognition methods. Thus, a novel objective assessment process based on convolutional neuronal networks is proposed in this paper that provides an end-to-end solution. This method could alleviate the problem and produced a classification accuracy of 91.25% for predicting the House-Brackmann degree on a given UPFP image dataset.

Keywords: Classification, Optical coherence tomography, Eye
Abstract: Age-related macular degeneration (AMD) is a major cause of irreversible blindness and loss of vision in people over 50 years old. Fluid (or cyst) regions such as intraretinal fluid (IRF), subretinal fluid (SRF), and sub-retinal pigment epithelium (sub-RPE), have different tissue appearance in Optical Coherence Tomography (OCT) images compared to normal retina tissue and are a defining feature of AMD. However, diagnosis of AMD requires an expert to go through every slice (B-scan) of the OCT volume to find whether it contains fluid. This is a tedious and time consuming task. In this paper, we proposed a new framework to help diagnosing AMD via automatically detecting the B-scan frames containing fluid regions. By converting each slice into a bag of features, we introduce multiple instance learning algorithm for the B-scan classification. Cross validation experiments show that the result of our framework have a good classification accuracy with F-measure above 0.85 and the multiple instance random forest algorithm we proposed outperforms other state-of-the-art algorithms.

Keywords: Image enhancement/restoration(noise and artifact reduction), Computed tomography (CT), Machine learning
Abstract: To reduce the potential radiation risk, low-dose CT has attracted much attention. However, simply lowering the radiation dose will lead to significant deterioration of the image quality. In this paper, we propose a noise reduction method for low-dose CT via deep neural network without accessing original projection data. A deep convolutional neural network is trained to transform low-dose CT images towards normal-dose CT images, patch by patch. Visual and quantitative evaluation demonstrates a competing performance of the proposed method.

Keywords: Computed tomography (CT), Abdomen, Classification
Abstract: Radiomics, an emerging field of quantitative imaging, encompasses a broad class of analytical techniques. Recent literature have interrogated associations between quantitatively derived GLCM-based texture features and clinical/pathology information using machine learning algorithms in many cancer settings, but often fail to elucidate the predictive power of these features. Moreover, for many cancers characterized by complex histopathological profiles, such as adrenocortical carcinoma, reducing the multivariate functional structure of GLCM to a set of summary statistics is potentially reductive, masking the patterns that distinguish malignancy from benignity. We develop a Bayesian probabilistic framework for predictive classification of lesion types, based on the entire GLCM. Our method, which uses a spatial Gaussian random field to model dependencies among neighboring cells of the GLCMs, was applied in a cancer detection context to discriminant malignant from benign adrenal lesions using GLCMs arising from non-contrast CT scans. Our method is shown to yield improved predictive power both in simulations as well as the adrenal CT application when compared to state-of-the-art diagnostic algorithms that use GLCM derived features.

Keywords: Histopathology imaging (e.g. whole slide imaging), Breast, Computer-aided detection and diagnosis (CAD)
Abstract: Differential diagnosis of proliferative breast lesions into the benign form of usual ductal hyperplasia (UDH) or the malignant form of ductal carcinoma in situ (DCIS) remains a big challenge. The usual diagnostic step is to undertake a manual light microscopic examination of histopathologic biospy sections, paying particular attention to architectural patterns, their sizes and locations, in addition to descriptive morphometric properties of nuclear atypia. Imposing diagnostic boundaries on features that otherwise exist on a continuum going from benign to atypia to malignant is thus a challenge. Current computational pathology methods have focused primarily on nuclear atypia in drawing these boundaries. In this paper, we improve on these approaches by encoding for both cellular morphology and spatial architectural patterns. Using a publicly available breast lesion database consisting of UDH and three different grades of DCIS, we improve the classification accuracy by 10% over the state-of-the-art method for discriminating UDH and DCIS. For the four way classification of UDH and the three grades of DCIS, our method improves the results by 6% in accuracy, 8% in micro-AUC, and 19% in macro-AUC.

Keywords: Histopathology imaging (e.g. whole slide imaging), Computer-aided detection and diagnosis (CAD), Breast
Abstract: In the field of pathological image analysis, the generation of appropriate training data set is significant yet difficult. As a solution, this paper addresses a novel unsupervised region proposal method for histopathological whole slide image based on Selective Search. Specifically, the method utilizes multiple magnifications, modifies the similarity measure for grouping regions and proposes a new Nucleus-Cytoplasm color space to gain more region proposals. Additionally, a new retrieval acceleration method based on Latent Dirichlet Allocation and Supervised Hashing is raised. The experiments on a large-scale multi-class database of breast histopathological images show the effectiveness and efficiency of our methods.

Keywords: Histopathology imaging (e.g. whole slide imaging), Pattern recognition and classification, Gastrointestinal tract
Abstract: The development of reliable imaging biomarkers for the analysis of colorectal cancer (CRC) in hematoxylin and eosin (H&E) stained histopathology images requires an accurate and reproducible classification of the main tissue components in the image. In this paper, we propose a system for CRC tissue classification based on convolutional networks (ConvNets). We investigate the importance of stain normalization in tissue classification of CRC tissue samples in H&E-stained images. Furthermore, we report the performance of ConvNets on a cohort of rectal cancer samples and on an independent publicly available dataset of colorectal H&E images.

Keywords: Multi-modality fusion, Surgical guidance/navigation, Vessels
Abstract: Response to Cardiac Resynchronization Therapy (CRT) may be improved if coronary venous anatomy is fused with information on regional left ventricular (LV) function to guide LV lead placement. We propose a method to register a 3D model of the LV and its corresponding coronary venous tree (reconstructed from X-ray fluoroscopy) for CRT applications. A template coronary venous tree (CVT) centerline, learned from a training set, was utilized to find point-to-point correspondence between the LV surface and the coronary vessels. The method was validated retrospectively on a set of 15 CRT candidates. Two validation experiments were conducted; on a randomly relocated patient specific LV and on a randomly relocated average LV.

The absolute distance error between the ground truth and the registered LV was 3.4±1.0 mm for the case of patient specific LV and 4.2±1.0 mm for the case of average LV model. The probability of targeting the correct coronary branch, when using the registered LV compared to the ground truth, was 95.3±2.5 % for the case of patient specific LV and 94.5±3.5 % for the case of average LV model. The results suggest that the method can be used to provide an automatic registration of a 3D LV model to X-ray imaging with potential applications in image fusion to aid image-guided CRT implantation.

Keywords: Surgical guidance/navigation, Image registration, Medical robotics
Abstract: Rigid point-based registration technique is widely used in all aspects of image-guided surgery (IGS). Estimating target registration error (TRE) statistics considering a coordinate reference frame (CRF) is an essential issue for applications such as optical tracking and image registration. In this paper, we extend the TRE estimation algorithm relative to a CRF to general fiducial localization error (FLE) case. The Monte Carlo simulation results demonstrate the algorithm's effectiveness for both homogeneous and inhomogeneous FLE cases.

Keywords: Functional imaging (e.g. fMRI), fMRI analysis, Validation
Abstract: A comparative investigation of motion correction techniques was performed in high and low motion fMRI data sets. Independent component analysis performed best.

Keywords: Ultrasound, Elastography imaging, Heart
Abstract: This study mapped transverse shear strains in the interventricular septum (IVS) of an in vivo open-chest porcine heart. Ultrasound coherent diverging wave compounding was employed to improve the radiofrequency signal quality in full-view at 250 fps, and cardiac systolic strains were estimated using our previously developed ultrasound strain imaging method. The two estimated transverse shear strains ―radial-circumferential and radial-longitudinal ―were found different on the left and right sides of IVS, thus indicating distinct shearing behavior.

Keywords: Classification, Prostate, Histopathology imaging (e.g. whole slide imaging)
Abstract: In this study, we apply deep convolutional neural networks (DCNN) to analyze prostate tissue specimen images and to detect prostate cancers. Segmenting individual nuclei within a tissue image, nuclear centroids and size information is obtained and used to learn the high-level feature representation of nuclei via DCNN. Tissue microarrays are adopted to evaluate the performance of DCNN in detecting prostate cancers. An AUC of 0.97 is obtained.

Keywords: Virtual/augmented reality, Brain, Angiographic imaging
Abstract: Purpose: To demonstrate the feasibility of augmented reality (AR) by means of a new head-mounted device for improved visualization of medical image data.

Methods: Neurovascular image data was imported into the developer edition of a head-mounted AR device (Hololens, Microsoft). Image data was acquired from six patients with aneurysms of the internal carotid artery (ICA) that underwent an MRI examination and a diagnostic 3D digital angiographic subtraction (DSA) examination prior to treatment. 3D representations of the Circle –of-Willis were segmented from the MRI time-of-flight (TOF) and 3D DSA data. 3D aneurysm models were extracted. Computational fluid dynamics (CFD) simulations were carried out with a prototype device (Siemens AG) to calculate the intra-aneurysmal flow field. An image pipeline was developed for importing the image data into the AR device (figure 1) and for displaying them with the 3D Viewer app included with the AR device.

Results: Visualization succeeded for all imaging modalities as well as the simulated data (figure 1). 3D models were scaled and moved in the augmented space, which allowed viewing anatomical structures together with fluid streamlines from inside- out thereby greatly enhancing the understanding of anatomical and flow complexity. Visualization was restricted by size limitation of the imported scenes imposed by the AR device.

Conclusion: Visualization of complex 3D structures derived from medical image data succeeded with a head-mounted augmented reality device. Potential applications of this device include education and virtual planning. Acceptance would be increased by improved user interaction (gestures) and removing size limitations of the 3D models.

Keywords: Magnetic resonance imaging (MRI), Brain, Probabilistic and statistical models & methods
Abstract: Analysing brain tumor tissue composition can improve the handling of tumor growth and resistance to therapies. We show on a 6 time point dataset of 8 rats that multiparametric MRI can be exploited via statistical clustering to quantify intra-lesional heterogeneity in space and time.

Keywords: Functional imaging (e.g. fMRI), Diffusion weighted imaging, Brain
Abstract: The thalamus, situated on the top of the brainstem near the center of the brain, a central relay station that transfers information of the brain[1, 2]. Thalamus infarction is a common cerebral disorder, which impairs or destruct the function of thalamus and information transfer with the cortex. Recently functional tensor imaging (FTI) has been proposed to characterize fMRI signals in brain white matter by capturing directional variations in their temporal correlations[3]. This provides a new view to understand the pathophysiology of, and identify biomarker for, thalamus infarction. In this work, white matter pathways connecting thalamus were constructed to examine whether fMRI signals along them were disrupted by thalamus infarction and how well FTI characterizes the underline functional variation along these white matter pathways in the thalamus infarction patients. Data acquisition: All MRI data were obtained on a 3T MRI scanner (Magnetom Tim Trio; Siemens, Erlangen, Germany). Resting-state fMRI image data were acquired using gradient-echo, echo-planar imaging sequence with TR/TE=3000/30ms, data size=64×64, slice thickness=3mm, gap=0mm, slices=43, and 124 dynamics. Image preprocessing: fMRI data were corrected for slice timing and subject’s head motion, and spatially smoothed at FWHM=4mm. The corrected images were then normalized to MNI space. All these procedures were implemented with SPM software.Functional pathway characterization: Fiber pathways were constructed using probabilistic tracking, which begins from randomly sampled voxels within the thalamus. The fiber tracking was repeated 1,000,000 times and those connecting any of the 47 Brodmann areas were retained. Functional tensors of all voxels along these fibers were calculated, from which functional fractional anisotropy (FFA) was derived . FFA of 47 fiber bundles connecting the Brodmann areas in each hemisphere was compared between 8 patients and 15 healthy subjects using Student’s-t test. Three fiber bundles in the ispilateral hemisphere of the patient group exhibited significantly decreased FFA, as shown in Fig.1 below.Three white matter fiber bundles connecting thalamus in the ispilateral hemisphere demonstrate significant decrease of FFA for the thalamus infarcted patients. This finding indicates that disease such as thalamus infarction interrupt the function of relevant white matter pathway.

Keywords: Brain, Magnetic resonance imaging (MRI), Machine learning
Abstract: The classification of mild cognitive impairment (MCI) subtypes is considered important because MCI is a prodromal condition to dementia and early identification can enable corrective treatments when it becomes available. There are accepted diagnosis criteria for MCI that are operationalized variably, resulting in different progression rates to dementia from MCI. However, progression to dementia from MCI appears to be high compared to progression from normal cognition. MCI is divided into two major subtypes: (i) amnestic MCI (aMCI) where memory is impaired and (ii) non-amnestic MCI (naMCI) where one or more non-memory domains are impaired. A lot of work has been carried out in this area using magnetic resonance images (MRI) to discriminate between MCI subtypes, cognitively normal individuals and those who are demented. Conventional methods as well as deep learning methods have been explored. In this work, we propose the use of generative adversarial networks (GAN) for this purpose that addresses some drawbacks of the previous methods[5]. A major drawback when using deep learning techniques is the paucity of data. We propose to overcome this problem by using the generative component of GAN in combination with a convolutional neural network (CNN) to come up with new samples, thereby effectively enlarging the datasets we have. In addition, the proposed GAN can actually be used to improve the spatial resolution of the MR images without additional complications such as compromising the temporal resolution. We discuss the methods and materials in the next section and conclude this abstract with the last section.

Keywords: Nuclear imaging (e.g. PET, SPECT), Brain, Cells & molecules
Abstract: This study sought to examine the specificity of an EphA2-based theranostic within gliomas, by examining the spatial correlation of 64Cu-EphA2 and 18F-DOPA PET tracers. Performing two PET scans is not feasible, so scanning simultaneously using a dual-tracer imaging technique was investigated and utilized. To compensate for the long biological half-life of the 64Cu-EphA2 tracer, it was injected first and modelled using the dynamic signal prior to the second tracer being injected after an optimal interval. The extrapolated 64Cu-EphA2 signal was subtracted from the post-injection data. Correlations between the two tracers were significant (p<10-6): the biodistribution of 64Cu-EphA2 indicates a 64Cu-EphA2 therapeutic is feasible.

Keywords: Quantification and estimation, Image registration, Spine
Abstract: Spine Scoliosis is a serious disease nowadays especially adolescent idiopathic scoliosis (AIS) in teenagers. Their spine morphology is quite different from normal range in adult ones. To have a more comprehensive comparison and classification of scoliotic patients with the normal population, this project aimed to investigate normal range of transformation parameters between neighbouring vertebra along an adult normal population so to set up a ground standard for later comparison with scoliosis patients. Registration between neighbouring spine was adopted to obtain translation, rotation, scaling parameters as previous researches on robust registration of the spine images have been mature.

Keywords: Image segmentation, Magnetic resonance imaging (MRI), Brain
Abstract: Segmentation of brain magnetic resonance images is often done using a mixture model, with a Markov random field (MRF) to add spatial smoothness. The smoothing parameter is specified manually, but it is not clear a priori what it should be. Additionally, the smoothing is applied uniformly across the image. We propose an MRF with smoothing dependent on the current local classification, and parameters fitted by maximum pseudolikelihood. This unsupervised method determines the smoothing adaptively, and has desirable statistical and numerical properties. We demonstrate it on a standard dataset.

Keywords: Nuclear imaging (e.g. PET, SPECT), Lung, Image reconstruction - analytical & iterative methods
Abstract: Multiplexed PET (mPET) is a new imaging technique compatible with any existing PET scanner which allows simultaneous imaging of two PET tracers in vivo. In this work we present initial results of a new procedure to measure liquid absorption (LA) in the airways, a promising biomarker for evaluating treatment response in lung diseases. We performed dynamic PET imaging in the lungs of 10 pigs in a clinical PET scanner using inhaled 13N-NH3 and 76Br-Albumin as tracers. List mode data from the scanner was analyzed using a custom-developed software to classify events into double- (coming from both tracers) and triple-coincidences (mostly coming from the 76Br-labeled tracer). Both datasets were used to obtain separated dynamic images of each tracer, thus allowing measurement of time-activity curves in the lungs. LA was obtained based on the differences in the clearance of these tracers in the airways. Image quality and quantitative accuracy of this approach has been found to be much better than that achievable using existing approaches such as dual-tracer scintigraphy.

Keywords: Image segmentation, Magnetic resonance imaging (MRI)
Abstract: Alzheimer’s Disease (AD) is characterized by the presence of extracellular Aβ-amyloid (Aβ) plaques, and intracellular neurofibrillary tangles (NFT) of hyperphosphorylated tau, with associated neurodegeneration of the hippocampus, basal forebrain, and the cortex. Tau and Aβ pathologies can now be imaged in vivo with PET. In this study, we investigated if the presence of high tau cortical retention in cognitively normal elderly subjects is associated with early signs of neurodegeneration, as assessed with volumes of the posterior nucleus basalis of Meynert (NBM). We found that subjects with high tau burden on PET scanning had significantly smaller posterior NBM volumes.

Keywords: Image segmentation, Machine learning, Abdomen
Abstract: An automated segmentation method is presented for multi-organ segmentation in abdominal computed tomography (CT) images. We used a fully convolutional network (FCN) with the multi-planar reconstruction (MPR) approach. For axial, sagittal, and coronal planes, three FCNs were independently trained to produce semantic segmentation. And then, three distinct segmentation results were merged by applying majority voting. Our proposed method has been evaluated on a test dataset of 10 abdominal CT images and achieves a promising segmentation performance with dice similarity coefficients (DSCs) of 0.93, 0.91, 0.85, and 0.86 for liver, spleen, left- and right-kidney, respectively.

Keywords: Computer-aided detection and diagnosis (CAD), Machine learning, Computed tomography (CT)
Abstract: We propose a deep feature classification method for small renal masses in abdominal CT images to differentiate fat-poor angiomyolipoma from clear cell renal cell carcinoma. We extract deep features from pre-trained convolutional neural network model with the mass-size input patches, and train a random forest classifier to classify two types of masses. To train the classifier invariant to mass size, we propose to use mass-size input patches instead of conventional fixed-size patches. In experiments, our method not only outperformed hand-crafted features, but also improved the classification accuracy compared to deep features with fixed-size image patches.

Keywords: Image quality assessment, Brain, Diffusion weighted imaging
Abstract: Diffusion weighted MRI is affected by artefacts caused by movement between the odd and even subvolumes in an interleaved acquisition. These volumes need to be identified and removed from analysis. We use a registration-based approach to identify affected volumes, and demonstrate that a single metric calculated from all subjects acquired using the same protocol is sufficient for reliable identification.

Keywords: Magnetic resonance imaging (MRI), Brain, Image reconstruction - analytical & iterative methods
Abstract: The image processing technique for magnetic resonance image (MRI) is discussed in this paper. In magnetic resonance imaging equipment, the phase difference of a proton longitudinal relaxation signal is changed into a difference in temperature, and the estimation of internal temperature change inside of the body becomes possible. Although measurement values become more accurate as the intensity of static magnetic field is higher, this paper assumes the intensity of 0.3 [T] magnetic field which is comparatively low. Then the purpose of the paper is to develop an image processing procedure which can measure internal change of the temperature under the condition of low intensity of magnetic field. This paper is organized as follows. In the first step, the phantom models which show high signal in T1 weighted images are constructed, then estimated temperature change by MRI and actual measurement by a fiber-optic thermometer are compared. The error range of the temperature is examined experimentally, which lead to a suitable parameter settings of the instrument. The second step, internal temperature measurements are calculated for a brain. The temperature at both sides of prefrontal cortex are measured, during the subject performs paced auditory serial addition test (PASAT) which is developed by Gronwall to evaluate and judge attention disorders in neuropsychology. The temperature rise in brain, caused by heavy load of the test, are measured and evaluated from the standpoint of neuropsychology. The third step, image processing procedure for temperature measurement by 0.3 [T] MRI are proposed and considered for practical use. The theme of the paper belongs to preliminary studies and future development is expected in medical scene.

Keywords: Image archiving, Magnetic resonance imaging (MRI)
Abstract: Data curation is an important first step in large-scale medical imaging studies. Manual curation is both time consuming and error prone. We have developed a flexible python framework composed of plugins for data curation. Our framework incorporates a machine readable protocol description to match and check that the data conform to expectation. By configuring different plugins and parameters, we can achieve project specific behavior while reusing components and achieving standardization.

Keywords: Computational Imaging, Compressive sensing & sampling, Image reconstruction - analytical & iterative methods
Abstract: Sparse volumetric reconstruction aims to recover an unknown signal of interest from a limited portion of its transform spectrum using a nonlinear sparsity-promoting iterative algorithm. A typical iterative sparse recovery algorithm can be regarded as estimation of a signal from a degraded observation at each iteration. Existing methods commonly address such degradations by modeling them as independent and identically distributed (i.i.d.) noise, leveraging a denoiser to alleviate the signal degradations at each iteration -- a strategy becoming popular as the plug&play approach. The i.i.d. noise modeling neglects any spatial correlation in the noise. In contrast to white noise, correlated noise can lead to disproportions in the magnitude of errors across the data spectrum, to an extent that i.i.d. denoisers may not effectively discern between the true signal and noise in regularization via shrinkage. Existence of strong spatial correlation in volumetric residuals at each iteration motivates us to model such degradations as spatially correlated noise. We demonstrate the effectiveness of modeling spatial correlation of the degradation within individual iterations through the reconstruction of volumetric phantom data from noisy and incomplete Fourier-domain measurements, a standard setup often used as a proxy for many medical inverse imaging problems. We adaptively estimate the correlation model from the degraded data by calculating the power spectral density of the residuals, finding order-of-magnitude differences across the spectrum. The obtained results on iterative reconstruction of non-zero phase volumetric brain imagery from incomplete noisy radial measurements attest the practical advantage of assuming spatially correlated over i.i.d. residuals in iterative volumetric data reconstruction. In particular, the proposed modeling leads to higher quality reconstruction, with sharper recovery of fine structures.

Keywords: Nuclear imaging (e.g. PET, SPECT), Image reconstruction - analytical & iterative methods, Parallel computing
Abstract: Focused beamformers have been extensively used in phased- array signal processing, leading to simple and efficient imaging proce- dures, with high sensitivity and resolution. The beamshape acts as a spatial filter, scanning the intensity of the incoming signal for particular locations. We introduce beamforming in the context of Positron Emission Tomography (PET), and propose a new beamformer called Sinobeam. Inspired by the Flexibeam framework [1], we sample the beamform- ing weights from an analytically-specified beamforming function. Since the weights are data-independent, the resulting imaging algorithm is extremely efficient, while presenting better resolution and contrast than state of the art methods as demonstrated by simulation.

Keywords: Vessels, Optical coherence tomography, Image segmentation
Abstract: Choroid is an important retinal tissue which is mainly comprised of vessels. The segmentation and quantification of choroidal vessel can provide clinicians with accurate information which can assist clinical analysis, diagnosis and treatment. In this paper, an automated choroidal vessel segmentation and quantification method is developed for 3D 1050nm wide-view OCT images.

Keywords: Heart, Ultrasound
Abstract: We conducted study in which patients were recruited to undergo measurements of two echocardiographic markers: pulsed wave tissue Doppler, and speckle tracked velocities for septal and lateral mitral annuli. Evaluation of ventricular motion might need to include all available markers where feasible, and recognise that tissue Doppler shows the highest feasibility.

Keywords: Image segmentation, X-ray imaging, Lung
Abstract: In this study we present a lung field segmentation system using fully convolutional network architecture called “U-Net” which yielded promising results with mean Dice similarity index of 0.96±0.03.

Keywords: Diffusion weighted imaging, Magnetic resonance imaging (MRI), Brain
Abstract: Axonal degeneration is a key pathological driver of disability in multiple sclerosis (MS). Treatments aiming to reduce or reverse axonal degeneration in MS require sensitive and specific markers. Here we explore fibre-specific markers of axonal degeneration based on diffusion-weighted MRI metrics – fibre density and fibre bundle cross-section. MS patients with optic neuritis were compared to control subjects. We identified significant reductions to both fibre density and cross-section in the visual pathways of patients. These results indicate the pathological specificity of fibre density and cross-section measures in MS.

Keywords: Computer-aided detection and diagnosis (CAD), Vessels, Ultrasound
Abstract: Carotid adventitial vasa vasorum (AVV) are markers of subclinical atherosclerotic lesions. Early lesion identification is essential in reducing the high mortality risk from cardiovascular diseases (CVD). In this work we present an automated method to process contrast ultrasound images (CUI) of carotid, identify AVV, and estimate relative risk of CVD events.

Keywords: Brain, Magnetic resonance imaging (MRI), Other-method
Abstract: Here we measure regional variation in the ‘evolvability’ (capacity for adaptive evolution) of thickness in the human cerebral cortex. The evolvability of a trait is reflected in its absolute genetic variation (as opposed to heritability estimates, which are relative to total variation). This can be indexed by regional differences in the mean-standardised additive genetic variance, which have not previously been studied in the brain. We show that cortical evolvability increases along a ventral-dorsal axis.

Keywords: Nuclear imaging (e.g. PET, SPECT), Machine learning, image filtering (e.g. mathematical morphology, wavelets,...)
Abstract: Positron Emission Tomography (PET) using 18F-FDG is recognized as the modality of choice for lymphoma, due to its high sensitiv- ity and specificity. Its wider use for the detection of lesions, quantification of their metabolic activity and evaluation of response to treatment demands the development of accurate and reproducible quantitative image interpretation tools. An accurate tumour delineation remains a real challenge in PET, due to the limitations the modality suffers from, despite being essential for quantifying reliable changes in tumour tissues. Due to the spatial and spectral properties of PET images, most methods rely on intensity-based strate- gies. Recent methods also propose to integrate anatomical priors to improve the segmentation process. However, the current routinely- used approach remains a local relative thresholding and requires important user interaction, leading to a process that is not only user-dependent but very laborious in the case of lymphomas. In this paper, we propose to rely on hierarchical image models embedding multimodality PET/CT descriptors for a fully automated PET lesion detection / segmentation, performed via machine learning process. More precisely, we propose to perform random forest classification within the mixed spatial-spectral space of component-trees modeling PET/CT mages. This new approach, combining the strengths of machine learning and morphological hierarchy models leads to intelligent thresholding based on high-level PET/CT knowledge. We evaluate our approach on a database of multi-centric PET/CT images of patients treated for lymphoma, delineated by an expert. Our method proves good efficiency, with the detection of 92% of all lesions, and accurate segmentation results with mean sensitivity and specificity of 0.73 and 0.99 respectively, without any user interaction.

Keywords: Endoscopy, Gastrointestinal tract, Computer-aided detection and diagnosis (CAD)
Abstract: Wireless Capsule Endoscopy (WCE) is a novel diagnostic modality of endoscopic imaging which facilitates direct visualization of the gastrointestinal (GI) tract. Many computational methods that can automatically detect and/or characterize the abnormalities from WCE sequences are developed to support medical decision-making. This paper presents a new approach for automated segmentation of blood regions in WCE images via a deep learning strategy. The proposed method first classify the bleeding samples into active and inactive subgroups based on the statistical features derived from the histogram probability of the color space. Then for each subgroup, we highlight the blood regions via fully convolutional networks (FCNs). Experimental results on the clinical WCE dataset demonstrate the efficacy of our approach, which achieves comparable or better performance than the state-of-the-art methods.

Keywords: Cells & molecules, Pattern recognition and classification, Microscopy - Light, Confocal, Fluorescence
Abstract: Quantitative analysis of vesicle-plasma membrane fusion events in the fluorescence microscopy, has been proven to be important in the vesicle exocytosis study. In this paper, we present a framework to automatically detect fusion events. First, an iterative searching algorithm is developed to extract image patch sequences containing potential events. Then, we propose an event image to integrate the critical image patches of a candidate event into a single-image joint representation as the input to Convolutional Neural Networks (CNNs). According to the duration of candidate events, we design three CNN architectures to automatically learn features for the fusion event classification. Compared on 9 challenging datasets, our proposed method showed very competitive performance and outperformed two state-of-the-arts.

Keywords: Pattern recognition and classification, Bone, X-ray imaging
Abstract: A panoramic radiography image provides not only details of teeth but also rich information about trabecular bone. Recent studies have addressed the correlation between trabecular bone structure and osteoporosis. In this paper, we collect a dataset containing 40 images from 40 different subjects, and construct a new methodology based on a two-stage classification framework that combines multiple trabecular bone regions of interest (ROIs) for osteoporosis prescreening. In the first stage, different support vector machines (SVMs) are adopted to describe different information of different ROIs. In the second stage, the output probabilities of the first stage are effectively combined by using an additional linear SVM model to make a final prediction. Based on our two stage model, we test the performance of different image features by using leave-one-out cross-valuation and analysis of variance rules. The results suggest that the proposed method with the HOG (histogram of oriented gradients) feature achieves the best overall accuracy.

Keywords: Magnetic resonance imaging (MRI), Brain, Image segmentation
Abstract: Automated segmentation of brain structures from MR images is an important practice in many neuroimage studies. In this paper, we explore the utilization of a multi-view ensemble approach that rely on neural networks to combine multiple decisions maps in achieving accurate hippocampus segmentations. Constructed under a general ConvNet structure, our Ensemble-Net networks explore different convolution configurations to capture the complementary information residing in the multiple label probabilities produced by our U-Seg-Net (a modified U-Net) segmentation neural network. T1-weighted MRI scans and the associated Hippocampal masks of 110 healthy subjects from the ADNI were used as both the training and testing data. The combined U-Seg-Net + Ensemble-Net framework achieves above 89% Dice coefficient on ADNI hippocampus segmentation.

Keywords: Image enhancement/restoration(noise and artifact reduction), Magnetic resonance imaging (MRI), Brain
Abstract: Example-based single image super-resolution (SR) has recently shown outcomes with high reconstruction performance. Several methods based on neural networks have successfully introduced techniques into SR problem. In this paper, we propose a three-dimensional (3D) convolutional neural network to generate high-resolution (HR) brain image from its input low-resolution (LR) with the help of patches of other HR brain images. Our work demonstrates the need of fitting data and network parameters for 3D brain MRI.

Keywords: Other-method, Ultrasound
Abstract: Ultrasound speckle tracking provides a robust motion estimation of fine tissue displacements along the beam direction. Extensions to 2-D have been proposed in recent years. Due to relatively coarse lateral sampling, several solutions relied on lateral interpolation in order to achieve subsample accuracy. We introduce a new multi-dimensional speckle tracking method (MDST) with subsample accuracy in all dimensions. The proposed algorithm is based on solving a least squares problem to estimate the coefficients of a second order polynomial expansion to fit the magnitude of the two dimensional complex normalized correlation of the generalized analytic signal in the vicinity of the true peak. The generalization method utilizes the Riesz transform which is the multidimensional Hilbert transform. The displacement is estimated from acquired successive radio-frequency data frames of the region of interest. Field II simulation of flow data in a channel with a bench mark known displacement is generated to validate the accuracy of the method. In addition, the new MDST method is applied to imaging data from a flow phantom (ATS Model 524) to estimate the flow motion and pulsating channel wall. Simulations and experimental results demonstrate the effectiveness of the proposed technique.

Keywords: Probabilistic and statistical models & methods, Microscopy - Light, Confocal, Fluorescence, In-vivo cellular and molecular imaging
Abstract: Microtubules play an essential role in many cellular processes whose disrupted functioning is associated with devastating human diseases such as cancer. The discovery and testing of microtubule targeting drugs often involve time-lapse fluorescence microscopy imaging of microtubule plus-end binding proteins and require highly accurate estimation of their dynamic behavior. Although many methods exist nowadays to perform fully automatic particle tracking in such images, their accuracy and precision are inevitably limited due to noise and other imaging artifacts, and this negatively affects the estimation of parameters such as microtubule growth speed. Here we propose a new approach to estimate such parameters based on Gaussian processes. It naturally deals with measurement noise and can be easily initialized from the data itself. Experimental results on both synthetic and real data demonstrate that our approach indeed yields more accurate estimates of microtubule dynamics.

Keywords: Motion compensation and analysis, Classification, Microscopy - Light, Confocal, Fluorescence
Abstract: Many bio-medical applications involve the analysis of sequences for motion characterization. In this article, we consider sequences where a particular type of motion (e.g. a blood flow) is associated with a particular area (e.g. an artery) but several kinds of motion may exist in the same sequences (e.g. there may be several blood vessels present). The characterization of this type of motion typically involves first finding areas where motion is present, followed by an analysis of these motions: speed, regularity, frequency, etc. In this article, we propose a methodology called "area-of-motion characterization" suitable for simultaneously detecting and characterizing areas where motion is present in a sequence. We can then classify this motion into consistent areas using unsupervised learning and produce directly usable metrics for various applications. We illustrate this methodology for the analysis of cilia motion on ex-vivo human samples, and we apply and validate the same methodology for blood flow analysis in fish embryo.

Keywords: Classification, Modeling - Anatomical, physiological and pathological
Abstract: Obstructive sleep apnea (OSA) is a medical condition in which the airway is repetitively obstructed and resulting in sleep disruption. Previous research has shown that this condition may be the cause or the result of the craniofacial structure, and that specific facial features such as ‘face width’ or ‘eye width’ are correlated with the risk of OSA. In this study we developed two automatic image processing systems that processed facial images and determined the likelihood of a subject having OSA, based on a dataset of photographs from 376 apnea and control subjects. In our first approach, an algorithm was developed to calculate craniofacial photographic features that were previously shown to be useful for OSA discrimination. These features were processed with a logistic classifier and the resulting system achieved an accuracy of 70% in discriminating patients with clinically significant OSA from controls. In our second approach, neural network was designed to automatically process the frontal and profile photographs and classify the patient as a normal or OSA. It achieved an accuracy of 61.8%.

Keywords: Shape analysis, Image registration, Atlases
Abstract: We propose a new formulation of the active surface model in 3D. Instead of aligning a shape dictionary through the similarity transform, we consider more flexible affine transformations and introduce an alignment method that is unbiased in the sense that it implicitly constructs a common reference shape. Our formulation is expressed in the continuous domain and we provide an algorithm to exactly implement the framework using spline-based parametric surfaces. We test our model on real 3D MRI data. A comparison with the classical active shape model shows that our method allows us to capture shape variability in a dictionary in a more precise way.

Keywords: Image segmentation, Atlases, Machine learning
Abstract: Deformable registration based multi-atlas segmentation has been successfully applied in a broad range of anatomy segmentation applications. However, the excellent performance comes with a high computational burden due to the requirement for deformable image registration and voxel-wise label fusion. To address this problem, we conduct an experimental study to investigate trade-off between computational cost and performance by first applying multi-atlas segmentation in coarse spatial resolution and then refining the results by learning-based error correction in the native image space. In a cardiac CT segmentation application, our experiments show that the new combination scheme can significantly reduce computational cost without losing accuracy.

Keywords: Angiographic imaging, Motion compensation and analysis, Vessels
Abstract: Non-recurrent intra-scan motion, such as respiration, corrupts rotational coronary angiography acquisitions and inhibits uncompensated 3D reconstruction. Therefore, state-of-the-art algorithms that rely on 3D/2D registration of initial reconstructions to the projection data are unfavorable as prior models of sufficient quality cannot be obtained. To overcome this limitation, we propose a compensation method that optimizes a task-based autofocus measure using graphical model-based optimization. The proposed algorithm is validated on two numerical phantom data sets and a clinical scan. In the phantom studies, we found a reduction of the root-mean-square error between the true and estimated motion pattern of 82 +- 2% when the proposed method was used, yielding residual errors well below the voxel size. For the clinical data set, we observed a substantially increased amount of voxels with low reprojection errors indicating superior image quality. Our results are promising and suggest that the proposed method effectively handles non-recurrent motion while overcoming the need for prior reconstructions.

Keywords: Motion compensation and analysis, Magnetic resonance imaging (MRI), Nuclear imaging (e.g. PET, SPECT)
Abstract: Head movement is a major issue in dynamic PET imaging. A simultaneous MR-PET scanner is capable of acquiring both MR and PET data concurrently, which enables opportunities to use MR information for PET motion correction. Here we propose an MR-based method to detect head motion and to correct motion artefacts during PET image reconstruction. The method is based on co-registration of multiple MR contrasts to extract motion parameters. The motion parameters are then used to guide the Multiple Acquisition Frame (MAF) algorithm to bin the PET list-mode data into multiple frames whenever significant motion occurs. Furthermore, motion parameters are used to re-align the PET attenuation u-map to each frame prior to the image reconstruction. Finally, PET images are reconstructed for each frame and combined to produce a final image. Using both phantom and in-vivo human data, we show that this method can significantly increase image quality and reduce motion artefacts.

Keywords: Image enhancement/restoration(noise and artifact reduction), Computational Imaging, Computed tomography (CT)
Abstract: Streaking artifacts caused by metallic objects severely affect the visual quality of CT images, resulting in medical mis-diagnosis. Commonly used approaches for metal artifact reduction usually consist of interpolation and iterative methods. The former one tends to lose image quality by introducing extra artifacts, while the latter is more computational expensive. This paper proposes a new approach based on the Euler's elastica inpainting technique, which can preserve sharp edges and curvature when reconstructing the sinogram image, resulting in better quality in the restored CT image. Results of quantitative and qualitative experiments on both simulated phantoms and clinical CT images demonstrate that our method can suppress metal artifacts significantly.

Keywords: Ultrasound, Image enhancement/restoration(noise and artifact reduction), Vessels
Abstract: We present a modification to time-delay and phase-shift estimators used in ultrasound imaging methods. In this modification, we selectively eliminate covariance terms from the estimator corresponding to channel signal pairs that have low coherence. We demonstrate this modification in simulation and phantom experiments in the application of color flow imaging. We demonstrate this technique in application to color flow imaging in simulations and a tissue-mimicking phantom. In these experiments, we show that elimination of the low coherence terms reduces velocity estimation jitter (standard deviation of the error).

Keywords: Image enhancement/restoration(noise and artifact reduction), Image registration, Animal models and imaging
Abstract: Functional optical imaging (OI) of intrinsic signals (like blood oxygenation coupled reflection changes) and of extrinsic properties of voltage sensitive probes (like voltage-sensitive dyes (VSD)) forms a group of invasive neuroimaging techniques that possess up to date the highest temporal and spatial resolution on a meso- to macroscopic scale.

The main challenge for OI-based functional studies is a very low signal to noise ratio (SNR). We propose a flow guided, selective averaging approach for a multi-trial experimental design. We use an optic flow based energy measure for a weighted averaging across the trials. Our approach is translation and illumination invariant and suited to remove periodic artefacts with less required samples for averaging.

For benchmarking of OI denoising strategies, we present a novel, semi-synthetic dataset, that simulates a continuous recording of intrinsic signals over ten trials at one second each with alternating baseline and stimulus condition.

Keywords: Retinal imaging, Image segmentation, Vessels
Abstract: Automatic segmentation of retinal blood vessels from fundus images plays an important role in the computer aided diagnosis of retinal diseases. The task of blood vessel segmentation is challenging due to the extreme variations in morphology of the vessels against noisy background. In this paper, we formulate the segmentation task as a multi-label inference task and utilize the implicit advantages of the combination of convolutional neural networks and structured prediction. Our proposed convolutional neural network based model achieves strong performance and significantly outperforms the state of-the-art for automatic retinal blood vessel segmentation on DRIVE dataset with 95.33% accuracy and 0.974 AUC score.

Keywords: Microscopy - Light, Confocal, Fluorescence, Skin, Image segmentation
Abstract: Reflectance confocal microscopy (RCM) is a powerful tool to visualize the skin layers at cellular resolution. The dermal epidermal junction (DEJ) is a thin complex 3D structure. It appears as a low-contrasted structure in confocal en-face sections, which is difficult to recognize visually, leading to uncertainty in the classification. In this article, we propose an automated method for segmenting the DEJ with reduced uncertainty. The proposed approach relies on a 3D Conditional Random Field to model the skin biological properties and impose regularization constraints. We improve the restitution of the epidermal and dermal labels while reducing the thickness of the uncertainty area in a coherent biological way from 16.9 µm (ground-truth) to 10.3 µm.

Keywords: Skin, Image segmentation
Abstract: Accurate skin lesion segmentation is an important yet challenging problem for medical image analysis. The skin lesion segmentation is subject to variety of challenges such as the significant pattern and colour diversity found within the lesions, presence of various artifacts, etc. In this paper, we present two fully convolutional networks with several side outputs to take advantage of discriminative capability of features learned at intermediate layers with varying resolutions and scales for the lesion segmentation. More specifically, we integrate fine and coarse prediction scores of the side-layers which allows our framework to not only output accurate probability map for the lesion, but also extract fine lesion boundary details such as the fuzzy border, which further improves the lesion segmentation. Quantitative evaluation is performed on the 2016 International Symposium on Biomedical Imaging(ISBI 2016) dataset, which shows our proposed approach compares favorably with state-of-the-art skin segmentation methods.

Keywords: Image segmentation, Probabilistic and statistical models & methods, Liver
Abstract: Medical image segmentation plays an important role in digital medical research, therapy planning, and computer aided diagnosis. However, the existence of noise and low contrast make automatic liver segmentation remains an open challenge. In this work we focus on a novel variational semi-automatic liver segmentation method. First, we used the signed distance functions (SDF) representing pattern shapes to build statistical shape model. Then global Gaussian fitting energy and enforced local feature fitting energy were established to guide the PCA-based topological transformation. We used the unconstrained shape coefficients and geometric transformation parameters to make the proposed method robust in a wide variety of pathological cases. Experiments on two public available datasets demonstrated that the proposed liver segmentation method achieves competitive results to that of the state-of-the-art.

Keywords: Classification, Pattern recognition and classification, Image registration
Abstract: Detection of calcified plaques in coronary arteries is helpful in cardiovascular disease risk assessment. This is often performed by radiologists on computed tomography (CT) images. We work towards an automatic solution for calcium detection in CT images. Most of previous work in this area combines CT and CTA for this purpose to facilitate the localization of the coronary arteries. Given the cost and dose advantages of using only CT scan compared to using both CT and CTA, we propose a solution for automatic calcium assessment in CT. We model the whole chest including all heart chambers and main arteries. Instead of localizing calcium candidates with respect to the coronary artery alone, we assess their position with respect to eight other anatomies, segmented from CT images using joint atlas label fusion methodology. This comprehensive spatial information together with other types of features such as shape, size and texture of each calcium candidate is used with a random forest classifier trained on 104 patients to detect coronary calcification. The results show that our method has a precision of 95.1% and a recall of 89.0% in classifying calcium candidates found based on thresholding. In the patient level, using this method, all the test patients with true calcification were detected as positive, yielding a patient level sensitivity of 100%. Among the test patients without calcification, 44 out of 56 patients resulted in no calcium finding, yielding a patient level specificity of 78.6%. We quantified the whole heart Agatston score for the manual and the automatically detected calcium on the 22 diseased test cases, and found a Pearson correlation coefficient of 0.98. These results show that our proposed framework can reliably detect calcification using CT data.

Keywords: image filtering (e.g. mathematical morphology, wavelets,...), X-ray imaging, Bone
Abstract: The mastoid of human temporal bone contains numerous air cells connected to each others. In order to gain further knowledge about these air cells, a more compact representation is needed to obtain an estimate of the size distribution of these cells. Already existing skeletonization methods often fail in producing a faithful skeleton mostly due to noise hampering the binary representation of the data. This paper proposes a different approach by extracting geometrical information embedded in the Euclidean distance transform of a volume via a structure tensor analysis based on quadrature lters, from which a secondary structure tensor allows the extraction of surface skeleton along with a curve skeleton from its eigenvalues. Preliminary results obtained on a X-ray micro-CT scans of a human temporal bone show very promising results.

Keywords: Elastography imaging, Ultrasound, Vessels
Abstract: Ultrasound shear wave elastography (SWE) is an emerging technique for characterizing local arterial stiffness – a known indicator for vascular health. However, the implications due to vascular anatomy and tissue environment are still relatively under-examined. Using polyvinyl alcohol (PVA) based tissue mimicking phantoms, this study assessed the current signal processing framework in demonstrating the challenges due to the wave dispersion (at the medium thicknesses smaller than the shear wavelength) and wave interference at the interface of different media which cause biased stiffness estimations. Hence, 5% PVA and 10% PVA phantoms of varying thicknesses (from 1 to 10 mm) were imaged when placed in water and in 5% PVA and 10% PVA phantoms. Our results confirmed that shear wave propagation was thickness dependent (315% underestimation in 10% PVA). The shear wave velocity was shown to be influenced by the surrounding media with a 150% overestimation in 5% PVA surrounded by 10% PVA. It also demonstrated a key limitation of arterial SWE in that the current phase velocity estimation does not provide accurate SWV estimation, requiring optimization for addressing wave interference.

Keywords: Machine learning, Fetus, Ultrasound
Abstract: First trimester placental volume measured with 3D ultrasound has been shown to be correlated to adverse pregnancy outcomes This could potentially be used as a screening test to predict the ``at risk'' pregnancy. However, manual segmentation whilst accurate is very time consuming. Semi-automated methods provide close agreement to manual segmentation but remain significantly operator dependant. To generate a screening tool fully automated placental segmentation is required. In this paper a previously published deep convolutional neural network, Deep Medic, was trained using the output of the semi-automated random walker method as the ground truth. A set of 300 ultrasound volumes was used to train, validate and test the neural network. Dice similarity coefficients from the neural network had a median value of 0.73. This work shows the feasibility of applying convolutional neural networks as a technique for automatic segmentation of the placenta using 3D ultrasound.

Keywords: Ultrasound, Image segmentation, Machine learning
Abstract: Endometrium assessment via thickness measurement is commonly performed in routine gynecological ultrasound examination for assessing the reproductive health of patients undergoing fertility related treatments and endometrium cancer screening in women with post-menopausal bleeding. This paper introduces a fully automated technique for endometrium thickness measurement from three-dimensional transvaginal ultrasound (TVUS) images. The algorithm combines the robustness of deep neural networks with the more interpretable level set method for segmentation. We propose a hybrid variational curve propagation model which embeds a deep-learned endometrium probability map in the segmentation energy functional. This solution provides approximately 30% performance improvement over a contemporary supervised learning method on a database of 59 TVUS images and the thickness measurement is found to be within ±2mm of the manual measurement in 87% of the cases.

Keywords: Machine learning, Classification, Skin
Abstract: Existing powerful neural networks assume that the output classes are mutually exclusive and equally important. Many datasets of medical images do not satisfy these conditions. For example, some skin disease datasets have images labelled as coarse-grained class (such as Benign) and fine-grained labels (such as Blue Nevus) from the same class. Conventional neural network can not leverage such additional data for training. Moreover, in the clinical decision making, some classes (such as skin cancer or Melanoma) often carry more importance than other lesion types. We propose a novel Tree-Loss function for training and fine-tuning a neural network classifier using all available labelled images. The key step is the definition of the class taxonomy tree, which is used to describe the relations between labels. The tree can be also adjusted to reflect the desired importance of each class. These steps can be performed by a domain expert without detailed knowledge of machine learning techniques. The experiments demonstrate the improved performance compared with the conventional approach even without using additional data.

Keywords: Image segmentation, Skin
Abstract: Skin melanoma is one of the highly addressed health problems in many countries. Dermatologists and clinicians diagnose melanoma using clinical assessment tools such as ABCD. However, computer vision tools have been introduced to assist in quantitative analysis of skin lesions. Deep learning is one of the trending machine learning techniques that have been successfully utilized to solve many difficult computer vision tasks. We proposed using a hybrid method that utilizes two popular deep learning methods: convolutional and recurrent neural networks. The proposed method was trained using 900 images and tested on 375 images. Images were obtained from ``Skin Lesion Analysis Toward Melanoma Detection" challenge which was hosted by ISBI 2016 conference. We achieved segmentation average accuracy of 97% and Jaccard index of 93%. Results were compared with other state-of-the-art methods, including winner of ISBI 2016 challenge for skin melanoma segmentation, along with the same evaluation criteria.

Keywords: Skin, Machine learning, Classification
Abstract: Knowledge transfer impacts the performance of deep learning --- the state of the art for image classification tasks, including automated melanoma screening. Deep learning’s greed for large amounts of training data poses a challenge for medical tasks, which we can alleviate by recycling knowledge from models trained on different tasks, in a scheme called transfer learning. Although much of the best art on automated melanoma screening employs some form of transfer learning, a systematic evaluation was missing. Here we investigate the presence of transfer, from which task the transfer is sourced, and the application of fine tuning (i.e., retraining of the deep learning model after transfer). We also test the impact of picking deeper (and more expensive) models. Our results favor deeper models, pre-trained over ImageNet, with fine-tuning, reaching an AUC of 80.7% and 84.5% for the two skin-lesion datasets evaluated.

Keywords: Skin, Microscopy - Light, Confocal, Fluorescence, Classification
Abstract: Dermoscopy image is usually used in early diagnosis of malignant melanoma. The diagnosis accuracy by visual inspection is highly relied on the dermatologist’s clinical experience. Due to the inaccuracy, subjectivity, and poor reproducibility of human judgement, an automatic recognition algorithm of dermoscopy image is highly desired. In this work, we present a hybrid classification framework for dermoscopy image assessment by combining deep convolutional neural network (CNN), Fisher vector (FV) and support vector machine (SVM). Specifically, the deep representations of subimages at various locations of a rescaled dermoscopy image are first extracted via a natural image dataset pre-trained on CNN. Then we adopt an orderless visual statistics based FV encoding methods to aggregate these features to build more invariant representations. Finally, the FV encoded representations are classified for diagnosis using a linear SVM. Compared with traditional low-level visual features based recognition approaches, our scheme is simpler and requires no complex preprocessing. Furthermore, the orderless representations are less sensitive to geometric deformation. We evaluate our proposed method on the ISBI 2016 Skin lesion challenge dataset and promising results are obtained. Also, we achieve consistent improvement in accuracy even without fine-tuning the CNN.

Keywords: Brain, Computed tomography (CT), Angiographic imaging
Abstract: Stroke is the leading cause of disabilty worldwide. Recently a new treatment, intra-arterial thrombectomy (IAT), has greatly improved the outcome of patients suffering from ischemic stroke. However, the eficacy of IAT is extremely time dependent. This presentation wil try to give an overview of the oppertunities that exist for computer aided image anlysis to increase the efficacy of IAT by speeding up CT based diagnosis, CT angiography (CTA) image analysis as well as aiding in treatment guidance and evaluation.

Keywords: Computed tomography (CT), Perfusion imaging, Brain
Abstract: -

Keywords: Computer-aided detection and diagnosis (CAD), Image segmentation, Pattern recognition and classification
Abstract: In the New Eng J Med, 2015, results of the first randomized clinical trial indicating that patient outcome in acute stroke can be improved with mechanical thrombectomy using newly developed stents (Berkhemer et al. 2015). This result was included in the “Top 10 cardiovascular disease research advances of 2015’ by the American Heart Association. Based on these results, novel studies are being initiated to determine which stroke patients benefit most from which available treatment under a given set of circumstances; CT imaging plays a crucial role here, and there is an urgent need for tools to support analysis of the CT data. In this presentation we will discuss how to develop CT imaging biomarkers that can be relevant in the management of stroke patients, and we review some current developments in this field.

Keywords: Brain, Image-guided treatment, Angiographic imaging
Abstract: The recent trials on the efficacy of thrombectomy for stroke treatment, such as MR CLEAN, EXTEND-IA, ESCAPE, SWIFT-PRIME, REVASCAT [1-5], have clearly demonstrated the improved patient outcome for interventional treatment through thrombectomy, provided that the onset was within a certain time frame. This has led to a turning point in the way the care flow of the hyper-acute phase will be set up in the near future. In this communication we will sketch the views of an manufacturer of medical devices and the experiences of a clinical center with respect to acute stroke treatment.

Keywords: Breast, Image segmentation, Magnetic resonance imaging (MRI)
Abstract: We introduce a new fully automated breast mass segmentation method from dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI). The method is based on globally optimal inference in a continuous space (GOCS) using a shape prior computed from a semantic segmentation produced by a deep learning (DL) model. We propose this approach because the limited amount of annotated training samples does not allow the implementation of a robust DL model that could produce accurate segmentation results on its own. Furthermore, GOCS does not need precise initialisation compared to locally optimal methods on a continuous space (e.g., Mumford-Shah based level set methods); also, GOCS has smaller memory complexity compared to globally optimal inference on a discrete space (e.g., graph cuts). Experimental results show that the proposed method produces the current state-of-the-art mass segmentation (from DCE-MRI) results, achieving a mean Dice coefficient of 0.77 for the test set.

Keywords: Machine learning, Breast, X-ray imaging
Abstract: In this paper, we propose a multi-view deep residual neural network (mResNet) for the fully automated classification of mammograms as either malignant or normal/benign. Specifically, our mResNet approach consists of an ensemble of deep residual networks (ResNet), which have six input images, including the unregistered craniocaudal (CC) and mediolateral oblique (MLO) mammogram views as well as the automatically produced binary segmentation maps of the masses and micro-calcifications in each view. We then form the mRes- Net by concatenating the outputs of each ResNet at the second to last layer, followed by a final, fully connected, layer. The resulting mResNet is trained in an end-to-end fashion to produce a case-based mammogram classifier that has the potential to be used in breast screening programs. We empirically show on the publicly available INbreast dataset, that the proposed mResNet classifies mammograms into malignant or normal/benign with an AUC of 0.8.

Keywords: Magnetic resonance imaging (MRI), Image segmentation, Machine learning
Abstract: We present a learning based fully automatic method to detect and segment the prostate in T2 weighted MR scans. It consists of a localization stage which uses a learned global context to detect the prostate location. This is followed by a segmentation stage which uses a learned local context using prostatic segment specific discriminative classifiers, to compute the probability of a point being on the prostatic boundary. The final segmentation is obtained by via min-cut on a sparse spherical graph, centered at detected prostate location, with edge weight computed from the probability for the edge to intersect the prostate boundary. The method was submitted to the Prostate MR Segmentation (PROMISE) challenge. We obtain a mean/median DICE score of 86.1/87.7 % and a mean run time of 3s on a commodity PC. With the final stage comprising graph cuts on a sparse graph, a benefit of this work is ability to perform real-time edits after automatic segmentation in a manner that combines user-edits with learned information.

Keywords: Image segmentation, Spine, Magnetic resonance imaging (MRI)
Abstract: We present a new surface energy potential for the segmentation of cylindrical objects in 3D medical imaging using parametric spline active contours (a.k.a. spline-snakes). Our energy formulation is based on an optimal steerable surface detector. Thus, we combine the concept of steerability with spline-snakes that have open topology for semi-automatic segmentation. We show that the proposed energy yields segmentation results that are more robust to noise compared to classical gradient-based surface energies. We finally validate our model by segmenting the aorta on a cohort of 14 real 3D MRI images, and also provide an example of spinal cord segmentation using the same tool.

Keywords: Magnetic resonance imaging (MRI), Bone, Image segmentation
Abstract: In this paper we address the problem of bone segmentation in MRI images of children, in the region of the pelvis. To cope with the complex structure of the bones in this region and their changing topology during growth, we propose a method relying on 3D bone templates. These models are built from 3D CT images. For a given MRI volume, the closest template is chosen and registered on the MRI data. This leads to an initial segmentation which is then refined using a deformable model approach, where the regularization parameters depend on the local curvature, and the landmarks used during the registration are fixed anchors during the deformation. This approach was successfully applied to 15 MRI volumes of children between 1 and 18 years old, with an average accuracy in terms of medium distance of MD=1.17+-0.29 mm and Dice Index of DC=0.81+-0.04.

Keywords: Image segmentation, Machine learning, Microscopy - Electron
Abstract: We present an interactive approach to train a deep neural network pixel classifier for the segmentation of neuronal structures. An interactive training scheme reduces the extremely tedious manual annotation task that is typically required for deep networks to perform well on image segmentation problems. Our proposed method employs a feedback loop that captures sparse annotations using a graphical user interface, trains a deep neural network based on recent and past annotations, and displays the prediction output to users in almost real-time. Our implementation of the algorithm also allows multiple users to provide annotations in parallel and receive feedback from the same classifier. Quick feedback on classifier performance in an interactive setting enables users to identify and label examples that are more important than others for segmentation purposes. Our experiments show that an interactively-trained pixel classifier produces better region segmentation results on Electron Microscopy (EM) images than those generated by a network of the same architecture trained offline on exhaustive ground-truth labels.

Keywords: Microscopy - Light, Confocal, Fluorescence, Brain, image filtering (e.g. mathematical morphology, wavelets,...)
Abstract: With the advancement of high throughput and high resolution volumetric brain imaging, there is an unmet need to trace dense neuron fibers and study long-range neuron connectivity. An initial pipeline is described for processing cellular level neuronal fiber data acquired by a new super resolution imaging method called Magnified Analysis of the Proteome (MAP). First, a multiscale vessel enhancement filter is applied to segment fibers of different diameters. Morphological operations are then employed to extract the fiber centerlines, from which a 3D connectivity map is computed. Applying this approach to an initial data set yielded 2% equal error rate for segmentation and 92% accuracy for end-to-end fiber tracing (22 out of 24 hand-traced fibers). Future work calls for scaling up the algorithm to process much larger brain datasets (terabytes and above) and performing graph-based long-range connectivity analysis. This work has the potential to extend our knowledge on brain networks at the cellular level.

Keywords: Image segmentation, Machine learning
Abstract: We propose a novel multiple-input multiple-output convolution neural network (MIMO-CNN) for cell segmentation in fluorescence microscopy images. The proposed network trains the network parameters using multiple resolutions of the input image, connects the intermediate layers for better localization and context and generates the output using multi-resolution deconvolution filters. The MIMO-CNN allows us to deal with variable intensity cell boundaries and highly variable cell size in the mouse pancreatic tissue. The results show that our method outperforms state-of-the-art deep learning based approaches for segmentation.

Keywords: Microscopy - Light, Confocal, Fluorescence, Tissue, Machine learning
Abstract: Cell detection in microscopy images is a common and challenging task. We propose a new approach for mitotic cell detection in histopathology images, which is based on a Deep Residual Network architecture combined with Hough voting. We propose a voting layer for neural networks and introduce a novel loss function. Our approach is learned from scratch using cell centroids and the original images. We benchmarked our approach on the challenging AMIDA13 dataset containing histology images of invasive breast carcinoma. It turned out that our approach achieved competitive results.

Keywords: Histopathology imaging (e.g. whole slide imaging), Image registration, Lung
Abstract: Quantifying T cells inside tumorous tissue can help identifying immune profiles in order to improve prognosis and possibly develop immunotherapy. However, to identify T cells and cancerous cells in these two consecutive staining slides is challenging: the tissue preparation introduces the problem of alignment on large size images with poor visual common information. This work presents a framework for aligning whole slide images by extracting their common information and performing elastic registration based on the B-splines to solve this problem. Experiments show good results on our images even if some developments are still needed. This preliminary work is publicly available as part of our open-source Icy platform.

Keywords: Breast, X-ray imaging, Ultrasound
Abstract: Current evidence from statistical data suggests that breast cancer is responsible for 23% of all cancer cases and 14% of cancer related deaths amongst women worldwide [1]. Breast imaging and the analysis of breast images represent effective tools in the reduction of morbidity and mortality associated with breast cancer, contributing to the early detection, assessment and diagnosis of breast cancer, image-guided biopsy and treatment planning and response monitoring [2]. The analysis of breast images is still mostly done manually, but the use of computer-aided detection/diagnosis (CAD) systems as a second reader has been shown to help radiologists make final patient management decisions [3]. The majority of breast image analysis CAD systems have been developed to work with several imaging modalities, such as: mammography, tomosynthesis, computed tomography (CT), ultrasound, and magnetic resonance imaging (MRI). However, recent advances in the field are focused on the analysis of multimodal breast images, the use of deep learning methods, and the biomechanical modelling of the breast. This special session targets presentations on this topic, where we invited worldwide renowned speakers to present recent works addressing challenging problems in breast image analysis. This is an extremely relevant topic to ISBI given the relatively large number of recent publications in this particular field at all major medical image analysis conferences (ISBI, MICCAI, IPMI). Additionally, breast image analysis has been addressed by the field at least for the last two decades, making it one of the most studied problems in the field.

Keywords: Computational Imaging, Multi-modality fusion, Breast
Abstract: Breast conservative surgery (BCS) combined with radiotherapy has become the treatment of choice for the majority of women presenting with early breast cancer. However, few works have addressed the choice between BCS techniques or mastectomy based on Tumour Resection/Breast Volume Ratio (TRBVR) and expected cosmetic results. In this work, we introduce the concept of virtual tools for surgical planning of BCS by integrating, validating, and demonstrating a patient specific system for the planning, evaluation and quantification of the TRBVR related with the aesthetic outcome after BCS, using both 3D external shape of the breast and simple measures of the tumour.

Keywords: Breast, Magnetic resonance imaging (MRI), Modeling - Anatomical, physiological and pathological
Abstract: In New Zealand, approximately 3,000 women are diagnosed with breast cancer each year, and 600 die from the disease each year. We are investigating the use of biomechanical modelling to help clinicians address some of the challenges associated with the diagnosis and treatment of the disease [1]. Our main goal is to develop an automated pipeline to predict the supine shape of the breast from prone MR images to help clinicians pinpoint the location of tumours during breast cancer treatment procedures.

The first step in this pipeline involves using novel machine learning algorithms to automatically segment the breast and its and internal tissues from diagnostic prone breast MR images [2]. Personalised biomechanical models are then automatically generated from the segmented image data using a combination of nonlinear geometric fitting and statistical shape analysis techniques [3]. Once the biomechanical model of the prone breast has been constructed, large deformation mechanics and finite element modelling (FEM) is used to simulate the change in shape that occur when the breast is re-positioned to the supine position. The final step involves applying the deformation field obtained from the biomechanics solution to the prone MR image to simulate a supine MR image to help visualise the position of the breast tissues in the supine position. The accuracy of the model has been assessed in pilot studies by using non-rigid registration techniques to compare model predictions of the supine shape with the supine shape observed in actual supine MRI [4]. This pipeline is currently being implemented at Auckland City Hospital in New Zealand and we are currently developing a web-based GUI to allow our clinical collaborators to easily interpret the new information provided by the models and to easily integrate the tool into their clinical workflow.

References: [1] TP Babarenda Gamage et al., in Patient-Specific Modeling in Tomorrow's Medicine, Ed. A Gefen, Springer. p. 379-412, 2012.

[2] H Baluwala et al., in Proceedings of The Third MICCAI International Workshop on Breast Image Analysis. p. 113-120, 2015.

[3] DTK Malcolm et al., in Computational Biomechanics for Medicine, Springer. p. 39-49, 2016.

[4] TP Babarenda Gamage, MICCAI Computational Biomechanics for Medicine XI Workshop, October 2016.

Keywords: Breast, Pattern recognition and classification, Computer-aided detection and diagnosis (CAD)
Abstract: Texture has played a role in computer-aided understanding of breast images across all modalities (mammography, ultrasound (US), magnetic resonance imaging (MRI), etc.). Here, a case is made that texture analysis is natural in analyzing breast images. Several methods for texture analysis are discussed and a distinction is made between quantifying specific textures based on prior knowledge and discovering texture patterns from data. Many of the issues considered apply to other medical image analysis problems and to image analysis outside biology and medicine as well.

Keywords: Genes, Connectivity analysis, Diffusion weighted imaging
Abstract: We present the largest population-based heritability study of the human brain structural connectome, including a pathology-sensitive extension, the disconnectome. The disconnectome maps the effect of white matter lesions throughout the brain. The connectome and disconnectome were generated from diffusion-weighted images of 3255 unrelated subjects from the Rotterdam Study aged between 45 and 99 years. Graph theory measures were derived for both the connectome and disconnectome. Genotypes were used to derive genetic relationship matrices between individuals for heritability analyses. High measures of heritability, from 33% to 51%, were found across all connectivity measures. The disconnectome showed more significantly heritable connectivity measures than the connectome, suggesting that the new proposed measure may reveal additional or complementary information about the genetic architecture of the human brain.

Keywords: Connectivity analysis, Brain, Functional imaging (e.g. fMRI)
Abstract: Data-driven brain parcellations aim to provide a more accurate representation of an individual's functional connectivity, since they are able to capture individual variability that arises due to development or disease. This renders comparisons between the emerging brain connectivity networks more challenging, since correspondences between their elements are not preserved. Unveiling these correspondences is of major importance to keep track of local functional connectivity changes. We propose a novel method based on graph edit distance for the comparison of brain graphs directly in their domain that can accurately reflect similarities between individual networks, while providing the network element correspondences. This method is validated on a dataset of 116 twin subjects provided by the Human Connectome Project.

Keywords: fMRI analysis, Machine learning, Brain
Abstract: Dynamic functional connectivity (dFC) analysis aims at understanding how interactions across the brain resting-state networks (RSNs) evolve over time. Here, we introduce a novel methodological framework operating at the level of RSN activity time courses. Through the use of coupled hidden Markov models (CHMMs), we model cross-network couplings, i.e. the ability of one RSN to influence state transitions of the others. Because such modulatory influences are not expected across all possible pairs of RSNs, we combine this modeling strategy with L1 regularisation to derive a sparse set of cross-network modulatory coefficients. As a validation of this framework, we first demonstrate the ability of the sparse CHMM approach to disentangle intrinsic state transition probabilities from external modulatory influences on an artificially generated dataset. We then perform preliminary analyses on a real resting-state dataset, using RSN activity time courses derived from a state-of-the-art deconvolution technique as inputs to our framework, and shed light on several significant cross-network couplings across major RSNs.

Keywords: Brain, Functional imaging (e.g. fMRI), Connectivity analysis
Abstract: Prior work has reported that brain functional networks can be utilized to differentiate healthy subjects and patients with mental disorder. Group independent component analysis (GICA) is a widely-used data-driven method for extracting brain functional networks from resting-state functional magnetic resonance imaging (fMRI) data of multiple subjects. GICA approaches estimate the group-level independent components first, then back-reconstruct the subject-specific networks and their associated time courses based on the group-level independent components. To estimate the subject-specific networks, previous studies have employed PCA-based, regression-based (e.g. dual regression or spatio-temporal regression (STR)) and group information guided ICA (GIG-ICA) methods, among which dual regression and GIG-ICA can be used to yield the subject-specific networks for additional subjects. However, it is largely unknown which GICA method is more sensitive to subtle group differences between controls and patients. This paper aims to evaluate the efficacy of identifying biomarkers from the subject-specific networks and time courses estimated from STR and GIG-ICA using fMRI data of healthy controls (HCs) and schizophrenia patients (SZs). Regarding the measures from functional network maps, GIG-ICA revealed markedly greater differences between HCs and SZs than STR. Furthermore, the interaction among networks (i.e. functional network connectivity) also showed more group differences using GIG-ICA method, compared to STR. In summary, our work suggests that while both methods provide similar overall conclusions, GIG-ICA that estimates individually tuned components based on higher order statistics is more sensitive to group differences and biomarker detection than STR.

Keywords: Connectivity analysis, Brain, Diffusion weighted imaging
Abstract: Brain connectivity is increasingly being studied using connectomes. Typical structural connectome definitions do not directly take white matter pathology into account. Presumably, pathology impedes signal transmission along fibres, leading to a reduction in function. In order to directly study disconnection and localize pathology within the connectome, we present the disconnectome, which only considers fibres that intersect with white matter pathology. To show the potential of the disconnectome in brain studies, we showed in a cohort of 4199 adults with varying loads of white matter lesions (WMLs) that: (1) Disconnection is not a function of streamline density; (2) Hubs are more affected by WMLs than peripheral nodes; (3) Connections between hubs are more severely and frequently affected by WMLs than other connection types; and (4) Connections between region clusters are often more severely affected than those within clusters.

Keywords: Image segmentation, Machine learning, Whole-body
Abstract: Classic Random Regression Forests (RRFs) used for multi-organ localization describe the random process of multivariate regression by storing the histograms of offset vectors along each bounding wall direction per leaf node. On the one hand, the RAM and storage requirements of classic RRFs may become exorbitantly high when such a RRF consists of many leaf nodes, but on the other hand, a large number of leaf nodes are required for better localization. We introduce Light Random Regression Forests (LRRFs) which eliminate the need to describe the random process by formulating the localization prediction based on the random variables that describe the random process. Consequently, LRRFs with the same localization capabilities require less RAM and storage space compared to classic RRFs. LRRF comprising 4 trees with 17 decision levels is approximately 9 times faster, takes 10 times less RAM, and uses 30 times less storage space compared to a similar classic RRF.

Keywords: Computed tomography (CT), Lung, Classification
Abstract: Pulmonary emphysema overlaps considerably with chronic obstructive pulmonary disease (COPD), and is traditionally subcategorized into three subtypes: centrilobular emphysema (CLE), panlobular emphysema (PLE) and paraseptal emphysema (PSE). Automated classification methods based on supervised learning are generally based upon the current definition of emphysema subtypes, while unsupervised learning of texture patterns enables the objective discovery of possible new radiological emphysema subtypes. In this work, we use a variant of the Latent Dirichlet Allocation (LDA) model to discover lung macroscopic patterns (LMPs) in an unsupervised way from lung regions that encode emphysematous areas. We evaluate the possible utility of the LMPs as potential novel emphysema subtypes via measuring their level of reproducibility when varying the learning set and by their ability to predict traditional radiological emphysema subtypes. Experimental results show that our algorithm can discover highly reproducible LMPs, that predict traditional emphysema subtypes.

Keywords: Lung, Computed tomography (CT), Machine learning
Abstract: We propose a new computer-aided detection system that uses 3D convolutional neural networks (CNN) for detecting lung nodules in low dose computed tomography. The system leverages both a priori knowledge about lung nodules and confounding anatomical structures and data-driven machine-learned features and classifier. Specifically, we generate nodule candidates using a local geometric-model-based filter and further reduce the structure variability by estimating the local orientation. The nodule candidates in the form of 3D cubes are fed into a deep 3D convolutional neural network that is trained to differentiate nodule and non-nodule inputs. We use data augmentation techniques to generate a large number of training examples and use regularization techniques to avoid overfitting. On a set of 99 CT scans, the proposed system achieved state-of-the-art performance and significantly outperformed a similar hybrid system that uses conventional shallow learning. The experimental results showed benefits of using a priori models to reduce the problem space for data-driven machine learning of complex deep neural networks. The results also showed the advantages of 3D CNN over 2D CNN in volumetric medical image analysis.

Keywords: Vessels, Classification, Machine learning
Abstract: Artery-vein classification on pulmonary computed tomography (CT) images is becoming of high interest in the scientific community due to the prevalence of pulmonary vascular disease that affects arteries and veins through different mechanisms. In this work, we present a novel approach to automatically segment and classify vessels from chest CT images. We use a scale-space particle segmentation to isolate vessels, and combine a convolutional neural network (CNN) to graph-cut (GC) to classify the single particles. Information about proximity of arteries to airways is learned by the network by means of a bronchus enhanced image. The methodology is evaluated on the superior and inferior lobes of the right lung of twenty clinical cases. Comparison with manual classification and a Random Forests (RF) classifier is performed. The algorithm achieves an overall accuracy of 87% when compared to manual reference, which is higher than the 73% accuracy achieved by RF.

Keywords: Machine learning, Vessels, Computer-aided detection and diagnosis (CAD)
Abstract: We propose an automated platform for extra-coronary calcification detection on low dose CT scans. We utilize faster regional convolutional neural networks (R-CNN) to directly detect calcifications at the lesion-level without performing vessel extraction. To segment detected calcifications at the voxel-level, we employ holistically nested edge detection (HED). CT scans of 112 vasculitis patients and 3219 images with labeled calcifications were used to develop and evaluate our method. By employing a two-class faster R-CNN, the average precision (AP) increased from 49.2% to 84.4% for calcification detection. In addition, sensitivity of 85.0% at 1 false positive per image was observed. The Dice Similarity Coefficient (DSC) for calcification segmentation using HED (0.83±0.08) was significantly better (p<<0.01) than the traditional threshold-based method (0.59±0.26).

Keywords: Cells & molecules, Image segmentation, Microscopy - Light, Confocal, Fluorescence
Abstract: We present a new method for cell segmentation which combines a marked point process model with a combinatorics-based method of finding global optima. The method employs an energy term that assesses possible segmentations by their fidelity to both local image information and a simple model of cell interaction, and we use a randomized iterative reweighting technique for its minimization. Our approach was successfully applied to cell microscopy images of varying difficulty and experimentally compared with both a standard segmentation method as well as a method based on Multiple Birth and Cut. The proposed method is found to improve upon previous approaches.

Keywords: Nerves, Classification, Shape analysis
Abstract: One of the most exciting open problems in biological image analysis is the categorization of neurons based on observed morphology. Complex arborization patterns, varied orientations, different sizes, nonidentical alignments of local branches act as barriers to a systematic and quantitative analysis of neurons. This open problem demands a refined approach to capture salient and robust, global and local features to compare digitally-traced 3D neuron atlases. We propose NeuroSoL to consolidate global as well as local morphological and geometrical features by leveraging a specialized graph framework. Specifically, without imposing any restriction on the structure and connectivity, we use fully-sorted and mixed-sorted Laplacians of the neuron graph and the corresponding complementary graph. For meaningful clustering of neurons, we also develop a similarity metric after retrieving optimal local alignment between the feature descriptors of two neurons. Results on neuron datasets show the efficacy of our approach over state-of-the-art techniques.

Keywords: Probabilistic and statistical models & methods, Cells & molecules, Microscopy - Light, Confocal, Fluorescence
Abstract: We propose a novel label fusion technique as well as a crowdsourcing protocol to efficiently obtain accurate epithelial cell segmentations from non-expert crowd workers. Our label fusion technique simultaneously estimates the true segmentation, the performance levels of individual crowd workers, and an image segmentation model in the form of a pairwise Markov random field. We term our approach image-aware STAPLE (iaSTAPLE) since our image segmentation model seamlessly integrates into the well-known and widely used STAPLE approach. In an evaluation on a light microscopy dataset containing more than 5000 membrane labeled epithelial cells of a fly wing, we show that iaSTAPLE outperforms STAPLE in terms of segmentation accuracy as well as in terms of the accuracy of estimated crowd worker performance levels, and is able to correctly segment 99% of all cells when compared to expert segmentations. These results show that iaSTAPLE is a highly useful tool for crowd sourcing image segmentation.

Keywords: Image segmentation, Cells & molecules, Microscopy - Light, Confocal, Fluorescence
Abstract: Cervical nuclei carry substantial diagnostic information for cervical cancer. Therefore, in automation-assisted reading of cervical cytology, automated and accurate segmentation of nuclei is essential. This paper proposes a novel approach for segmentation of cervical nuclei that combines fully convolutional networks (FCN) and graph-based approach (FCNG). FCN is trained to learn the nucleus high-level features to generate a nucleus label mask and a nucleus probabilistic map. The mask is used to construct a graph by image transforming. The map is formulated into the graph cost function in addition to the properties of the nucleus border and nucleus region. The prior constraints regarding the context of nucleus-cytoplasm position are also utilized to modify the local cost functions. The globally optimal path in the constructed graph is identified by dynamic programming. Validation of our method was performed on cell nuclei from Herlev Pap smear dataset. Our method shows a Zijdenbos similarity index (ZSI) of 0.92 +-0.09, compared to the best state-of-the-art approach of 0.89+-0.15. The nucleus areas measured by our method correlated strongly with the independent standard (r2 = 0.91).

Keywords: Image segmentation, Graphical models & methods, Animal models and imaging
Abstract: We introduce a novel segmentation method for time-lapse image stacks of neurites based on the co-segmentation principle. Our method aggregates information from multiple stacks to improve the segmentation task, using a neurite model and a tree similarity term. The neurite model takes into account branching characteristics, such as local shape smoothness and continuity, while the tree similarity term exploits the local branch dynamics across image stacks. Our approach improves accuracy in ambiguous regions, handling successfully out-of-focus effects and branching bifurcations. We validated our method using Drosophila sensory neuron datasets and made comparisons with existing methods.