Carole H. Sudre

Generalised Dice Overlap as a Deep Learning Loss Function for Highly Unbalanced Segmentations

Deep-learning has proved in recent years to be a powerful tool for image analysis and is now widely used to segment both 2D and 3D medical images. Deep-learning segmentation frameworks rely not only on the choice of network architecture but also on the choice of loss function. When the segmentation process targets rare observations, a severe class imbalance is likely to occur between candidate labels, thus resulting in sub-optimal performance. In order to mitigate this issue, strategies such as the weighted cross-entropy function, the sensitivity function or the Dice loss function, have been proposed. In this work, we investigate the behavior of these loss functions and their sensitivity to learning rate tuning in the presence of different rates of label imbalance across 2D and 3D segmentation tasks. We also propose to use the class re-balancing properties of the Generalized Dice overlap, a known metric for segmentation assessment, as a robust and accurate deep-learning loss function for unbalanced tasks.

Longitudinal multiple sclerosis lesion segmentation: Resource and challenge

Abstract In conjunction with the ISBI 2015 conference, we organized a longitudinal lesion segmentation challenge providing training and test data to registered participants. The training data consisted of five subjects with a mean of 4.4 time‐points, and test data of fourteen subjects with a mean of 4.4 time‐points. All 82 data sets had the white matter lesions associated with multiple sclerosis delineated by two human expert raters. Eleven teams submitted results using state‐of‐the‐art lesion segmentation algorithms to the challenge, with ten teams presenting their results at the conference. We present a quantitative evaluation comparing the consistency of the two raters as well as exploring the performance of the eleven submitted results in addition to three other lesion segmentation algorithms. The challenge presented three unique opportunities: (1) the sharing of a rich data set; (2) collaboration and comparison of the various avenues of research being pursued in the community; and (3) a review and refinement of the evaluation metrics currently in use. We report on the performance of the challenge participants, as well as the construction and evaluation of a consensus delineation. The image data and manual delineations will continue to be available for download, through an evaluation website2 as a resource for future researchers in the area. This data resource provides a platform to compare existing methods in a fair and consistent manner to each other and multiple manual raters. HighlightsPublic lesion data base of 21 training data sets and 61 testing data sets.Fully automated evaluation website.Comparison between 14 state‐of‐the‐art algorithms and 2 manual delineators.

Bayesian Model Selection for Pathological Neuroimaging Data Applied to White Matter Lesion Segmentation

In neuroimaging studies, pathologies can present themselves as abnormal intensity patterns. Thus, solutions for detecting abnormal intensities are currently under investigation. As each patient is unique, an unbiased and biologically plausible model of pathological data would have to be able to adapt to the subjects individual presentation. Such a model would provide the means for a better understanding of the underlying biological processes and improve ones ability to define pathologically meaningful imaging biomarkers. With this aim in mind, this work proposes a hierarchical fully unsupervised model selection framework for neuroimaging data which enables the distinction between different types of abnormal image patterns without pathological a priori knowledge. Its application on simulated and clinical data demonstrated the ability to detect abnormal intensity clusters, resulting in a competitive to improved behavior in white matter lesion segmentation when compared to three other freely-available automated methods.

NiftyNet: a deep-learning platform for medical imaging

Highlights • An open-source platform is implemented based on TensorFlow APIs for deep learning in medical imaging domain.• A modular implementation of the typical medical imaging machine learning pipeline facilitates (1) warm starts with established pre-trained networks, (2) adapting existing neural network architectures to new problems, and (3) rapid prototyping of new solutions.• Three deep-learning applications, including segmentation, regression, image generation and representation learning, are presented as concrete examples illustrating the platform’s key features.

White matter hyperintensities are associated with disproportionate progressive hippocampal atrophy.

This study investigates relationships between white matter hyperintensity (WMH) volume, cerebrospinal fluid (CSF) Alzheimers disease (AD) pathology markers, and brain and hippocampal volume loss. Subjects included 198 controls, 345 mild cognitive impairment (MCI), and 154 AD subjects with serial volumetric 1.5‐T MRI. CSF Aβ42 and total tau were measured (n = 353). Brain and hippocampal loss were quantified from serial MRI using the boundary shift integral (BSI). Multiple linear regression models assessed the relationships between WMHs and hippocampal and brain atrophy rates. Models were refitted adjusting for (a) concurrent brain/hippocampal atrophy rates and (b) CSF Aβ42 and tau in subjects with CSF data. WMH burden was positively associated with hippocampal atrophy rate in controls (P = 0.002) and MCI subjects (P = 0.03), and with brain atrophy rate in controls (P = 0.03). The associations with hippocampal atrophy rate remained following adjustment for concurrent brain atrophy rate in controls and MCIs, and for CSF biomarkers in controls (P = 0.007). These novel results suggest that vascular damage alongside AD pathology is associated with disproportionately greater hippocampal atrophy in nondemented older adults.

Relationship of grey and white matter abnormalities with distance from the surface of the brain in multiple sclerosis

Objective To assess the association between proximity to the inner (ventricular and aqueductal) and outer (pial) surfaces of the brain and the distribution of normal appearing white matter (NAWM) and grey matter (GM) abnormalities, and white matter (WM) lesions, in multiple sclerosis (MS). Methods 67 people with relapse-onset MS and 30 healthy controls were included in the study. Volumetric T1 images and high-resolution (1 mm3) magnetisation transfer ratio (MTR) images were acquired and segmented into 12 bands between the inner and outer surfaces of the brain. The first and last bands were discarded to limit partial volume effects with cerebrospinal fluid. MTR values were computed for all bands in supratentorial NAWM, cerebellar NAWM and brainstem NA tissue, and deep and cortical GM. Band WM lesion volumes were also measured. Results Proximity to the ventricular surfaces was associated with progressively lower MTR values in the MS group but not in controls in supratentorial and cerebellar NAWM, brainstem NA and in deep and cortical GM. The density of WM lesions was associated with proximity to the ventricles only in the supratentorial compartment, and no link was found with distance from the pial surfaces. Conclusions In MS, MTR abnormalities in NAWM and GM are related to distance from the inner and outer surfaces of the brain, and this suggests that there is a common factor underlying their spatial distribution. A similar pattern was not found for WM lesions, raising the possibility that different factors promote their formation.

Template-Based Multimodal Joint Generative Model of Brain Data.

The advent of large of multi-modal imaging databases opens up the opportunity to learn how local intensity patterns covariate between multiple modalities. These models can then be used to describe expected intensities in an unseen image modalities given one or multiple observations, or to detect deviations (e.g. pathology) from the expected intensity patterns. In this work, we propose a template-based multi-modal generative mixture-model of imaging data and apply it to the problems of inlier/outlier pattern classification and image synthesis. Results on synthetic and patient data demonstrate that the proposed method is able to synthesise unseen data and accurately localise pathological regions, even in the presence of large abnormalities. It also demonstrates that the proposed model can provide accurate and uncertainty-aware intensity estimates of expected imaging patterns.

White matter hyperintensities are seen only in GRN mutation carriers in the GENFI cohort

Genetic frontotemporal dementia is most commonly caused by mutations in the progranulin (GRN), microtubule-associated protein tau (MAPT) and chromosome 9 open reading frame 72 (C9orf72) genes. Previous small studies have reported the presence of cerebral white matter hyperintensities (WMH) in genetic FTD but this has not been systematically studied across the different mutations. In this study WMH were assessed in 180 participants from the Genetic FTD Initiative (GENFI) with 3D T1- and T2-weighed magnetic resonance images: 43 symptomatic (7 GRN, 13 MAPT and 23 C9orf72), 61 presymptomatic mutation carriers (25 GRN, 8 MAPT and 28 C9orf72) and 76 mutation negative non-carrier family members. An automatic detection and quantification algorithm was developed for determining load, location and appearance of WMH. Significant differences were seen only in the symptomatic GRN group compared with the other groups with no differences in the MAPT or C9orf72 groups: increased global load of WMH was seen, with WMH located in the frontal and occipital lobes more so than the parietal lobes, and nearer to the ventricles rather than juxtacortical. Although no differences were seen in the presymptomatic group as a whole, in the GRN cohort only there was an association of increased WMH volume with expected years from symptom onset. The appearance of the WMH was also different in the GRN group compared with the other groups, with the lesions in the GRN group being more similar to each other. The presence of WMH in those with progranulin deficiency may be related to the known role of progranulin in neuroinflammation, although other roles are also proposed including an effect on blood-brain barrier permeability and the cerebral vasculature. Future studies will be useful to investigate the longitudinal evolution of WMH and their potential use as a biomarker as well as post-mortem studies investigating the histopathological nature of the lesions.

Longitudinal segmentation of age-related white matter hyperintensities

HIGHLIGHTSA longitudinal WMH segmentation algorithm is proposed.Time points segmentation are constrained by an average data model.It is shown to be unbiased to time point ordering.A longitudinal lesion simulator is developed for validation.The longitudinal solution is more robust compared to cross‐sectional methods. ABSTRACT Although white matter hyperintensities evolve in the course of ageing, few solutions exist to consider the lesion segmentation problem longitudinally. Based on an existing automatic lesion segmentation algorithm, a longitudinal extension is proposed. For evaluation purposes, a longitudinal lesion simulator is created allowing for the comparison between the longitudinal and the cross‐sectional version in various situations of lesion load progression. Finally, applied to clinical data, the proposed framework demonstrates an increased robustness compared to available cross‐sectional methods and findings are aligned with previously reported clinical patterns.

Bayesian Model Selection for Pathological Data

The detection of abnormal intensities in brain images caused by the presence of pathologies is currently under great scrutiny. Selecting appropriate models for pathological data is of critical importance for an unbiased and biologically plausible model fit, which in itself enables a better understanding of the underlying data and biological processes. Besides, it impacts on ones ability to extract pathologically meaningful imaging biomarkers. With this aim in mind, this work proposes a fully unsupervised hierarchical model selection framework for neuroimaging data which permits the stratification of different types of abnormal image atterns without prior knowledge about the subjects pathological status.