Knowledge distillation with ensembles of convolutional neural networks for medical image segmentation.

Purpose: Ensembles of convolutional neural networks (CNNs) often outperform a single CNN in medical image segmentation tasks, but inference is computationally more expensive and makes ensembles unattractive for some applications. We compared the performance of differently constructed ensembles with the performance of CNNs derived from these ensembles using knowledge distillation, a technique for reducing the footprint of large models such as ensembles. Approach: We investigated two different types of ensembles, namely, diverse ensembles of networks with three different architectures and two different loss-functions, and uniform ensembles of networks with the same architecture but initialized with different random seeds. For each ensemble, additionally, a single student network was trained to mimic the class probabilities predicted by the teacher model, the ensemble. We evaluated the performance of each network, the ensembles, and the corresponding distilled networks across three different publicly available datasets. These included chest computed tomography scans with four annotated organs of interest, brain magnetic resonance imaging (MRI) with six annotated brain structures, and cardiac cine-MRI with three annotated heart structures. Results: Both uniform and diverse ensembles obtained better results than any of the individual networks in the ensemble. Furthermore, applying knowledge distillation resulted in a single network that was smaller and faster without compromising performance compared with the ensemble it learned from. The distilled networks significantly outperformed the same network trained with reference segmentation instead of knowledge distillation. Conclusion: Knowledge distillation can compress segmentation ensembles of uniform or diverse composition into a single CNN while maintaining the performance of the ensemble.

A Simulation Toolkit for Testing the Sensitivity and Accuracy of Corticometry Pipelines.

In recent years, the replicability of neuroimaging findings has become an important concern to the research community. Neuroimaging pipelines consist of myriad numerical procedures, which can have a cumulative effect on the accuracy of findings. To address this problem, we propose a method for simulating artificial lesions in the brain in order to estimate the sensitivity and specificity of lesion detection, using different automated corticometry pipelines. We have applied this method to different versions of two widely used neuroimaging pipelines (CIVET and FreeSurfer), in terms of coefficients of variation; sensitivity and specificity of detecting lesions in 4 different regions of interest in the cortex, while introducing variations to the lesion size, the blurring kernel used prior to statistical analyses, and different thickness metrics (in CIVET). These variations are tested in a between-subject design (in two random groups, with and without lesions, using T1-weigted MRIs of 152 individuals from the International Consortium of Brain Mapping (ICBM) dataset) and in a within-subject pre-/post-lesion design [using 21 T1-Weighted MRIs of a single adult individual, scanned in the Infant Brain Imaging Study (IBIS)]. The simulation method is sensitive to partial volume effect and lesion size. Comparisons between pipelines illustrate the ability of this method to uncover differences in sensitivity and specificity of lesion detection. We propose that this method be adopted in the workflow of software development and release.

LAMA: automated image analysis for the developmental phenotyping of mouse embryos.

Advanced 3D imaging modalities, such as micro-computed tomography (micro-CT), have been incorporated into the high-throughput embryo pipeline of the International Mouse Phenotyping Consortium (IMPC). This project generates large volumes of raw data that cannot be immediately exploited without significant resources of personnel and expertise. Thus, rapid automated annotation is crucial to ensure that 3D imaging data can be integrated with other multi-dimensional phenotyping data. We present an automated computational mouse embryo phenotyping pipeline that harnesses the large amount of wild-type control data available in the IMPC embryo pipeline in order to address issues of low mutant sample number as well as incomplete penetrance and variable expressivity. We also investigate the effect of developmental substage on automated phenotyping results. Designed primarily for developmental biologists, our software performs image pre-processing, registration, statistical analysis and segmentation of embryo images. We also present a novel anatomical E14.5 embryo atlas average and, using it with LAMA, show that we can uncover known and novel dysmorphology from two IMPC knockout lines.

Progressive neuroanatomical changes caused by Grin1 loss-of-function mutation.

NMDA receptor dysfunction is central to the encephalopathies caused by missense mutations in the NMDA receptor subunit genes. Missense variants of GRIN1, GRIN2A, and GRIN2B cause similar syndromes with varying severity of intellectual impairment, autism, epilepsy, and motor dysfunction. To gain insight into possible biomarkers of NMDAR hypofunction, we asked whether a loss-of-function variant in the Grin1 gene would cause structural changes in the brain that could be detected by MRI. We also studied the developmental trajectory of these changes to determine whether structural changes coincided with reported cognitive impairments in the mice. We performed magnetic resonance imaging in male Grin1-/- knockdown mice (GluN1KD) that were three, six, or twelve weeks old. Deformation-based morphometry was used to assess neuroanatomical differences. Volumetric reductions were detected in substantia nigra and striatum of GluN1KD mice at all ages. Changes in limbic structures were only evident at six weeks of age. Reductions in white matter volumes were first evident at three weeks, and additional deficits were detected at six and twelve weeks. FluoroJade immunofluorescence revealed degenerating neurons in twelve-week old GluN1KD mice. We conclude that Grin1 loss-of-function mutations cause volume reductions in dopaminergic structures early in development, while changes to limbic and white matter structures are delayed and are more pronounced in post-adolescent ages. The evidence of degenerating neurons in the mature brain indicates an ongoing process of cell loss as a consequence of NMDAR hypofunction.

MRI to Assess Neurological Function.

This article describes a detailed set of protocols for mouse brain imaging using MRI. We focus primarily on measuring changes in neuroanatomy, and provide both instructions for mouse preparation and details on image acquisition, image processing, and statistics. Practical details as well as theoretical considerations are provided. © 2018 by John Wiley & Sons, Inc.

Structural covariance of brain region volumes is associated with both structural connectivity and transcriptomic similarity.

An organizational pattern seen in the brain, termed structural covariance, is the statistical association of pairs of brain regions in their anatomical properties. These associations, measured across a population as covariances or correlations usually in cortical thickness or volume, are thought to reflect genetic and environmental underpinnings. Here, we examine the biological basis of structural volume covariance in the mouse brain. We first examined large scale associations between brain region volumes using an atlas-based approach that parcellated the entire mouse brain into 318 regions over which correlations in volume were assessed, for volumes obtained from 153 mouse brain images via high-resolution MRI. We then used a seed-based approach and determined, for 108 different seed regions across the brain and using mouse gene expression and connectivity data from the Allen Institute for Brain Science, the variation in structural covariance data that could be explained by distance to seed, transcriptomic similarity to seed, and connectivity to seed. We found that overall, correlations in structure volumes hierarchically clustered into distinct anatomical systems, similar to findings from other studies and similar to other types of networks in the brain, including structural connectivity and transcriptomic similarity networks. Across seeds, this structural covariance was significantly explained by distance (17% of the variation, up to a maximum of 49% for structural covariance to the visceral area of the cortex), transcriptomic similarity (13% of the variation, up to maximum of 28% for structural covariance to the primary visual area) and connectivity (15% of the variation, up to a maximum of 36% for structural covariance to the intermediate reticular nucleus in the medulla) of covarying structures. Together, distance, connectivity, and transcriptomic similarity explained 37% of structural covariance, up to a maximum of 63% for structural covariance to the visceral area. Additionally, this pattern of explained variation differed spatially across the brain, with transcriptomic similarity playing a larger role in the cortex than subcortex, while connectivity explains structural covariance best in parts of the cortex, midbrain, and hindbrain. These results suggest that both gene expression and connectivity underlie structural volume covariance, albeit to different extents depending on brain region, and this relationship is modulated by distance.

P63 expression plays a role in developmental rate, embryo size, and local morphogenesis.

BACKGROUND: The p63 gene is integral to the development of many body parts including limb, palate, teeth, and urogenital tract. Loss of p63 expression may alter developmental rate, which is crucial to normal morphogenesis. To validate a novel, unbiased embryo phenotyping software tool, we tested whether delayed development contributes to the pathological phenotype of a p63 mouse mutant (p63-/- ). We quantified dysmorphology in p63-/- embryos and tested for universal growth delay relative to wild-type (WT) embryos. Fixed embryos (n = 6; p63-/- ) aged day (E) 15.5 were micro-CT scanned and quantitatively analyzed using a digital WT atlas that defined volumetric differences between p63-/- and WT embryos. RESULTS: p63-/- embryos showed a growth delay of approximately 22 hr (0.9 days). Among the E15.5 mutants, overall size was closest to WT E14.6 mice but shape was closest to WT E14.0. The atlas clearly identified in p63-/- embryos malformations of epithelial derivatives including limbs, tail, urogenital structures, brain, face, and tooth. CONCLUSIONS: The software atlas technique described the p63-/- phenotype as a combination of developmental delay (i.e., heterochrony) and malformation (i.e., pathological shape; failed organogenesis). This study identifies for the first time global and local roles for p63 in prenatal growth and development. Developmental Dynamics 247:779-787, 2018. © 2018 Wiley Periodicals, Inc.

Mouse models of 17q21.31 microdeletion and microduplication syndromes highlight the importance of Kansl1 for cognition.

Koolen-de Vries syndrome (KdVS) is a multi-system disorder characterized by intellectual disability, friendly behavior, and congenital malformations. The syndrome is caused either by microdeletions in the 17q21.31 chromosomal region or by variants in the KANSL1 gene. The reciprocal 17q21.31 microduplication syndrome is associated with psychomotor delay, and reduced social interaction. To investigate the pathophysiology of 17q21.31 microdeletion and microduplication syndromes, we generated three mouse models: 1) the deletion (Del/+); or 2) the reciprocal duplication (Dup/+) of the 17q21.31 syntenic region; and 3) a heterozygous Kansl1 (Kans1+/-) model. We found altered weight, general activity, social behaviors, object recognition, and fear conditioning memory associated with craniofacial and brain structural changes observed in both Del/+ and Dup/+ animals. By investigating hippocampus function, we showed synaptic transmission defects in Del/+ and Dup/+ mice. Mutant mice with a heterozygous loss-of-function mutation in Kansl1 displayed similar behavioral and anatomical phenotypes compared to Del/+ mice with the exception of sociability phenotypes. Genes controlling chromatin organization, synaptic transmission and neurogenesis were upregulated in the hippocampus of Del/+ and Kansl1+/- animals. Our results demonstrate the implication of KANSL1 in the manifestation of KdVS phenotypes and extend substantially our knowledge about biological processes affected by these mutations. Clear differences in social behavior and gene expression profiles between Del/+ and Kansl1+/- mice suggested potential roles of other genes affected by the 17q21.31 deletion. Together, these novel mouse models provide new genetic tools valuable for the development of therapeutic approaches.

4D atlas of the mouse embryo for precise morphological staging.

After more than a century of research, the mouse remains the gold-standard model system, for it recapitulates human development and disease and is quickly and highly tractable to genetic manipulations. Fundamental to the power and success of using a mouse model is the ability to stage embryonic mouse development accurately. Past staging systems were limited by the technologies of the day, such that only surface features, visible with a light microscope, could be recognized and used to define stages. With the advent of high-throughput 3D imaging tools that capture embryo morphology in microscopic detail, we now present the first 4D atlas staging system for mouse embryonic development using optical projection tomography and image registration methods. By tracking 3D trajectories of every anatomical point in the mouse embryo from E11.5 to E14.0, we established the first 4D atlas compiled from ex vivo 3D mouse embryo reference images. The resulting 4D atlas comprises 51 interpolated 3D images in this gestational range, resulting in a temporal resolution of 72 min. From this 4D atlas, any mouse embryo image can be subsequently compared and staged at the global, voxel and/or structural level. Assigning an embryonic stage to each point in anatomy allows for unprecedented quantitative analysis of developmental asynchrony among different anatomical structures in the same mouse embryo. This comprehensive developmental data set offers developmental biologists a new, powerful staging system that can identify and compare differences in developmental timing in wild-type embryos and shows promise for localizing deviations in mutant development.

Pydpiper: a flexible toolkit for constructing novel registration pipelines.

Using neuroimaging technologies to elucidate the relationship between genotype and phenotype and brain and behavior will be a key contribution to biomedical research in the twenty-first century. Among the many methods for analyzing neuroimaging data, image registration deserves particular attention due to its wide range of applications. Finding strategies to register together many images and analyze the differences between them can be a challenge, particularly given that different experimental designs require different registration strategies. Moreover, writing software that can handle different types of image registration pipelines in a flexible, reusable and extensible way can be challenging. In response to this challenge, we have created Pydpiper, a neuroimaging registration toolkit written in Python. Pydpiper is an open-source, freely available software package that provides multiple modules for various image registration applications. Pydpiper offers five key innovations. Specifically: (1) a robust file handling class that allows access to outputs from all stages of registration at any point in the pipeline; (2) the ability of the framework to eliminate duplicate stages; (3) reusable, easy to subclass modules; (4) a development toolkit written for non-developers; (5) four complete applications that run complex image registration pipelines "out-of-the-box." In this paper, we will discuss both the general Pydpiper framework and the various ways in which component modules can be pieced together to easily create new registration pipelines. This will include a discussion of the core principles motivating code development and a comparison of Pydpiper with other available toolkits. We also provide a comprehensive, line-by-line example to orient users with limited programming knowledge and highlight some of the most useful features of Pydpiper. In addition, we will present the four current applications of the code.

Mapping registration sensitivity in MR mouse brain images.

Nonlinear registration algorithms provide a way to estimate structural (brain) differences based on magnetic resonance images. Their ability to align images of different individuals and across modalities has been well-researched, but the bounds of their sensitivity with respect to the recovery of salient morphological differences between groups are unclear. Here we develop a novel approach to simulate deformations on MR brain images to evaluate the ability of two registration algorithms to extract structural differences corresponding to biologically plausible atrophy and expansion. We show that at a neuroanatomical level registration accuracy is influenced by the size and compactness of structures, but do so differently depending on how much change is simulated. The size of structures has a small influence on the recovered accuracy. There is a trend for larger structures to be recovered more accurately, which becomes only significant as the amount of simulated change is large. More compact structures can be recovered more accurately regardless of the amount of simulated change. Both tested algorithms underestimate the full extent of the simulated atrophy and expansion. Finally we show that when multiple comparisons are corrected for at a voxelwise level, a very low rate of false positives is obtained. More interesting is that true positive rates average around 40%, indicating that the simulated changes are not fully recovered. Simulation experiments were run using two fundamentally different registration algorithms and we identified the same results, suggesting that our findings are generalizable across different classes of nonlinear registration algorithms.

Performing label-fusion-based segmentation using multiple automatically generated templates.

Classically, model-based segmentation procedures match magnetic resonance imaging (MRI) volumes to an expertly labeled atlas using nonlinear registration. The accuracy of these techniques are limited due to atlas biases, misregistration, and resampling error. Multi-atlas-based approaches are used as a remedy and involve matching each subject to a number of manually labeled templates. This approach yields numerous independent segmentations that are fused using a voxel-by-voxel label-voting procedure. In this article, we demonstrate how the multi-atlas approach can be extended to work with input atlases that are unique and extremely time consuming to construct by generating a library of multiple automatically generated templates of different brains (MAGeT Brain). We demonstrate the efficacy of our method for the mouse and human using two different nonlinear registration algorithms (ANIMAL and ANTs). The input atlases consist a high-resolution mouse brain atlas and an atlas of the human basal ganglia and thalamus derived from serial histological data. MAGeT Brain segmentation improves the identification of the mouse anterior commissure (mean Dice Kappa values (κ = 0.801), but may be encountering a ceiling effect for hippocampal segmentations. Applying MAGeT Brain to human subcortical structures improves segmentation accuracy for all structures compared to regular model-based techniques (κ = 0.845, 0.752, and 0.861 for the striatum, globus pallidus, and thalamus, respectively). Experiments performed with three manually derived input templates suggest that MAGeT Brain can approach or exceed the accuracy of multi-atlas label-fusion segmentation (κ = 0.894, 0.815, and 0.895 for the striatum, globus pallidus, and thalamus, respectively).