Fetal development of functional thalamocortical and cortico-cortical connectivity.

Measuring and understanding functional fetal brain development in utero is critical for the study of the developmental foundations of our cognitive abilities, possible early detection of disorders, and their prevention. Thalamocortical connections are an intricate component of shaping the cortical layout, but so far, only ex-vivo studies provide evidence of how axons enter the sub-plate and cortex during this highly dynamic phase. Evidence for normal in-utero development of the functional thalamocortical connectome in humans is missing. Here, we modeled fetal functional thalamocortical connectome development using in-utero functional magnetic resonance imaging in fetuses observed from 19th to 40th weeks of gestation (GW). We observed a peak increase of thalamocortical functional connectivity strength between 29th and 31st GW, right before axons establish synapses in the cortex. The cortico-cortical connectivity increases in a similar time window, and exhibits significant functional laterality in temporal-superior, -medial, and -inferior areas. Homologous regions exhibit overall similar mirrored connectivity profiles, but this similarity decreases during gestation giving way to a more diverse cortical interconnectedness. Our results complement the understanding of structural development of the human connectome and may serve as the basis for the investigation of disease and deviations from a normal developmental trajectory of connectivity development.

Accurate Bayesian segmentation of thalamic nuclei using diffusion MRI and an improved histological atlas.

The human thalamus is a highly connected brain structure, which is key for the control of numerous functions and is involved in several neurological disorders. Recently, neuroimaging studies have increasingly focused on the volume and connectivity of the specific nuclei comprising this structure, rather than looking at the thalamus as a whole. However, accurate identification of cytoarchitectonically designed histological nuclei on standard in vivo structural MRI is hampered by the lack of image contrast that can be used to distinguish nuclei from each other and from surrounding white matter tracts. While diffusion MRI may offer such contrast, it has lower resolution and lacks some boundaries visible in structural imaging. In this work, we present a Bayesian segmentation algorithm for the thalamus. This algorithm combines prior information from a probabilistic atlas with likelihood models for both structural and diffusion MRI, allowing segmentation of 25 thalamic labels per hemisphere informed by both modalities. We present an improved probabilistic atlas, incorporating thalamic nuclei identified from histology and 45 white matter tracts surrounding the thalamus identified in ultra-high gradient strength diffusion imaging. We present a family of likelihood models for diffusion tensor imaging, ensuring compatibility with the vast majority of neuroimaging datasets that include diffusion MRI data. The use of these diffusion likelihood models greatly improves identification of nuclear groups versus segmentation based solely on structural MRI. Dice comparison of 5 manually identifiable groups of nuclei to ground truth segmentations show improvements of up to 10 percentage points. Additionally, our chosen model shows a high degree of reliability, with median test-retest Dice scores above 0.85 for four out of five nuclei groups, whilst also offering improved detection of differential thalamic involvement in Alzheimer's disease (AUROC 81.98%). The probabilistic atlas and segmentation tool will be made publicly available as part of the neuroimaging package FreeSurfer (https://freesurfer.net/fswiki/ThalamicNucleiDTI).

Cross-Sectional Observational Study of Typical in utero Fetal Movements Using Machine Learning.

Early variations of fetal movements are the hallmark of a healthy developing central nervous system. However, there are no automatic methods to quantify the complex 3D motion of the developing fetus in utero. The aim of this prospective study was to use machine learning (ML) on in utero MRI to perform quantitative kinematic analysis of fetal limb movement, assessing the impact of maternal, placental, and fetal factors. In this cross-sectional, observational study, we used 76 sets of fetal (24-40 gestational weeks [GW]) blood oxygenation level-dependent (BOLD) MRI scans of 52 women (18-45 years old) during typical pregnancies. Pregnant women were scanned for 5-10 min while breathing room air (21% O2) and for 5-10 min while breathing 100% FiO2 in supine and/or lateral position. BOLD acquisition time was 20 min in total with effective temporal resolution approximately 3 s. To quantify upper and lower limb kinematics, we used a 3D convolutional neural network previously trained to track fetal key points (wrists, elbows, shoulders, ankles, knees, hips) on similar BOLD time series. Tracking was visually assessed, errors were manually corrected, and the absolute movement time (AMT) for each joint was calculated. To identify variables that had a significant association with AMT, we constructed a mixed-model ANOVA with interaction terms. Fetuses showed significantly longer duration of limb movements during maternal hyperoxia. We also found a significant centrifugal increase of AMT across limbs and significantly longer AMT of upper extremities <31 GW and longer AMT of lower extremities >35 GW. In conclusion, using ML we successfully quantified complex 3D fetal limb motion in utero and across gestation, showing maternal factors (hyperoxia) and fetal factors (gestational age, joint) that impact movement. Quantification of fetal motion on MRI is a potential new biomarker of fetal health and neuromuscular development.

Automated detection and reacquisition of motion-degraded images in fetal HASTE imaging at 3 T.

PURPOSE: Fetal brain Magnetic Resonance Imaging suffers from unpredictable and unconstrained fetal motion that causes severe image artifacts even with half-Fourier single-shot fast spin echo (HASTE) readouts. This work presents the implementation of a closed-loop pipeline that automatically detects and reacquires HASTE images that were degraded by fetal motion without any human interaction. METHODS: A convolutional neural network that performs automatic image quality assessment (IQA) was run on an external GPU-equipped computer that was connected to the internal network of the MRI scanner. The modified HASTE pulse sequence sent each image to the external computer, where the IQA convolutional neural network evaluated it, and then the IQA score was sent back to the sequence. At the end of the HASTE stack, the IQA scores from all the slices were sorted, and only slices with the lowest scores (corresponding to the slices with worst image quality) were reacquired. RESULTS: The closed-loop HASTE acquisition framework was tested on 10 pregnant mothers, for a total of 73 acquisitions of our modified HASTE sequence. The IQA convolutional neural network, which was successfully employed by our modified sequence in real time, achieved an accuracy of 85.2% and area under the receiver operator characteristic of 0.899. CONCLUSION: The proposed acquisition/reconstruction pipeline was shown to successfully identify and automatically reacquire only the motion degraded fetal brain HASTE slices in the prescribed stack. This minimizes the overall time spent on HASTE acquisitions by avoiding the need to repeat the entire stack if only few slices in the stack are motion-degraded.

Joint super-resolution and synthesis of 1 mm isotropic MP-RAGE volumes from clinical MRI exams with scans of different orientation, resolution and contrast.

Most existing algorithms for automatic 3D morphometry of human brain MRI scans are designed for data with near-isotropic voxels at approximately 1 mm resolution, and frequently have contrast constraints as well-typically requiring T1-weighted images (e.g., MP-RAGE scans). This limitation prevents the analysis of millions of MRI scans acquired with large inter-slice spacing in clinical settings every year. In turn, the inability to quantitatively analyze these scans hinders the adoption of quantitative neuro imaging in healthcare, and also precludes research studies that could attain huge sample sizes and hence greatly improve our understanding of the human brain. Recent advances in convolutional neural networks (CNNs) are producing outstanding results in super-resolution and contrast synthesis of MRI. However, these approaches are very sensitive to the specific combination of contrast, resolution and orientation of the input images, and thus do not generalize to diverse clinical acquisition protocols - even within sites. In this article, we present SynthSR, a method to train a CNN that receives one or more scans with spaced slices, acquired with different contrast, resolution and orientation, and produces an isotropic scan of canonical contrast (typically a 1 mm MP-RAGE). The presented method does not require any preprocessing, beyond rigid coregistration of the input scans. Crucially, SynthSR trains on synthetic input images generated from 3D segmentations, and can thus be used to train CNNs for any combination of contrasts, resolutions and orientations without high-resolution real images of the input contrasts. We test the images generated with SynthSR in an array of common downstream analyses, and show that they can be reliably used for subcortical segmentation and volumetry, image registration (e.g., for tensor-based morphometry), and, if some image quality requirements are met, even cortical thickness morphometry. The source code is publicly available at https://github.com/BBillot/SynthSR.

Deformable MRI-Ultrasound registration using correlation-based attribute matching for brain shift correction: Accuracy and generality in multi-site data.

Intraoperative tissue deformation, known as brain shift, decreases the benefit of using preoperative images to guide neurosurgery. Non-rigid registration of preoperative magnetic resonance (MR) to intraoperative ultrasound (iUS) has been proposed as a means to compensate for brain shift. We focus on the initial registration from MR to predurotomy iUS. We present a method that builds on previous work to address the need for accuracy and generality of MR-iUS registration algorithms in multi-site clinical data. High-dimensional texture attributes were used instead of image intensities for image registration and the standard difference-based attribute matching was replaced with correlation-based attribute matching. A strategy that deals explicitly with the large field-of-view mismatch between MR and iUS images was proposed. Key parameters were optimized across independent MR-iUS brain tumor datasets acquired at 3 institutions, with a total of 43 tumor patients and 758 reference landmarks for evaluating the accuracy of the proposed algorithm. Despite differences in imaging protocols, patient demographics and landmark distributions, the algorithm is able to reduce landmark errors prior to registration in three data sets (5.37±4.27, 4.18±1.97 and 6.18±3.38 mm, respectively) to a consistently low level (2.28±0.71, 2.08±0.37 and 2.24±0.78 mm, respectively). This algorithm was tested against 15 other algorithms and it is competitive with the state-of-the-art on multiple datasets. We show that the algorithm has one of the lowest errors in all datasets (accuracy), and this is achieved while sticking to a fixed set of parameters for multi-site data (generality). In contrast, other algorithms/tools of similar performance need per-dataset parameter tuning (high accuracy but lower generality), and those that stick to fixed parameters have larger errors or inconsistent performance (generality but not the top accuracy). Landmark errors were further characterized according to brain regions and tumor types, a topic so far missing in the literature.

Spatiotemporal alignment of in utero BOLD-MRI series.

PURPOSE: To present a method for spatiotemporal alignment of in-utero magnetic resonance imaging (MRI) time series acquired during maternal hyperoxia for enabling improved quantitative tracking of blood oxygen level-dependent (BOLD) signal changes that characterize oxygen transport through the placenta to fetal organs. MATERIALS AND METHODS: The proposed pipeline for spatiotemporal alignment of images acquired with a single-shot gradient echo echo-planar imaging includes 1) signal nonuniformity correction, 2) intravolume motion correction based on nonrigid registration, 3) correction of motion and nonrigid deformations across volumes, and 4) detection of the outlier volumes to be discarded from subsequent analysis. BOLD MRI time series collected from 10 pregnant women during 3T scans were analyzed using this pipeline. To assess pipeline performance, signal fluctuations between consecutive timepoints were examined. In addition, volume overlap and distance between manual region of interest (ROI) delineations in a subset of frames and the delineations obtained through propagation of the ROIs from the reference frame were used to quantify alignment accuracy. A previously demonstrated rigid registration approach was used for comparison. RESULTS: The proposed pipeline improved anatomical alignment of placenta and fetal organs over the state-of-the-art rigid motion correction methods. In particular, unexpected temporal signal fluctuations during the first normoxia period were significantly decreased (P < 0.01) and volume overlap and distance between region boundaries measures were significantly improved (P < 0.01). CONCLUSION: The proposed approach to align MRI time series enables more accurate quantitative studies of placental function by improving spatiotemporal alignment across placenta and fetal organs. LEVEL OF EVIDENCE: 1 Technical Efficacy: Stage 1 J. MAGN. RESON. IMAGING 2017;46:403-412.

Identifying Shared Brain Networks in Individuals by Decoupling Functional and Anatomical Variability.

The connectivity architecture of the human brain varies across individuals. Mapping functional anatomy at the individual level is challenging, but critical for basic neuroscience research and clinical intervention. Using resting-state functional connectivity, we parcellated functional systems in an "embedding space" based on functional characteristics common across the population, while simultaneously accounting for individual variability in the cortical distribution of functional units. The functional connectivity patterns observed in resting-state data were mapped in the embedding space and the maps were aligned across individuals. A clustering algorithm was performed on the aligned embedding maps and the resulting clusters were transformed back to the unique anatomical space of each individual. This novel approach identified functional systems that were reproducible within subjects, but were distributed across different anatomical locations in different subjects. Using this approach for intersubject alignment improved the predictability of individual differences in language laterality when compared with anatomical alignment alone. Our results further revealed that the strength of association between function and macroanatomy varied across the cortex, which was strong in unimodal sensorimotor networks, but weak in association networks.

BrainPrint: a discriminative characterization of brain morphology.

We introduce BrainPrint, a compact and discriminative representation of brain morphology. BrainPrint captures shape information of an ensemble of cortical and subcortical structures by solving the eigenvalue problem of the 2D and 3D Laplace-Beltrami operator on triangular (boundary) and tetrahedral (volumetric) meshes. This discriminative characterization enables new ways to study the similarity between brains; the focus can either be on a specific brain structure of interest or on the overall brain similarity. We highlight four applications for BrainPrint in this article: (i) subject identification, (ii) age and sex prediction, (iii) brain asymmetry analysis, and (iv) potential genetic influences on brain morphology. The properties of BrainPrint require the derivation of new algorithms to account for the heterogeneous mix of brain structures with varying discriminative power. We conduct experiments on three datasets, including over 3000 MRI scans from the ADNI database, 436 MRI scans from the OASIS dataset, and 236 MRI scans from the VETSA twin study. All processing steps for obtaining the compact representation are fully automated, making this processing framework particularly attractive for handling large datasets.

Multi-modal robust inverse-consistent linear registration.

Registration performance can significantly deteriorate when image regions do not comply with model assumptions. Robust estimation improves registration accuracy by reducing or ignoring the contribution of voxels with large intensity differences, but existing approaches are limited to monomodal registration. In this work, we propose a robust and inverse-consistent technique for cross-modal, affine image registration. The algorithm is derived from a contextual framework of image registration. The key idea is to use a modality invariant representation of images based on local entropy estimation, and to incorporate a heteroskedastic noise model. This noise model allows us to draw the analogy to iteratively reweighted least squares estimation and to leverage existing weighting functions to account for differences in local information content in multimodal registration. Furthermore, we use the nonparametric windows density estimator to reliably calculate entropy of small image patches. Finally, we derive the Gauss-Newton update and show that it is equivalent to the efficient second-order minimization for the fully symmetric registration approach. We illustrate excellent performance of the proposed methods on datasets containing outliers for alignment of brain tumor, full head, and histology images.

An image analysis toolbox for high-throughput C. elegans assays.

We present a toolbox for high-throughput screening of image-based Caenorhabditis elegans phenotypes. The image analysis algorithms measure morphological phenotypes in individual worms and are effective for a variety of assays and imaging systems. This WormToolbox is available through the open-source CellProfiler project and enables objective scoring of whole-worm high-throughput image-based assays of C. elegans for the study of diverse biological pathways that are relevant to human disease.

Coping with confounds in multivoxel pattern analysis: what should we do about reaction time differences? A comment on Todd, Nystrom & Cohen 2013.

Multivoxel pattern analysis (MVPA) is a sensitive and increasingly popular method for examining differences between neural activation patterns that cannot be detected using classical mass-univariate analysis. Recently, Todd et al. ("Confounds in multivariate pattern analysis: Theory and rule representation case study", 2013, NeuroImage 77: 157-165) highlighted a potential problem for these methods: high sensitivity to confounds at the level of individual participants due to the use of directionless summary statistics. Unlike traditional mass-univariate analyses where confounding activation differences in opposite directions tend to approximately average out at group level, group level MVPA results may be driven by any activation differences that can be discriminated in individual participants. In Todd et al.'s empirical data, factoring out differences in reaction time (RT) reduced a classifier's ability to distinguish patterns of activation pertaining to two task rules. This raises two significant questions for the field: to what extent have previous multivoxel discriminations in the literature been driven by RT differences, and by what methods should future studies take RT and other confounds into account? We build on the work of Todd et al. and compare two different approaches to remove the effect of RT in MVPA. We show that in our empirical data, in contrast to that of Todd et al., the effect of RT on rule decoding is negligible, and results were not affected by the specific details of RT modelling. We discuss the meaning of and sensitivity for confounds in traditional and multivoxel approaches to fMRI analysis. We observe that the increased sensitivity of MVPA comes at a price of reduced specificity, meaning that these methods in particular call for careful consideration of what differs between our conditions of interest. We conclude that the additional complexity of the experimental design, analysis and interpretation needed for MVPA is still not a reason to favour a less sensitive approach.

Decoupling function and anatomy in atlases of functional connectivity patterns: language mapping in tumor patients.

In this paper we construct an atlas that summarizes functional connectivity characteristics of a cognitive process from a population of individuals. The atlas encodes functional connectivity structure in a low-dimensional embedding space that is derived from a diffusion process on a graph that represents correlations of fMRI time courses. The functional atlas is decoupled from the anatomical space, and thus can represent functional networks with variable spatial distribution in a population. In practice the atlas is represented by a common prior distribution for the embedded fMRI signals of all subjects. We derive an algorithm for fitting this generative model to the observed data in a population. Our results in a language fMRI study demonstrate that the method identifies coherent and functionally equivalent regions across subjects. The method also successfully maps functional networks from a healthy population used as a training set to individuals whose language networks are affected by tumors.

Segmentation of anatomical layers and imaging artifacts in intravascular polarization sensitive optical coherence tomography using attending physician and boundary cardinality losses.

Intravascular ultrasound and optical coherence tomography are widely available for assessing coronary stenoses and provide critical information to optimize percutaneous coronary intervention. Intravascular polarization-sensitive optical coherence tomography (PS-OCT) measures the polarization state of the light scattered by the vessel wall in addition to conventional cross-sectional images of subsurface microstructure. This affords reconstruction of tissue polarization properties and reveals improved contrast between the layers of the vessel wall along with insight into collagen and smooth muscle content. Here, we propose a convolutional neural network model, optimized using two new loss terms (Boundary Cardinality and Attending Physician), that takes advantage of the additional polarization contrast and classifies the lumen, intima, and media layers in addition to guidewire and plaque shadows. Our model segments the media boundaries through fibrotic plaques and continues to estimate the outer media boundary behind shadows of lipid-rich plaques. We demonstrate that our multi-class classification model outperforms existing methods that exclusively use conventional OCT data, predominantly segment the lumen, and consider subsurface layers at most in regions of minimal disease. Segmentation of all anatomical layers throughout diseased vessels may facilitate stent sizing and will enable automated characterization of plaque polarization properties for investigation of the natural history and significance of coronary atheromas.

SE(3)-Equivariant and Noise-Invariant 3D Rigid Motion Tracking in Brain MRI.

Rigid motion tracking is paramount in many medical imaging applications where movements need to be detected, corrected, or accounted for. Modern strategies rely on convolutional neural networks (CNN) and pose this problem as rigid registration. Yet, CNNs do not exploit natural symmetries in this task, as they are equivariant to translations (their outputs shift with their inputs) but not to rotations. Here we propose EquiTrack, the first method that uses recent steerable SE(3)-equivariant CNNs (E-CNN) for motion tracking. While steerable E-CNNs can extract corresponding features across different poses, testing them on noisy medical images reveals that they do not have enough learning capacity to learn noise invariance. Thus, we introduce a hybrid architecture that pairs a denoiser with an E-CNN to decouple the processing of anatomically irrelevant intensity features from the extraction of equivariant spatial features. Rigid transforms are then estimated in closed-form. EquiTrack outperforms state-of-the-art learning and optimisation methods for motion tracking in adult brain MRI and fetal MRI time series. Our code is available at https://github.com/BBillot/EquiTrack.

Spatial-Intensity Transforms for Medical Image-to-Image Translation.

Image-to-image translation has seen major advances in computer vision but can be difficult to apply to medical images, where imaging artifacts and data scarcity degrade the performance of conditional generative adversarial networks. We develop the spatial-intensity transform (SIT) to improve output image quality while closely matching the target domain. SIT constrains the generator to a smooth spatial transform (diffeomorphism) composed with sparse intensity changes. SIT is a lightweight, modular network component that is effective on various architectures and training schemes. Relative to unconstrained baselines, this technique significantly improves image fidelity, and our models generalize robustly to different scanners. Additionally, SIT provides a disentangled view of anatomical and textural changes for each translation, making it easier to interpret the model's predictions in terms of physiological phenomena. We demonstrate SIT on two tasks: predicting longitudinal brain MRIs in patients with various stages of neurodegeneration, and visualizing changes with age and stroke severity in clinical brain scans of stroke patients. On the first task, our model accurately forecasts brain aging trajectories without supervised training on paired scans. On the second task, it captures associations between ventricle expansion and aging, as well as between white matter hyperintensities and stroke severity. As conditional generative models become increasingly versatile tools for visualization and forecasting, our approach demonstrates a simple and powerful technique for improving robustness, which is critical for translation to clinical settings. Source code is available at github.com/clintonjwang/spatial-intensity-transforms.

NeSVoR: Implicit Neural Representation for Slice-to-Volume Reconstruction in MRI.

Reconstructing 3D MR volumes from multiple motion-corrupted stacks of 2D slices has shown promise in imaging of moving subjects, e. g., fetal MRI. However, existing slice-to-volume reconstruction methods are time-consuming, especially when a high-resolution volume is desired. Moreover, they are still vulnerable to severe subject motion and when image artifacts are present in acquired slices. In this work, we present NeSVoR, a resolution-agnostic slice-to-volume reconstruction method, which models the underlying volume as a continuous function of spatial coordinates with implicit neural representation. To improve robustness to subject motion and other image artifacts, we adopt a continuous and comprehensive slice acquisition model that takes into account rigid inter-slice motion, point spread function, and bias fields. NeSVoR also estimates pixel-wise and slice-wise variances of image noise and enables removal of outliers during reconstruction and visualization of uncertainty. Extensive experiments are performed on both simulated and in vivo data to evaluate the proposed method. Results show that NeSVoR achieves state-of-the-art reconstruction quality while providing two to ten-fold acceleration in reconstruction times over the state-of-the-art algorithms.

Supervision by Denoising.

Learning-based image reconstruction models, such as those based on the U-Net, require a large set of labeled images if good generalization is to be guaranteed. In some imaging domains, however, labeled data with pixel- or voxel-level label accuracy are scarce due to the cost of acquiring them. This problem is exacerbated further in domains like medical imaging, where there is no single ground truth label, resulting in large amounts of repeat variability in the labels. Therefore, training reconstruction networks to generalize better by learning from both labeled and unlabeled examples (called semi-supervised learning) is problem of practical and theoretical interest. However, traditional semi-supervised learning methods for image reconstruction often necessitate handcrafting a differentiable regularizer specific to some given imaging problem, which can be extremely time-consuming. In this work, we propose "supervision by denoising" (SUD), a framework to supervise reconstruction models using their own denoised output as labels. SUD unifies stochastic averaging and spatial denoising techniques under a spatio-temporal denoising framework and alternates denoising and model weight update steps in an optimization framework for semi-supervision. As example applications, we apply SUD to two problems from biomedical imaging-anatomical brain reconstruction (3D) and cortical parcellation (2D)-to demonstrate a significant improvement in reconstruction over supervised-only and ensembling baselines. Our code available at https://github.com/seannz/sud.

SynthSR: A public AI tool to turn heterogeneous clinical brain scans into high-resolution T1-weighted images for 3D morphometry.

Every year, millions of brain magnetic resonance imaging (MRI) scans are acquired in hospitals across the world. These have the potential to revolutionize our understanding of many neurological diseases, but their morphometric analysis has not yet been possible due to their anisotropic resolution. We present an artificial intelligence technique, "SynthSR," that takes clinical brain MRI scans with any MR contrast (T1, T2, etc.), orientation (axial/coronal/sagittal), and resolution and turns them into high-resolution T1 scans that are usable by virtually all existing human neuroimaging tools. We present results on segmentation, registration, and atlasing of >10,000 scans of controls and patients with brain tumors, strokes, and Alzheimer's disease. SynthSR yields morphometric results that are very highly correlated with what one would have obtained with high-resolution T1 scans. SynthSR allows sample sizes that have the potential to overcome the power limitations of prospective research studies and shed new light on the healthy and diseased human brain.

Domain-agnostic segmentation of thalamic nuclei from joint structural and diffusion MRI.

The human thalamus is a highly connected subcortical grey-matter structure within the brain. It comprises dozens of nuclei with different function and connectivity, which are affected differently by disease. For this reason, there is growing interest in studying the thalamic nuclei in vivo with MRI. Tools are available to segment the thalamus from 1 mm T1 scans, but the contrast of the lateral and internal boundaries is too faint to produce reliable segmentations. Some tools have attempted to incorporate information from diffusion MRI in the segmentation to refine these boundaries, but do not generalise well across diffusion MRI acquisitions. Here we present the first CNN that can segment thalamic nuclei from T1 and diffusion data of any resolution without retraining or fine tuning. Our method builds on a public histological atlas of the thalamic nuclei and silver standard segmentations on high-quality diffusion data obtained with a recent Bayesian adaptive segmentation tool. We combine these with an approximate degradation model for fast domain randomisation during training. Our CNN produces a segmentation at 0.7 mm isotropic resolution, irrespective of the resolution of the input. Moreover, it uses a parsimonious model of the diffusion signal at each voxel (fractional anisotropy and principal eigenvector) that is compatible with virtually any set of directions and b-values, including huge amounts of legacy data. We show results of our proposed method on three heterogeneous datasets acquired on dozens of different scanners. An implementation of the method is publicly available at https://freesurfer.net/fswiki/ThalamicNucleiDTI.