Multi-Head Graph Convolutional Network for Structural Connectome Classification.

We tackle classification based on brain connectivity derived from diffusion magnetic resonance images. We propose a machine-learning model inspired by graph convolutional networks (GCNs), which takes a brain-connectivity input graph and processes the data separately through a parallel GCN mechanism with multiple heads. The proposed network is a simple design that employs different heads involving graph convolutions focused on edges and nodes, thoroughly capturing representations from the input data. To test the ability of our model to extract complementary and representative features from brain connectivity data, we chose the task of sex classification. This quantifies the degree to which the connectome varies depending on the sex, which is important for improving our understanding of health and disease in both sexes. We show experiments on two publicly available datasets: PREVENT-AD (347 subjects) and OASIS3 (771 subjects). The proposed model demonstrates the highest performance compared to the existing machine-learning algorithms we tested, including classical methods and (graph and non-graph) deep learning. We provide a detailed analysis of each component of our model.

BOOSTING SKULL-STRIPPING PERFORMANCE FOR PEDIATRIC BRAIN IMAGES.

Skull-stripping is the removal of background and non-brain anatomical features from brain images. While many skull-stripping tools exist, few target pediatric populations. With the emergence of multi-institutional pediatric data acquisition efforts to broaden the understanding of perinatal brain development, it is essential to develop robust and well-tested tools ready for the relevant data processing. However, the broad range of neuroanatomical variation in the developing brain, combined with additional challenges such as high motion levels, as well as shoulder and chest signal in the images, leaves many adult-specific tools ill-suited for pediatric skull-stripping. Building on an existing framework for robust and accurate skull-stripping, we propose developmental SynthStrip (d-SynthStrip), a skull-stripping model tailored to pediatric images. This framework exposes networks to highly variable images synthesized from label maps. Our model substantially outperforms pediatric baselines across scan types and age cohorts. In addition, the <1-minute runtime of our tool compares favorably to the fastest baselines. We distribute our model at https://w3id.org/synthstrip.

Scaling behaviours of deep learning and linear algorithms for the prediction of stroke severity.

Deep learning has allowed for remarkable progress in many medical scenarios. Deep learning prediction models often require 105-107 examples. It is currently unknown whether deep learning can also enhance predictions of symptoms post-stroke in real-world samples of stroke patients that are often several magnitudes smaller. Such stroke outcome predictions however could be particularly instrumental in guiding acute clinical and rehabilitation care decisions. We here compared the capacities of classically used linear and novel deep learning algorithms in their prediction of stroke severity. Our analyses relied on a total of 1430 patients assembled from the MRI-Genetics Interface Exploration collaboration and a Massachusetts General Hospital-based study. The outcome of interest was National Institutes of Health Stroke Scale-based stroke severity in the acute phase after ischaemic stroke onset, which we predict by means of MRI-derived lesion location. We automatically derived lesion segmentations from diffusion-weighted clinical MRI scans, performed spatial normalization and included a principal component analysis step, retaining 95% of the variance of the original data. We then repeatedly separated a train, validation and test set to investigate the effects of sample size; we subsampled the train set to 100, 300 and 900 and trained the algorithms to predict the stroke severity score for each sample size with regularized linear regression and an eight-layered neural network. We selected hyperparameters on the validation set. We evaluated model performance based on the explained variance (R2) in the test set. While linear regression performed significantly better for a sample size of 100 patients, deep learning started to significantly outperform linear regression when trained on 900 patients. Average prediction performance improved by ∼20% when increasing the sample size 9× [maximum for 100 patients: 0.279 ± 0.005 (R2, 95% confidence interval), 900 patients: 0.337 ± 0.006]. In summary, for sample sizes of 900 patients, deep learning showed a higher prediction performance than typically employed linear methods. These findings suggest the existence of non-linear relationships between lesion location and stroke severity that can be utilized for an improved prediction performance for larger sample sizes.

Machine learning of dissection photographs and surface scanning for quantitative 3D neuropathology.

We present open-source tools for 3D analysis of photographs of dissected slices of human brains, which are routinely acquired in brain banks but seldom used for quantitative analysis. Our tools can: (i) 3D reconstruct a volume from the photographs and, optionally, a surface scan; and (ii) produce a high-resolution 3D segmentation into 11 brain regions per hemisphere (22 in total), independently of the slice thickness. Our tools can be used as a substitute for ex vivo magnetic resonance imaging (MRI), which requires access to an MRI scanner, ex vivo scanning expertise, and considerable financial resources. We tested our tools on synthetic and real data from two NIH Alzheimer's Disease Research Centers. The results show that our methodology yields accurate 3D reconstructions, segmentations, and volumetric measurements that are highly correlated to those from MRI. Our method also detects expected differences between post mortem confirmed Alzheimer's disease cases and controls. The tools are available in our widespread neuroimaging suite "FreeSurfer" ( https://surfer.nmr.mgh.harvard.edu/fswiki/PhotoTools ).

A cellular resolution atlas of Broca's area.

Brain cells are arranged in laminar, nuclear, or columnar structures, spanning a range of scales. Here, we construct a reliable cell census in the frontal lobe of human cerebral cortex at micrometer resolution in a magnetic resonance imaging (MRI)-referenced system using innovative imaging and analysis methodologies. MRI establishes a macroscopic reference coordinate system of laminar and cytoarchitectural boundaries. Cell counting is obtained with a digital stereological approach on the 3D reconstruction at cellular resolution from a custom-made inverted confocal light-sheet fluorescence microscope (LSFM). Mesoscale optical coherence tomography enables the registration of the distorted histological cell typing obtained with LSFM to the MRI-based atlas coordinate system. The outcome is an integrated high-resolution cellular census of Broca's area in a human postmortem specimen, within a whole-brain reference space atlas.

Joint cortical registration of geometry and function using semi-supervised learning.

Brain surface-based image registration, an important component of brain image analysis, establishes spatial correspondence between cortical surfaces. Existing iterative and learning-based approaches focus on accurate registration of folding patterns of the cerebral cortex, and assume that geometry predicts function and thus functional areas will also be well aligned. However, structure/functional variability of anatomically corresponding areas across subjects has been widely reported. In this work, we introduce a learning-based cortical registration framework, JOSA, which jointly aligns folding patterns and functional maps while simultaneously learning an optimal atlas. We demonstrate that JOSA can substantially improve registration performance in both anatomical and functional domains over existing methods. By employing a semi-supervised training strategy, the proposed framework obviates the need for functional data during inference, enabling its use in broad neuroscientific domains where functional data may not be observed. The source code of JOSA will be released to the public at https://voxelmorph.net.

Can we predict motion artifacts in clinical MRI before the scan completes?

Subject motion can cause artifacts in clinical MRI, frequently necessitating repeat scans. We propose to alleviate this inefficiency by predicting artifact scores from partial multi-shot multi-slice acquisitions, which may guide the operator in aborting corrupted scans early.

A robust and interpretable deep learning framework for multi-modal registration via keypoints.

We present KeyMorph, a deep learning-based image registration framework that relies on automatically detecting corresponding keypoints. State-of-the-art deep learning methods for registration often are not robust to large misalignments, are not interpretable, and do not incorporate the symmetries of the problem. In addition, most models produce only a single prediction at test-time. Our core insight which addresses these shortcomings is that corresponding keypoints between images can be used to obtain the optimal transformation via a differentiable closed-form expression. We use this observation to drive the end-to-end learning of keypoints tailored for the registration task, and without knowledge of ground-truth keypoints. This framework not only leads to substantially more robust registration but also yields better interpretability, since the keypoints reveal which parts of the image are driving the final alignment. Moreover, KeyMorph can be designed to be equivariant under image translations and/or symmetric with respect to the input image ordering. Finally, we show how multiple deformation fields can be computed efficiently and in closed-form at test time corresponding to different transformation variants. We demonstrate the proposed framework in solving 3D affine and spline-based registration of multi-modal brain MRI scans. In particular, we show registration accuracy that surpasses current state-of-the-art methods, especially in the context of large displacements. Our code is available at https://github.com/alanqrwang/keymorph.

Clinical evaluation of k-space correlation informed motion artifact detection in segmented multi-slice MRI.

Motion artifacts can negatively impact diagnosis, patient experience, and radiology workflow especially when a patient recall is required. Detecting motion artifacts while the patient is still in the scanner could potentially improve workflow and reduce costs by enabling immediate corrective action. We demonstrate in a clinical k-space dataset that using cross-correlation between adjacent phase-encoding lines can detect motion artifacts directly from raw k-space in multi-shot multi-slice scans. We train a split-attention residual network to examine the performance in predicting motion artifact severity. The network is trained on simulated data and tested on real clinical data.

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.

Multi-Head Graph Convolutional Network for Structural Connectome Classification.

We tackle classification based on brain connectivity derived from diffusion magnetic resonance images. We propose a machine-learning model inspired by graph convolutional networks (GCNs), which takes a brain connectivity input graph and processes the data separately through a parallel GCN mechanism with multiple heads. The proposed network is a simple design that employs different heads involving graph convolutions focused on edges and nodes, capturing representations from the input data thoroughly. To test the ability of our model to extract complementary and representative features from brain connectivity data, we chose the task of sex classification. This quantifies the degree to which the connectome varies depending on the sex, which is important for improving our understanding of health and disease in both sexes. We show experiments on two publicly available datasets: PREVENT-AD (347 subjects) and OASIS3 (771 subjects). The proposed model demonstrates the highest performance compared to the existing machine-learning algorithms we tested, including classical methods and (graph and non-graph) deep learning. We provide a detailed analysis of each component of our model.

SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining.

Despite advances in data augmentation and transfer learning, convolutional neural networks (CNNs) difficultly generalise to unseen domains. When segmenting brain scans, CNNs are highly sensitive to changes in resolution and contrast: even within the same MRI modality, performance can decrease across datasets. Here we introduce SynthSeg, the first segmentation CNN robust against changes in contrast and resolution. SynthSeg is trained with synthetic data sampled from a generative model conditioned on segmentations. Crucially, we adopt a domain randomisation strategy where we fully randomise the contrast and resolution of the synthetic training data. Consequently, SynthSeg can segment real scans from a wide range of target domains without retraining or fine-tuning, which enables straightforward analysis of huge amounts of heterogeneous clinical data. Because SynthSeg only requires segmentations to be trained (no images), it can learn from labels obtained by automated methods on diverse populations (e.g., ageing and diseased), thus achieving robustness to a wide range of morphological variability. We demonstrate SynthSeg on 5,000 scans of six modalities (including CT) and ten resolutions, where it exhibits unparallelled generalisation compared with supervised CNNs, state-of-the-art domain adaptation, and Bayesian segmentation. Finally, we demonstrate the generalisability of SynthSeg by applying it to cardiac MRI and CT scans.

Hyper-convolutions via implicit kernels for medical image analysis.

The convolutional neural network (CNN) is one of the most commonly used architectures for computer vision tasks. The key building block of a CNN is the convolutional kernel that aggregates information from the pixel neighborhood and shares weights across all pixels. A standard CNN's capacity, and thus its performance, is directly related to the number of learnable kernel weights, which is determined by the number of channels and the kernel size (support). In this paper, we present the hyper-convolution, a novel building block that implicitly encodes the convolutional kernel using spatial coordinates. Unlike a regular convolutional kernel, whose weights are independently learned, hyper-convolution kernel weights are correlated through an encoder that maps spatial coordinates to their corresponding values. Hyper-convolutions decouple kernel size from the total number of learnable parameters, enabling a more flexible architecture design. We demonstrate in our experiments that replacing regular convolutions with hyper-convolutions can improve performance with less parameters, and increase robustness against noise. We provide our code here: https://github.com/tym002/Hyper-Convolution.

The relevance of rich club regions for functional outcome post-stroke is enhanced in women.

This study aimed to investigate the influence of stroke lesions in predefined highly interconnected (rich-club) brain regions on functional outcome post-stroke, determine their spatial specificity and explore the effects of biological sex on their relevance. We analyzed MRI data recorded at index stroke and ~3-months modified Rankin Scale (mRS) data from patients with acute ischemic stroke enrolled in the multisite MRI-GENIE study. Spatially normalized structural stroke lesions were parcellated into 108 atlas-defined bilateral (sub)cortical brain regions. Unfavorable outcome (mRS > 2) was modeled in a Bayesian logistic regression framework. Effects of individual brain regions were captured as two compound effects for (i) six bilateral rich club and (ii) all further non-rich club regions. In spatial specificity analyses, we randomized the split into "rich club" and "non-rich club" regions and compared the effect of the actual rich club regions to the distribution of effects from 1000 combinations of six random regions. In sex-specific analyses, we introduced an additional hierarchical level in our model structure to compare male and female-specific rich club effects. A total of 822 patients (age: 64.7[15.0], 39% women) were analyzed. Rich club regions had substantial relevance in explaining unfavorable functional outcome (mean of posterior distribution: 0.08, area under the curve: 0.8). In particular, the rich club-combination had a higher relevance than 98.4% of random constellations. Rich club regions were substantially more important in explaining long-term outcome in women than in men. All in all, lesions in rich club regions were associated with increased odds of unfavorable outcome. These effects were spatially specific and more pronounced in women.

Radiomics-Derived Brain Age Predicts Functional Outcome After Acute Ischemic Stroke.

BACKGROUND AND OBJECTIVES: While chronological age is one of the most influential determinants of poststroke outcomes, little is known of the impact of neuroimaging-derived biological "brain age." We hypothesized that radiomics analyses of T2-FLAIR images texture would provide brain age estimates and that advanced brain age of patients with stroke will be associated with cardiovascular risk factors and worse functional outcomes. METHODS: We extracted radiomics from T2-FLAIR images acquired during acute stroke clinical evaluation. Brain age was determined from brain parenchyma radiomics using an ElasticNet linear regression model. Subsequently, relative brain age (RBA), which expresses brain age in comparison with chronological age-matched peers, was estimated. Finally, we built a linear regression model of RBA using clinical cardiovascular characteristics as inputs and a logistic regression model of favorable functional outcomes taking RBA as input. RESULTS: We reviewed 4,163 patients from a large multisite ischemic stroke cohort (mean age = 62.8 years, 42.0% female patients). T2-FLAIR radiomics predicted chronological ages (mean absolute error = 6.9 years, r = 0.81). After adjustment for covariates, RBA was higher and therefore described older-appearing brains in patients with hypertension, diabetes mellitus, a history of smoking, and a history of a prior stroke. In multivariate analyses, age, RBA, NIHSS, and a history of prior stroke were all significantly associated with functional outcome (respective adjusted odds ratios: 0.58, 0.76, 0.48, 0.55; all p-values < 0.001). Moreover, the negative effect of RBA on outcome was especially pronounced in minor strokes. DISCUSSION: T2-FLAIR radiomics can be used to predict brain age and derive RBA. Older-appearing brains, characterized by a higher RBA, reflect cardiovascular risk factor accumulation and are linked to worse outcomes after stroke.

SuperWarp: Supervised Learning and Warping on U-Net for Invariant Subvoxel-Precise Registration.

In recent years, learning-based image registration methods have gradually moved away from direct supervision with target warps to instead use self-supervision, with excellent results in several registration benchmarks. These approaches utilize a loss function that penalizes the intensity differences between the fixed and moving images, along with a suitable regularizer on the deformation. However, since images typically have large untextured regions, merely maximizing similarity between the two images is not sufficient to recover the true deformation. This problem is exacerbated by texture in other regions, which introduces severe non-convexity into the landscape of the training objective and ultimately leads to overfitting. In this paper, we argue that the relative failure of supervised registration approaches can in part be blamed on the use of regular U-Nets, which are jointly tasked with feature extraction, feature matching and deformation estimation. Here, we introduce a simple but crucial modification to the U-Net that disentangles feature extraction and matching from deformation prediction, allowing the U-Net to warp the features, across levels, as the deformation field is evolved. With this modification, direct supervision using target warps begins to outperform self-supervision approaches that require segmentations, presenting new directions for registration when images do not have segmentations. We hope that our findings in this preliminary workshop paper will re-ignite research interest in supervised image registration techniques. Our code is publicly available from http://github.com/balbasty/superwarp.

Joint Frequency and Image Space Learning for MRI Reconstruction and Analysis.

We propose neural network layers that explicitly combine frequency and image feature representations and show that they can be used as a versatile building block for reconstruction from frequency space data. Our work is motivated by the challenges arising in MRI acquisition where the signal is a corrupted Fourier transform of the desired image. The proposed joint learning schemes enable both correction of artifacts native to the frequency space and manipulation of image space representations to reconstruct coherent image structures at every layer of the network. This is in contrast to most current deep learning approaches for image reconstruction that treat frequency and image space features separately and often operate exclusively in one of the two spaces. We demonstrate the advantages of joint convolutional learning for a variety of tasks, including motion correction, denoising, reconstruction from undersampled acquisitions, and combined undersampling and motion correction on simulated and real world multicoil MRI data. The joint models produce consistently high quality output images across all tasks and datasets. When integrated into a state of the art unrolled optimization network with physics-inspired data consistency constraints for undersampled reconstruction, the proposed architectures significantly improve the optimization landscape, which yields an order of magnitude reduction of training time. This result suggests that joint representations are particularly well suited for MRI signals in deep learning networks. Our code and pretrained models are publicly available at https://github.com/nalinimsingh/interlacer.

Deep profiling of multiple ischemic lesions in a large, multi-center cohort: Frequency, spatial distribution, and associations to clinical characteristics.

Background purpose: A substantial number of patients with acute ischemic stroke (AIS) experience multiple acute lesions (MAL). We here aimed to scrutinize MAL in a large radiologically deep-phenotyped cohort. Materials and methods: Analyses relied upon imaging and clinical data from the international MRI-GENIE study. Imaging data comprised both Fluid-attenuated inversion recovery (FLAIR) for white matter hyperintensity (WMH) burden estimation and diffusion-weighted imaging (DWI) sequences for the assessment of acute stroke lesions. The initial step featured the systematic evaluation of occurrences of MAL within one and several vascular supply territories. Associations between MAL and important imaging and clinical characteristics were subsequently determined. The interaction effect between single and multiple lesion status and lesion volume was estimated by means of Bayesian hierarchical regression modeling for both stroke severity and functional outcome. Results: We analyzed 2,466 patients (age = 63.4 ± 14.8, 39% women), 49.7% of which presented with a single lesion. Another 37.4% experienced MAL in a single vascular territory, while 12.9% featured lesions in multiple vascular territories. Within most territories, MAL occurred as frequently as single lesions (ratio ∼1:1). Only the brainstem region comprised fewer patients with MAL (ratio 1:4). Patients with MAL presented with a significantly higher lesion volume and acute NIHSS (7.7 vs. 1.7 ml and 4 vs. 3, p FDR < 0.001). In contrast, patients with a single lesion were characterized by a significantly higher WMH burden (6.1 vs. 5.3 ml, p FDR = 0.048). Functional outcome did not differ significantly between patients with single versus multiple lesions. Bayesian analyses suggested that the association between lesion volume and stroke severity between single and multiple lesions was the same in case of anterior circulation stroke. In case of posterior circulation stroke, lesion volume was linked to a higher NIHSS only among those with MAL. Conclusion: Multiple lesions, especially those within one vascular territory, occurred more frequently than previously reported. Overall, multiple lesions were distinctly linked to a higher acute stroke severity, a higher total DWI lesion volume and a lower WMH lesion volume. In posterior circulation stroke, lesion volume was linked to a higher stroke severity in multiple lesions only.

Learning the Effect of Registration Hyperparameters with HyperMorph.

We introduce HyperMorph, a framework that facilitates efficient hyperparameter tuning in learning-based deformable image registration. Classical registration algorithms perform an iterative pair-wise optimization to compute a deformation field that aligns two images. Recent learning-based approaches leverage large image datasets to learn a function that rapidly estimates a deformation for a given image pair. In both strategies, the accuracy of the resulting spatial correspondences is strongly influenced by the choice of certain hyperparameter values. However, an effective hyperparameter search consumes substantial time and human effort as it often involves training multiple models for different fixed hyperparameter values and may lead to suboptimal registration. We propose an amortized hyperparameter learning strategy to alleviate this burden by learning the impact of hyperparameters on deformation fields. We design a meta network, or hypernetwork, that predicts the parameters of a registration network for input hyperparameters, thereby comprising a single model that generates the optimal deformation field corresponding to given hyperparameter values. This strategy enables fast, high-resolution hyperparameter search at test-time, reducing the inefficiency of traditional approaches while increasing flexibility. We also demonstrate additional benefits of HyperMorph, including enhanced robustness to model initialization and the ability to rapidly identify optimal hyperparameter values specific to a dataset, image contrast, task, or even anatomical region, all without the need to retrain models. We make our code publicly available at http://hypermorph.voxelmorph.net.

Association of Stroke Lesion Pattern and White Matter Hyperintensity Burden With Stroke Severity and Outcome.

BACKGROUND AND OBJECTIVES: To examine whether high white matter hyperintensity (WMH) burden is associated with greater stroke severity and worse functional outcomes in lesion pattern-specific ways. METHODS: MR neuroimaging and NIH Stroke Scale data at index stroke and the modified Rankin Scale (mRS) score at 3-6 months after stroke were obtained from the MRI-Genetics Interface Exploration study of patients with acute ischemic stroke (AIS). Individual WMH volume was automatically derived from fluid-attenuated inversion recovery images. Stroke lesions were automatically segmented from diffusion-weighted imaging (DWI) images, parcellated into atlas-defined brain regions and further condensed to 10 lesion patterns via machine learning-based dimensionality reduction. Stroke lesion effects on AIS severity and unfavorable outcomes (mRS score >2) were modeled within purpose-built Bayesian linear and logistic regression frameworks. Interaction effects between stroke lesions and a high vs low WMH burden were integrated via hierarchical model structures. Models were adjusted for age, age2, sex, total DWI lesion and WMH volumes, and comorbidities. Data were split into derivation and validation cohorts. RESULTS: A total of 928 patients with AIS contributed to acute stroke severity analyses (age: 64.8 [14.5] years, 40% women) and 698 patients to long-term functional outcome analyses (age: 65.9 [14.7] years, 41% women). Stroke severity was mainly explained by lesions focused on bilateral subcortical and left hemispherically pronounced cortical regions across patients with both a high and low WMH burden. Lesions centered on left-hemispheric insular, opercular, and inferior frontal regions and lesions affecting right-hemispheric temporoparietal regions had more pronounced effects on stroke severity in case of high compared with low WMH burden. Unfavorable outcomes were predominantly explained by lesions in bilateral subcortical regions. In difference to the lesion location-specific WMH effects on stroke severity, higher WMH burden increased the odds of unfavorable outcomes independent of lesion location. DISCUSSION: Higher WMH burden may be associated with an increased stroke severity in case of stroke lesions involving left-hemispheric insular, opercular, and inferior frontal regions (potentially linked to language functions) and right-hemispheric temporoparietal regions (potentially linked to attention). Our findings suggest that patients with specific constellations of WMH burden and lesion locations may have greater benefits from acute recanalization treatments. Future clinical studies are warranted to systematically assess this assumption and guide more tailored treatment decisions.