BrainMass: Advancing Brain Network Analysis for Diagnosis with Large-scale Self-Supervised Learning.

Foundation models pretrained on large-scale datasets via self-supervised learning demonstrate exceptional versatility across various tasks. Due to the heterogeneity and hard-to-collect medical data, this approach is especially beneficial for medical image analysis and neuroscience research, as it streamlines broad downstream tasks without the need for numerous costly annotations. However, there has been limited investigation into brain network foundation models, limiting their adaptability and generalizability for broad neuroscience studies. In this study, we aim to bridge this gap. In particular, (1) we curated a comprehensive dataset by collating images from 30 datasets, which comprises 70,781 samples of 46,686 participants. Moreover, we introduce pseudo-functional connectivity (pFC) to further generates millions of augmented brain networks by randomly dropping certain timepoints of the BOLD signal. (2) We propose the BrainMass framework for brain network self-supervised learning via mask modeling and feature alignment. BrainMass employs Mask-ROI Modeling (MRM) to bolster intra-network dependencies and regional specificity. Furthermore, Latent Representation Alignment (LRA) module is utilized to regularize augmented brain networks of the same participant with similar topological properties to yield similar latent representations by aligning their latent embeddings. Extensive experiments on eight internal tasks and seven external brain disorder diagnosis tasks show BrainMass's superior performance, highlighting its significant generalizability and adaptability. Nonetheless, BrainMass demonstrates powerful few/zero-shot learning abilities and exhibits meaningful interpretation to various diseases, showcasing its potential use for clinical applications.

Mapping Multi-Modal Brain Connectome for Brain Disorder Diagnosis via Cross-Modal Mutual Learning.

Recently, the study of multi-modal brain connectome has recorded a tremendous increase and facilitated the diagnosis of brain disorders. In this paradigm, functional and structural networks, e.g., functional and structural connectivity derived from fMRI and DTI, are in some manner interacted but are not necessarily linearly related. Accordingly, there remains a great challenge to leverage complementary information for brain connectome analysis. Recently, Graph Convolutional Networks (GNN) have been widely applied to the fusion of multi-modal brain connectome. However, most existing GNN methods fail to couple inter-modal relationships. In this regard, we propose a Cross-modal Graph Neural Network (Cross-GNN) that captures inter-modal dependencies through dynamic graph learning and mutual learning. Specifically, the inter-modal representations are attentively coupled into a compositional space for reasoning inter-modal dependencies. Additionally, we investigate mutual learning in explicit and implicit ways: (1) Cross-modal representations are obtained by cross-embedding explicitly based on the inter-modal correspondence matrix. (2) We propose a cross-modal distillation method to implicitly regularize latent representations with cross-modal semantic contexts. We carry out statistical analysis on the attentively learned correspondence matrices to evaluate inter-modal relationships for associating disease biomarkers. Our extensive experiments on three datasets demonstrate the superiority of our proposed method for disease diagnosis with promising prediction performance and multi-modal connectome biomarker location.

MR-based spatiotemporal anisotropic atrophy evaluation of hippocampus in Alzheimer's disease progression by multiscale skeletal representation.

Increasing evidence has shown a higher sensitivity of Alzheimer's disease (AD) progression by local hippocampal atrophy rather than the whole volume. However, existing morphological methods based on subfield-volume or surface in imaging studies are not capable to describe the comprehensive process of hippocampal atrophy as sensitive as histological findings. To map histological distinctive measurements onto medical magnetic resonance (MR) images, we propose a multiscale skeletal representation (m-s-rep) to quantify focal hippocampal atrophy during AD progression in longitudinal cohorts from the Alzheimer's Disease Neuroimaging Initiative (ADNI). The m-s-rep captures large-to-small-scale hippocampal morphology by spoke interpolation over label projection on skeletal models. To enhance morphological correspondence within subjects, we align the longitudinal m-s-reps by surface-based transformations from baseline to subsequent timepoints. Cross-sectional and longitudinal measurements derived from m-s-rep are statistically analyzed to comprehensively evaluate the bilateral hippocampal atrophy. Our findings reveal that during the early AD progression, atrophy primarily affects the lateral-medial extent of the hippocampus, with a difference of 1.8 mm in lateral-medial width in 2 years preceding conversion (p < .001), and the medial head exhibits a maximum difference of 3.05%/year in local atrophy rate (p = .011) compared to controls. Moreover, progressive mild cognitive impairment (pMCI) exhibits more severe and widespread atrophy in the head and body compared to stable mild cognitive impairment (sMCI), with a maximum difference of 1.21 mm in thickness in the medial head 1 year preceding conversion (p = .012). In summary, our proposed method can quantitatively measure the hippocampal morphological changes on 3T MR images, potentially assisting the pre-diagnosis and prognosis of AD.

CReg-KD: Model refinement via confidence regularized knowledge distillation for brain imaging.

One of the core challenges of deep learning in medical image analysis is data insufficiency, especially for 3D brain imaging, which may lead to model over-fitting and poor generalization. Regularization strategies such as knowledge distillation are powerful tools to mitigate the issue by penalizing predictive distributions and introducing additional knowledge to reinforce the training process. In this paper, we revisit knowledge distillation as a regularization paradigm by penalizing attentive output distributions and intermediate representations. In particular, we propose a Confidence Regularized Knowledge Distillation (CReg-KD) framework, which adaptively transfers knowledge for distillation in light of knowledge confidence. Two strategies are advocated to regularize the global and local dependencies between teacher and student knowledge. In detail, a gated distillation mechanism is proposed to soften the transferred knowledge globally by utilizing the teacher loss as a confidence score. Moreover, the intermediate representations are attentively and locally refined with key semantic context to mimic meaningful features. To demonstrate the superiority of our proposed framework, we evaluated the framework on two brain imaging analysis tasks (i.e. Alzheimer's Disease classification and brain age estimation based on T1-weighted MRI) on the Alzheimer's Disease Neuroimaging Initiative dataset including 902 subjects and a cohort of 3655 subjects from 4 public datasets. Extensive experimental results show that CReg-KD achieves consistent improvements over the baseline teacher model and outperforms other state-of-the-art knowledge distillation approaches, manifesting that CReg-KD as a powerful medical image analysis tool in terms of both promising prediction performance and generalizability.

A deep connectome learning network using graph convolution for connectome-disease association study.

Multivariate analysis approaches provide insights into the identification of phenotype associations in brain connectome data. In recent years, deep learning methods including convolutional neural network (CNN) and graph neural network (GNN), have shifted the development of connectome-wide association studies (CWAS) and made breakthroughs for connectome representation learning by leveraging deep embedded features. However, most existing studies remain limited by potentially ignoring the exploration of region-specific features, which play a key role in distinguishing brain disorders with high intra-class variations, such as autism spectrum disorder (ASD), and attention deficit hyperactivity disorder (ADHD). Here, we propose a multivariate distance-based connectome network (MDCN) that addresses the local specificity problem by efficient parcellation-wise learning, as well as associating population and parcellation dependencies to map individual differences. The approach incorporating an explainable method, parcellation-wise gradient and class activation map (p-GradCAM), is feasible for identifying individual patterns of interest and pinpointing connectome associations with diseases. We demonstrate the utility of our method on two largely aggregated multicenter public datasets by distinguishing ASD and ADHD from healthy controls and assessing their associations with underlying diseases. Extensive experiments have demonstrated the superiority of MDCN in classification and interpretation, where MDCN outperformed competitive state-of-the-art methods and achieved a high proportion of overlap with previous findings. As a CWAS-guided deep learning method, our proposed MDCN framework may narrow the bridge between deep learning and CWAS approaches, and provide new insights for connectome-wide association studies.

Spatial-temporal patterns of brain disconnectome in Alzheimer's disease.

Mounting evidences have shown that progression of white matter hyperintensities (WMHs) with vascular origin might cause cognitive dysfunction symptoms through their effects on brain networks. However, the vulnerability of specific neural connection related to WMHs in Alzheimer's disease (AD) still remains unclear. In this study, we established an atlas-guided computational framework based on brain disconnectome to assess the spatial-temporal patterns of WMH-related structural disconnectivity within a longitudinal investigation. Alzheimer's Disease Neuroimaging Initiative (ADNI) database was adopted with 91, 90 and 44 subjects including in cognitive normal aging, stable and progressive mild cognitive impairment (MCI), respectively. The parcel-wise disconnectome was computed by indirect mapping of individual WMHs onto population-averaged tractography atlas. By performing chi-square test, we discovered a spatial-temporal pattern of brain disconnectome along AD evolution. When applied such pattern as predictor, our models achieved highest mean accuracy of 0.82, mean sensitivity of 0.86, mean specificity of 0.82 and mean area under the receiver operating characteristic curve (AUC) of 0.91 for predicting conversion from MCI to dementia, which outperformed methods utilizing lesion volume as predictors. Our analysis suggests that brain WMH-related structural disconnectome contributes to AD evolution mainly through attacking connections between: (1) parahippocampal gyrus and superior frontal gyrus, orbital gyrus, and lateral occipital cortex; and (2) hippocampus and cingulate gyrus, which are also vulnerable to Aß and tau confirmed by other researches. All the results further indicate that a synergistic relationship exists between multiple contributors of AD as they attack similar brain connectivity at the prodromal stage of disease.

Three-Dimensional Digital Reconstruction of the Cerebellar Cortex: Lobule Thickness, Surface Area Measurements, and Layer Architecture.

The cerebellum is ontogenetically one of the first structures to develop in the central nervous system; nevertheless, it has been only recently reconsidered for its significant neurobiological, functional, and clinical relevance in humans. Thus, it has been a relatively under-studied compared to the cerebrum. Currently, non-invasive imaging modalities can barely reach the necessary resolution to unfold its entire, convoluted surface, while only histological analyses can reveal local information at the micrometer scale. Herein, we used the BigBrain dataset to generate area and point-wise thickness measurements for all layers of the cerebellar cortex and for each lobule in particular. We found that the overall surface area of the cerebellar granular layer (including Purkinje cells) was 1,732 cm2 and the molecular layer was 1,945 cm2. The average thickness of the granular layer is 0.88 mm (± 0.83) and that of the molecular layer is 0.32 mm (± 0.08). The cerebellum (both granular and molecular layers) is thicker at the depth of the sulci and thinner at the crowns of the gyri. Globally, the granular layer is thicker in the lateral-posterior-inferior region than the medial-superior regions. The characterization of individual layers in the cerebellum achieved herein represents a stepping-stone for investigations interrelating structural and functional connectivity with cerebellar architectonics using neuroimaging, which is a matter of considerable relevance in basic and clinical neuroscience. Furthermore, these data provide templates for the construction of cerebellar topographic maps and the precise localization of structural and functional alterations in diseases affecting the cerebellum.

Structural network alterations in patients with nasopharyngeal carcinoma after radiotherapy: A 1-year longitudinal study.

This longitudinal study explored the changed patterns of structural brain network after radiotherapy (RT) in patients with nasopharyngeal carcinoma (NPC). Diffusion tensor imaging (DTI) data were gathered from 35 patients with NPC at four time points: before RT (baseline), 0∼3 (acute), 6 (early delayed), and 12 months (late-delayed) after RT. The graph theory was used to characterize the dynamic topological properties after RT and the significant changes were detected over time at the global, regional and modular levels. Significantly altered regional metrics (nodal efficiency and degree centrality) were distributed in the prefrontal, temporal, parietal, frontal, and subcortical regions. The module, that exhibited a significantly altered within-module connectivity, had a high overlap with the default mode network (DMN). In addition, the global, regional and modular metrics showed a tendency of progressive decrease at the acute and early delayed stages, and a partial/full recovery at the late-delayed stage. This changed pattern illustrated that the radiation-induced brain damage began at the acute reaction stage and were aggravated at the early-delayed stage, and then partially recovered at the late-delayed stage. Furthermore, the spearman's correlations between the abnormal nodal metrics and temporal dose were calculated and high correlations were found at the temporal (MTG.R and HES.L), subcortical (INS.R), prefrontal (ORBinf.L and ACG.L), and parietal (IPL.R) indicating that these regions were more sensitive to dose and should be mainly considered in radiotherapy treatment plan.

Ensemble learning via supervision augmentation for white matter hyperintensity segmentation.

Since the ambiguous boundary of the lesion and inter-observer variability, white matter hyperintensity segmentation annotations are inherently noisy and uncertain. On the other hand, the high capacity of deep neural networks (DNN) enables them to overfit labels with noise and uncertainty, which may lead to biased models with weak generalization ability. This challenge has been addressed by leveraging multiple annotations per image. However, multiple annotations are often not available in a real-world scenario. To mitigate the issue, this paper proposes a supervision augmentation method (SA) and combines it with ensemble learning (SA-EN) to improve the generalization ability of the model. SA can obtain diverse supervision information by estimating the uncertainty of annotation in a real-world scenario that per image have only one ambiguous annotation. Then different base learners in EN are trained with diverse supervision information. The experimental results on two white matter hyperintensity segmentation datasets demonstrate that SA-EN gets the optimal accuracy compared with other state-of-the-art ensemble methods. SA-EN is more effective on small datasets, which is more suitable for medical image segmentation with few annotations. A quantitative study is presented to show the effect of ensemble size and the effectiveness of the ensemble model. Furthermore, SA-EN can capture two types of uncertainty, aleatoric uncertainty modeled in SA and epistemic uncertainty modeled in EN.

Brain age vector: A measure of brain aging with enhanced neurodegenerative disorder specificity.

Neuroimaging-driven brain age estimation has become popular in measuring brain aging and identifying neurodegenerations. However, the single estimated brain age (gap) compromises regional variations of brain aging, losing spatial specificity across diseases which is valuable for early screening. In this study, we combined brain age modeling with Shapley Additive Explanations to measure brain aging as a feature contribution vector underlying spatial pathological aging mechanism. Specifically, we regressed age with volumetric brain features using machine learning to construct the brain age model, and model-agnostic Shapley values were calculated to attribute regional brain aging for each subject's age estimation, forming the brain age vector. Spatial specificity of the brain age vector was evaluated among groups of normal aging, prodromal Parkinson disease (PD), stable mild cognitive impairment (sMCI), and progressive mild cognitive impairment (pMCI). Machine learning methods were adopted to examine the discriminability of the brain age vector in early disease screening, compared with the other two brain aging metrics (single brain age gap, regional brain age gaps) and brain volumes. Results showed that the proposed brain age vector accurately reflected disorder-specific abnormal aging patterns related to the medial temporal and the striatum for prodromal AD (sMCI vs. pMCI) and PD (healthy controls [HC] vs. prodromal PD), respectively, and demonstrated outstanding performance in early disease screening, with area under the curves of 83.39% and 72.28% in detecting pMCI and prodromal PD, respectively. In conclusion, the proposed brain age vector effectively improves spatial specificity of brain aging measurement and enables individual screening of neurodegenerative diseases.

MRF-Net: A multi-branch residual fusion network for fast and accurate whole-brain MRI segmentation.

Whole-brain segmentation from T1-weighted magnetic resonance imaging (MRI) is an essential prerequisite for brain structural analysis, e.g., locating morphometric changes for brain aging analysis. Traditional neuroimaging analysis pipelines are implemented based on registration methods, which involve time-consuming optimization steps. Recent related deep learning methods speed up the segmentation pipeline but are limited to distinguishing fuzzy boundaries, especially encountering the multi-grained whole-brain segmentation task, where there exists high variability in size and shape among various anatomical regions. In this article, we propose a deep learning-based network, termed Multi-branch Residual Fusion Network, for the whole brain segmentation, which is capable of segmenting the whole brain into 136 parcels in seconds, outperforming the existing state-of-the-art networks. To tackle the multi-grained regions, the multi-branch cross-attention module (MCAM) is proposed to relate and aggregate the dependencies among multi-grained contextual information. Moreover, we propose a residual error fusion module (REFM) to improve the network's representations fuzzy boundaries. Evaluations of two datasets demonstrate the reliability and generalization ability of our method for the whole brain segmentation, indicating that our method represents a rapid and efficient segmentation tool for neuroimage analysis.

Clinical and Preclinical Neuroimaging Changes in Spinocerebellar Ataxia Type 12: A Study of Three Chinese Pedigrees.

BACKGROUND: Spinocerebellar ataxia type 12 (SCA12) is a rare SCA subtype with unclear clinical and imaging features. Also, the radiological changes in prodromal and early stages remain unknown. METHODS: Ten symptomatic and two pre-symptomatic cases from three Chinese pedigrees received clinical assessments and imaging studies including routine magnetic resonance imaging (MRI), diffusion kurtosis imaging (DKI), and positron emission tomography (PET) using 18F-flurodeoxyglucose (FDG) to investigate glucose metabolism in brain and 18F-vesicle monoamine transporter 2 (VMAT2) to inspect the integrity of the dopaminergic neuron. Seventy-two healthy individuals were recruited as controls in the quantitative FDG-PET analysis. Imaging parameters were compared between symptomatic and presymptomatic cases with different disease durations. RESULTS: Patients displayed prominent action tremor, moderate ataxia, and subtle parkinsonism with poor levodopa-response. MRI showed extensive but heterogeneous cerebral atrophy, which was most evident in the frontoparietal lobes. Cerebellar atrophy was apparent in later stages. DKI detected impaired fibers in the cerebellar peduncles. In both symptomatic and pre-symptomatic cases, PET-CT showed an earlier FDG decline than atrophic changes in multiple regions, and the frontoparietal lobes were the earliest and most severe. However, the VMAT2 density were normal in the putamen and caudate nucleus of most cases (7/8). CONCLUSIONS: We first found that hypometabolism in the cerebral cortex, but not cerebellum, is an early and prominent change in SCA12. The integrity of presynaptic dopaminergic neurons remains largely spared during the whole disease process.

A Multi-Atlas-Based [18F]9-Fluoropropyl-(+)-Dihydrotetrabenazine Positron Emission Tomography Image Segmentation Method for Parkinson's Disease Quantification.

Objectives: [18F]9-fluoropropyl-(+)-dihydrotetrabenazine ([18F]-FP-DTBZ) positron emission tomography (PET) provides reliable information for the diagnosis of Parkinson's disease (PD). In this study, we proposed a multi-atlas-based [18F]-FP-DTBZ PET image segmentation method for PD quantification assessment. Methods: A total of 99 subjects from Xuanwu Hospital of Capital Medical University were included in this study, and both brain PET and magnetic resonance (MR) scans were conducted. Data from 20 subjects were used to generate atlases, based on which a multi-atlas-based [18F]-FP-DTBZ PET segmentation method was developed especially for striatum and its subregions. The proposed method was compared with the template-based method through striatal subregion parcellation performance and the standard uptake value ratio (SUVR) quantification accuracy. Discriminant analysis between healthy controls (HCs) and PD patients was further performed. Results: Segmentation results of the multi-atlas-based method showed better consistency than the template-based method with the ground truth, yielding a dice coefficient of 0.81 over 0.73 on the full striatum. The SUVRs calculated by the multi-atlas-based method had an average interclass correlation coefficient (ICC) of 0.953 with the standardized result, whereas the template-based method only reached 0.815. The SUVRs of HCs were generally higher than that of patients with PD and showed significant differences in all of the striatal subregions (all p < 0.001). The median and posterior putamen performed best in discriminating patients with PD from HCs. Conclusion: The proposed multi-atlas-based [18F]-FP-DTBZ PET image segmentation method achieved better performance than the template-based method, indicating great potential in improving accuracy and efficiency for PD diagnosis in clinical routine.

Coupling of brain activity and structural network in multiple sclerosis: A graph frequency analysis study.

The brain activities and the underlying wiring diagrams are vulnerable in multiple sclerosis (MS). Also, it remains unknown whether the complex coupling between these functional and structural brain properties would be affected. To address this issue, we adopted graph frequency analysis to quantify the high-order structural-functional interactions based on a combination of brain diffusion and functional MRI data. The structural-functional decoupling index was proposed to measure how much brain regional functional activity with different graph frequency was organized atop the underlying wiring diagram in MS. The identified patterns in MS included (1) disruption of inherent structural-functional coupling in the somatomotor network (ß = 0.05, p = 0.03), and (2) excessive decrease of decoupling in the subcortical (ß = -0.10, p = 0.02), visual (ß = -0.04, p = 0.01), and dorsal attention networks (ß = -0.12, p = 0.03). Besides, this structural-functional coupling signature in the somatomotor network was associated with cognitive worsening of MS patients (ß = -24.31, p = 0.006). Overall, our study unveiled a unique signature of brain structural-functional reorganization in MS.

Alteration of brain structural connectivity in progression of Parkinson's disease: A connectome-wide network analysis.

Pinpointing the brain dysconnectivity in idiopathic rapid eye movement sleep behaviour disorder (iRBD) can facilitate preventing the conversion of Parkinson's disease (PD) from prodromal phase. Recent neuroimage investigations reported disruptive brain white matter connectivity in both iRBD and PD, respectively. However, the intrinsic process of the human brain structural network evolving from iRBD to PD still remains largely unknown. To address this issue, 151 participants including iRBD, PD and age-matched normal controls were recruited to receive diffusion MRI scans and neuropsychological examinations. The connectome-wide association analysis was performed to detect reorganization of brain structural network along with PD progression. Eight brain seed regions in both cortical and subcortical areas demonstrated significant structural pattern changes along with the progression of PD. Applying machine learning on the key connectivity related to these seed regions demonstrated better classification accuracy compared to conventional network-based statistic. Our study shows that connectome-wide association analysis reveals the underlying structural connectivity patterns related to the progression of PD, and provide a promising distinct capability to predict prodromal PD patients.

An anatomical knowledge-based MRI deep learning pipeline for white matter hyperintensity quantification associated with cognitive impairment.

Recent studies have confirmed that white matter hyperintensities (WMHs) accumulated in strategic brain regions can predict cognitive impairments associated with Alzheimer's disease (AD). The knowledge of white matter anatomy facilitates lesion-symptom mapping associated with cognition, and provides important spatial information for lesion segmentation algorithms. However, deep learning-based methods in the white matter hyperintensity (WMH) segmentation realm do not take full advantage of anatomical knowledge in decision-making and lesion localization processes. In this paper, we proposed an anatomical knowledge-based MRI deep learning pipeline (AU-Net), handcrafted anatomical-based spatial features developed from brain atlas were integrated with a well-designed U-Net configuration to simultaneously segment and locate WMHs. Manually annotated data from WMH segmentation challenge were used for the evaluation. We then applied this pipeline to investigate the association between WMH burden and cognition within another publicly available database. The results showed that AU-Net significantly improved segmentation performance compared with methods that did not incorporate anatomical knowledge (p < 0.05), and achieved a 14-17% increase based on area under the curve (AUC) in the cohort with mild WMH burden. Different configurations for incorporating anatomical knowledge were evaluated, the proposed stage-wise AU-Net-two-step method achieved the best performance (Dice: 0.86, modified Hausdorff distance: 3.06 mm), which was comparable with the state-of-the-art method (Dice: 0.87, modified Hausdorff distance: 3.62 mm). We observed different WMH accumulation patterns associated with normal aging and cognitive impairments. Furthermore, the characteristics of individual WMH lesions located in strategic regions (frontal and parietal subcortical white matter, as well as corpus callosum) were significantly correlated with cognition after adjusting for total lesion volumes. Our findings suggest that AU-Net is a reliable method to segment and quantify brain WMHs in elderly cohorts, and is valuable in individual cognition evaluation.

Multimodal MRI assessment for first episode psychosis: A major change in the thalamus and an efficient stratification of a subgroup.

Multi-institutional brain imaging studies have emerged to resolve conflicting results among individual studies. However, adjusting multiple variables at the technical and cohort levels is challenging. Therefore, it is important to explore approaches that provide meaningful results from relatively small samples at institutional levels. We studied 87 first episode psychosis (FEP) patients and 62 healthy subjects by combining supervised integrated factor analysis (SIFA) with a novel pipeline for automated structure-based analysis, an efficient and comprehensive method for dimensional data reduction that our group recently established. We integrated multiple MRI features (volume, DTI indices, resting state fMRI-rsfMRI) in the whole brain of each participant in an unbiased manner. The automated structure-based analysis showed widespread DTI abnormalities in FEP and rs-fMRI differences between FEP and healthy subjects mostly centered in thalamus. The combination of multiple modalities with SIFA was more efficient than the use of single modalities to stratify a subgroup of FEP (individuals with schizophrenia or schizoaffective disorder) that had more robust deficits from the overall FEP group. The information from multiple MRI modalities and analytical methods highlighted the thalamus as significantly abnormal in FEP. This study serves as a proof-of-concept for the potential of this methodology to reveal disease underpins and to stratify populations into more homogeneous sub-groups.