Biology subjects

Chattopadhyay, T.

Publications and source records attributed to Chattopadhyay, T..

At least 19 recordsLinked to original sources

bioRxiv2026

Personalized Prediction of Differences in fMRI Connectivity Estimates Across Preprocessing Pipelines

Longitudinal functional MRI of the human brain is valuable for tracking changes in functional connectivity (FC) associated with disease progression or treatment response. Even so, measured FC values depend on preprocessing, and nuisance regression and temporal filtering affect individuals differently. We examined whether we could use earlier scans to predict how preprocessing would affect a later scan. In 69 ADNI (Alzheimers Disease Neuroimaging Initiative) participants scanned three times, we measured how nuisance regression and filtering, alone and in sequence, changed the measured FC. A ridge regression model combined these responses with baseline FC from the third scan to predict the effect of reversing the processing order. Just four coefficients were shared across 79,800 connections. In leave-one-participant-out testing, mean squared error was 0.012, compared with 0.017 for predicting the same average response for everyone. It outperformed the best of seven alternative predictions chosen for each participant (0.013; normalized gain 0.069, 95% bootstrap interval 0.006-0.134), with lower error in 44 of 69 participants. About half the benefit of using each persons earlier responses, rather than the average across participants, came from which connections changed. This advantage persisted after accounting for acquisition variables. The ADNI-trained model reduced mean error in ten Midnight Scan Club participants without refitting. These results support using personal preprocessing history to predict changes in connectivity estimates on newly acquired scans.

neuroscience↗

bioRxiv2026

Temporal Structure in Resting-State fMRI under Marginal and Cross-Spectral Surrogate Controls

Resting-state functional MRI is commonly analyzed using correlations between regional signals and their frequency spectra, which characterize average connectivity and spectral properties but may miss other aspects of temporal organization. Measures of temporal complexity offer complementary information about the recurrence of signal patterns, how fluctuations vary across timescales, and signal roughness. These properties may reveal disease-related alterations in brain activity, offering opportunities for biomarker discovery and insights into mechanisms of neurological and psychiatric disorders. Here we examined sample entropy (SampEn), multifractal de-trended fluctuation analysis (MFDFA), and Higuchi curve shapes in 732 scans from 270 ADNI participants, including cognitively normal individuals, people with mild cognitive impairment, and people with dementia. Using the seven Yeo resting-state networks (Yeo-7), we compared recordings with randomized signals preserving either each networks amplitude distribution and approximate spectrum (IAAFT), or the complete cross-spectrum across networks. If these properties fully explained temporal complexity, real recordings should resemble their randomized references. We therefore counted how often each real recording was the strongest outlier among itself and 20 references. This occurred in 21.7%, 12.7%, and 44.2% of 5,124 network-visits for SampEn, MFDFA, and Higuchi, versus 5.5%, 5.1%, and 6.3% in surrogate-only controls. Cross-spectrum-preserving references yielded 39.8%, 17.5%, and 49.8%. Simulations showed that changes in sampling or signal processing, including abrupt changes in signal amplitude, can produce similar differences without changing the underlying dynamics. The measures showed repeatability across visits but did not improve retrospective prediction of changes in Clinical Dementia Rating Sum of Boxes (CDR-SB) scores. These findings identify temporal features not reproduced by the tested spectral controls and provide a basis for evaluating their potential as disease biomarkers.

neuroscience↗

bioRxiv2026

Evaluating State-Transition Models in Short Resting-State fMRI with Continuous-Dynamics Controls

Hidden Markov models (HMMs) are widely used to describe changes in resting-state functional MRI as transitions between a finite set of inferred brain states. However, continuously varying brain activity can produce similar apparent transitions, raising questions about their biological validity. We tested whether discrete and continuous brain dynamics could be distinguished by predicting subsequent fMRI measurements, and if temporal filtering affects this distinction. We compared hidden Markov models (HMMs) with vector autoregressive (VAR) models, which predict signals from preceding measurements without assuming discrete states. Each time point contained signals from the seven Yeo functional networks. Both models used the same preceding observations to predict the next, with accuracy measured by the log probability density assigned to the observed signals. Across 864 simulated conditions, HMMs performed better in 139/144 unfiltered discrete-state cases, whereas VAR performed better in 67/72 continuous cases. Changing only the filter during testing increased the number of continuous cases favoring HMMs from 2/72 with fixed filtering to 68/72, despite unchanged underlying dynamics. We also analyzed resting-state fMRI from 270 ADNI participants (90 cognitively normal, 90 with mild cognitive impairment, and 90 with dementia). VAR outperformed HMM in 160/184 datasets with at least 180 time points and 60/86 shorter recordings. In simulations, HMMs inferred directional state transitions even from smoothly evolving activity without discrete switching. Thus, apparent fMRI state transitions do not necessarily indicate discrete underlying dynamics. Their interpretation requires comparison with continuous models and attention to temporal filtering.

neuroscience↗

bioRxiv2026

Dementia Severity Is Associated with Resting-State fMRI Network Responses to Nuisance Regression

Apparent disease effects in resting-state fMRI may be confounded by the effects of preprocessing, making it essential to understand whether nuisance removal affects patients differently across levels of clinical severity. Yet few studies directly examine whether sensitivity to preprocessing is itself clinically informative. We investigated this in 201 ADNI participants spanning normal cognition, mild cognitive impairment, and dementia. We quantified the coordinated temporal structure of the seven Yeo resting-state networks using a delay-based concentration statistic, comparing estimates after regression of 24 motion parameters and linear trends with estimates obtained after additionally removing five anatomical CompCor components. Adding these components generally reduced the measured temporal concentration, but this reduction was smaller in participants with greater dementia severity (+0.472 percentage points per CDR-SB point). In contrast, the same network measures showed little association with severity before adding CompCor. To determine whether this finding depended on how temporal trends were removed, we repeated the analysis with identical discrete cosine regressors added to both preprocessing designs. The clinical association remained similar (+0.464). Controlled simulations showed that temporal concentration increased with the strength of delayed coupling between networks, even when their individual statistical properties and instantaneous covariance remained unchanged. However, changes in observation amplitude and contamination of nuisance regressors also altered the response to preprocessing without changing the underlying neural signal. These findings suggest that disease-related differences in fMRI may partly reflect differential sensitivity to preprocessing. Systematic sensitivity analyses, including multiverse comparisons across preprocessing strategies, may therefore be essential to distinguish genuine disease effects from preprocessing-dependent associations.

neuroscience↗

bioRxiv2026

Temporal Variability in Brain Connectivity for Estimating Two-Year Clinical Severity in Mild Cognitive Impairment

Resting-state functional MRI measures coordinated activity across brain networks and may reveal early changes relevant to cognitive decline. Conventional static functional connectivity analyses often compute time-series correlations across an entire scan, missing informative fluctuations over shorter timescales. We investigated whether these fluctuations, in fMRI scans collected at baseline, improve estimation of clinical severity two years later in mild cognitive impairment (MCI). In 48 ADNI participants with MCI, we summarized temporal variability using the standard deviation of windowed correlations between 21 pairs of brain networks. Using repeated nested cross-validation, the mean absolute error in predicting follow-up Clinical Dementia Rating Sum of Boxes (CDR-SB) was 1.62 points with clinical covariates alone, 1.66 after adding static connectivity, and 1.51 after adding connectivity variability. With additional motion covariates, the corresponding errors were 1.70, 1.73, and 1.59. Connectivity variability outperformed both conventional models in all three cross-validation repetitions, while randomized signals preserving the complete cross-spectrum did not reproduce the improvement. Perhaps surprisingly though, most of the gain occurred in participants without subsequent clinical worsening, and results were sensitive to the choice of temporal windows. These preliminary findings suggest that compact measures of connectivity variability may capture clinically relevant information beyond conventional measures of static functional connectivity, motivating larger studies to evaluate their value for predicting future cognitive impairment.

neuroscience↗

bioRxiv2026

Understanding Performance Plateaus and Representation Collapse in 3D MRI Transformers Trained to Classify Brain Disease

In medical imaging AI, we commonly assume that diagnostic classifiers will increase in accuracy if they are trained on more data, but in reality, their performance often plateaus. To understand this behavior, we ran a spectral analysis of 3D vision transformers, a common supervised deep learning model that can be trained to classify disease based on 3D brain MRI. For simpler classification methods such as linear discriminant analysis, the classifier performance AUC(N), when trained on a sample of size N, is known to depend on the encoder's eigenspectrum. Each class-discriminative direction in the latent space contributes only after the sample size N reaches a threshold, so the spectrum influences how performance scales with training data. Building on this theory, we show that supervised transformers compress almost all of their discriminative signal into one principal component (EVR@1 = 0.90--0.96): after this representation collapse happens, no further directions cross the threshold as training data increases, and accuracy plateaus. In principle, this plateau in performance could be due to either the encoder's learned feature space or the difficulty of the task. To distinguish these cases, we trained three types of supervised 3D transformers (ViT3D, MINiT, NIT) and, as a non-collapsing alternative, a frozen self-supervised encoder (DINO), to classify people with dementia or mild cognitive impairment (MCI) versus cognitively normal (CN) individuals from 3D brain MRI, using data from the Alzheimer's Disease Neuroimaging Initiative (ADNI). We froze each encoder and applied it to a new task, so that only the task difficulty varied. An easy task encoder (Dementia vs CN), transferred to the harder task (MCI vs CN), did not reproduce the plateau (MINiT threshold 9.2), but full retraining did (2758.5). The plateau is therefore due to the learned encoder, rather than task difficulty. Ranking the feature directions by the proportion of the disease signal they contain, we were able to identify a sign of representation collapse, i.e., a single dominant spike in the latent feature eigenspectrum. This spike was 11--26 times larger than predicted from the rest of the spectrum for the disease-trained (supervised) encoders, but was absent for DINO (0.66). To predict how transformer accuracy scales with the number of training scans, a large spike in the encoder spectrum suggests that the representation has collapsed and additional scans will not help; a small spike means that the signal is spread across many encoder latent space directions, and more training data should boost the classifier performance.

neuroscience↗

bioRxiv2026

Distributed Genetic Effects on Human Brain Structure Emerge Across Multiple Spatial Scales

Genome-wide association studies (GWAS) have identified hundreds of common genetic variants associated with regional brain volumes, enabling the construction of polygenic scores (PGS) that summarize genetic predisposition for variation in specific neuroanatomical traits. To investigate how these genetic influences are exerted spatially throughout the brain, we computed PGS for ten brain volume phenotypes, including nine major subcortical structures and intracranial volume. Each locus was weighted by its estimated GWAS effect size on regional volume in the original GWAS. In an independent, non-overlapping sample of 2,830 UK Biobank participants, we performed whole-brain voxel-based morphometry (VBM) analyses of 3D volumetric brain MRI to reveal voxel-wise associations between each PGS and modulated gray matter volume (GMV). To probe genetic effects across multiple spatial scales, analyses were repeated across Gaussian smoothing kernels ranging from 2-mm to 12-mm full-width at half-maximum (FWHM). Several PGS demonstrated highly significant associations with GMV, including localized effects in the hippocampus, amygdala, thalamus, and basal ganglia, whereas the brainstem PGS showed more widespread associations throughout the brain. For most of the PGS, the fraction of voxels surviving the false discovery rate (FDR) correction increased with increasing FWHM. Peak voxel-wise significance was often strongest at intermediate smoothing levels. Hippocampal significance maps showed progressively larger regions of significant signal at higher smoothing levels, and subsampling showed that detectable signal remained present even with substantial reductions in sample size. These findings suggest that genetic influences on brain morphology are expressed across multiple spatial scales, with consequences that may help to guide the design of deep learning methods to discover genomic loci associated with brain structure and brain diseases.

neuroscience↗

bioRxiv2026

When does more data help? Spectral Geometry and Scaling Laws in MRI Transformers

Scaling laws describe how model performance improves as the amount of training data increases, and recent theories such as the zeta law suggest that scaling behavior is influenced by the eigenspectrum of the models latent representation. Here, we evaluated whether the distribution of discriminative signals across spectral modes predicts the future scaling behavior, for MRI transformers trained for disease classification. We trained three supervised 3D vision transformers (ViT3D, MINiT, and NIT) for Alzheimers disease classification using 2,822 training scans from the Alzheimers Disease Neuroimaging Initiative (ADNI); we compared their encoder spectra with that of a frozen self-supervised DINO ViT-B/16 encoder adapted to 3D MRI. The supervised models learned highly concentrated representations, with 90-96% of CLS-token variance captured by a single principal component, whereas DINO distributed signal across many latent directions. Via spectral expansion of the Mahalanobis signal, we found that supervised training concentrated disease information into a single dominant mode, while self-supervised training produced a richer spectral geometry with higher effective rank and discoverability. This led to different scaling behavior: supervised models exhibited flatter AUC(N) curves, yet DINO continued to improve as sample size increased, gaining 11.0 percentage points from N=50 to N=2,822. Overall, the spectral distribution of the discriminative signal, for these different encoder types, influenced how much performance remained discoverable as sample size increased. Distributed representations may retain signal across many latent modes and continue to improve with additional data, whereas concentrated representations tend to exhaust most of the discoverable signal at much lower sample sizes.

neuroscience↗

bioRxiv2026

Mapping Focal and Generalized Effects of Common Genetic Variants on Human Brain Structure

Genome-wide association studies (GWAS) have advanced the quest to understand how specific genetic variants influence human brain structure and function. Recent work has identified hundreds of common variants associated with subcortical brain volumes, sparking interest in how these genetic markers overlap across brain networks. While this can be estimated by hierarchical clustering of the genetic correlation matrix to identify modular patterns of shared architecture, no brain-wide maps of these effects are available. To address this, we computed polygenic scores (PGS) from loci associated with ten brain volume regions of interest (ROIs): nine major subcortical structures and intracranial volume, with each locus weighted by its association with regional volume. In an independent sample from the discovery GWAS, we performed large-scale segmentation of 3D volumetric T1-weighted MRI scans using voxel-based morphometry (VBM) to map 3D profile of regions where gray matter volume (GMV) was associated with each PGS. We found statistically significant, localized effects for PGS defined for the amygdala, thalamus, and basal ganglia, but PGS for brainstem volume was associated with widespread differences throughout the brain. These brain-wide maps reveal patterns consistent with both localized and distributed genetic influences, offering a novel approach to interpret the genomic architecture of brain structure.

neuroscience↗

bioRxiv2026

Predicting Autopsy-Confirmed Neuropathology across Clinical, Neuroimaging, and CSF Biomarkers using Machine Learning

Accurate in vivo prediction of neuropathology is critical for advancing diagnosis and treatment of Alzheimers disease and related dementias (ADRDs). As many individuals with ADRDs have mixed pathologies ({beta}-amyloid, pathologic tau, cerebrovascular disease, vascular brain injury, pathologic TDP-43, hippocampal sclerosis, Lewy bodies), there is interest in determining how accurately we can infer these pathologic changes from clinical data, biofluid assays (e.g., CSF), and neuroimaging. Here we evaluated automated machine learning models trained on data curated by the AD Sequencing Project Phenotype Harmonization Consortium (N=7,894 individuals), to predict 26 autopsy-confirmed neuropathological outcomes. Predictors included in vivo clinical and cognitive composite scores, brain measures from 3D structural MRI and diffusion tensor imaging, image-derived measures of white matter hyperintensities (WMH), and CSF biomarkers. Predictive models were trained using ensemble learning with stratified cross-validation. We assessed performance using Spearmans rank correlation and Matthews correlation coefficient, to accommodate co-occurring pathologic changes. The added value of neuroimaging and CSF versus clinical features alone was quantified. Braak stage was among the most consistently predicted outcomes. CSF biomarkers best predicted {beta}-amyloid and tau pathology, but diffusion MRI metrics best captured vascular brain injury and white matter injury, and outperformed clinical and cognitive measures and anatomical MRI in predicting Lewy body disease. Anatomical measures from structural MRI outperformed standard clinical assessments in assessing neurodegeneration and hippocampal sclerosis, and WMH complemented cognitive measures in predicting TDP-43 pathology. These results establish a baseline for comparing modalities for inferring neuropathology.

neuroscience↗

bioRxiv2025

Generating diffusion MRI scalar maps from T1-weighted images using Reversible GANs

Diffusion tensor imaging (DTI) provides valuable insights into brain tissue microstructure, but acquiring high-quality DTI data is time-intensive and not always feasible. To mitigate data scarcity and enhance accessibility, we investigate the generation of synthetic DTI scalar maps--specifically mean diffusivity (MD)--from structural 3D volumetric T1-weighted brain MRI using a reversible generative adversarial network (RevGAN). Unlike conventional pipelines requiring multiple steps, our approach enables a single-step translation from T1 to diffusion-derived measures. We assess the quality and utility of the synthetic maps in two downstream tasks: sex classification and Alzheimers disease classification. Performance comparisons between models trained on real and synthetic DTI maps demonstrate that RevGAN-generated images retain meaningful microstructural features and offer competitive accuracy, underscoring their potential for data augmentation and analysis in neuroimaging workflows. We also examine how well models trained on these data generalize to a new population dataset from India (NIMHANS cohort).

neuroscience↗

bioRxiv2025

Using Diffusion Transformers to Generate Synthetic Diffusion Scalar Maps for Data Augmentation

Generation of high-quality synthetic brain MRI data could be beneficial for advancing neuroimaging research, particularly when access to large-scale, labeled datasets is limited. In this work, we leverage a pretrained Diffusion Transformer (DiT) architecture to synthesize 3D mean diffusivity (MD) scalar maps from the Cam-CAN dataset. To adapt the DiT model--originally trained on 2D natural images--for 3D neuroimaging data, we implemented a preprocessing strategy that tiles 2D slices from 3D volumes into composite 2D images, enabling effective finetuning. The quality of the generated synthetic images was evaluated using Multi-Scale Structural Similarity (MS-SSIM) and Maximum Mean Discrepancy (MMD) metrics, demonstrating high fidelity and anatomical coherence. To assess the utility of synthetic data in downstream tasks, we conducted transfer learning experiments for dementia classification on the ADNI dataset. A sex classification model, trained on both real and synthetic Cam-CAN data, was repurposed for this task, showing that synthetic samples can enhance model performance. These results highlight the potential of diffusion-based generative models for augmenting neuroimaging datasets and supporting clinical applications.

neuroscience↗

bioRxiv2025

Evaluation of Deep Learning Algorithms to Predict Multiple Dementia-Related Neuropathologies from Brain MRI, Clinical and Genetic Data

Alzheimers disease and related dementias (ADRD) involve overlapping neurodegenerative and vascular pathologies--such as amyloid-{beta} (A{beta}), tau, cerebral amyloid angiopathy (CAA), TDP-43, and alpha-synuclein--that complicate diagnosis and treatment. While PET and CSF biomarkers are useful for detecting A{beta} and tau, they are invasive, expensive, and not widely available. In contrast, magnetic resonance imaging (MRI) is non-invasive and widely accessible, offering an opportunity for pathology prediction when combined with deep learning. Most prior studies have focused on single-pathology detection, but there remains a need for models that can jointly predict multiple co-occurring pathologies. In this work, we evaluate deep learning models that integrate structural MRI with demographic, clinical, and genetic data to classify six autopsy-confirmed neuropathologies: A{beta}, tau, CAA, TDP-43, hippocampal sclerosis, and dementia with Lewy bodies. We compare our hybrid deep learning model to AutoGluon, an automated machine learning framework. Our findings support the potential of multimodal AI to enable non-invasive, comprehensive neuropathological profiling in ADRD.

neuroscience↗

bioRxiv2025

Synthetic Diffusion Tensor Imaging Maps Generated by 2D and 3D Probabilistic Diffusion Models: Evaluation and Applications

Diffusion tensor imaging (DTI) is a key neuroimaging modality for assessing brain tissue microstructure, yet high-quality acquisitions are costly, time-intensive, and prone to artifacts. To address data scarcity and privacy concerns - and to augment the available data for training deep learning methods - synthetic DTI generation has gained interest. Specifically, denoising diffusion probabilistic models (DDPMs) have emerged as a promising approach due to their superior fidelity, diversity, controllability, and stability compared to generative adversarial networks (GANs) and variational autoencoders (VAEs). In this work, we evaluate the quality, fidelity and added value for downstream applications of synthetic DTI mean diffusivity (MD) maps generated by 2D slice-wise and 3D volume-wise DDPMs. We evaluate their computational efficiency and utility for data augmentation in two downstream tasks: sex classification and dementia classification using 2D and 3D convolutional neural networks (CNNs). Our findings show that 3D synthesis outperforms 2D slice-wise generation in downstream tasks. We present a benchmark analysis of synthetic diffusion-weighted imaging approaches, highlighting key trade-offs in image quality, diversity, efficiency, and downstream performance.

neuroscience↗

bioRxiv2025

Deep Learning to Predict Future Cognitive Decline: A Multimodal Approach Using Brain MRI and Clinical Data

Predicting the trajectory of clinical decline in aging individuals is a pressing challenge, especially for people with mild cognitive impairment, Alzheimers disease, Parkinsons disease, or vascular dementia. Accurate predictions can guide treatment decisions, identify risk factors, and optimize clinical trials. In this study, we compared two deep learning approaches for forecasting changes, over a 2-year interval, in the Clinical Dementia Rating scale sum of boxes score (sobCDR). This is a key metric in dementia research, and scores range from 0 (no impairment) to 18 (severe impairment). To predict decline, we trained a hybrid convolutional neural network that integrates 3D T1-weighted brain MRI scans with tabular clinical and demographic features (including age, sex, body mass index (BMI), and baseline sobCDR). We benchmarked its performance against AutoGluon, an automated multimodal machine learning framework that selects an appropriate neural network architecture. Our results demonstrate the importance of combining image and tabular data in predictive modeling for clinical applications. Deep learning algorithms can fuse image-based brain signatures and tabular clinical data, with potential for personalized prognostics in aging and dementia.

neuroscience↗

bioRxiv2024

Lifespan Normative Modeling of Brain Microstructure

Normative models of brain metrics based on large populations could be extremely valuable for detecting brain abnormalities in patients with a variety of disorders, including degenerative, psychiatric and neurodevelopmental conditions, but no such models exist for the brains white matter (WM) microstructure. Here we present a large-scale normative model of brain WM microstructure - based on 19 international diffusion MRI datasets covering almost the entire lifespan (totaling N=54,583 individuals; age: 4-91 years). We extracted regional diffusion tensor imaging (DTI) metrics using a standardized analysis and quality control protocol and used hierarchical Bayesian regression (HBR) to model the statistical distribution of derived WM metrics as a function of age and sex. We extracted the average lifespan trajectories and corresponding centile curves for each WM region. We illustrate the utility of the method by applying it to detect and visualize profiles of WM microstructural deviations in a variety of contexts: in mild cognitive impairment, Alzheimers disease, and 22q11.2 deletion syndrome - a neurogenetic condition that markedly increases risk for schizophrenia. The resulting large-scale model provides a common reference to identify disease effects on the brains microstructure in individuals or groups, and to compare disorders, and discover factors affecting WM abnormalities. The derived normative models are a valuable resource publicly available to the community, adaptable and extendable to future datasets as the available data expands.

neuroscience↗

bioRxiv2024

Evaluating Synthetic Diffusion MRI Maps created with Diffusion Denoising Probabilistic Models

Generative AI models, such as Stable Diffusion, DALL-E, and MidJourney, have recently gained widespread attention as they can generate high-quality synthetic images by learning the distribution of complex, high-dimensional image data. These models are now being adapted for medical and neuroimaging data, where AI-based tasks such as diagnostic classification and predictive modeling typically use deep learning methods, such as convolutional neural networks (CNNs) and vision transformers (ViTs), with interpretability enhancements. In our study, we trained latent diffusion models (LDM) and denoising diffusion probabilistic models (DDPM) specifically to generate synthetic diffusion tensor imaging (DTI) maps. We developed models that generate synthetic DTI maps of mean diffusivity by training on real 3D DTI scans, and evaluating realism and diversity of the synthetic data using maximum mean discrepancy (MMD) and multi-scale structural similarity index (MS-SSIM). We also assess the performance of a 3D CNN-based sex classifier, by training on combinations of real and synthetic DTIs, to check if performance improved when adding the synthetic scans during training, as a form of data augmentation. Our approach efficiently produces realistic and diverse synthetic data, potentially helping to create interpretable AI-driven maps for neuroscience research and clinical diagnostics.

neuroscience↗

bioRxiv2024

Comparison of Explainable AI Models for MRI-based Alzheimer's Disease Classification

Deep learning models based on convolutional neural networks (CNNs) have been used to classify Alzheimers disease or infer dementia severity from 3D T1-weighted brain MRI scans. Here, we examine the value of adding occlusion sensitivity analysis (OSA) and gradient-weighted class activation mapping (Grad-CAM) to these models to make the results more interpretable. Much research in this area focuses on specific datasets such as the Alzheimers Disease Neuroimaging Initiative (ADNI) or National Alzheimers Coordinating Center (NACC), which assess people of North American, predominantly European ancestry, so we examine how well models trained on these data generalize to a new population dataset from India (NIMHANS cohort). We also evaluate the benefit of using a combined dataset to train the CNN models. Our experiments show feature localization consistent with knowledge of AD from other methods. OSA and Grad-CAM resolve features at different scales to help interpret diagnostic inferences made by CNNs.

neuroscience↗