Digital pathology of the living brain: a voxel-level spatio-temporal network for explainable ADHD diagnosis from raw rs-fMRI across multiple scanner sites
By
Punna Rao Vuyyuru
Sathya Babu Korra
Srinivas Naik Nenavath
June 30, 2026
Clinical Scorecard: Voxel-Level Spatio-Temporal Network Analysis for Objective ADHD Diagnosis Using Raw rs-fMRI Data Across Multiple Scanner Locations
At a Glance
Category Detail
Condition Attention-Deficit/Hyperactivity Disorder (ADHD) Key Mechanisms Voxel-level spatiotemporal analysis of resting-state fMRI data Target Population Children and adolescents with ADHD Care Setting Neurodevelopmental disorder diagnosis using functional neuroimaging
Key Highlights
Achieved 98.7% accuracy with five-fold cross-validation. Utilized a two-stage processing pipeline to retain raw BOLD signal. Demonstrated strong within-cohort performance and competitive cross-site generalizability. HiResCAM provided voxel-level interpretability, highlighting the right caudate nucleus. Addressed the sustainability problem in rs-fMRI-based diagnosis. Guideline-Based Recommendations
Diagnosis
Utilize resting-state fMRI for objective ADHD identification. Management
Implement VoxSTNet framework for ADHD diagnosis in clinical settings. Monitoring & Follow-up
Use voxel-level saliency maps for tracking diagnostic relevance. Risks
Consider inter-scanner variability in BOLD signal intensity. Patient & Prescribing Data
760 subjects (300 ADHD and 460 controls) from six acquisition sites.
Direct voxel-level modeling improves performance over traditional derivative-feature approaches.
Clinical Best Practices
Employ subject-wise z-score normalization to mitigate scanner-specific variations. Utilize Leave-One-Site-Out cross-validation for robust performance evaluation. Related Resources & Content
