Digital pathology of the living brain: a voxel-level spatio-temporal network for explainable ADHD diagnosis from raw rs-fMRI across multiple scanner sites.

Introduction: Attention-Deficit/Hyperactivity Disorder (ADHD) is one of the most prevalent neurodevelopmental disorders, affecting approximately 5-7% of children and adolescents worldwide. Clinical diagnosis currently relies on behavioral assessments that are susceptible to subjectivity and inter-rater variability. Resting-state functional magnetic resonance imaging (rs-fMRI) offers a promising avenue for objective ADHD identification; however, most existing approaches depend on derivative feature representations, such as functional connectivity (FC), fractional amplitude of low-frequency fluctuations (fALFF), or regional homogeneity (ReHo), which substantially compress the original blood-oxygen-level-dependent (BOLD) signal prior to model training. Furthermore, limited cross-site generalizability and insufficient voxel-level interpretability remain barriers to clinical translation. Methods: We propose VoxSTNet (Voxel-level Spatiotemporal Network), an explainable and telepathology-ready framework that operates directly on four-dimensional rs-fMRI BOLD volumes. A two-stage processing pipeline preserves the complete raw BOLD signal while reducing computational burden through moderate compression. Subject-wise z-score normalization mitigates scanner-specific intensity variations without introducing fold leakage. A time-distributed three-dimensional convolutional neural network (3D-CNN) coupled with a gated recurrent unit (GRU) captures spatiotemporal representations, while HiResCAM provides voxel-level interpretability. Experiments were conducted on the ADHD-200 dataset comprising 760 subjects (300 ADHD and 460 controls) from six acquisition sites. Performance was evaluated using Leave-One-Site-Out (LOSO) cross-validation as the primary assessment and five-fold cross-validation as a secondary analysis. Results: Five-fold cross-validation achieved an accuracy of 98.7 ± 0.4%, sensitivity of 98.2%, specificity of 99.1%, and area under the receiver operating characteristic curve (AUC) of 99.4% (95% confidence interval [CI]: 97.9-99.5%). Under the more stringent LOSO protocol, the model achieved a mean accuracy of 78.4% (95% CI: 75.1-81.7%). A controlled data-selection analysis demonstrated that retaining raw voxel-level information improved performance relative to derivative-feature baselines. HiResCAM saliency maps consistently highlighted the right caudate nucleus across validation subjects (mean Dice coefficient = 0.61 ± 0.08; Wilcoxon Discussion: VoxSTNet demonstrates that direct voxel-level modeling of rs-fMRI can achieve strong within-cohort performance while maintaining competitive cross-site generalizability. The identified saliency patterns align with established ADHD-related neurobiological findings, supporting the model's interpretability. Future work will focus on harmonization and domain-generalization strategies to further improve cross-site deployment performance.