A unified deep-learning-heuristic framework for MRI-based classification of Alzheimer’s, epilepsy, and aneurysms

MRI-based diagnosis of neurological disorders typically requires processing entire image volumes despite substantial inter-slice redundancy, resulting in unnecessary computational overhead. We present a computational framework that automates frame selection for three brain disorders, namely Alzheimer’s disease subtypes, epilepsy, and intracranial aneurysms, by identifying diagnostically informative slices. The framework combines EfficientNetB0 and ResNet50V2 with precision-constrained heuristic decision rules to prioritize frames while maintaining diagnostic performance. Evaluated on 1526 patient scans across six datasets, the approach achieved 100% subject-level accuracy on internal held-out test sets (n = 17–42 subjects per dataset) while processing only 11% of available frames, representing an 89% reduction in computational load. On frame-level evaluation, baseline CNN accuracies on the held-out test set ranged from 40.62% (ADNI-192) to 95.00% (Epilepsy) at the best single frame. The heuristic module further improved subject-level performance by 5.00–59.38 percentage-points over standard aggregation. External validation across many clinical locations and scanner suppliers is needed to establish generalizability before clinical adoption.