NeuroFM: A Neuropathology-optimized Foundation Model For Enhanced Diagnostic Accuracy In Alzheimer's Disease And Related Dementias

Accurate neuropathological diagnosis of Alzheimer's disease (AD) and related dementias requires recognizing subtle morphological patterns across brain regions. This expertise is scarce and prone to interobserver variability. Current AI pathology foundation models1 are trained primarily on non-neural, neoplastic surgical tissue, likely limiting their performance on CNS disease and leaving a gap in computational neuropathology. We introduce NeuroFM, a domain-specialized foundation model trained on ~1 billion tiles from 585,657 whole-slide images (80% brain tissue), using a DINOv2 self-supervised architecture with GMA attention. NeuroFM outperformed five general-purpose pathology models on 12 of 59 downstream neuropathology tasks, establishing a new benchmark for AI-assisted dementia diagnosis.