A research team says its AI framework can identify Alzheimer's disease from MRI scans more accurately than conventional methods — a claim that, if it holds up in clinical settings, could matter for early dementia screening.
NeuroBridge combines self-supervised pretraining on large MRI datasets with three simultaneous learning objectives: hippocampal segmentation, hippocampal atrophy classification, and image reconstruction. The model then uses gated fusion fine-tuning to blend those signals before making a diagnosis. Tested on the ADNI and OASIS cohorts — two widely used neurodegenerative disease datasets — it hit 88.17% accuracy distinguishing Alzheimer's patients from cognitively normal controls in ADNI, and 82.78% on OASIS. The biggest improvements over baseline came in the harder cases: mild cognitive impairment and mixed-diagnosis scenarios.
MCI is where early intervention could actually change outcomes, so gains there carry more clinical weight than headline accuracy on clear-cut cases. The cross-cohort generalization result is also notable — models that ace one dataset and fall apart on another are a recurring problem in medical AI, and NeuroBridge held up reasonably well across both.
The framework is still a research prototype evaluated on curated cohorts, not a tool cleared for clinical use. Medical AI has a long history of promising benchmark numbers that soften considerably when exposed to the messier reality of hospital data pipelines and population diversity.