USC researchers publish an Alzheimer's model built for missing scans
MEMOIR-VLM combines MRI, diffusion imaging and clinical scores, but its strongest diagnostic signal came from cognitive tests.
By Ryan Merket · Published
Primary source: X
Why it matters
MEMOIR-VLM tests whether one model can handle the incomplete scans common in practice. Its own results show the limit of the pitch: cognitive scores drive most diagnostic performance, and the language model did not beat nearest-neighbor retrieval.
Tamoghna Chattopadhyay (@TamoghnaChatto2) and colleagues at the University of Southern California published MEMOIR-VLM on October 1st, a research model designed to classify Alzheimer's disease using whatever combination of brain scans and clinical data is available. The paper's headline result was 91.3% balanced accuracy distinguishing cognitively normal people from people with dementia on held-out research data. Its three-category accuracy, including mild cognitive impairment, was 68.2%.…