A new study says the tool you use to carve up a brain scan matters as much as the algorithm that reads it for Alzheimer's disease.
Researchers built a benchmark that separates two steps usually judged together: parcellation (mapping brain regions) and classification (deciding if a scan shows Alzheimer's). They tested two fast deep-learning parcellation tools, SynthSeg+ and OpenMAP-T1, against the FreeSurfer clinical baseline, then ran the results through four different classifier types - clinical threshold rules, supervised feedforward networks, ensemble methods, and foundation models prompted with zero or few examples. The team also compared hard versus soft volumetry strategies. Everything was tested on the OASIS-1 dataset, with results reported using bootstrap confidence intervals rather than single-point accuracy scores.
That confidence-interval discipline is the real story. Most Alzheimer's detection papers report a single pipeline's accuracy and call it a result, treating brain segmentation as settled infrastructure rather than a variable worth testing on its own. This benchmark argues that choice of parcellation tool, and even how volumes get counted, changes downstream diagnostic performance enough to matter for anyone deploying these systems clinically.
Foundation models get name-checked here too, prompted zero- and few-shot rather than trained specifically for this task - a sign that even niche medical-imaging benchmarks now feel obligated to test the general-purpose model against the purpose-built one.