Segmentation-Guided Development of Visual Classification Criteria for Alzheimer's Disease

Alzheimer's disease (AD) causes progressive structural brain changes that precede clinical symptoms by years. Detecting these changes using structural MRI remains challenging, especially in early stages and when relying on visual interpretation alone. Automated semantic segmentation methods offer anatomical precision and objective measurements, but their outputs are rarely used to support human visual assessment. In this study, we explored whether such segmentation outputs can be used to guide a non-expert investigator in developing and applying interpretable diagnostic criteria. We used images from the Alzheimer's Disease Neuroimaging Initiative (ADNI) and implemented a structured, segmentation-informed workflow in which a novice with no prior training in radiology or neuroanatomy developed classification rules based on visual appearance and volumetric readouts through three guided pilot phases. In a fourth phase, the investigator applied these criteria to an independent subset of ADNI images while blinded to the respective ADNI participants' diagnostic labels. Using an anatomical segmentation model (MAPER) with training data from a pre-release version of the Hammers Adult Brain Atlas Database (120 brain regions), the investigator focused on the piriform cortex (PC). The choice of PC was context-driven, reflecting an ongoing quantitative study of PC volume. A binary classification (AD-like versus CN-like) rule based on PC volume (< or > 430 mm^3), supported by assessments of PC shape and global atrophy, yielded an accuracy of 0.71 across 200 cases spanning four diagnostic groups. Accuracy increased to 0.77 when the analysis was restricted to CN and AD cases (with intermediate pathology (MCI) excluded). These results show that segmentation-guided visual workflows can enable non-experts to apply anatomically grounded classification criteria with moderate accuracy. Our framework can be expanded to other regions and promises to be useful for generating interpretable models, for supporting explainable AI, and for accelerating the acquisition of diagnostic skills.