FDG-PET Image Classification in Alzheimer's Disease: from Traditional Visual Analysis to Advanced Transfer Learning.
Level 4 - case-series / case-control
Retrospective cross-sectional diagnostic classification study with an imperfect reference standard
PubMed 40385367 · doi:10.1007/s13139-025-00908-2
What was done
Baseline 18F-FDG-PET scans from 794 participants with probable Alzheimer's disease enrolled in two Phase III clinical trials were visually classified into typical Alzheimer's disease (temporoparietal hypometabolism) or mixed patterns (patchy hypometabolism in frontal and cerebellar regions alongside temporoparietal hypometabolism). Differences were evaluated using region-of-interest Standardized Uptake Value Ratio (SUVR) analysis and automated via transfer learning on a subset of 100 scans (50 typical, 50 mixed) using visual classification as the reference standard.
What was found
Visual reading categorized 533 of 794 participants (284 female) as typical Alzheimer's disease and 261 (154 female) as mixed. In the machine learning subset, one cross-validation loop achieved a sensitivity of 94.73%, specificity of 95.23%, and accuracy of 95%. Across 5-fold cross-validation, average classification accuracy was 97.5%. The abstract reported no numeric SUVR values.
Why it matters
Transfer learning can reproduce expert visual interpretation of complex hypometabolic patterns on FDG-PET in clinical trial datasets.
Limits
The reference standard was subjective visual interpretation rather than neuropathological confirmation or multi-modal biomarker validation. The machine learning model was trained and tested on only 100 scans without validation in an independent external cohort.
Cited by
- supports Poor glucose utilization in the temporal and parietal regions on PET scan is a hallmark of Alzheimer's disease.