Hybrid CNN-Transformer Framework for Molecular Imaging-based Early Alzheimer's Disease Detection

Hybrid convolutional neural network (CNN)-transformer models are increasingly proposed for molecular imaging, but scan-level leakage and targets derived from the same image-quantification pipeline can overstate performance. We tested whether an image-only hybrid model could classify florbetapir PET amyloid positivity under participant-disjoint validation.

The Dallas Lifespan Brain Study contributed 551 longitudinal PET scans from 295 participants; 152 scans (27.6%) were amyloid-positive using the released global standardized uptake value ratio (SUVR) threshold of 1.09. Raw PET volumes were rigidly registered without label access and represented by 24 axial slices. CNN-GAP, slice-transformer, and age/sex baselines were evaluated using nested five-fold stratified group cross-validation. Calibration and operating thresholds were learned only within validation partitions, and 95% confidence intervals used 2,000 participant-cluster bootstrap resamples.

The hybrid achieved ROC-AUC 0.496 (95% CI 0.427-0.569), PR-AUC 0.274, balanced accuracy 0.490, and Matthews correlation coefficient -0.017. It did not outperform CNN-GAP (paired AUC difference -0.006, 95% CI -0.078 to 0.065) and performed below the age/sex baseline (AUC 0.690; hybrid-minus-baseline difference -0.194, 95% CI -0.287 to -0.100). Baseline-only and stable-label sensitivity analyses yielded hybrid AUCs of 0.529 and 0.502.

Raw registered PET slices did not provide stable participant-generalizable classification of an SUVR-derived amyloid target. Participant-level separation, validated tracer quantification, demographic baselines, and external validation remain necessary before clinical interpretation.