UNICAL: Uncertainty-Aware Multimodal Learning with MRI-Derived Synthetic Amyloid PET for Alzheimer's Disease Classification

Limited availability of amyloid positron emission tomography (PET) motivates imaging-informatics methods that derive complementary molecular-imaging representations from routinely acquired magnetic resonance imaging (MRI). We developed UNICAL, an uncertainty-aware multimodal framework for Alzheimer’s disease-spectrum versus cognitively normal classification that combines a self-supervised three-dimensional MRI encoder, Contourlet Wavelet Stationary Radiomics (CoWS-Rad) engineered MRI features, and MRI-derived synthetic 18F-florbetapir (AV45) PET. CoCoLIT generated four stochastic PET realizations per MRI; their mean and voxelwise variance were used for uncertainty gating before probability-anchored transformer fusion. Synthetic PET was validated in 228 paired OASIS-3 MRI–AV45 examinations, achieving a mean per-examination voxelwise Pearson correlation of r = 0.864 and strong cortical-composite SUVR agreement with acquired PET (Pearson r = 0.954). In the OASIS-3 development cohort (813 sessions from 811 participants), participant-grouped five-fold cross-validation repeated across 20 seeds yielded AUROC 0.868 and expected calibration error (ECE) 0.061. The OASIS-3 development endpoint used the manuscript’s AD-spectrum reference standard, whereas ADNI3 used AD-versus-CN research-group labels, providing cross-dataset evaluation on a related but non-identical prediction endpoint. Without retraining, feature refitting, threshold optimization, or recalibration, the frozen framework was evaluated in an independent matched ADNI3 cohort of 261 participants (87 AD , 174 CN), achieving AUROC 0.836 (95% CI 0.791–0.876). ECE increased to 0.089 under external transfer, indicating that calibration was more sensitive to domain shift than rank discrimination. Ablation analysis and K-sample sensitivity supported the contributions of uncertainty gating, adaptive weighting, modality dropout, residual correction, and the K = 4 synthesis configuration. UNICAL therefore provides an uncertainty-aware MRI-derived multimodal representation for AD-spectrum classification that retains substantial cross-dataset discrimination while remaining distinct from acquired molecular imaging and not replacing amyloid PET as a clinical biomarker.