CFML: Cross-Modal Feature-Gated Multi-Task Learning with T-Hybrid Loss for Alzheimer’s Diagnosis and MMSE Prediction

Abstract

Alzheimer’s disease (AD) is a progressive neurode-generative disorder for which early and accurate diagnosis is critical for clinical management. Effective diagnosis depends on both disease staging and continuous cognitive monitoring through the Mini-Mental State Examination (MMSE), yet most existing deep learning-based methods predict these two targets with separate models, and the few joint multi-task approaches struggle to balance classification and regression during training. To overcome these limitations, we propose CFML, a multitask architecture that integrates 3D MRI and demographic metadata through cross-modal fusion and attention-based feature gating, producing task-tailored representations for joint AD stage classification and MMSE regression. Within CFML, we introduce the T-Hybrid Loss, a dual-branch adaptive task-weighting strategy that fuses loss-aware and gradient-aware signals through a dispersion-driven coefficient, simultaneously correcting cross-task scale mismatch and gradient dominance. Comprehensive evaluations on ADNI-1, ADNI-2, and AIBL show that CFML achieves 97.32% accuracy with 1.43 (RMSE) MMSE on ADNI-1 and 97.35% accuracy with 1.45 (RMSE) MMSE in the combined three-class AD/CN/MCI setting, outperforming recent single-task and multi-task baselines.