Biomarker-Conditioned Vision Transformers with Deformable Biomarker Attention for Alzheimer's Disease Classification.
Authors
Affiliations (5)
Affiliations (5)
- School of Computer and Information Sciences, The University of Newcastle, Newcastle, NSW, 2308, Australia. [email protected].
- Department of Computer Science, College of Khurma, Taif University, 21944, Taif, Saudi Arabia. [email protected].
- School of Computer and Information Sciences, The University of Newcastle, Newcastle, NSW, 2308, Australia.
- School of Computer and Information Sciences, The University of Newcastle, Newcastle, NSW, 2308, Australia. [email protected].
- Centre for Artificial Intelligence Research and Optimisation, Business and Hospitality Vertical, Torrens University, Surry Hills, NSW, 2010, Australia. [email protected].
Abstract
Alzheimer's disease (AD) classification from structural magnetic resonance imaging (MRI) remains challenging, particularly when distinguishing mild cognitive impairment (MCI) from both cognitively normal (CN) ageing and established AD. Multimodal approaches that combine imaging with clinical information are promising, but most confine the influence of clinical variables to the classifier head, so biomarkers cannot shape spatial feature extraction inside the imaging encoder. We propose a bimodal deep learning framework operating on two input sources: 3D T1-weighted structural MRI and tabular clinical assessment scores. A biomarker encoder maps the clinical scores into a conditioning representation that modulates vision transformer patch tokens through spatial rescaling and cross-attention throughout feature extraction. Within selected transformer layers, deformable biomarker attention (DBA) acts as an internal cross-modal modulation pathway rather than a third input modality, enabling sparse biomarker-guided spatial sampling from the 3D MRI feature grid. The framework was evaluated on ADNI data under strict subject-wise train, validation and test splits, so that repeated scans from the same individual never crossed partitions. On the held-out test set, the model achieved 95.68% accuracy, 95.39% macroprecision, 94.65% macrorecall and 95.01% macro F1 score, exceeding all comparison methods reproduced under the same protocol. Prediction uncertainty was higher for misclassified cases, and rejecting the most uncertain cases raised retained set accuracy to 98.31% at 85.03% coverage. Performance degraded gradually rather than catastrophically when clinical scores were with-held, falling to 93.20% accuracy with all five biomarkers imputed. These results indicate that within-encoder biomarker conditioning improves AD classification and yields uncertainty-aware predictions that may support decision-making in clinical research settings.