Biomarker-Conditioned Vision Transformers with Deformable Biomarker Attention for Alzheimer's Disease Classification

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate Alzheimer's disease (AD) classification, especially differentiating mild cognitive impairment (MCI) from cognitively normal (CN) and AD, using structural MRI is challenging.
• Current multimodal approaches often limit clinical data's influence to the classifier, preventing it from shaping imaging feature extraction.

Purpose of the Study:

• To develop and evaluate a bimodal deep learning framework that integrates 3D structural MRI and clinical assessment scores for improved AD classification.
• To enable clinical biomarkers to actively modulate spatial feature extraction within the imaging encoder.

Main Methods:

• A bimodal deep learning framework utilizing 3D T1-weighted MRI and tabular clinical scores.
• A biomarker encoder generating a conditioning representation to modulate vision transformer patch tokens via spatial rescaling and cross-attention.
• Deformable biomarker attention (DBA) for internal cross-modal modulation and sparse, biomarker-guided spatial sampling.

Main Results:

• Achieved 95.68% accuracy, 95.39% macroprecision, 94.65% macrorecall, and 95.01% macro F1 score on the ADNI dataset.
• Uncertainty-aware predictions were observed, with higher uncertainty in misclassified cases, improving accuracy to 98.31% when uncertain predictions were rejected.
• Model performance showed gradual degradation (93.20% accuracy) when clinical scores were imputed, indicating robustness.

Conclusions:

• Within-encoder biomarker conditioning significantly enhances Alzheimer's disease classification accuracy.
• The proposed framework provides uncertainty-aware predictions, potentially aiding clinical decision-making and research.
• This approach offers a more effective integration of multimodal data for neurodegenerative disease classification.