Neuro-Duo-MIL: A Dual-Branch DINOv3 Framework for Leakage-Aware Alzheimer’s Disease Classification Using 2D MRI and Clinical Data
Undergraduate Thesis
- Axial and coronal MRI slices fused through separate DINOv3 ViT backbones with bidirectional cross-view attention and gated-attention MIL pooling, combined with tokenized clinical covariates via cross-modal fusion.
- Achieved 94% test accuracy on CN/MCI/AD classification on ADNI-1 using strict subject-level splits.
- Accepted at the 41st International Conference on Image and Vision Computing New Zealand.