Deep learning improves Alzheimer’s diagnosis by focusing on brain regions
Anatomical Grounding and Leakage-Aware Multimodal Contrastive Learning for Alzheimer's Disease Classification from Structural MRI
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Alzheimer’s disease diagnosis from MRI scans can sometimes rely on irrelevant parts of the images or on clinical data that already hints at the diagnosis. The authors showed that training their image analysis model to focus on specific brain areas linked to Alzheimer’s helps improve accuracy. They also found that combining brain scan features with certain clinical data can boost diagnosis but may reflect information already used to label patients. Their approach uses new methods to ensure the model pays attention to meaningful brain structures for better Alzheimer's staging.
What this means in practice
- •For clinical imaging teams: Use anatomically focused deep learning models to improve MRI-based Alzheimer’s disease staging accuracy by reducing reliance on irrelevant image regions.
- •For hospital data teams: Integrate cognitive scores and regional brain volume data with MRI features to enhance Alzheimer’s diagnostic models in clinical data systems.
Tested on one dataset.
Authors
Paul-Gabriel Nicolae, Irina Georgiana Mocanu
Abstract
Deep networks trained on structural MRI for Alzheimer's disease (AD) staging often reach reasonable accuracy while attending to anatomically irrelevant regions, and multimodal models that add clinical tables frequently rely on variables that were used to assign the diagnostic label in the first place. We study both issues with a deliberately lightweight slice-based encoder (ResNet18 with a one-layer Transformer over slices) on 1,075 baseline T1-weighted scans from ADNI-1. First, we use FastSurfer segmentations as an anatomical reference: YOLOv8 models trained on segmentation-derived labels localize Alzheimer-relevant structures with mAP_50 above 0.96, and a Grad-CAM comparison shows that the image-only classifier frequently attends to the skull, orbits and background. Second, we adapt a CLIP-style image - tabular contrastive framework and organize ADNIMERGE variables along a label-leakage spectrum. Fusion with cognitive scores yields 87.3% three-way accuracy, which we treat as a leakage-driven upper bound rather than an imaging result; fusion with regional volumes yields 73.0%. We observe that the choice of contrastive target changes what the image encoder learns: on MCI vs. CN, the image-only head reaches 52.4% when the encoder is aligned to cognitive scores and 73.8\% when aligned to volumes, although no tabular input is used at inference. Third, restricting the input to a per-subject crop of the medial temporal lobe raises image-only three-way accuracy from 58.7% to 65.1%. All results come from single runs on a small balanced test set, and we report confidence intervals and the protocol differences that prevent direct comparison with published numbers.