Classification d’images IRM cérébrales fondée sur un mécanisme d’attention invariant à l’échelle

Abstract

This thesis addresses the fine-grained classification of structural brain MRI while maintaining anatomical interpretability. We compare four complementary model families : (1) a hierarchical SIFT–MLP pipeline that classifies 3D keypoints and aggregates subject-level votes; (2) end-to end 3D CNNs trained on native T1-weighted MRI volumes; (3) a shallow CNN operating on 3D SIFT descriptor maps; and (4) a hybrid 3D-SIFT–Attention model that fuses scale-invariant local morphology with global context via cross-attention. Furthermore, we propose an Integrated Gradients–Guided Keypoint Attribution (IG-KPA) method, which aligns attributions with SIFT keypoints to enhance explainability. Our hybrid model improves classification accuracy in all cases, compared to standard networks or keypoints alone, for ADNI, HCP, OASIS datasets and sex, Alzheimer’s and age classification. State-of-the-art sex classification accuracy of 94% is obtained from rigidly aligned data, however size differences between males and females is a significant confound. After removing size differences via deformable registration, sex classification accuracy appears limited to 92%, consistent with other work, indicating that the brain MRIs of 8% of people appear more similar to those of the opposite sex. Moreover, interpretability analyses consistently localize signal to neuroanatomically plausible regions, with bilateral deep thalamo-capsular and periventricular territories for sex classification, and predominantly ventricular signatures for age and Alzheimer’s disease. Overall, the results underscore the advantage of attention mechanisms that merge stable, scale-invariant local cues with broader spatial context in fine-grained brain image classification tasks.

Date	11 Dec 2025
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Matthew Toews (Supervisor)