Discriminative Feature Learning for Lacune Detection in 2D T2-FLAIR Images using Supervised Contrastive Learning

Motivation: Current deep learning approaches struggle to distinguish lacunar infarcts from similar structures in cerebral small vessel disease. Goal(s): To develop an automated method that can effectively differentiate lacunes from similar-appearing structures (perivascular spaces, vessels) in imbalanced datasets. Approach: We trained a ResNet-34 encoder using supervised contrastive learning for enhanced feature discrimination, with Attention U-Net as the downstream segmentation network, using 427 FLAIR scans with expert annotations. Results: This work demonstrates an effective encoder training strategy for distinguishing small lesions like lacunes in cerebral small vessel disease through enhanced feature discrimination, potentially reducing both radiological interpretation time and inter-reader variability. Impact: This work demonstrates an effective encoder training strategy for distinguishing small lesions like lacunes in cerebral small vessel disease through enhanced feature discrimination, potentially reducing both radiological interpretation time and inter-reader variability.