Self-Supervised Contrastive Learning with Attention Refinement for Chest X-Ray Disease Classification

This work combines momentum-based contrastive self-supervised (MoCo v2) pretraining with a CBAM-augmented ResNet-50 classifier to investigate their joint and individual contributions to four-class chest disease classification. A controlled ablation isolates the effect of pretraining strategy by holding the architecture, loss function, and training schedule fixed between the self-supervised and ImageNet-initialized variants.