Semi-Supervised Contrastive Learning for Breast Cancer Diagnosis in Ultrasound Imaging: a Unified Framework with Auxiliary System

To address the domain gap between natural and medical images, this study proposes a semi-supervised contrastive learning framework based on SimCLR for model pre-training. By fine-tuning the pre-trained ResNet-18 model on the BUSI dataset (780 breast ultrasound images), we achieve an average IoU of 70.27 % for segmentation and 73.13% classification accuracy. Additionally, a clinical auxiliary diagnosis system integrating tumor prediction, segmentation, batch processing, and confidence visualization is developed to enhance diagnostic efficiency.