DRSeg: a weakly supervised framework for breast ultrasound image segmentation

Automatic segmentation of breast tumors in ultrasound images can assist physicians in making accurate and effective decisions. However, fully supervised learning methods typically require large amounts of pixel-level annotations, which are often difficult to obtain. Weakly supervised segmentation using image-level labels offers a promising alternative, but generating high-quality pseudo-segmentation labels remains a major challenge. In this study, we propose DRSeg, a weakly supervised segmentation framework for breast ultrasound (BUS) images that explicitly addresses the variability of pseudo-label quality. DRSeg augments a standard CAM-based pipeline with four components: (1) Class Activation Map (CAM) generation and refinement; (2) a Dual-ROI selection algorithm that identifies images with stable CAM localizations for reliable pseudo-label generation; (3) pseudo-label generation using the Segment Anything Model; and (4) training a segmentation model with the Mean Teacher strategy, leveraging both pseudo-labeled and non-pseudo-labeled images. Extensive experiments on two public BUS datasets, BUSI and BLU, demonstrate the effectiveness of the proposed DRSeg framework and its individual components. Using Swin Transformer V2 as the backbone, DRSeg achieves an IoU of 59.35% and an F1 score of 69.27% on BUSI. Using ResNet-50 as the backbone, DRSeg achieves an IoU of 51.79% and an F1 score of 60.00% on BLU, outperforming six existing weakly supervised methods. Overall, DRSeg reduces reliance on pixel-wise annotations while achieving competitive segmentation performance across two public BUS datasets and two backbones.