DRSeg: a weakly supervised framework for breast ultrasound image segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate breast tumor segmentation in ultrasound images aids clinical decisions.
• Fully supervised methods demand extensive pixel-level annotations, which are labor-intensive.
• Weakly supervised segmentation using image-level labels is a viable alternative, but pseudo-label quality is a hurdle.

Purpose of the Study:

• To propose DRSeg, a weakly supervised segmentation framework for breast ultrasound images.
• To address the challenge of variable pseudo-label quality in automated segmentation.
• To reduce the dependency on pixel-wise annotations while maintaining high segmentation performance.

Main Methods:

• DRSeg employs a Class Activation Map (CAM)-based pipeline with refinements.
• A Dual-Region of Interest (ROI) selection algorithm identifies stable CAM localizations for reliable pseudo-label generation.
• Pseudo-labels are generated using the Segment Anything Model, and a Mean Teacher strategy trains the segmentation model.

Main Results:

• DRSeg demonstrated effectiveness on the BUSI and BLU datasets.
• With Swin Transformer V2, DRSeg achieved 59.35% IoU and 69.27% F1 score on BUSI.
• With ResNet-50, DRSeg achieved 51.79% IoU and 60.00% F1 score on BLU, outperforming existing methods.

Conclusions:

• DRSeg effectively handles pseudo-label quality variations in weakly supervised breast ultrasound segmentation.
• The framework achieves competitive segmentation performance, comparable to fully supervised methods.
• DRSeg offers a practical solution for automated breast tumor segmentation with reduced annotation burden.