From automatic detection to segmentation: A label-efficient SAM pipeline with feature-guided star-shape priors for breast ultrasound

Accurate segmentation of breast lesions in ultrasound images is important for quantitative lesion assessment, but remains difficult because of low contrast, noise, and artifacts. The Segment Anything Model (SAM) provides a flexible prompt-based segmentation framework, yet its performance on ultrasound images depends strongly on prompt quality and may require further boundary refinement. We propose a label-efficient automated pipeline that trains only an object detector with bounding-box annotations, keeps SAM frozen, and uses pixel-level masks only for benchmark evaluation. Detector-derived boxes are refined with a Chan–Vese (CV) model, and foreground/background point prompts are generated automatically without manual clicks or ground-truth prompts. After SAM prediction, a CV-inspired adaptive energy functional constructed on frozen SAM feature maps is jointly optimized with a multi-center star-shape constraint. This design uses SAM’s pretrained zero-shot segmentation capability without target-domain mask fine-tuning and retains an explicit, interpretable inference-time refinement process. The method was evaluated on three public breast ultrasound datasets, with a dermoscopic skin-lesion dataset included only as auxiliary out-of-domain validation. Relative to the best listed SAM-family baseline, the Dice improvements were 2.60, 1.93, and 1.07 percentage points on BUSI, BUS-UCLM, and UDIAT, respectively. Code is available at https://github.com/JeffJFZ/SAMwithStarShape .