SAM Driven Robust Breast Lesion Segmentation via Adversarial Prompt Engineering

Early breast cancer detection is vital for improved prognosis. Traditional ultrasound breast cancer segmentation methods suffer from complexity, extensive data requirements, and poor generalization. This work applies the Perturbed prompts Segment Anything Model (PP-SAM), initially for polyp segmentation, to breast cancer lesion segmentation, addressing these limitations. PP-SAM employs variable bounding box prompt perturbations during fine-tuning, enhancing generalization and robustness, especially with limited data. This reduces reliance on extensive labeled datasets, a key bottleneck. PP-SAM’s efficacy was evaluated on the public Breast Ultrasound Images (BUSI) dataset. Results show PP-SAM’s high data efficiency and performance: 0-shot fine-tuning yielded a 56.77% Dice Score (DSC), improving to 74.42% (1-shot) and 85.7% (200-shots). This few-shot performance surpasses existing BUSI dataset state-of-the-art models, highlighting PP-SAM as a potent, data-efficient foundational model for clinical breast ultrasound segmentation.