Segment Anything Small for ultrasound: enhanced segmentation of small anatomical structures using iterative point prompts and image transformations

Purpose: Accurate segmentation of anatomical structures in ultrasound (US) images, particularly small ones, is challenging due to noise and variability in imaging conditions (e.g. probe position, patient, anatomy, tissue, pathology variety, etc.). In this paper, we introduce a novel approach that enhances the segmentation of small anatomical structures by integrating advanced image transformations with iterative segmentation using point prompts, named Segment Anything Small (SAS). Method: SAS employs a dual transformation strategy: transforming images to simulate different organ sizes and adjusting the region of interest (ROI) to represent varying pixel values and noise levels typical of US. These transformations improve the model’s robustness to noise and variability, ensuring effective segmentation of small structures without compromising the accuracy of larger ones. Results: Experimental results demonstrate improvement in Dice scores up to 0.37, with an average gain of 0.19 (±0.03 std) points. Additionally, our iterative point prompts provide precise control and adaptive refinement, achieving comparable performance to bounding box prompts with just 2 points. Conclusions: By combining transformative image techniques with point prompts, the SAS method offers a simple, versatile and effective solution for achieving detailed and accurate segmentations across diverse US imaging applications.