A SAM-Guided ROI Quality Assessment Framework for Robust Breast Ultrasound Lesion Segmentation

Abstract

Breast ultrasound images have the following challenges; speckle noise, low contrast, irregular lesion borders, large variability among patients, all of which make segmentation difficult. However, these studies also demonstrate several disadvantages including producing many false positives, failing to accurately define object boundaries and having little interpretability. While SAM produces excellent candidate region proposals, it has not been shown to be effective when directly applied to ultrasound images because of the unique characteristics of ultrasound images.The purpose of this research is to develop a SAM-Guided ROI Quality Aware Segmentation Framework for segmenting lesions in breast ultrasound images. This framework will include four interconnected components: ROI Proposal, ROI Quality Assessment, Segmentation, and Boundary Refining. Once the ROI has been selected, the ROI will be segmented using an Attention U-Net Deep Learning Network. Finally, a Boundary Refine component will refine the border of the lesion and remove false positives from the ROI. Results obtained from experiments performed on the BUSI breast ultrasound database were evaluated using Dice Score and Intersection Over Union (IoU). Results indicated that integration of foundation model based ROI reasoning with deep learning segmentation networks improved both the location and quality of lesion segmentation, while reducing boundary error.