An Edge-aware Prompt-enhanced SAM for Ultrasound Image Segmentation

Jul 16, 2026

Upload by Eden

Ultrasound image segmentation is vital for outlining anatomical structures and lesions, enabling reliable diagnosis. Existing use of the Segment Anything Model (SAM) is limited by weak boundary delineation on ultrasound, caused by speckle noise, low contrast, and blurred contours. EP-SAM introduces an edge-aware, prompt-enhanced adaptation: multi-block feature extraction improves coarse-to-fine semantics, while edge-aware supervision strengthens contour robustness. High-quality prompts guide the model to regions of interest, and experiments on multiple benchmarks show consistent performance gains over SAM-based approaches.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages6

Reading Time15 min

Abstract

Index Terms

Introduction

An Edge-aware Prompt-enhanced SAM for Ultrasound Image Segmentation Wenhao Li 1 , Fangyi Liu2* , Bo Du2* 1 Key Laboratory of Aerospace Information Security and Trusted Computing, Ministry of Education, School of Cyber Science and Engineering, Wuhan University, Wuhan, China 2 National Engineering Research Center for Multimedia Software, Institute of Artificial Intelligence, School of Computer Science and Hubei Key Laboratory of Multimedia and Network Communication Engineering, Wuhan University, Wuhan, China [2024202210100@whu.edu.cn](2024202210100@whu.edu.cn), [fangyiliu@whu.edu.cn](fangyiliu@whu.edu.cn), [dubo@whu.edu.cn](dubo@whu.edu.cn) arXiv :2607 .07240v 1 [ cs .CV] 8 Jul 2026 Abstract—Ultrasound image segmentation is essential for delineating anatomical structures and lesions, providing the foundation for accurate diagnosis. While the Segment Anything Model (SAM) has demonstrated remarkable success on natural images, its performance on ultrasound data is often hindered by poor boundary delineation. To address this limitation, we propose EP-SAM, an edge-aware and prompt-enhanced adaptation of SAM. Specifically, we leverage multi-block feature extraction from the image encoder to enrich coarse-to-fine semantic representations, while edge-aware supervision of the image encoder improves robustness to contour ambiguity and speckle noise. By integrating these complementary cues, EP-SAM generates high-quality prompts that effectively guide the model toward target regions of interest. Experimental results on multiple benchmarks demonstrate that EP-SAM consistently outperforms existing SAM-based methods. Index Terms—Ultrasound image segmentation, SAM, edgeaware, prompt-enhanced I. INTRODUCTION Ultrasound medical image segmentation is crucial for clinical diagnosis, treatment planning, and biomedical research [20] . Owing to its radiation-free, portable, and costeffective nature, ultrasound is widely adopted for point-of-care applications. However, its real-time, reflection-based imaging mechanism introduces speckle noise, low contrast, and blurred anatomical boundaries, which degrade image quality and obscure structural details [32] . These inherent challenges not only complicate accurate segmentation but also hinder the direct application of deep learning models and prompt-based strategies originally designed for natural images. Recently, the Segment Anything Model (SAM) has demonstrated remarkable segmentation capabilities. This breakthrough has further opened new opportunities for advancing ultrasound image segmentation and has motivated a growing body of SAM-based methods, which can be broadly categorized as follows. (1) Decoder-level adaptation focuses on refining the segmentation head [21] by fine-tuning themask decoder to better align with medical domain data. (2) Encoder-level adaptation modifies the image encoder to enhance feature representation [27], [31], and (3) Promptlevel adaptation optimizes or learns new prompts to guide the *Corresponding authors | Prompt （a） | Prompt Encoder Prompt （b） | | --- | --- | Image Encoder Prompt Encoder A Mask Decoder Mask Decoder （c） Image Encoder Prompt Encoder Mask Decoder Learnable Frozen A Adapter Fig. 1: Comparison of SAM-based adaptations for medical image segmentation. Unlike other prompting strategies, our EP-SAM enhances the synergy of individual components and explicitly alleviates boundary ambiguity. segmentation results without modifying SAM’s weights [28] . However, as shown in Figure 1(a), most existing adaptations treat the constituent components of SAM as isolated modules, resulting in a lack of synergy between them. This architecture often fails to leverage the rich semantic information available during the encoding stage to guide the subsequent prompting process. To address this, self-prompting methods have emerged as a promising solution. As illustrated in Figure 1(b), they

FAQs

What problem does EP-SAM address in ultrasound image segmentation?

Ultrasound boundaries are often blurred and corrupted by speckle noise, which makes SAM struggle with accurate contour delineation and leads to edge drift. EP-SAM targets this boundary ambiguity with edge-aware guidance.

How does EP-SAM improve SAM performance on ultrasound images?

EP-SAM uses multi-block feature extraction to enrich coarse-to-fine semantic representations. It also applies edge-aware supervision via an Edge-Aware Module to improve robustness to contour ambiguity and speckle noise.

What are the two main modules in EP-SAM?

EP-SAM consists of an Edge-Aware Module (EAM) that extracts fine-grained boundary cues for supervision and a Prompt Enhanced Module (PEM) that fuses intermediate features with boundary priors to produce boundary-informed initial masks and prompts.

Jul 16, 2026

Upload by Eden