PD05.07. Enhancing SAM-Med3D with Prototype-Contrast Boundary Optimization for Esophageal Squamous Cell Carcinoma Segmentation

Abstract

Esophageal Cancer: Other Precise segmentation of esophageal squamous cell carcinoma (ESCC) lesions on CT imaging is essential for treatment planning. While the Segment Anything Model for 3D medical images (SAM-Med3D) shows promise, its performance in delineating tumor boundaries remains suboptimal. We propose a prototype-contrast boundary optimization strategy to enhance segmentation accuracy. In this retrospective multicenter study, CT volumetric data from 310 ESCC patients who underwent esophagectomy were collected from two centers (Center A: n=161, July 2018–February 2023; Center B: n=149). Data from Center A were used for training and internal validation, while Center B served as the external testing cohort. We evaluated SAM-Med3D for ESCC lesion segmentation and compared it with state-of-the-art methods including nnU-Net, SegVol, and MedSAM. To address SAM-Med3D’s limitation in boundary delineation, we introduced a prototype-contrast boundary pixel clustering optimization. This approach employs contrastive learning to refine boundary pixel classification by pulling features toward correct class prototypes while pushing them away from incorrect ones. Performance was assessed using Dice Similarity Coefficient (DSC) and 95th percentile Hausdorff Distance (HD95). The optimized SAM-Med3D achieved a DSC of 76.23% on internal validation and 73.57% on external testing, representing a 3.11% improvement over the original SAM-Med3D (internal DSC: 73.12%; external DSC: 70.46%). Our method outperformed all baseline approaches: nnU-Net (internal/external DSC: 72.45%/69.18%), SegVol (70.83%/67.52%), and MedSAM (71.26%/68.34%). The optimized model also achieved the lowest HD95 values (internal: 6.48 mm; external: 7.81 mm), indicating superior boundary delineation accuracy. The performance advantage was consistent across both centers, demonstrating robust generalizability. The prototype-contrast boundary pixel clustering optimization effectively enhances SAM-Med3D’s capability for ESCC lesion segmentation, achieving superior performance over nnU-Net, SegVol, and MedSAM across multicenter datasets. This approach addresses the critical challenge of accurate tumor boundary delineation and demonstrates strong generalizability, providing a valuable tool for preoperative assessment and surgical planning in ESCC.