Scholarlink

<p>Three-dimensional segmentation of abdominal organs plays a fundamental role in quantitative medical image analysis, yet it remains difficult due to low tissue contrast, complex anatomical variability, and ambiguous boundaries in CT and MRI scans. To address these challenges, we introduce SCC-Seg, a shape-constrained collaborative framework that integrates semantic features with geometric priors for robust boundary reconstruction. The framework incorporates three components: a Boundary Initialization Module for contour extraction, a Contextual Convolutional Network (CCN) for boundary-aware probability mapping, and a Keypoint-based Shape Optimization Module with a neighborhood-constrained graph convolutional network (NC-GCN) for fine-grained refinement, guided by a joint loss that enforces semantic and geometric consistency. Experiments on the FLARE2021 and MSD datasets demonstrate that SCC-Seg achieves competitive Dice performance with improved boundary preservation, obtaining 90.73% Dice / 3.82 mm HD95 on FLARE2021 and 96.13% Dice / 0.20 mm HD95 on spleen segmentation. Notably, SCC-Seg provides enhanced HD95 stability on challenging organs such as the pancreas, reflecting its ability to maintain anatomically consistent boundaries under low-contrast conditions. These results highlight the effectiveness and broad applicability of SCC-Seg for accurate and efficient multi-organ segmentation in clinical practice.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SCC-Seg: A shape-constrained collaborative deep learning framework for accurate abdominal organ segmentation

  • 期刊论文

  • 发表日期：

    2026年2月19日

查看全文
  • Zihao Lv,
  • Yuanzhi Cheng

摘要

Three-dimensional segmentation of abdominal organs plays a fundamental role in quantitative medical image analysis, yet it remains difficult due to low tissue contrast, complex anatomical variability, and ambiguous boundaries in CT and MRI scans. To address these challenges, we introduce SCC-Seg, a shape-constrained collaborative framework that integrates semantic features with geometric priors for robust boundary reconstruction. The framework incorporates three components: a Boundary Initialization Module for contour extraction, a Contextual Convolutional Network (CCN) for boundary-aware probability mapping, and a Keypoint-based Shape Optimization Module with a neighborhood-constrained graph convolutional network (NC-GCN) for fine-grained refinement, guided by a joint loss that enforces semantic and geometric consistency. Experiments on the FLARE2021 and MSD datasets demonstrate that SCC-Seg achieves competitive Dice performance with improved boundary preservation, obtaining 90.73% Dice / 3.82 mm HD95 on FLARE2021 and 96.13% Dice / 0.20 mm HD95 on spleen segmentation. Notably, SCC-Seg provides enhanced HD95 stability on challenging organs such as the pancreas, reflecting its ability to maintain anatomically consistent boundaries under low-contrast conditions. These results highlight the effectiveness and broad applicability of SCC-Seg for accurate and efficient multi-organ segmentation in clinical practice.