Scholarlink

<p>Automatic segmentation of gastric tumors in computed tomography (CT) images enhances radiological assessments by precisely delineating tumor boundaries, thereby improving diagnosis, treatment planning, and monitoring. However, gastric tumor segmentation remains challenging due to inter-class imbalance, where tumors occupy a small fraction of the image, and high inter-class similarity, as tumors and surrounding tissues exhibit comparable pixel-level appearances. Few-shot semantic segmentation offers a potential solution by reducing dependence on extensive annotations, yet struggles with the inherent complexities of medical imaging. To address these limitations, we propose a dual contrastive learning framework that effectively leverages information from non-target-class images. Our method comprises: (1) an instance contrastive learning module that enhances feature discrimination by mitigating interference from non-tumor regions, and (2) a pixel contrastive learning module that refines tumor boundary delineation. Experimental evaluations on an internal CT dataset demonstrate that our approach outperforms existing methods. Furthermore, we validate its generalizability by applying it to liver tumor segmentation using the LiTS17 dataset.</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

Dual Contrastive Learning for Few-Shot Gastric Cancer CT Images Segmentation

  • 期刊论文

  • 发表日期：

    2025年5月28日

查看全文
  • Junxian Bao,
  • Simin Gao,
  • Weijie Shen,
  • Lihua Hang,
  • Keyang Cheng

摘要

Automatic segmentation of gastric tumors in computed tomography (CT) images enhances radiological assessments by precisely delineating tumor boundaries, thereby improving diagnosis, treatment planning, and monitoring. However, gastric tumor segmentation remains challenging due to inter-class imbalance, where tumors occupy a small fraction of the image, and high inter-class similarity, as tumors and surrounding tissues exhibit comparable pixel-level appearances. Few-shot semantic segmentation offers a potential solution by reducing dependence on extensive annotations, yet struggles with the inherent complexities of medical imaging. To address these limitations, we propose a dual contrastive learning framework that effectively leverages information from non-target-class images. Our method comprises: (1) an instance contrastive learning module that enhances feature discrimination by mitigating interference from non-tumor regions, and (2) a pixel contrastive learning module that refines tumor boundary delineation. Experimental evaluations on an internal CT dataset demonstrate that our approach outperforms existing methods. Furthermore, we validate its generalizability by applying it to liver tumor segmentation using the LiTS17 dataset.