Scholarlink

<p>Fully supervised deep learning methods have achieved remarkable success in medical image segmentation but remain heavily dependent on large-scale, manually annotated datasets. This is particularly challenging for liver tumor segmentation due to significant heterogeneity in tumor morphology, including number, size, and texture variations. These complexities make manual annotation difficult and limit the effectiveness of deep learning models in capturing intricate features, particularly due to their inability to fully incorporate domain expertise in modeling tumor-specific characteristics. To address these challenges, we seek to explore an emerging topic in this field: clinical knowledge-inspired learning. We propose MedCraft, a novel framework for annotation-free 3D liver tumor segmentation for computed tomography (CT) images. Our approach synthesizes tumors through a clinical knowledge-inspired pipeline that systematically incorporates medical expertise through tumor localization, texture synthesis, morphological modeling, and integration refinement. The pipeline integrates clinical knowledge and morphological image processing techniques to generate realistic tumors and corresponding annotations. Experimental results demonstrate that our synthetic tumor data enables practical model training without real tumor annotations and leads to comparable or superior performance in several key metrics compared to deep learning models trained on real tumor data. The proposed approach provides a new perspective for addressing the challenges of medical image annotation and segmentation. Our code is available at: <a href="https://github.com/koukihk/MedCraft-Public">https://github.com/koukihk/MedCraft-Public</a>.</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

The gift of clinical knowledge: annotation-free liver tumor segmentation via knowledge-driven synthesis

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Keyi Zhong,
  • Feng Ouyang,
  • Peter Xiaoping Liu,
  • Xuhui Huang,
  • Jianfeng Xu,
  • Huan Wan,
  • Xin Wei

摘要

Fully supervised deep learning methods have achieved remarkable success in medical image segmentation but remain heavily dependent on large-scale, manually annotated datasets. This is particularly challenging for liver tumor segmentation due to significant heterogeneity in tumor morphology, including number, size, and texture variations. These complexities make manual annotation difficult and limit the effectiveness of deep learning models in capturing intricate features, particularly due to their inability to fully incorporate domain expertise in modeling tumor-specific characteristics. To address these challenges, we seek to explore an emerging topic in this field: clinical knowledge-inspired learning. We propose MedCraft, a novel framework for annotation-free 3D liver tumor segmentation for computed tomography (CT) images. Our approach synthesizes tumors through a clinical knowledge-inspired pipeline that systematically incorporates medical expertise through tumor localization, texture synthesis, morphological modeling, and integration refinement. The pipeline integrates clinical knowledge and morphological image processing techniques to generate realistic tumors and corresponding annotations. Experimental results demonstrate that our synthetic tumor data enables practical model training without real tumor annotations and leads to comparable or superior performance in several key metrics compared to deep learning models trained on real tumor data. The proposed approach provides a new perspective for addressing the challenges of medical image annotation and segmentation. Our code is available at: https://github.com/koukihk/MedCraft-Public.