Scholarlink

<p>In-depth analysis of the morphological characteristics of lung tumors in computed tomography (CT) images is crucial for achieving accurate tumor segmentation. While deep learning models based on artificial intelligence have demonstrated excellent segmentation performance, the lack of high-quality training data limits their reliability and applicability in this field. This study constructed a CT image dataset for lung tumor segmentation, comprising 730 cases of early-stage lung cancer and 449 cases of advanced-stage lung cancer. The data were segmented by experienced radiologists and radiation oncologists, and the patient information was anonymized. We proposed a segmentation model based on Segment Anything Model 2 (SAM2) with adapter fine-tuning and contrast-awareness. The Dice coefficients on the proposed dataset reached 91.64% and 89.51%, respectively, demonstrating the feasibility of training the segmentation model using the constructed dataset. This dataset can be used for the development, comparison, and validation of lung tumor CT image segmentation methods, contributing to the promotion of reproducible research and application exploration in the field of medical image segmentation.</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

Computed tomography lung tumor dataset for deep learning to facilitate medical image segmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月28日

查看全文
  • Hao Zhou,
  • Chang Chen,
  • Jiahao Li,
  • Tao Tao,
  • Xuhong Min,
  • Qilong Song

摘要

In-depth analysis of the morphological characteristics of lung tumors in computed tomography (CT) images is crucial for achieving accurate tumor segmentation. While deep learning models based on artificial intelligence have demonstrated excellent segmentation performance, the lack of high-quality training data limits their reliability and applicability in this field. This study constructed a CT image dataset for lung tumor segmentation, comprising 730 cases of early-stage lung cancer and 449 cases of advanced-stage lung cancer. The data were segmented by experienced radiologists and radiation oncologists, and the patient information was anonymized. We proposed a segmentation model based on Segment Anything Model 2 (SAM2) with adapter fine-tuning and contrast-awareness. The Dice coefficients on the proposed dataset reached 91.64% and 89.51%, respectively, demonstrating the feasibility of training the segmentation model using the constructed dataset. This dataset can be used for the development, comparison, and validation of lung tumor CT image segmentation methods, contributing to the promotion of reproducible research and application exploration in the field of medical image segmentation.