Scholarlink

Background <p>Esophageal cancer tumors exhibit complex and variable distribution. Due to differences in clinical experience, junior oncologists often show less accuracy in gross tumor volume (GTV) delineation compared to their senior counterparts. Deep learning methods have the potential to assist junior oncologists in improving the accuracy of GTV delineation. This study proposes a human–AI collaborative workflow for esophageal cancer radiotherapy and explores its clinical feasibility.</p> Methods <p>A retrospective collection of 730 esophageal cancer radiotherapy cases from our center was divided into training and testing sets at a 4:1 ratio. The human–AI collaborative workflow involves the following steps: (1) developing an automatic GTV delineation model based on planning CT images and obtaining an AI-prompt; (2) junior oncologists manually refining the GTVs generated by the model to serve as human-guided prompt; (3) using both prompts in combination with planning CT images to build two automatic refinement models; and (4) junior oncologists comparing the refined GTVs obtained by the above models, ultimately submitting the finalized GTVs for expert review by senior oncologists. Three junior oncologists from our center independently completed the manual tasks in steps 2 and 4, and the GTVs derived from the new approach were compared for accuracy against the ground truth delineated by senior oncologists.</p> Results <p>The results from the testing set indicated that after manual refinement by the three junior oncologists, the Dice similarity coefficient (DSC) between the automatic GTV delineation (step 1) and the ground truth increased from an initial value of 0.6538 ± 0.2022 to 0.7109 ± 0.1958, 0.7253 ± 0.1632, and 0.7236 ± 0.1965. Furthermore, by combining manual refinement, automatic refinement, and decision-making, the accuracy of the GTV delineation further improved to 0.7552 ± 0.1366, 0.7671 ± 0.1190, and 0.7757 ± 0.1354. These results demonstrate that the new approach significantly enhances the accuracy and stability of esophageal cancer GTV delineation.</p> Conclusion <p>With the support of the human–AI collaborative workflow, junior oncologists can delineate esophageal cancer GTV contours with improved accuracy, providing a reliable foundation for the integration of AI in clinical radiotherapy. The study also demonstrates that, in the current trend toward radiotherapy automation, clinical oncologists remain an indispensable and vital part of the process.</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

A human–AI collaborative workflow for improved gross tumor volume delineation in esophageal cancer radiotherapy

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月9日

查看全文
  • Hongfei Sun,
  • Jie Li,
  • Yufen Liu,
  • Fan Meng,
  • Ziqi An,
  • Yixin Cui,
  • Bo Lv,
  • Guangjin Chai,
  • Lecheng Jia,
  • Zihan Shi,
  • Meixi Liu,
  • Bingqi Zhu,
  • Jie Gong,
  • Lina Zhao

摘要

Background

Esophageal cancer tumors exhibit complex and variable distribution. Due to differences in clinical experience, junior oncologists often show less accuracy in gross tumor volume (GTV) delineation compared to their senior counterparts. Deep learning methods have the potential to assist junior oncologists in improving the accuracy of GTV delineation. This study proposes a human–AI collaborative workflow for esophageal cancer radiotherapy and explores its clinical feasibility.

Methods

A retrospective collection of 730 esophageal cancer radiotherapy cases from our center was divided into training and testing sets at a 4:1 ratio. The human–AI collaborative workflow involves the following steps: (1) developing an automatic GTV delineation model based on planning CT images and obtaining an AI-prompt; (2) junior oncologists manually refining the GTVs generated by the model to serve as human-guided prompt; (3) using both prompts in combination with planning CT images to build two automatic refinement models; and (4) junior oncologists comparing the refined GTVs obtained by the above models, ultimately submitting the finalized GTVs for expert review by senior oncologists. Three junior oncologists from our center independently completed the manual tasks in steps 2 and 4, and the GTVs derived from the new approach were compared for accuracy against the ground truth delineated by senior oncologists.

Results

The results from the testing set indicated that after manual refinement by the three junior oncologists, the Dice similarity coefficient (DSC) between the automatic GTV delineation (step 1) and the ground truth increased from an initial value of 0.6538 ± 0.2022 to 0.7109 ± 0.1958, 0.7253 ± 0.1632, and 0.7236 ± 0.1965. Furthermore, by combining manual refinement, automatic refinement, and decision-making, the accuracy of the GTV delineation further improved to 0.7552 ± 0.1366, 0.7671 ± 0.1190, and 0.7757 ± 0.1354. These results demonstrate that the new approach significantly enhances the accuracy and stability of esophageal cancer GTV delineation.

Conclusion

With the support of the human–AI collaborative workflow, junior oncologists can delineate esophageal cancer GTV contours with improved accuracy, providing a reliable foundation for the integration of AI in clinical radiotherapy. The study also demonstrates that, in the current trend toward radiotherapy automation, clinical oncologists remain an indispensable and vital part of the process.