文献速递乳腺癌放疗放疗物理与技术

用于局部区域乳腺癌放疗的深度学习分割模型的训练、验证与临床实施

Training, validation, and clinical implementation of a deep-learning segmentation model for radiotherapy of loco-regional breast cancer.
PubMed 文献 · Radiother Oncol · 2022-05-23

在挪威两个放疗中心的170例左侧乳腺癌病例上训练了覆盖7个临床靶区和11个危及器官的深度学习分割模型,另用30例验证并评估几何相似度、定性可用性及剂量学参数,以人工观察者间变异为基准。模型在多数结构上几何相似度显著优于人工观察者间变异,定性评估中靶区14%无需或仅需微小修正、71%需微小修正,危及器官72%无需修正、26%需微小修正;验证结果促使决定临床实施,并将在适用性、标准化和效率方面进行监测。

要点

  • 模型在多数结构上的几何相似度显著优于人工观察者间变异。
  • 定性评估中靶区14%无需或仅需微小修正、71%需微小修正,危及器官72%无需修正、26%需微小修正。
  • 基于D98>95%的剂量覆盖在乳腺靶区100%满足、淋巴结靶区89%满足。

核心结论

  • 验证结果令人信服,促使决定进行临床实施。
  • 临床使用将在适用性、标准化和效率方面进行监测。

研究设计与样本

  • 在挪威两个放疗中心的170例左侧乳腺癌病例上训练了针对7个临床靶区和11个危及器官的深度学习分割模型。
  • 另用30例患者病例进行验证,评估Dice相似系数和Hausdorff距离、按临床可用性进行定性评分以及相关剂量学参数。
  • 同时评估人工观察者间变异并作为基准,靶区勾画遵循ESTRO指南。

主要发现

  • 基于几何相似度指标,模型在多数结构上显著优于人工观察者间变异。
  • 定性评估中,靶区14%无需或仅需微小修正、71%需微小修正;危及器官72%无需修正、26%需微小修正。
  • 靶区15%需重大修正,危及器官2%需重大修正;最常发生的修正位于结构的头侧和尾侧部分。
  • 基于D98>95%的剂量覆盖在乳腺靶区100%满足、淋巴结靶区89%满足。
  • 危及器官剂量参数未发现具有临床相关性的差异;模型已嵌入商用治疗计划系统,生成结构耗时1.5分钟。

关键数值与推荐

  • 靶区勾画遵循ESTRO指南。
  • 定量评估采用Dice相似系数和Hausdorff距离。
  • 定性评分按临床可用性分为无需修正、微小修正和重大修正。
  • 剂量覆盖以D98>95%为满足标准。

临床怎么用

  • 原文未据此给出推荐。

局限性

  • 原文未报告局限性。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

AIM: To train and validate a comprehensive deep-learning (DL) segmentation model for loco-regional breast cancer with the aim of clinical implementation. METHODS: DL segmentation models for 7 clinical target volumes (CTVs) and 11 organs at risk (OARs) were trained on 170 left-sided breast cancer cases from two radiotherapy centres in Norway. Another 30 patient cases were used for validation, which included the evaluation of Dice similarity coefficient and Hausdorff distance, qualitative scoring according to clinical usability, and relevant dosimetric parameters. The manual inter-observer variation (IOV) was also evaluated and served as a benchmark. Delineation of the target volumes followed the ESTRO guidelines. RESULTS: Based on the geometric similarity metrics, the model performed significantly better than IOV for most structures. Qualitatively, no or only minor corrections were required for 14% and 71% of the CTVs and 72% and 26% of the OARs, respectively. Major corrections were required for 15% of the CTVs and 2% of the OARs. The most frequent corrections occurred in the cranial and caudal parts of the structures. The dose coverage, based on D98 > 95%, was fulfilled for 100% and 89% of the breast and lymph node CTVs, respectively. No differences in OAR dose parameters were considered clinically relevant. The model was implemented in a commercial treatment planning system, which generates the structures in 1.5 min. CONCLUSION: Convincing results from the validation led to the decision of clinical implementation. The clinical use will be monitored regarding applicability, standardization and efficiency.

来源:Radiother Oncol,查看原文。版权归原作者所有。

引用

  1. Almberg SS, Lervåg C, Frengen J, et al. Training, validation, and clinical implementation of a deep-learning segmentation model for radiotherapy of loco-regional breast cancer. Radiother Oncol. 2022. PMID: 35618100. doi:10.1016/j.radonc.2022.05.018

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。