Scholarlink

<p>Accurate delineation of the gross tumor volume (GTV) is an important component of glioblastoma radiotherapy planning, as contouring variability may influence target definition and subsequent treatment workflows. Manual contouring remains time-consuming and prone to substantial inter-observer variability. This retrospective cohort study evaluated the segmentation performance and boundary accuracy of a deep learning–based automated GTV segmentation framework using multimodal magnetic resonance imaging (MRI) within a radiotherapy-informed GTV segmentation evaluation framework. Multimodal MRI data, including T1-weighted (T1W), T1 contrast-enhanced (T1C), T2-weighted (T2W), and FLAIR sequences, were obtained from 195 patients in the BraTS 2024 dataset, acquired at standardized clinical field strengths. One hundred forty-one (141) cases were used for model training with five-fold cross-validation, and an independent test set of 54 patients was reserved for final evaluation. The GTV annotations of the training cohort were reviewed and refined by two radiation oncologists. The nnU-Net v2 architecture was trained with standardized preprocessing and data augmentation. Segmentation accuracy was assessed using volumetric overlap metrics (Dice similarity coefficient [DSC], Intersection-over-Union [IoU]), surface-based metrics (Hausdorff distance [HD], 95th percentile HD [HD95%], mean surface distance [MSD]), voxel-wise classification measures, and an adapted Gamma index (2&#xa0;mm/2%) as a complementary metric for quantifying spatial agreement. Statistical analysis included descriptive statistics, bootstrap confidence interval estimation, and agreement analysis. In the independent test cohort, the automated model achieved a mean DSC of 0.944 ± 0.032 and an IoU of 0.894 ± 0.045, with high boundary accuracy demonstrated by a mean HD95% of 2.99 ± 3.37&#xa0;mm and an MSD of 0.405 ± 0.255&#xa0;mm. The mean Gamma pass rate was 0.93 ± 0.04, providing an additional measure of spatial correspondence with expert-defined contours. Based on these results, the proposed nnU-Net v2 framework demonstrated high volumetric agreement and boundary consistency with expert-defined contours. These findings support its potential for automated GTV segmentation research in the context of radiotherapy applications; however, further validation using radiotherapy-specific datasets, external cohorts, and treatment-planning studies is required before its clinical utility can be established.</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

Deep learning-based boundary-accurate GTV segmentation in glioblastoma using multimodal MRI: a radiotherapy-oriented evaluation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月23日

查看全文
  • Amirmohammad Soltaninejad,
  • Daryoush Shahbazi-Gahrouei,
  • Amir Khorasani,
  • Simin Hemati

摘要

Accurate delineation of the gross tumor volume (GTV) is an important component of glioblastoma radiotherapy planning, as contouring variability may influence target definition and subsequent treatment workflows. Manual contouring remains time-consuming and prone to substantial inter-observer variability. This retrospective cohort study evaluated the segmentation performance and boundary accuracy of a deep learning–based automated GTV segmentation framework using multimodal magnetic resonance imaging (MRI) within a radiotherapy-informed GTV segmentation evaluation framework. Multimodal MRI data, including T1-weighted (T1W), T1 contrast-enhanced (T1C), T2-weighted (T2W), and FLAIR sequences, were obtained from 195 patients in the BraTS 2024 dataset, acquired at standardized clinical field strengths. One hundred forty-one (141) cases were used for model training with five-fold cross-validation, and an independent test set of 54 patients was reserved for final evaluation. The GTV annotations of the training cohort were reviewed and refined by two radiation oncologists. The nnU-Net v2 architecture was trained with standardized preprocessing and data augmentation. Segmentation accuracy was assessed using volumetric overlap metrics (Dice similarity coefficient [DSC], Intersection-over-Union [IoU]), surface-based metrics (Hausdorff distance [HD], 95th percentile HD [HD95%], mean surface distance [MSD]), voxel-wise classification measures, and an adapted Gamma index (2 mm/2%) as a complementary metric for quantifying spatial agreement. Statistical analysis included descriptive statistics, bootstrap confidence interval estimation, and agreement analysis. In the independent test cohort, the automated model achieved a mean DSC of 0.944 ± 0.032 and an IoU of 0.894 ± 0.045, with high boundary accuracy demonstrated by a mean HD95% of 2.99 ± 3.37 mm and an MSD of 0.405 ± 0.255 mm. The mean Gamma pass rate was 0.93 ± 0.04, providing an additional measure of spatial correspondence with expert-defined contours. Based on these results, the proposed nnU-Net v2 framework demonstrated high volumetric agreement and boundary consistency with expert-defined contours. These findings support its potential for automated GTV segmentation research in the context of radiotherapy applications; however, further validation using radiotherapy-specific datasets, external cohorts, and treatment-planning studies is required before its clinical utility can be established.