MRI-based gross tumor volume delineation in high-grade glioma using a deep convolutional neural network.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiation Physics, The University of Texas MD Anderson Cancer Center, Houston, TX, USA.
- Department of Radiation Oncology, Cancer Institute, School of Medicine, Tehran University of Medical Sciences, Tehran, Iran. Electronic address: [email protected].
- Department of Public Health Sciences, Penn State College of Medicine, Hershey, PA, USA.
- Radiotherapy Oncology Department, Shohada-e Tajrish Educational Hospital, School of Medicine, Shahid Beheshti University of Medical Sciences, Tehran, Iran.
- Department of Radiation Oncology, Huntsman Cancer Institute, University of Utah, Salt Lake City, UT, USA. Electronic address: [email protected].
Abstract
In brain radiotherapy, accurate target volume delineation is a key step in treatment planning. However, manual contour delineation can be time-consuming, labor-intensive, and subject to inter-physician variability. Therefore, this study aimed to develop a deep learning (DL)-assisted, MRI-based auto-contouring model for gross tumor volume (GTV) delineation in radiotherapy for high-grade glioma (HGG). In addition to geometric evaluation, the dosimetric impact of using DL-based segmentations versus manually delineated contours for treatment planning was evaluated. A 2D ResU-Net deep convolutional neural network (CNN) was trained on 469 HGG patients from the Brain Tumor Segmentation (BraTS) datasets. Quantitative similarity metrics, including the Dice similarity coefficient (DSC) and Hausdorff distance (HD), were used for geometric evaluation. To evaluate the network's performance, the algorithm was tested on an independent clinical dataset (n = 17 patients). An experienced radiation oncologist manually delineated GTVs on MRI scans of these 17 HGG patients as reference contours. Subsequently, CT-MRI rigid brain registrations were performed in treatment planning software. Quantitative similarity metrics, such as the DSC and HD, were used for geometric evaluation. Additionally, dose-volume parameters were analyzed. The ResU-Net achieved a DSC score of 90.0% with an HD of 1.53 mm on the clinical test dataset. There was no significant (P = 0.801) difference in the volumes of the structure of interest (i.e., GTV) between the manual and auto-segmented contours. No significant differences in dose-volume parameters were found between the auto-segmented and manual contours. A 2D deep CNN model was developed for MRI-based auto-segmentation of HGG. Our data demonstrated the feasibility of DL-assisted auto-contouring. While DL-based auto-contouring cannot replace clinical experts, it can serve as a valuable contouring assessment tool, especially in medical educational centers.