深度学习在子宫内膜癌术后临床靶区自动分割中的应用

Objective To evaluate the application of three deep learning algorithms in automatic segmentation of clinical target volumes (CTVs) in high-dose-rate brachytherapy after surgery for endometrial carcinoma. Methods A dataset comprising computed tomography scans from 306 post-surgery patients with endometrial carcinoma was divided into three subsets: 246 cases for training, 30 cases for validation, and 30 cases for testing. Three deep convolutional neural network models, 3D U-Net, 3D Res U-Net, and V-Net, were compared for CTV segmentation. Several commonly used quantitative metrics were employed, i.e., Dice similarity coefficient, Hausdorff distance, 95th percentile of Hausdorff distance, and Intersection over Union. Results During the testing phase, CTV segmentation with 3D U-Net, 3D Res U-Net, and V-Net showed a mean Dice similarity coefficient of 0.90 ± 0.07, 0.95 ± 0.06, and 0.95 ± 0.06, a mean Hausdorff distance of 2.51 ± 1.70, 0.96 ± 1.01, and 0.98 ± 0.95 mm, a mean 95th percentile of Hausdorff distance of 1.33 ± 1.02, 0.65 ± 0.91, and 0.40 ± 0.72 mm, and a mean Intersection over Union of 0.85 ± 0.11, 0.91 ± 0.09, and 0.92 ± 0.09, respectively. Segmentation based on V-Net was similarly to that performed by experienced radiation oncologists. The CTV segmentation time was < 3.2 s, which could save the work time of clinicians. Conclusion V-Net is better than other models in CTV segmentation as indicated by quantitative metrics and clinician assessment. Additionally, the method is highly consistent with the ground truth, reducing inter-doctor variability and treatment time.