An automatic delineation self-configuring nnU-Net for HR-CTV in cervical brachytherapy

Abstract

Objective: To evaluate the performance of nnU-Net in automatic delineation of HR-CTV (High-Risk Clinical Target Volume) for CT-based brachytherapy in cervical cancer. It was compared with MONAI's 3D SegResNet, which serves as a strong baseline commonly used for pelvic soft tissue segmentation, in order to provide quantitative evidence for clinical AI-assisted tumor radiotherapy target delineation. Methods: A total of 388 patients undergoing CT-guided HDR (High-Dose-Rate) brachytherapy were retrospectively enrolled and divided into 271 training, 58 validation, and 59 testing sets using stratified randomization (seed 42). nnUNet performs 5-fold cross-validation on the merged 329-case development set, with 1000 rounds per fold, and uses the average of the five-fold softmax probabilities as the final output during inference.As a control, the same data was used to train a built-in 3D SegResNet baseline in MONAI 1.5.2 (DiceCE loss, 300 epochs, with 58 checkpoints selected from the validation set).The other 59 test sets did not participate in the training and model selection process. Dice, IoU, HD95, and Surface Dice @1/2/3 mm were calculated for the two methods using the same evaluation script. Differences between groups were tested using the Wilcoxon signed-rank test and the paired t-test, and Cohen's d effect size was recorded. Results: On the test set of 59 cases, the Dice, HD95, and Surface Dice @1 mm of SegResNet were 0.762 ± 0.094, 9.38 ± 7.02 mm, and 0.367 ± 0.139, respectively; whereas the corresponding indicators for the nnU-Net method with 5-fold cross-validation were 0.813 ± 0.084, 7.36 ± 6.69 mm, and 0.478 ± 0.187. The nnU-Net method was significantly better than SegResNet in terms of Dice, IoU, HD95, and Surface Dice @1 mm (p < 0.001).The Cohen's d for Dice and IoU were 0.78 and 0.80, respectively, indicating a "large effect". Among the 59 patients in the test set, the nnU-Net method outperformed in 48 cases, with 11 cases slightly lower, and a maximum regression of 0.107. Conclusions: Based on 388 cases of single-center data, the nnU-Net method can stably provide automatic delineation results for HR-CTV with a Dice score of approximately 0.81, an HD95 of about 7 mm, and a Surface Dice @ 2 mm of around 0.59. It has obvious and clinically meaningful advantages over the commonly used strong baseline of SegResNet for pelvic segmentation (Dice +5%, HD95 reduced by 2 mm, |d| ≥ 0.5).