Self-configuring nnU-Net for automatic delineation of the organs at risk and target in high-dose rate cervical brachytherapy, a low/middle-income country's experience

Abstract

Background: The high-dose rate (HDR) brachytherapy treatment planning workflow for cervical cancer is a labor-intensive, time-consuming, and expertise-driven process. These issues are amplified in low/middle-income countries with large deficits in experienced healthcare professionals. Automation has the ability to substantially reduce bottlenecks in the planning process but often require a high level of expertise to develop. Purpose: To implement the out of the box self-configuring nnU-Net package for the auto-segmentation of the organs at risk (OARs) and high-risk CTV (HR CTV) for Ring-Tandem (R-T) HDR cervical brachytherapy treatment planning. Methods: The computed tomography (CT) scans of 100 previously treated patients were used to train and test three different nnU-Net configurations (2D, 3DFR, and 3DCasc). The performance of the models was evaluated by calculating the Sørensen-dice similarity coefficient, Hausdorff distance (HD), 95th percentile Hausdorff distance, mean surface distance (MSD), and precision score for 20 test patients. The dosimetric accuracy between the manual and predicted contours was assessed by looking at the various dose volume histogram (DVH) parameters and volume differences. Three different radiation oncologists (ROs) scored the predicted bladder, rectum, and HR CTV contours generated by the best performing model. The manual contouring, prediction, and editing times were recorded. Results: The mean DSC, HD, HD95, MSD and precision scores for our best performing model (3DFR) were 0.92/7.5 mm/3.0 mm/ 0.8 mm/0.91 for the bladder, 0.84/13.8 mm/5.3 mm/1.4 mm/0.84 for the rectum, and 0.81/8.5 mm/6.0 mm/2.2 mm/0.80 for the HR CTV. Mean dose differences (D2cc/90%) and volume differences were 0.08 Gy/1.3 cm3 for the bladder, 0.02 Gy/0.7 cm3 for the rectum, and 0.33 Gy/1.5 cm3 for the HR CTV. On average, 65% of the generated contours were clinically acceptable, 33% requiring minor edits, 2% required major edits, and no contours were rejected. Average manual contouring time was 14.0 min, while the average prediction and editing times were 1.6 and 2.1 min, respectively. Conclusion: Our best performing model (3DFR) provided fast accurate auto generated OARs and HR CTV contours with a large clinical acceptance rate.