An nnUNet model for deep-learning based auto-segmentation of target volume in high-dose-rate brachytherapy for cervical cancer

Summary: Brachytherapy treatment planning is a labor-intensive, time-consuming process with frequent practitioner dependent differences in contouring of target volumes. Auto-segmentation may standardize target volumes while improving efficiency in the planning workflow. The purpose of this project is to create and evaluate a nnUNet model for the auto-segmentation of high-risk CTV (HRCTV) in interstitial, tandem-and-ring (T&R), and tandem-and-ovoid (T&O) high-dose-rate (HDR) cervical brachytherapy treatment planning. Here we test the feasibility of training this model with a limited patient dataset. The dataset used in this study detailed 30 cervical cancer patients previously treated with HDR brachytherapy at our institution including planning computed tomography (CT) scans and organ-at-risk (OAR) contours of the bladder, urethra, small bowel, and sigmoid colon. nnUNet is an auto-segmentation framework which has become the benchmark for auto-contouring tasks and is able to train 2d and 3d U-Net configurations. Provided the CT and OAR contours as input, the model outputted the HRCTV. The training set included 19 cases (7 interstitial, 6 T&R, 6 T&O), and the test set included 11 cases (6 interstitial, 4 T&R, 1 T&O). Five-fold cross-validation was used. Models with 2d and 3d full resolution (3DFR) U-Net configurations were trained. The output of the model was compared to clinically approved contours used when treating these cases by calculating the Dice similarity coefficient (DSC) and mean surface distance (MSD), two commonly used metrics when determining the difference between similar contours. Our best model was the 3DFR configuration of nnUNet which showed 0.78 average DSC +/- 0.08 STDDEV, and 2.8 mm average MSD +/- 1.5 mm STDDEV on our test set. Despite a small number of cases, nnUNet is able to create a promising starting point in autosegmentation of HRCTV in brachytherapy for cervical cancer. Future studies with increased training data are underway based on these results. Samuel L. Chen (he/him/his), MD (Presenting Author) - UAB Department of Radiation Oncology; Roman Travis, MD (Co-Author) - University of Alabama at Birmingham Department of Radiation Oncology; Eric Simiele, Ph.D. (Co-Author) - University of Alabama at Birmingham Department of Radiation Oncology; Samantha Simiele, Ph.D., DABR (Co-Author) - University of Alabama Department of Radiation Oncology; Sui Shen (he/him/his), Ph.D. (Co-Author) - UAB Department of Radiation Oncology; Samuel Marcrom, MD (Co-Author) - UAB Radiation Oncology; Carlos Cardenas (he/him/his), Associate Professor (Co-Author) - UAB Department of Radiation Oncology