2887 - Evaluation of Deep-Learning-Based Commercial Contouring Software for CT-Based Gynecological Brachytherapy
Presenter(s)
H. J. Yang1, J. Patrick2, J. Vickress2, D. P. D'Souza2, V. Velker3, L. Mendez2, M. T. M. Starling2, A. Fenster4, and D. Hoover2; 1Schulich School of Medicine and Dentistry, London, ON, Canada, 2London Health Sciences Centre, London, ON, Canada, 3Department of Radiation Oncology, Western University, London, ON, Canada, 4Robarts Research Institute, London, ON, Canada
Purpose/Objective(s): High dose rate (HDR) brachytherapy (BT) is important for managing gynecological cancers. As part of BT planning, surrounding organs at risk must be segmented, which is time-consuming and prone to interobserver variability. AI-based segmentation tools have previously shown enhanced efficiency for external beam radiotherapy; however, this can be more challenging in BT due to the presence of applicators. A commercial deep-learning-based software has recently been released for auto-segmentation in HDR gynecological BT. We hypothesize that this AI tool will have an acceptable performance at contouring organs at risk in CT-based gynecological BT, as compared to the manual approach, while reducing overall contouring time.
Materials/Methods: This software uses deep convolution neural network models trained from a multi-institutional dataset with a variety of BT applicators. We collected CT images from 137 patients treated with BT (19.5-28 Gy in 3-4 fractions) at our institution from January 2018 to December 2022. Of these, 107 de-identified patients were provided to the software company to be included in the training dataset, while 30 patients were held back to test model performance. Clinical and AI contours for bladder, bowel, rectum, and sigmoid were obtained. Five patients were randomly selected from the test set and manually re-contoured retrospectively by four radiation oncologists. Contouring was then repeated 2 weeks later with the AI contours as the starting point (“AI-assisted” approach). Comparisons amongst clinical, AI, AI-assisted, and manual retrospective contours were made using Dice similarity coefficient (DSC), mean distance to agreement (MDA), and unsigned D2cc difference. Two-tailed t test and Wilcoxon signed-rank test were used with statistical significance defined at a p-value of 0.05.
Results: Between clinical and AI contours, DSC was 0.92, 0.78, 0.62, 0.66, while MDA was 1.05, 1.98, 8.5, 4.9 mm for bladder, rectum, sigmoid, and bowel, respectively. Rectum and sigmoid had the lowest mean unsigned D2cc difference of 0.21 Gy/fraction between clinical and AI contours, while bowel had the largest mean difference of 0.38 Gy/fraction. The agreement between fully automated AI and clinical contours was generally not different compared to agreement between AI-assisted and clinical contours. AI-assisted interobserver agreement was better than manual interobserver agreement for all organs and metrics. The mean time to contour all organs for manual and AI-assisted approaches was 14.1±4.6 and 7.7±3.4 minutes per patient (p<0.001), respectively.
Conclusion: The agreement between AI or AI-assisted contours against the clinical contours was similar to manual interobserver agreement. Implementation of the AI-assisted contouring approach could enhance clinical workflow by decreasing both contouring time and interobserver variability.