2725 - Enhanced Deep Learning Segmentation for MRI-Guided HDR Brachytherapy in Cervical Cancer
Presenter(s)
W. Cho1, C. Kim2, J. H. Jeong3, S. H. Seo1,4, H. Park1, and J. S. Kim5; 1Oncosoft Inc., Seoul, Korea, Republic of (South), 2National Cancer Center, Goyang, Gyeonggi, Korea, Republic of (South), 3Proton Therapy Center, National Cancer Center, Goyang, Korea, Republic of (South), 4Department of Clinical Medicine, Sungkyunkwan University School of Medicine, Suwon, Korea, Republic of (South), 5Oncosoft, Seoul, Korea, Republic of (South)
Purpose/Objective(s): We propose a newly designed segmentation model that substantially improves the accuracy and efficiency of contouring target volumes and organs at risk (OARs) in MRI-guided HDR brachytherapy for cervical cancer. Our aim is to automate the segmentation of the gross tumor volume (GTV), high-risk clinical target volume (HR-CTV), intermediate-risk clinical target volume (IR-CTV), and key OARs (anus, bladder, bowel, urethra, sigmoid, and rectum) to standardize contours and optimize clinical workflow.
Materials/Methods: We retrospectively analyzed T2-weighted MRI scans from 158 cervical cancer patients (2000–2023). Physician-delineated multilabel contours served as ground truth. Our model, based on nnU-Net, was modified from a multiclass to a multilabel paradigm with architecture and loss function adjustments. Enhanced pre- and post-processing improved performance. The model was trained using a batch-wise Adam optimizer with a linearly decaying learning rate. Data augmentation, including rotations, flips, and intensity variations, improved generalizability, while a smart ensemble combined predictions from multiple model variants. Performance was evaluated using Dice Similarity Coefficient (DSC), Mean Surface Distance (MSD), and 95th percentile Hausdorff Distance (HD95), reported as median (Q1–Q3).
Results: DSC for each contour was as follows: Anus 0.648 (0.550–0.754), Bladder 0.967 (0.951–0.974), Bowel 0.805 (0.772–0.846), GTV 0.530 (0.348–0.709), HR-CTV 0.757 (0.656–0.820), IR-CTV 0.807 (0.719–0.870), Rectum 0.786 (0.706–0.849), Sigmoid 0.666 (0.533–0.767), and Urethra 0.804 (0.752–0.840). MSD for each contour was: Anus 1.304 (0.721–2.123), Bladder 0.436 (0.287–0.868), Bowel 1.478 (0.992–2.714), GTV 1.999 (1.281–4.445), HR-CTV 1.422 (0.911–2.304), IR-CTV 1.555 (0.953–2.353), Rectum 1.026 (0.812–1.912), Sigmoid 3.953 (2.060–9.428), and Urethra 0.549 (0.374–0.808). HD95 for each contour was: Anus 5.651 (4.000–9.933), Bladder 1.414 (1.000–3.182), Bowel 10.368 (5.358–25.019), GTV 10.124 (5.074–16.028), HR-CTV 4.791 (3.122–9.894), IR-CTV 5.657 (3.000–9.499), Rectum 5.702 (4.000–10.250), Sigmoid 24.956 (11.947–69.541), and Urethra 2.236 (2.000–3.273).
Conclusion: This enhanced multilabel deep learning model significantly improves contouring precision and efficiency in MRI-guided HDR brachytherapy planning for cervical cancer. By standardizing delineations and reducing variability, the approach not only optimizes treatment planning but also holds promise for improving patient outcomes. Future multi-institutional prospective studies are warranted to validate these findings and support broader clinical implementation.