3752 - Two Stages Auto-Delineation of Cervical Cancer Radiotherapy Areas Using a Deep Learning-Based Approach
Presenter(s)
W. C. You1, C. Y. Lai2, Y. T. Wu2, K. L. Yang3, C. W. Jao2, C. Y. Lin2, Y. H. Lin4, Y. Y. Hsu4, M. S. Chi3, W. K. Lee5, C. C. Wen3, C. H. Hsu3, and Y. F. Lu6; 1Department of Post-Baccalaureate Medicine, National Chung Hsing University, Taichug City, Taiwan, 2Institute of Biophotonics, National Yang Ming Chiao Tung University, Taipei City, Taiwan, 3Department of Radiation Therapy and Oncology, Shin Kong Wu Ho-Su Memorial Hospital, Taipei, Taiwan, 4Department of Radiation Oncology, Taichung Veterans General Hospital, Taichung City, Taiwan, 5Brain Research Center, National Yang Ming Chiao Tung University, Taipei City, Taiwan, 6Department of Radiation Oncology, Taichung Veterans General Hospital, Taichung, Taiwan, Taichung City, Taiwan
Purpose/Objective(s): Radiotherapy is a crucial treatment modality for cervical cancer, utilizing ionizing radiation to control tumor progression. Accurate delineation of clinical target volumes (CTVs) and organs at risk (OARs) is essential to ensure tumor control while minimizing radiation-induced toxicity to healthy tissues. This study leverages deep learning algorithms to achieve automatic delineation of CTVs and OARs, incorporating anatomical information to enhance segmentation accuracy in cervical cancer radiotherapy planning.
Materials/Methods: This retrospective study utilized computed tomography (CT) scans from 100 cervical cancer patients treated between 2020 and 2023. The CTV and OAR annotations were manually delineated by experienced radiation oncologists and medical physicists. The CTVs in this study include the pelvic lymph nodes, common iliac lymph nodes, external iliac lymph nodes, internal iliac lymph nodes, presacral lymph nodes, obturator lymph nodes, and the uterus, while the OARs include the bladder and rectum. We adopted a two-stage deep learning strategy. The first stage is the 3D U-Net model was employed to detect the lowest point of the left femoral head to define the volume of interest (VOI), which corresponds to the pelvic cavity. Then we used the 3D U-Net for the auto-segmentation of CTVs and OARs based on the VOI area from the first stage in the second stage. Both models were optimized using the Dice loss function, trained for 500 epochs, and employed the Adam optimizer with a learning rate of . The quantitative evaluation tools for the segmentation results used in this study were the Dice similarity coefficient (DSC), precision, and the 95th percentile Hausdorff distance (HD95).
Results: Quantitative evaluation of the first stage confirmed that the lowest detected point of the left femoral head was identical to the ground truth annotation, achieving a perfect 100% overlap in the test dataset. In the second stage, the 3D U-Net segmentation model achieved an average DSC of 0.80 for the CTVs, with an average recall of 0.77, precision of 0.86, and HD95 of 2.27 mm. For the OARs, the model achieved an average DSC of 0.86, recall of 0.82, precision of 0.91, and HD95 of 1.54 mm.
Conclusion: This study demonstrates the feasibility of a two-stage deep learning approach for the automated delineation of cervical cancer radiotherapy areas. The proposed method effectively identifies the volume of interest (VOI), lowers computational demands and memory constraints, and improves the consistency and accuracy of clinical target volume (CTV) and organ at risk (OAR) segmentation.