3749 - Automatic Segmentation of Multi-Class Clinical Target Volumes in Prostate Cancer Using a 3D U-Net Model
Presenter(s)
C. Y. Lai1, Y. T. Wu1, W. C. You2, Y. N. Chang1, C. W. Jao1, C. Y. Lin1, Y. H. Lin3, Y. Y. Hsu3, M. S. Chi4, W. K. Lee5, C. C. Wen4, C. H. Hsu4, and K. L. Yang4,6; 1Institute of Biophotonics, National Yang Ming Chiao Tung University, Taipei City, Taiwan, 2Department of Radiation Oncology, Taichung Veterans General Hospital, Taichung, Taiwan, 3Department of Radiation Oncology, Taichung Veterans General Hospital, Taichung City, Taiwan, 4Department of Radiation Therapy and Oncology, Shin Kong Wu Ho-Su Memorial Hospital, Taipei, Taiwan, 5Brain Research Center, National Yang Ming Chiao Tung University, Taipei City, Taiwan, 6School of Medicine, Fu Jen Catholic University, New Taipei, Taiwan
Purpose/Objective(s):
Radiotherapy is a primary treatment for prostate cancer. Based on tumor location and invasion characteristics, prostate cancer is classified into different clinical target volume (CTV) delineations. This study aims to develop automatic segmentation models for three distinct CTV delineations in prostate cancer.Materials/Methods:
This retrospective study included 217 prostate cancer patients who underwent computed tomography (CT) imaging between 2013 and 2023. Patients were categorized into three groups based on CTV delineation: (1) prostate with proximal seminal vesicle, (2) prostate with the entire seminal vesicle, and (3) prostate bed (post-surgical). The ground truth for each patient was manually delineated by trained physicians. Three separate 3D U-Net models were trained for the automatic segmentation of these regions. Patients were randomly divided into training and test sets in an 8:2 ratio. The models were optimized using the Dice loss function, trained for 500 epochs, and employed the Adam optimizer with a learning rate of . Performance was evaluated using the Dice similarity coefficient (DSC), recall, precision, and the 95th percentile Hausdorff distance (HD95).Results:
Our model achieved a DSC and standard deviations (SD) of 0.81±0.08, recall of 0.86±0.12, precision of 0.79±0.12, and HD95 of 9.11±4.03 for delineation prostate with entire seminal vesicle. For the prostate with proximal seminal vesicle, the model achieved a DSC and SD of 0.72±0.14, recall of 0.67±0.21, precision of 0.84±0.07, and HD95 of 11.17±5.93. For the prostate bed, the model achieved a DSC and SD of 0.75±0.08, recall of 0.69±0.11, precision of 0.82±0.08, and HD95 of 3.72±3.33.Conclusion:
Our study demonstrates the feasibility of automated CTV delineation using a 3D U-Net model, achieving promising segmentation accuracy across different CTV types. These findings suggest the potential for clinical implementation to enhance radiotherapy planning for prostate cancer patients.