2266 - An Optimized Deep Learning CBCT Auto-Segmentation Method for Bladder-Filled Patients in Prostate Adaptive Radiation Therapy
Presenter(s)
Y. J. Wang1, W. W. Cheng1, C. C. Kuo1, and J. K. Wu2; 1Department of Radiation Oncology, Fu Jen Catholic University Hospital, Fu Jen Catholic University, New Taipei City, Taiwan, 2Division of Radiation Oncology, Department of Oncology, National Taiwan University Hospital, Taipei, Taiwan
Purpose/Objective(s): To introduce and evaluate a novel method integrating MATLAB-based image optimization with deep learning–based auto-segmentation of cone beam computed tomography (CBCT) images for bladder-filled patients undergoing prostate adaptive radiation therapy. This approach aims to overcome challenges of variable bladder filling and lower CBCT image quality, improving segmentation accuracy and clinical workflow efficiency.
Materials/Methods:
CBCT images were acquired using a low-dose protocol without filters and a larger field of view (FOV), often resulting in streak artifacts. Images were obtained using Truebeam STX, Edge, and Versa HD systems. CT imaging employed GE Discovery and Philips Big Bore CT systems. A MATLAB-driven pipeline smoothed and renormalized CBCT images, incorporating a near-fast non-local means algorithm for artifact reduction. This algorithm utilized 2D analysis of key regions followed by a 1D calculation process based on Euclidean distance between six defined kernels. A classical 2D U-Net deep learning model was trained on pre-processed CBCT images using the holdout method. Over 300 pre-processed CBCTs and >60 simulation CTs were used for training, with femoral heads as internal quality-control benchmarks. Performance was evaluated using various metrics including Dice Similarity Coefficient (DSC), Hausdorff distance (HD95), and Jaccard index. Clinically acceptable thresholds for bladder segmentation were set at DSC >0.85 and Jaccard >0.80.Results:
The MATLAB-based pre-processing pipeline significantly improved CBCT image quality, enabling high segmentation accuracy. Bladder segmentation achieved a mean DSC of 0.89 ± 0.17 (Jaccard 0.83 ± 0.17), while femoral heads showed robust agreement with mean DSC >0.95. Body contour DSC approached 0.99 ± 0.05. The model's performance was quantified through loss metrics, with bladder loss of 0.04 and femoral head loss of 0.03. Statistical analysis confirmed significant improvements in bladder segmentation compared to manual approaches and other auto-segmentation algorithms (p<0.05).Conclusion: The proposed MATLAB-optimized deep learning strategy substantially enhances CBCT segmentation accuracy for bladder-filled prostate cancer patients. By leveraging rigorous quality-control and combining large, modality-specific datasets, this method achieves clinically acceptable bladder delineations. The approach reduces operator workload, improves consistency, and shows potential for broader adoption in adaptive radiation therapy workflows. Future work will explore MRI data integration, advanced data augmentation strategies, and prospective trials to validate clinical outcomes.