Deep-Learning Based Automated Contouring, Dose-Prediction, and Planning in High-Dose-Rate Prostate Brachytherapy
C V Guthier1,2*, T C Harris1, D A O'Farrell3, S A Friesen4, M T King5, I M Buzurovic6, R A Cormack7, (1) Department of Radiation Oncology, Brigham and Womens Hospital/Dana-Farber Cancer Institute and Harvard Medical School, Boston MA (2) Artificial Intelligence in Medicine (AIM) Program at Harvard-MGB
Presentations
TU-I345-IePD-F2-4 (Tuesday, 7/12/2022) 3:45 PM - 4:15 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 2
Purpose: Transrectal ultrasound-guided (TRUS) High-Dose-Rate-brachytherapy (HDR-BT) is a widely used treatment modality for prostate cancer. Contouring based on TRUS, depends on users’ experience, and prolongs the treatment planning process. In this work we developed and validated a deep learning (DL)-based approach for automated contouring, dose prediction, and treatment planning.
Methods: A U-Net with ResNet encoders was trained to segment relevant structures from TRUS. The same architecture was used to predict isodose lines. The input for this prediction is the TRUS and segmented contours. Contours and dose-predictions are used to optimize a treatment plan with an in-house planning system. 150 patients treated with HDR-BT under TRUS guidance and manually segmented structures, and treatment plans were split into training (n=110), validation (n=20), and test (n=20) groups. Geometric and dosimetric validation was performed by comparing the prediction against the manual ground truth. Treatment plans were dosimetrically evaluated by comparing metrics and dose distribution. Dose similarity was evaluated via linear regression of the vectorized dose distributions.
Results: The average segmentation and dose prediction time was (0.98±0.53)seconds. Compared to 10-30 minutes for manual segmentation. The median dice for the prostate was 0.87(IQR=0.04), for urethra 0.82(IQR=0.08), rectum 0.89(IQR=0.07), and bladder 0.63(IQR=0.24). The predicted iso-dose lines showed a dice of 0.82(IQR=0.04). Difference in contouring yield a deviation in key metrics of 0.08%(IQR=4.10%). Based on a Wilcoxon signed rank test this difference is not statistically significant (p>.05). Comparing the planning strategy that utilizes dose-prediction to guide the optimization to our conventional planning yields similar (p>.05) metrics with a median deviation of -0.0%(IQR=3.1)% and dose R2=0.94(IQR=0.12).
Conclusion: The proposed strategy returned geometrically accurate contours and dose prediction which are used during inverse treatment planning to find acceptable treatment plans without manual interaction. The DL models encapsulate expert knowledge that can easily be transferred to allow equitable access of care.
Keywords
Brachytherapy, Inverse Planning, HDR
Taxonomy
TH- Brachytherapy: Treatment planning using machine learning/automation
Contact Email