Author: Mojtaba Behzadipour, Suman Gautam, Tianjun Ma, Ikchit Singh Sangha, Bongyong Song, William Song, Kumari Sunidhi 👨🔬
Affiliation: UC San Diego, Virginia Commonwealth University 🌍
Purpose: This study aims to develop a knowledge-based voxel-wise dose prediction system using a convolutional neural network (CNN) for high-dose-rate (HDR) prostate brachytherapy and to evaluate its predictive performance.
Methods: A 3D CNN U-NET model was employed to predict voxel-wise dose distributions based on input features including organs-at-risk (OARs), prostate, planning target volume (PTV), and source dwell positions. A dataset comprising 219 retrospective HDR prostate brachytherapy plans, treated with prescription doses ranging from 13.5 Gy per fraction, was utilized. The dataset was partitioned into training (153 cases), validation (38 cases), and testing (28 cases) cohorts. Binary masks of anatomical structures were consolidated into a single mask and used in conjunction with a PTV distance map as input. Data augmentation was applied to the training and validation sets to expand the dataset by a factor of four. Model performance was assessed using Dose-Volume Histogram (DVH) metrics relevant to brachytherapy plan quality (PTV D95%, urethra D0.1CC, bladder D2cc, rectum D2cc) with ΔDx = Dx-actual −Dx-predicted mean, standard deviation. Voxel-wise dose differences, including mean and standard deviation, within the dose range of 20–120% of the prescription dose were also evaluated.
Results: The predicted relative DVH metrics in the test set demonstrated strong agreement with the actual values: PTV mean ΔD95%±σ = -0.08±0.6 Gy, Urethra mean ΔD0.1CC±σ = 0.14±0.56 Gy, Rectum mean ΔD2CC±σ = -0.02±0.82 Gy, Bladder mean ΔD2CC±σ = -0.09±0.55. The mean absolute difference between predicted and actual dose distribution (mean ΔD ± σ) was found to be 0.26 ± 0.08 Gy.
Conclusion: The proposed 3D U-NET model demonstrates robust capability in predicting dose distributions and corresponding DVH metrics for specific anatomies and source dwell positions. These predictions can facilitate plan quality verification and assist in brachytherapy treatment planning.