Room: Stars at Night Ballroom 1
Purpose: It is desirable to predict a dose distribution as the reference for physician and dosimetrist to eliminate the iterative process in radiotherapy treatment planning. Anatomy-based model predicts a dose distribution without considering physicianâ€™s preference on organ dose trade-offs. Here we develop a deep learning based dose prediction model based on both patient anatomy and physician preferred dose volume constraints.
Methods: Dose volume constraints (DVCs) are firstly converted into dose volume histogram (DVH). Then a threshold-driven optimization for reference-based auto-planning algorithm (TORA) is utilized to optimize the treatment plan based on the DVH. The optimized dose distribution is used as the ground truth to train a U-Net with the vectorized DVH and patient organ masks as input channels.
Results: 20 IMRT prostate patients were used in this study, with 16 for training and 4 for testing. Each training patient has 20 plans of various organ dose trade-offs. The visualization of the results show that the predicted dose distributions match closely with the dose distributions generated by TORA for corresponding DVCs. The mean and max dose differences for all crucial structures are employed as a quantitative evaluation for testing. For the PTV and all OARs, the largest average error in mean dose is about 1.6% of the prescription dose, and the largest average error in maximum dose is about 1.7% of the prescription dose.
Conclusion: We have developed a deep learning model to predict dose distributions taking into account physicianâ€™s preference on organ dose trade-offs.