Room: AAPM ePoster Library
Purpose: Many publications have shown that deep learning can be used to generate clinically acceptable plans for radiation therapy treatment. Our previous work showed that for prostate and head and neck patients, plans could be generated within 5 percent of clinically delivered plans. Furthermore, we developed a DVH based loss function that could be used to directly map both the desired dose distribution and the DVH curves to the predicted plan. However, there is no methodology available to enforce hard constraints on a trained deep learning model. Enforcing constraints such as the maximum allowable hot-spot or maximum dose allowed to a specific organ at risk were not possible. We have developed a novel training method that can incorporate multiple hard constraints with user specified priority.
Methods: Our robust in-house model based on a HD Connected U-net was used as the foundational for this work. A custom loss function was written to approximate a hard dose constraint in the training process. This loss function allows the user to select the desired structure and assign a priority to the hard constraint. The model was then trained on 220, validated on 40 and tested on 40 VMAT head and neck patient plans.
Results: Our model was able to enforce the desired hard constraint without negatively effecting the remaining predicted dose distribution. This methodology will allow for our in-house deep learning models to be modified to meet physician and patient specific treatment goals. Percentage of voxels compliant for the PTV constraint went from 97.9 to 98.9% and compliance of the oral cavity dose went from 74.3 to 95.2%.
Conclusion: This methodology will allows for our in-house deep learning models to be modified to meet patient specific treatment goals, while still harnessing the power, speed and accuracy of a deep learning based dose distribution prediction.