Room: AAPM ePoster Library
Purpose: The quantity of time spent fine-tuning simple dose distribution calculations between dosimetrists and radiation oncologists has been identified as a consistent point of delay in our clinic's patient throughput. As such, this project was initiated to introduce a deep learning-based dose calculation tool for radiation oncologists to employ without the need for dosimetrists' input. The final product will be able to automatically generate high-quality treatment plans that are capable of meeting departmental standards. Starting with the automation of the simplest treatment plans may save substantial amounts of time, and also allow the radiation oncologists and dosimetrists to apply their efforts towards more complex clinical tasks that demand their full expertise.
Methods: Repurposing an open-source convolutional neural net, U-Net, we have created a model-generating algorithm designed to be trained on a cohort of 35 Whole Brain Irradiation (WBI) CT DICOM-RT anonymized patient datasets. It is capable of predicting a dose distribution using only a WBI scan with contoured PTV and OAR's of cord, L/R lens, and L/R eye as input. Generated dose distributions will be analyzed via paired Dice similarity coefficient (DSC) and Average Hausdorff Distance tests for agreement with their parent plans.
Results: As similar projects indicate an expected 0.9 DSC agreement between predicted and true isodose volumes, our algorithm is anticipated to achieve a comparable level of success.
Conclusion: We have found success in designing a novel pseudorandom slice selection component subprocess that increases the net “training value per slice" during model generation, partially compensating for our relatively slow processing capabilities (due to insufficient computational equipment and data scarcity) that have considerably hampered the project's overall progress. Once our results are generated and analyzed, our framework will be expanded to address a wide variety of clinical sites in the near future.