Room: Stars at Night Ballroom 1
Purpose: Treatment planning can be a time-consuming process in radiation oncology, due to the iterative nature of it. In this work we apply a deep learning method to radiation dose distribution for prostate cancer volumetric modulated arc therapy (VMAT) to assist physicians and dosimetrists reaching the desired treatment plan in an efficient matter.
Methods: We used a deep neural network to predict dose distributions for prostate VMAT treatment. Global and local features were captured by combining two state-of-art deep learning architectures. Integration of U-net and Dense net created an improved architecture called Hierarchically-Dense U-net (HD-U-net). Using HD-U-net allows us to achieve more efficient feature propagation. We applied a weighted loss function to improve OARs sparing. Treatment plans for 203 prostate cancer patients who underwent treatment using VMAT were used. The treatment plan data were grouped based on prescription dose. The data were divided into 151 training, 38 validation and 14 testing sets. From each treatment plan, we used PTVâ€™s contours and four OARs (bladder, rectum, femur heads) as input.
Results: The predicted dose distributions are comparable to the delivered clinical plans. The mean dose deviation for the target were within 0.9% (D95), 3% (D50) of the prescription dose. The mean dose difference for rectum, bladder, and femoral heads are within 3%, 4%, and 2%, respectively. The differences between dose volume histograms were negligible.
Conclusion: We demonstrated the feasibility of using deep learning to predict radiation dose distributions for prostate VMAT treatment. The predicted dose distributions could be used to improve the clinical work flow and efficiency while providing a benchmark and plan quality control.
Not Applicable / None Entered.