Room: Track 3
Purpose: Treatment planning in radiation oncology is a time-consuming process due to its iterative nature. In this work, we apply a deep learning method to predict optimal dose distributions for prostate cancer volumetric modulated arc therapy (VMAT), to ultimately transfer the knowledge and expertise of major academic institutions to those with limited resources through deep learning (DL) to provide a guide and benchmark for physicians and dosimetrists to improve the quality of treatment planning.
Methods: For 203 prostate cancer patients, the clinically delivered dose distributions were used to train the DL model. From each treatment plan, we use the contours for the PTV and four OARs (bladder, rectum, and femoral heads) as input of the model. The ability for transferring knowledge from one institution to another is tested by predicting dose distributions using a new dataset obtained from an institution different from the training institution, after adapting the model through transfer learning. The new dataset consists of 60 patients, where the plans were obtained through automated treatment planning; we used only 20 (16 training, four validation) for domain adaptation, and 40 cases for the test.
Results: The predicted dose distributions are comparable to the delivered clinical plans. The mean dose deviation for the target was within 0.34%(D95), 1.89 %(D50)of the prescription dose. The mean dose difference for rectum, bladder, and femoral heads, Dmean are within 0.02%, 0.95%, and0.96%, and Dmax are within 0.24%, 0.85%, and 0.70% respectively. The average PTV-homogeneity((D2-D98)/D50) was predicted to be 0.02, which matched the ground truth.
Conclusion: We demonstrated the feasibility of using transfer learning to predict radiation dose distributions for prostate VMAT treatments with minimal training dataset for the inter-institutional setting. The predicted dose distributions could be used to improve the clinical workflow and plan quality by providing a benchmark and plan quality control tool.