Three-dimensional radiotherapy dose prediction in patients with liver and pancreas cancer treated with the Viewray system.

Abstract

Listen

Translate article

612 Background: The radiotherapy (RT) planning process for patients with gastrointestinal tract cancers is regarded as a time-consuming and complicated task due to proximity of multiple organs at risk (OARs). Dose prediction model based on deep learning might be a solution to increase efficiency by reducing the time required to find the optimal plan. Dose distribution model was developed using a three-dimensional (3D) U-Net deep learning. Methods: Altogether RT plan of patients with liver cancer (N=40) and pancreatic cancer (N=26) were acquired from the MRIdian (Viewray Technologies, Oakwood Village, OH) treatment planning system from 2016 to 2018. Magnetic resonance (MR) images for RT, planning dose, and contoured RT structures were retrospectively collected. The 3D U-Net was employed, and transfer-learning was performed on the training set (N=60, 91.3%). The initial learning rate was 0.1, decreasing by 10% every 5 epochs, for a total of 200 epochs. A developed model was evaluated by the testing set (N=5, 7.7%). The difference between ground truth and predicted dose was compared by calculating normalized minimum, maximum, and mean doses of the planning target volume (PTV) and OARs. Results: Root-mean-square error (RMSE) of voxels between ground truth and predictive dose was 3.04 (range 2.29-3.79). Percentage of mean dose difference and standard deviation between clinical and prediction models were 1.75 ± 2.84%, 0.08 ± 0.11%, 0.13 ± 0.08%, 0.05 ± 0.03%, 0.07 ± 0.05% for PTV, duodenum, liver, kidney, and stomach, respectively. And the percentage of max dose difference between the two were 0.91 ± 0.01%, 0.34 ± 0.43%, 0.63 ± 0.39%, 0.25 ± 0.18%, 0.33 ± 0.30%, respectively. Overall, the predicted normalized target DVH metrics were within 2% of the ground truth plans. Conclusions: We developed a deep learning model for dose prediction in liver and pancreatic cancer using MR images acquired from MRIdian system. Larger-scale training and validation are required to prove the efficacy of the dose prediction model by the 3D U-Net.