Purpose: Adaptive radiation therapy requires ongoing dosimetric plan modifications during the treatment course, in response to daily anatomy changes. However, its clinical implementation for abdominal cancers is challenged by the time-consuming beam optimization and dose calculation. Here, to improve the efficiency of dosimatric plans, we aim to develop a deep-learning (DL) model of dose prediction for pancreatic stereotactic body radiation therapy (SBRT) based on segmented tumor and organs-at-risk (OARs).
Methods: The contours and clinically approved dose distributions for 80 pancreatic cancer patients were retrospectively solicited from our institutional database. The whole cohort was divided into a training set (80% patients), a validation set (10% patients) and a testing set (10% patients). A 9-channel 7-level U-net model was used in our study, as detailed in figure 1. The input to the network was the stacked contours of gross tumor volume (GTV), planning tumor volume (PTV), body, and 6 OARs (stomach, duodenum, small bowel, liver, kidneys, and spinal cord) from all training patients and the output is the corresponding dose distributions. Mean squared error and Adam algorithm were used as the loss function and optimizer in the training process.
Results: An example of predicted dose distribution on a test subject is shown in figure 2. The similarity between the original and predicted doses was quantified by Dice coefficients over 0% to 100% isodose volume. As shown in figure 3, DL predicted dose resulted in Dice coefficients of over 0.9 for the entire isodose range for the training set, and over 0.8 for the testing set.
Conclusion: We successfully implemented an U-net based dose prediction using the geometric contours of GTV and OARs. The derived model predicted dose distributions for pancreatic SBRT plans with high capacity. The proposed DL model may increase the efficiency of the future clinical routine of SBRT planning.