Predicting Gastrointestinal Organ Motion for SBRT Treatments of Pancreatic Cancer Using a Generative Deep Learning Model
Abstract
Purpose
As pancreatic cancer treatments are trending toward ablative regimes, accounting for the motion of organs-at-risk is crucial due to potential toxicities. MR-guided online adaptative radiotherapy is effective at mitigating organ motion, however, the availability is limited. The aim of this study is to develop a generative deep learning model to predict future abdominal anatomies in aid of treatment planning strategies facilitating anatomically robust treatment planning.
Methods
The model was trained on data from 107 pancreatic cancer patients, treated with MR-guided online adaptive radiotherapy to 50 Gy in 5 fractions. Patients were split into 86 for training, including 5 for validation, and 21 for testing. Each patient had a simulation MRI and five daily MRIs. The stomach, duodenum and bowels were contoured, and physician validated on each MRI. Based on a variational autoencoder architecture, a probabilistic generative deep learning daily anatomy model was built. The model learns the main modes of motion, then given a planning image it generates deformation vector fields that warp the image into possible patient-specific future anatomies. The model performance was assessed by comparing the training vs model generated distributions of change in volume and center-of-mass (COM), by computing the Wasserstein distance between the distributions.
Results
The model generated warped MR images with associated contours, resembling plausible patient anatomies. The Wasserstein distances between the training and generated volume change distributions were 0.28, 0.38, 0.28 and 0.34 versus 0.58, 0.39, 0.60, 0.67 for the center of mass change for the stomach, duodenum, colon, and small bowel respectively.
Conclusion
Future anatomies were predicted with volume and COM changes approaching the clinically observed changes. The complexity of abdominal motion poses a challenge; however, the current model performance shows a promising start into patient-specific organ range prediction that could be used as a successful tool for anatomically robust planning.