Author: Casey E. Bojechko, Lance C Moore 👨🔬
Affiliation: University of California, San Diego, University of California San Diego 🌍
Purpose: EPID images collected during treatment can serve as an in-vivo error detection mechanism. Previous works have shown that comparing in-vivo EPID images to AI-predicted EPID images for IMRT plans are able to detect errors that would otherwise be missed during normal patient specific QA. However, this method had not yet been extended to VMAT treatment plans. In this work, we present the first large-scale EPID image prediction model for VMAT radiotherapy using only pretreatment cone beam data and plan parameters.
Methods: The dataset consisted of 203,281 EPID images (70/15/15% train/validation/test) for VMAT treatment plans, with corresponding CBCTs from 2902 unique patients from our treatment center. These data were used train a bespoke transformer model based on the DiT model architecture. As a proof of model feasibility, we represent the continuous VMAT arc by sampling every 10 control points, generating masks to represent the MLC positions at that control point. For each of these control points, the on-treatment CBCT was registered to the planning CT and rotated around the isocenter to the required gantry angle, before a projection was used to project the rotated 3D image to a 2D plane. The masks and projected CBCTs were then encoded using a pre-trained variational auto-encoder to compress the images into latent space.
Results: Our results on the test set (>30,000 images) indicate that the model is reasonably accurate with an overall MAE of 4.81(+/- 3.24% maximum), and an average gamma pass rate of 60/79(+/-14/14%) for 3%/3mm and 5%/5mm respectively.
Conclusion: The results from this preliminary work suggest that this approach can be used to predict EPID images for VMAT radiotherapy plans with sufficient data. Future work will investigate refining the accuracy of the prediction model, with the end goal of clinical utilization of the model for error detection.