A Novel Method Combining Global and Local Assessments to Evaluate CBCT-Based Synthetic CTs

Deep learning models are increasingly used to generate synthetic images. Synthetic CTs (sCTs) generated from on-treatment cone-beam CTs (CBCTs) hold potential for adaptive radiotherapy, promising a high-quality representation of daily anatomy without requiring additional imaging or dose to the patient. However, validating sCT is very challenging as an accurate and appropriate ground truth is hard to come by in medical imaging. Current global metrics in the literature fail to provide a complete picture of how accurate synthetic images are. We introduce a novel method to evaluate sCTs utilising global error assessment and a local, voxel-wise statistical assessment of the sCT and the current ground truth, a deformably registered CT (dCT). Our methodology allows for the identification of individual cases where the sCT might offer an improved representation of the daily anatomy due to changes that occur over time, as well as showing regions where either the model or image registration under-performs. Our methodology can be used to guide future model development to improve the mapping between modalities, and also assist in deciphering when it is most appropriate to choose a sCT for image guided radiotherapy over the existing standard, the dCT.