How to Evaluate Generated skVCT from MVCT? Optimal Parameters and Clinical Validation

This study investigates the correlation of computer vision metrics with the clinical utility of domain-transferred synthetic medical images. Here, we use a Cycle-Consistency Generative Adversarial Network using softly paired data to create synthetic kilovoltage computed tomography (CT) from megavoltage computed tomography images, leveraging a unique database of 1109 head-and-neck CTs, an order of magnitude larger than previous works. The quality of synthetic images was quantified using standard metrics such as Structural Similarity, which we used to identify the best models in the clinical context of adaptative radiation therapy. Using the framework defined here, we identify the optimal parameter values of widely used metrics and validate our findings through a clinical evaluation using a panel of nine radiation oncologists who evaluated 4320 images.