Evaluating the dosimetric and positioning accuracy of a deep learning based synthetic-CT generation model for liver radiotherapy.

Abstract

Background and purpose: An MRI-only workflow requires synthetic computed tomography (sCT) images to enable dose calculation. This study evaluated the dosimetric and patient positioning accuracy of deep learning-generated sCT for liver radiotherapy. Methods and materials: sCT images were generated for eleven patients using a CycleGAN algorithm. Clinical volumetric modulated arc treatment plans (VMAT) were calculated on CT and recalculated on sCT, and dose differences were assessed using dose volume histogram (DVH). For position verification, the sCT images were validated as reference images to 4D cone beam computed tomography (4D CBCT) by calculating the translational and rotational differences between sCT and CT registrations to 4D CBCT. Results: The mean dose differences for the planning target volume (PTV) and organs at risk (OAR) between the CT and sCT plans were 0.0% and < 0.5% respectively. For positioning verification, the systematic translational and rotational differences were < 0.5 mm and < 0.5° respectively in all directions Conclusion: This is the first study to validate a sCT model for liver cancer in terms of both dosimetry and patient positioning, marking a significant step in demonstrating the feasibility of an MRI-only workflow for treating liver cancer. The generated sCT showed dosimetric differences within clinically acceptable levels and were successfully used as reference images for treatment position verification. This CycleGAN model is accessible through the research version of a commercial vendor, with potential for development as a clinical solution. .