Author: Ergun E. Ahunbay, Colette Gage, Abdul Kareem Parchur, Eric S. Paulson 👨🔬
Affiliation: Department of Radiation Oncology, Medical College of Wisconsin 🌍
Purpose: AI-generated synthetic CT (sCT) images address challenges with prior sCT approaches, including atlas- and threshold-based methods. Commercial AI-based sCT tools have been introduced. This work aimed to validate a commercial AI brain sCT model for clinical use in MR-only workflows.
Methods: A retrospective analysis was conducted on five brain cancer patients (mean age 51.2±15.8) who underwent 3T MRI and CT simulation for treatment planning. A deep learning-based algorithm within the Siemens RT Image Suite VB60 was used to generate axial sCT images from sagittal 3D T1w Dixon MR images. Structural Similarity Index (SSIM) maps and mean SSIM values were evaluated between digitally reconstructed radiographs (DRRs) of sCT and CT. Additionally, patients’ portal images were automatically registered with DRRs of both sCT and CT to assess alignment differences. Hounsfield unit (HU) differences between CT and the aligned sCT were assessed using Mean Absolute Error (MAE). RT plans were recalculated on sCT using the SciMoCa™ Monte Carlo engine. Dosimetric accuracy was analyzed using 3D Gamma passing rates with varying thresholds.
Results: The mean SSIM between DRRs of sCT and CT was ~0.9, demonstrating high structural similarity. No registration differences were detected when portal images were aligned to DRRs from sCT vs reference CT. MAE values of 19.3±2.3, 12.9±1.6, 345±27, and 144.5±8.9 HU were observed for brain, brainstem, bone, and external contours, respectively. The mean 3D gamma passing rates were 96.9%±1.4%, 97.7%± 1%, 99.3%±0.9%, 99.5%±0.6%, 99.8%±0.4%, and 99.9%±0.3% for gamma criteria of 1%/2mm, 1%/3mm, 2%/2mm, 2%/3mm, 3%/2mm, and 3%/3mm, respectively. Dosimetric differences were within ~2%.
Conclusion: The AI-generated brain sCT model demonstrated performance comparable to clinical CT, with high structural similarity and acceptable dosimetric accuracy. These findings support the use of AI based sCT in clinical use in MR-only workflows for brain radiotherapy.