Author: Silambarasan Anbumani, Nicolette O'Connell, Eenas A. Omari, Amanda Pan, Eric S. Paulson, Lindsay Puckett, Monica E. Shukla, Dan Thill, Jiaofeng Xu 👨🔬
Affiliation: Elekta Inc, Elekta Limited, Linac House, Department of Radiation Oncology, Medical College of Wisconsin 🌍
Purpose: Accurate electron density information from on-board imaging is essential for direct dose calculations in adaptive radiotherapy (ART). This study evaluates a deep learning model for thoracic synthetic CT (sCT) generation from CBCT, assessing HU and dose calculation accuracy, and auto-segmentation performance.
Methods: An enhanced Cycle Generative Adversarial Network (CycleGAN) model trained on over 80 paired CBCT thoracic images from two institutions was implemented in ADMIRE (Elekta, Stockholm). A retrospective study was conducted on twenty lung cancer patients using planning CT (refCT) and first-fraction CBCT images obtained from an Elekta Versa HD XVI system. Organs at risk (OAR), including aorta, esophagus, heart, lungs, spinal cord, and trachea, were contoured for each patient and utilized for refCT and sCT HU comparison. Auto-segmentation accuracy was assessed using Dice similarity coefficients(DSC) and mean distance to agreement(MDA). The performance was then compared to that of the refCT using a Wilcoxon signed-rank test (p<0.05). Dose distributions were compared using 3D global gamma analysis with 3%/2mm criteria.
Results: The absolute mean HU differences between sCT and refCT for aorta(10.1HU), bone(20.3HU), esophagus(40.8HU), heart(9.7HU), lung_left(19.1HU), lung_right(26.7HU), spinal cord(7.6HU), and trachea(43.8HU) were all within acceptable published tolerances. The mean dose distribution agreement was 99.1±1%. Auto-segmentation performance (DSC, MDA) for the sCT and refCT images were comparable, except for the esophagus and spinal cord. The significant difference in the esophagus and spinal cord were due to outliers stemming from poor CBCT image quality with observed artifacts.
Conclusion: Thoracic sCT images generated from daily CBCT through the cycleGAN model were comparable to refCT in terms of autosegmentation performance, and HU and dose calculation accuracy. This is a major step toward the use of sCT in adaptive radiotherapy for thoracic cancer patients. Future work will aim toward enhancing the sCT model to reduce the effects of CBCT artifacts on sCT image quality.