S. Anbumani1, E. S. Paulson1, J. Xu2, A. Pan2, D. Thill2, N. OConnell2, L. Puckett1, M. E. Shukla1, and E. A. Omari1; 1Department of Radiation Oncology, Medical College of Wisconsin, Milwaukee, WI, 2Elekta Limited, Linac House, Crawley, West Sussex, United Kingdom
Purpose/Objective(s): Despite several advantages of cone-beam computed tomography (CBCT) in image guided radiation therapy, its application in adaptive radiotherapy has been limited. Current CBCT image quality hinders accurate dose calculation and structure delineation. In this work, we evaluated an improved thoracic synthetic CT (sCT) deep learning model for CBCT images. Materials/
Methods: A enhanced Cycle Generative Adversarial Network (CycleGAN) model with structure constraints model trained on 80 thoracic datasets from two institutions was utilized to generate sCT images from CBCT. To enhance accuracy, the model incorporated a structural similarity index map (SSIM)-weighted L1-loss feature, facilitating pixel-level precision. Seven independent datasets for lung cancer patients were used to evaluate the model. Hounsfield unit (HU) comparisons on various organs at risk (OARs) between sCT and planning CT (ref CT) were examined to ensure dose accuracy within 1% based on established guidelines. An FDA cleared commercially available deep learning auto-segmentation algorithm (Contour Protégé AI, MIM® software) was used for structure delineation. Contour quality was assessed using Dice and Jaccard similarity coefficients, and mean distance to agreement (MDA). Dose was calculated using a Monte Carlo based algorithm and 3D global gamma analysis was performed on sCT dose against the original planned refCT dose with agreement indices of 2%/2mm (>90%) and 3%/3mm (>95%). Results: The model was executed in a research tool from a precision radiation medicine company installed on an Intel® Xeon® Gold 6132 Processor (2.6 GHz, 128 GB RAM) and took less than 30 seconds to generate sCT images. The median ±standard deviation HU values between the sCT and ref CT for thoracic structures were as follows: Aorta (5 ±1), Bone (10 ±13), Esophagus (7 ±3), heart (5 ±4), left lung (19 ±23), right lung (5±27) and spinal cord (6 ±1). The HU differences were within acceptable published tolerances, for instances the HU recommendations are ±50, ±170 and ±50 for lung, bone, and soft tissues, respectively as per European Society of Radiotherapy and Oncology (ESTRO). Mean ±standard deviation Dice, Jaccard, and MDA between the refCT and sCT were (0.84±0.09), (0.73±0.13), and (2.45±0.88) respectively, demonstrating robust auto-segmentation performance. All the cases passed the gamma analysis with mean gamma passing rates of 96±2% and 99±1% for 2%/2mm and 3%/3mm, respectively. Conclusion: The enhanced cycleGAN-based model generated high resolution sCT from CBCT, enabling accurate structure delineation and dose calculations. Improved sCT is critical for online adaptive radiotherapy and can eliminate the need for patient re-simulation for offline adaptive radiotherapy.
