1032 - Comparison of Enhanced Cone-Beam CT Using Next Generation Reconstruction with Deep Learning-Based Synthetic CT for Adaptive Radiotherapy in Pelvic Cancers

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): Suboptimal image quality presents a challenge for the use of cone-beam CT (CBCT) in adaptive radiotherapy (ART). It hinders accurate structure delineation and precise dose calculation. In this work, we evaluated enhanced CBCT images reconstructed using an AI-based scatter correction model and compared it with a deep learning (DL)-based synthetic CT (sCT) model for pelvic cancer patients. Materials/

Methods: Two advanced models were utilized for CBCT image enhancement: one employing deep AI-based scatter correction during CBCT reconstruction (recon CBCT), and the other leveraging a DL-based Generative Adversarial Network (CycleGAN) to produce sCT images. Ten independent datasets of pelvic RT patients (6 male, 4 female) were used for evaluation. A systematic evaluation of CBCT images generated by these models was conducted, including: 1) comparison of Hounsfield unit (HU) values to the planning CT (refCT), 2) analysis of line intensity profiles for corresponding regions across clinical CBCT, recon CBCT, sCT, and refCT images to highlight HU discrepancies, 3) assessment of organ at risks (OARs) DL-based auto-segmentation performance using commercial software, 4) evaluation of image quality metrics such as Pearson correlation coefficient (PCC) and Root Mean Square differences (RMSE) and 4) Monte Carlo-based dose calculation accuracy for 3D global gamma analysis for a criterion of 3%/3mm (> 95%) was performed by rigid registration to the refCT. In cases of large patients with truncated CBCT volumes, identical volumes on CBCT and refCT were utilized for analysis. Contour quality was compared to the refCT using Dice and Jaccard similarity coefficients and mean distance to agreement (MDA).
Results: The models were executed within 20 to 40 seconds. The HU values of various pelvis OARs delineated on recon CBCT and sCT closely matched the refCT values and fell within acceptable HU tolerances for bone and soft tissues based on published recommendations (AAPM/IAEA/ESTRO/IAPM). Line profile affirmed consistent intensity with the refCT throughout the anatomy for both the recon CBCT and sCT, confirming preserved CBCT anatomy. Additionally, comparable mean values for Dice (0.82 vs. 0.84), Jaccard (0.70 vs. 0.73), and MDA (3.05 vs. 2.41 mm) were obtained for recon CBCT and sCT, respectively, relative to refCT for bone, femoral heads, lymph nodes, and prostate, indicating robust auto-segmentation performance. While a better image quality in terms of PCC (recon CBCT vs. sCT: 0.74±0.05 vs. 0.85±0.03) and RMSE (recon CBCT vs. sCT: 554±29 vs 385±29) was observed for sCT images, both recon CBCT and sCT passed the gamma analysis (recon CBCT vs. sCT: 98±1% vs. 99±1%).
Conclusion: Enhanced reconstructed CBCT using AI based scatter correction is comparable to DL-based sCT. The improved preservation of anatomical structures, HU consistency, and dose accuracy in the pelvic region is promising for precise and efficient online ART on conventional c-arm linacs, which can eliminate or reduce the need for sCT.