2009 - Artifact Suppression and Soft-Tissue Enhancement in Synthetic CT for Thorax CBCT Images Using an Enhanced CycleGAN Model with Expanded Receptive Field

Purpose/Objective(s): High-quality synthetic CT images (sCT) generated from CBCT hold great potential for daily dose calculation in online adaptive radiotherapy (ART). However, hallucinations and metal artifacts continue to limit the applicability of sCT to a wide range of patients. The aim of this study was to evaluate image quality and dosimetric performance of an enhanced neural network-assisted approach to generating sCT images.

Materials/Methods: A modified Cycle Generative Adversarial Network (CycleGAN) model with additional constraints trained on over 80 paired CBCT thoracic images was utilized to generate sCT from kV CBCT images acquired on a precision radiation medicine company Versa HD and Infinity Linacs. To evaluate the model, ten independent thoracic datasets were considered, including cases with metal implants, aorta calcification, and pacemakers. Our evaluation included: (1) detailed qualitative assessment focusing on artifact presence and hallucinations, (2) quantitative analysis of Hounsfield Unit (HU) values by comparing sCT images with planning CT (RefCT) images, (3) Monte Carlo-based dose calculation followed by 3D global gamma analysis using 3%/2mm criteria, and (4) clinical evaluation of the utility of sCT images by analyzing the performance of organ-at-risk auto-segmentation using commercial software. Contour quality was compared to the ground truth contours generated based on RefCT contours using Dice similarity coefficients (DSC) and mean distance to agreement (MDA).

Results: The enhanced sCT model produced images with improved soft tissue contrast and reduced isolated bone artifacts. The absolute HU differences between sCT and RefCT images were: aorta (7.2 HU), bone (14.5 HU), esophagus (45.3 HU), heart (4.3 HU), L lung (17.6 HU), R lung (21.5 HU), spinal cord (16.8 HU), and trachea (39.3 HU). Auto-segmentation performance was robust for the heart (DSC: 0.91, MDA: 2.4 mm), lungs (DSC: 0.98, MDA: 5 mm), spinal cord (DSC: 0.92, MDA: 0.4 mm), and trachea (DSC: 0.92, MDA: 0.5 mm). Aorta (DSC: 0.63, MDA: 7.8 mm) and esophagus (DSC: 0.65, MDA: 6.6 mm) achieved limited autosegmentation performance due to poor visualization on clinical CBCT images, which translated to sub-optimal enhancement on sCT. The mean gamma passing rate was 99.8%.

Conclusion: The enhanced sCT model improved image quality by reducing metal artifacts and bone hallucinations while enhancing soft-tissue contrast. This advancement holds potential to support adaptive replanning on conventional c-arm Linacs.