Author: Mahya Ahmadzadeh, Nagarajan Kandasamy, Keyur Shah, Gregory C. Sharp, Santhosh Vadivel, John MacLaren Walsh 👨🔬
Affiliation: Electrical and Computer Engineering Department, Massachusetts General Hospital, Emory University, Drexel University 🌍
Purpose: In image-guided radiotherapy (IGRT), cone beam CTs (CBCTs) suffer from distortions that degrade registration with planning CTs. While CycleGANs can generate synthetic CTs (sCTs) from CBCTs, existing approaches hallucinate artifacts that limit their clinical utility.
This research assesses the sCT image quality benefits of skin-boundary masking, a replay-buffer, and early-stopping enhancements to the CycleGAN training process.
Methods: Three CBCT to sCT CycleGAN translation approaches are compared: (1) a baseline without masking, (2) mask based patch-weighting, and (3) a masks-enforcing method. Patch-weighting converts skin contours into weights that scale discriminator losses, while mask-enforcing applies skin contours to remove out-of-body image components at both the input and output of the generators. Training of the mask-enforcing CycleGAN also introduces a replay buffer and early-stopping.
sCT image quality under three approaches is evaluated on the pelvic reference dataset using the structural similarity index (SSIM), mean squared error (MSE), and peak signal-to-noise ratio (PSNR) measures to compare the sCT with a rigidly registered planning CT.
Results: The enforced masking approach achieves superior image quality metrics (SSIM: 93.7 ± 0.028, MSE: 0.003 ± 0.002, PSNR: 33.8 ± 3.40 dB) to mask-based patch-weighting (SSIM: 88.4 ± 0.030, MSE: 0.161 ± 0.036, PSNR: 30.8 ± 2.47) and no-masking (SSIM: 87.7 ± 0.026, MSE: 0.004 ± 0.002, PSNR: 30.3 ± 1.97). Likewise, it yields HU value histograms closest to reference CT images and visual inspection verifies it to suffer from fewer artifacts and better preserve anatomical detail.
Conclusion: The enforced masked approach, combined with a replay buffer and a new way of training, proved to be the most effective strategy for mask utilization during CycleGAN training.
Future research will assess these improved sCTs in segmentation and deformable registration tasks, with the goal of clinical integration to improve IGRT.