What question did this study set out to answer?

The study aims to enhance the perceptual realism of non-contrast cardiac CT images using a CycleGAN-based approach.

April 19, 2026BMC Medical Informatics and Decision Making0 citationsOpen Access

CycleGAN-based image-to-image translation for synthetic contrast enhancement in non-contrast cardiac CT: a ViT-CNN hybrid deep learning approach

DPDoyoung ParkNational Heart Centre Singapore JKJinsoo KimGachon University XRXucong RuanChangi General Hospital

Key Points

The study aims to enhance the perceptual realism of non-contrast cardiac CT images using a CycleGAN-based approach.
Developed RICAUNet-v2, a ViT-CNN hybrid architecture for NCCT-to-CECT translation.
Compared RICAUNet-v2 against several architectures using metrics like MAE, PSNR, SSIM, and FID.
Utilized a blinded reader study for clinical evaluation of perceptual realism.
Conducted ablation studies to analyze the impact of input types and sampling methods.
CycleGAN-RICAUNet-v2 achieved the lowest FID score of 0.0639, indicating superior perceptual realism.
Pix2pix-RICAUNet-v2 had the best MAE at 92.72 and SSIM at 0.5380.
MaskGAN recorded the highest PSNR of 13.50.
The clinical evaluation flagged 100% precision but only 15% recall, highlighting potential bias in misclassification.

Abstract

Contrast-enhanced computed tomography (CECT) is essential for visualizing cardiac structures and stenosis; however, it is unsuitable for patients with contrast agent allergies. To address this, we propose RICAUNet-v2, a ViT-CNN hybrid architecture designed for CycleGAN-based similarity-guided unpaired NCCT-to-CECT translation. RICAUNet-v2 integrates RICA blocks and Vision Transformer (ViT) layers to capture local and global features. We compared our architecture against UNet, RICAU-Net, MaskGAN, and a supervised pix2pix-RICAUNet-v2 using metrics including Mean Absolute Error (MAE), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Fréchet Inception Distance (FID). Statistical significance was assessed using the Wilcoxon signed-rank test, and a blinded reader study conducted by two clinicians evaluated perceptual realism. CycleGAN-RICAUNet-v2 achieved the lowest FID score (0.0639). Although pix2pix-RICAUNet-v2 attained the best MAE (92.72) and SSIM (0.5380), and MaskGAN achieved the highest PSNR (13.50), the superior FID score of CycleGAN-RICAUNet-v2 indicates better perceptual realism for unpaired translation. Ablation studies confirmed that using three-channel windowed inputs and similarity-guided sampling reduced the FID compared to single-channel input and unpaired cross-patient sampling (FID of 0.0802). Clinical evaluation demonstrated 100% precision and 15% recall, with a Cohen’s kappa of 0.40, suggesting that most misclassifications of real images as synthetic may be due to conservative or expectation bias rather than a lack of realism. RICAUNet-v2 improves perceptual realism in similarity-guided unpaired NCCT-to-CECT translation by integrating ViT and RICA blocks. Although further clinically relevant validation of diagnostic utility is required, the proposed architecture demonstrates potential for contrast-free cardiac CT enhancement. Not applicable.

Ask AI

Helpful

Bookmark

View Full Paper