FreqRefine-DDPM: Virtual Contrast-Enhanced CT Synthesis from Non-Contrast CT via Frequency-Domain Enhanced Diffusion Model with Adaptive Weighted Refinement
Purpose/Objective(s):
Contrast-enhanced computed tomography (CECT) is essential for clinical diagnosis, yet the administration of iodinated contrast agents poses risks including allergic reactions and nephrotoxicity. This study aims to develop a deep learning framework capable of synthesizing virtual contrast-enhanced CT images directly from non-contrast CT (NCCT) scans, thereby reducing patient exposure to contrast agents while preserving diagnostic image quality.
Materials/Methods:
A total of 210 nasopharyngeal cases are enrolled in this study and divide into three datasets: 178 for training, 11 for validation and 21for testing. We propose FreqRefine-DDPM, a cascaded framework that integrates a denoising diffusion probabilistic model (DDPM) with a simple U-Net refinement network. The DDPM serves as the generative backbone to learn the global mapping from NCCT to CECT, while the subsequent U-Net further refines the DDPM output to enhance fine-grained details.
Results:
The proposed method was evaluated on 21 test cases, achieving a MAE of 12.42 ± 4.65 HU, a peak signal-to-noise ratio (PSNR) of 30.96 ± 3.22 dB, and a structural similarity index (SSIM) of 0.9620 ± 0.0167.
Conclusion:
The FreqRefine-DDPM framework effectively generates virtual contrast-enhanced CT images from non-contrast scans with promising quantitative performance. The integration of Otsu-based adaptive weighted loss and frequency-domain bottleneck processing contributes to improved reconstruction accuracy, particularly in clinically relevant enhancement regions. This approach offers a potential pathway toward reducing contrast agent dependency in clinical CT imaging.
Keywords:
virtual contrast-enhanced CT; denoising diffusion probabilistic model; frequency-domain feature enhancement;
Figure 1. CECT prediction result of a random selected patient in test dataset (Line 1: Model input; Line 2: Ground truth; Line 3:Freq-Refine DDPM prediction; Line 4: Ablation study of DDPM prediction. Red line: GTVnd contour)
