3DAD: Super-Resolution Image Synthesis from Anisotropic CT Images Using a Three-Dimensional Adversarial Diffusion Model.

High-resolution thin-slice computed tomography (CT) images are often compressed into lower-quality thick-slice images for long-term storage, necessitating synthesis for medical diagnosis. In this paper, we propose a novel 3D adversarial diffusion model (3DAD) for high-fidelity synthesis of thin-slice CT from compressed thick-slice CT. 3DAD is composed of a generator and a discriminator for synthesizing denoised thin-slice images from random noise and source images and distinguishing between noised samples from real and denoised synthetic thin-slice images. Specific models were trained on two-slice to six-slice scenarios for abdominal data, using thick-slice CT compressed from real thin-slice CT as the source. 3DAD was evaluated at the time of HCC diagnosis, at the observation and patient levels, using real thin-slice and synthetic thin-slice CT, with DeLong's test to compare the similarity of receiver operating characteristic (ROC) curves. We further evaluated 3DAD on real-world data with both thin and thick images, with the synthetic image quality assessed by radiologists and in radiomics feature analysis. Based on the external dataset with 548 samples, the achieved mean squared error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM) values were 81.374, 29.478, and 0.916, respectively, for the five-slice scenarios at the portal venous phase. The Areas Under Curves (AUCs) achieved were 0.896 on synthetic thin-slice images compared with 0.889 on real thin-slice images at the observation level (