Toward protocol simplification: Deep learning-based image synthesis in three-phase CT urography

The purpose of this study was to develop and evaluate a method for synthesizing 3D urothelial phase images in CTU examinations from the dual inputs of non-contrast and excretory phase images using a diffusion model integrated with a Swin transformer-based deep learning approach.

This retrospective single-center study included 335 patients who underwent three-phase CTU (mean age±SD = 63 ± 15 years; 179/156 males/females). The three phases for each patient were aligned with a deformable registration algorithm. The cohort was split 80/10/10 into training/validation/testing sets. A custom deep learning model coined dsSNICT (diffusion model with swin transformer for synthetic images in CTU) was developed and implemented to synthesize the urothelial phase images. Performance was assessed using peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), mean absolute error (MAE), and Fréchet video distance (FVD). Qualitative evaluation of the synthesized images was performed by two fellowship-trained abdominal radiologists on a Likert scale from 1 to 5.

The synthetic urothelial phase images generated by the dsSNICT model achieved high PSNR (26.2 ± 4.4 dB), SSIM (0.84 ± 0.069), MAE (12.8 ± 5.2 HU), and FVD (1373). Two radiologists provided average scores of 3.5 for ground truth images and 3.4 for synthetic images, with no significant difference in scores between the two sets of images (p-value = 0.5).

dsSNICT may enable a 33% reduction in radiation dose for CTU without compromising image quality, and can also be used to salvage urothelial phase images affected by poor contrast timing or motion artifact, thereby improving CTU safety and diagnostic utility.