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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Computed tomography urography (CTU) is crucial for diagnosing urinary tract conditions.
• Current CTU protocols often involve multiple phases, potentially increasing radiation exposure.
• Synthesizing specific phases could optimize imaging protocols.

Purpose of the Study:

• To develop and evaluate a deep learning method for synthesizing 3D urothelial phase images in CTU.
• To utilize non-contrast and excretory phase images as dual inputs.
• To employ a diffusion model integrated with a Swin transformer architecture.

Main Methods:

• A retrospective study of 335 patients undergoing three-phase CTU.
• Development of the dsSNICT (diffusion model with swin transformer for synthetic images in CTU) deep learning model.
• Performance evaluation using quantitative metrics (PSNR, SSIM, MAE, FVD) and qualitative radiologist assessment.

Main Results:

• The dsSNICT model generated synthetic urothelial phase images with high PSNR (26.2 dB), SSIM (0.84), and acceptable MAE (12.8 HU).
• Quantitative metrics like Fréchet video distance (FVD) were also assessed.
• Radiologist evaluation showed no significant difference between synthetic and ground truth images (average Likert score 3.4 vs. 3.5).

Conclusions:

• The dsSNICT model can synthesize high-quality 3D urothelial phase images for CTU.
• This method holds potential for a 33% reduction in radiation dose.
• It can also salvage images compromised by timing or motion artifacts, enhancing CTU safety and diagnostic value.