Enhancement of artery visualization in contrast-enhanced cerebral MR angiography using generative neural networks

Time-resolved contrast-enhanced magnetic resonance angiography (TR-MRA) is an established technique used to capture the movement of a contrast bolus through the vascular system. However, in the cerebral vasculature, which has a short arteriovenous transit time, venous overlay can occur even in the image of the peak arterial phase, making it challenging to obtain pure artery-only angiograms. This study aimed to develop conditional generative neural networks to generate pure angiograms with enhanced arterial contrast using clinical cerebral TR-MRA data. To achieve this, we proposed a preprocessing algorithm that synthesized angiograms with optimal arterial contrast by utilizing contrast dynamics. This synthetic image and temporal maximum intensity projection (MIP) of TR-MRA served as the target and source images, respectively, for training the conditional generative adversarial network (GAN) and denoising diffusion probabilistic model (DDPM). The results showed that the conditional DDPM achieved substantially higher arterial contrast compared to raw TR-MRA data, as confirmed by the relative artery–vein contrast ratio (0.9889 ± 0.0049 vs. 0.8265 ± 0.0502) and the artery–muscle contrast ratio (0.9825 ± 0.0038 vs. 0.8806 ± 0.0225). The conditional DDPM outperformed conditional GAN in the quality of artery visualizations, as confirmed by the qualitative visibility score (4.07 ± 0.49 vs. 3.82 ± 0.70). Furthermore, we demonstrated the feasibility of applying the trained generative models to single-phase contrast-enhanced MRA images with high spatial resolution.