Utilizing conditional generative adversarial network to generate head MRA based on nonvascular sequences: comparative study of single-modality and multi-modality methods.

Authors

Abstract

To develop a deep learning model synthesizing head MRA images from the preferred nonvascular sequences (T1W, T2W, and FLAIR) and evaluate the diagnostic performance of synthetic MRA (syn-MRA). This retrospective study included geriatric inpatients who underwent multimodal MRI (including MRA and nonvascular sequences) at our institution between January 2022 and March 2023. Three single-modality conditional generative adversarial network (cGAN) models (T1W, T2W, and FLAIR-based) and one multi-modality model (MIX model) were constructed. Quantitative metrics were used to evaluate image quality. Two radiologists independently assessed image visual quality using Likert scales. Two senior neuroradiologists evaluated structured reports and diagnostic confidence. This study ultimately included 140 patients: 98 training, 14 validation, 28 testing. The MIX model outperformed single-modality models (test set: structural similarity index measure 0.880, peak signal-to-noise ratio 33.18 dB, root mean squared error 0.022). Syn-MRA from all models demonstrated significantly higher signal-to-noise ratio and contrast-to-noise ratio compared to real MRA (<i>p</i> < 0.001), improved vascular signal uniformity (<i>p</i> > 0.05), but more pronounced venous contamination (<i>p</i> < 0.001). The MIX model achieved comparable overall image quality (median: 5 vs. 5, <i>p</i> = 0.344), vessel sharpness (median: 5 vs. 5, <i>p</i> = 0.145), and diagnostic confidence (median: 5 vs. 5, <i>p</i> = 0.102) relative to real MRA. MIX model syn-MRA showed 92.2% diagnostic accuracy (vessel level). The MIX model outperformed single-modality models, with image quality and diagnostic performance comparable to real MRA.

Tags

Topics