Purpose: To improve the quality of a fast multi-contrast MR protocol acquisition using deep learning.

Materials and methods: 350 patients (age: 64 ± 17 yrs; 155 male) underwent both a fast brain MR multi-contrast sequence (NeuroMix), capturing five contrasts in a single 2.5-minute sequence, and conventional high-resolution imaging. This retrospective study was approved by the Stanford IRB (eProtocol 26147, IRB registration 6208). The paired images were used to enhance resolution and image quality using a Swin U-Net Transformer (SwinUNETR) approach, focusing on T1-weighted (T1w), T2-weighted (T2w), and T2 FLAIR images (NeuroMix-DL). Evaluation included standard image quality metrics, such as the root mean squared error (RMSE) and a clinical quality assessment using a five-point image quality scale (1 = poor, 5 = excellent). A pairwise t-test was calculated to evaluate the values of the qualitative and quantitative metrics across the various image processing approaches.

Results: We found significant improvement in image quality after applying the trained SwinUNETR, with RMSE reductions of 42 ± 3%, 33 ± 2%, and 33 ± 9% for T1w, T2w, and FLAIR images, respectively (p < 0.001 for all) compared to original NeuroMix images, using conventional sequences as reference. The clinical readers found higher image quality scores for NeuroMix-DL images compared to the original NeuroMix images (12 ± 8%, 17 ± 11%, and 15 ± 6% for T1w, T2w, and FLAIR images, respectively). Visual quality assessment demonstrated improvements in prevalent artifacts, including motion, herringbone artifact, inhomogeneity artifact, and RF overflow.

Conclusion: The SwinUNETR model offers a viable approach for improving the quality of fast multi-contrast MR images while effectively mitigating artifacts, improving the cost-benefit ratio of MRI.