Double-contrast, multi-slice deep learning model achieves high-fidelity brain MRI reconstruction from 30% k-space sampling

A multi-slice deep learning model using alternating masks and double-contrast (T1w/T2w) input reconstructed brain MR images from 30% k-space data with a PSNR of 44.4 and SSIM of 0.987. The U-net approach outperformed two generative adversarial network models (DAGAN, SARA-GAN) wi…

• The proposed method uses two complementary undersampling masks alternating across slices to leverage adjacent-slice and cross-contrast (T1w/T2w) information for k-space estimation.
• The final reconstruction employs a two-channel U-net that takes the estimated T1w and T2w images as input to produce the final T1w image.
• Key limitation is the lack of external validation; performance was assessed on a single dataset and compared to only two existing models, with no reader study for diagnostic quality.

RadPigeon summaries are original and for information only. They are not clinical advice.