Abnormal brain regions degrade faster with higher acceleration in DL-reconstructed MRI

Accelerated brain MRI with deep learning reconstruction showed disproportionate quality loss in regions with abnormalities above 4-fold acceleration—all image-quality metrics worsened faster than normal regions (adjusted P<0.001).

• The study used raw k-space data from 5847 brain MRI examinations; the test set was 1001 radiologist-annotated cases, with 3715 for training and 1131 for validation.
• Images were retrospectively undersampled at acceleration factors 2 to 12 and reconstructed with an end-to-end variational network.
• Above 4-fold acceleration, abnormal regions showed a sharper rise in NMSE and faster decline in PSNR/SSIM, with region-by-acceleration interaction significant for all metrics (all adjusted P<0.001).

Automated summary

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