The Challenge of Rendering Abnormalities Accurately Using Deep Learning-Based Reconstruction in Accelerated Brain...

Deep learning (DL)-based reconstruction methods can substantially accelerate MRI acquisition while preserving image quality relative to traditional techniques. However, concerns remain regarding their ability to accurately reconstruct abnormalities at higher acceleration factors. This study quantitatively assesses the impact of higher-order acceleration on image quality in DL-reconstructed brain MRI in normal and abnormal cases, emphasizing the utmost need to render abnormal findings accurately.

Raw k-space image data from 5847 brain MRI examinations (fastMRI) were used. A radiologist-annotated subset of 1001 examinations (fastMRI+) served as the test set; the remainder were split into training (n=3715) and validation (n=1131) sets. Images were retrospectively undersampled (acceleration factors 2 to 12) and reconstructed using an end-to-end variational network. Image quality metrics, including normalized mean squared error (NMSE), structural similarity index measure (SSIM), and peak signal-to-noise ratio (PSNR), were computed within bounding-box annotations containing abnormalities and in size-matched normal control regions, and at the slice-level. The region-by-acceleration interaction was tested using linear mixed-effects models with a subject-level random intercept.

All metrics demonstrated progressive degradation with increasing acceleration across slice-level and bounding-box-level analyses. At lower acceleration rates, bounding boxes with and without abnormalities showed comparable image quality. However, above 4-fold acceleration, NMSE within abnormal regions exhibited a notably sharper increase with rising acceleration compared to normal regions, while abnormal regions demonstrated a significantly faster decline in PSNR and SSIM. This nonsymmetric information loss was particularly evident for NMSE at higher acceleration rates. The region-by-acceleration interaction was statistically significant for all 3 metrics at both levels (all adjusted P<0.001), with the effect largest and most consistent for NMSE.

Our findings demonstrate a disproportionate, pathology-localized degradation of reconstruction quality as acceleration increases. This emphasizes the importance of rigorous clinical assessment of image quality in DL-MRI reconstruction, particularly at higher undersampling rates, highlighting the need for reconstruction strategies tailored to maintaining diagnostically critical features.