Deep Learning-Based Image Quality Enhancement Combining Denoising and Super-Resolution for Fat-Suppressed T2-Weighted Breast MRI: A Qualitative and Quantitative Evaluation

Introduction Fat-suppressed T2-weighted breast MRI faces a trade-off among signal-to-noise ratio (SNR), resolution, and scan time. This study evaluated the intrinsic effect of a commercial deep learning reconstruction (DLR) method, combining denoising and super-resolution, on image quality by comparing it with conventional reconstruction (Conv) generated from identical raw data. Materials and methods For this retrospective study, 49 women who underwent 3-T breast MRI were included. From the same k-space data, Conv and DLR images were produced. Qualitative assessment involved two blinded readers qualitatively scoring five image quality parameters. For quantitative analysis (n = 44 analyzable cases), regions of interest were placed to define the SNR within the pectoralis major and the contrast ratio (CR) between muscle and fat. Results Qualitatively, DLR yielded higher scores for contrast, noise, and depiction of breast parenchyma for both readers (all p < 0.001). Signal uniformity improved modestly for one reader. Artifact ratings were mixed: one reader favored DLR (p = 0.002), whereas the other showed no significant difference (p = 0.670). Inter-reader agreement was good to very good for most parameters (kappa = 0.75-0.83), but moderate for artifacts (kappa = 0.42). Quantitatively, DLR increased the SNR by approximately 31% (median 5.00 vs. 3.79; p < 0.001), while the CR changed minimally (median 0.56 vs. 0.53; p = 0.029). Conclusion These findings indicate that DLR enhances perceived conspicuity and SNR via denoising and sharpening while preserving intrinsic tissue contrast. Applying DLR without altering acquisition parameters intrinsically improves image quality, supporting future protocol optimization toward shorter scan times or higher resolution.