Deep Learning-Based Image Quality Enhancement Combining Denoising and Super-Resolution for Fat-Suppressed T2-Weighted...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Fat-suppressed T2-weighted breast MRI faces challenges balancing signal-to-noise ratio (SNR), resolution, and scan time.
• Deep learning reconstruction (DLR) offers a potential solution by combining denoising and super-resolution techniques.

Purpose of the Study:

• To evaluate the intrinsic impact of a commercial DLR method on breast MRI image quality compared to conventional reconstruction (Conv).
• To assess both qualitative and quantitative image quality metrics.

Main Methods:

• Retrospective analysis of 49 patients undergoing 3-T breast MRI.
• Generation of both Conv and DLR images from identical k-space data.
• Qualitative scoring by two blinded readers and quantitative analysis of SNR and contrast ratio (CR).

Main Results:

• DLR significantly improved qualitative scores for contrast, noise, and breast parenchyma depiction (p < 0.001).
• Quantitative analysis showed a 31% median increase in SNR for DLR images (p < 0.001).
• CR changed minimally with DLR (p = 0.029), indicating preserved intrinsic tissue contrast.

Conclusions:

• DLR intrinsically enhances breast MRI image quality, improving perceived conspicuity and SNR through denoising and sharpening.
• The application of DLR without altering acquisition parameters supports future optimization for shorter scan times or higher resolution.