Structure-preserving Image-quality Enhancement for 3D Synthetic FLAIR Using a 3D U-Net with Content and Style Losses

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Synthetic MRI offers multi-contrast imaging from a single acquisition.
• Conventional FLAIR images are crucial for diagnosing demyelinating diseases.
• Synthetic FLAIR often exhibits lower image quality compared to conventional FLAIR.

Purpose of the Study:

• To enhance the image quality of 3D synthetic FLAIR using deep learning.
• To maintain the scan-time efficiency of synthetic MRI.
• To improve diagnostic accuracy in neuroradiologic assessments.

Main Methods:

• Trained a deep learning model (U-Net-based attention network) on 55 adult patients with inflammatory demyelinating diseases.
• Utilized five 3D-QALAS source images to generate synthetic FLAIR, with conventional FLAIR as reference.
• Employed a combined loss function (MSE, content, style) and assessed image quality, lesion overlap (Dice coefficient), and radiologist ratings.

Main Results:

• Deep learning-generated FLAIR showed superior agreement with reference FLAIR (P < 0.001).
• Improved lesion overlap (0.642 vs. 0.487) and visibility of focal lesions.
• While image quality improved, reader scores remained lower than conventional FLAIR (P < 0.001).

Conclusions:

• Deep learning effectively improved 3D synthetic FLAIR image quality from 3D-QALAS source images.
• The enhanced synthetic FLAIR shows potential for increased clinical utility in neuroradiology.
• This approach preserves the benefits of rapid, multi-contrast imaging with synthetic MRI.