Deep-Learning-Based Contrast Synthesis From MRF Parameter Maps in the Knee Joint.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Quantitative MRI

Background:

• Magnetic resonance fingerprinting (MRF) accelerates quantitative MRI data acquisition.
• MRF typically does not produce contrast-weighted images needed by radiologists, limiting scan time improvements.
• Synthesizing contrast-weighted images from MRF data could significantly decrease overall imaging time.

Purpose of the Study:

• To enhance the clinical utility of MRF by synthesizing contrast-weighted MR images.
• Utilized U-net deep learning models trained for image synthesis.
• Employed L1 and perceptual loss functions, and their combinations, for training.

Main Methods:

• Retrospective analysis of knee joint MRI data from 184 subjects.
• Employed a 3T multislice-MRF sequence.
• U-net models were trained using L1 and perceptual loss functions; synthetic images were evaluated by radiologists and quantitative metrics.

Main Results:

• Deep learning models successfully synthesized conventional MRI images with high quality (Likert scores 3-4).
• Qualitative assessment showed combined L1 and perceptual loss, or perceptual loss alone, yielded the best results.
• Quantitative metrics favored pure L1 loss, while qualitative assessment indicated poorer quality with L1 loss alone.

Conclusions:

• Deep learning-based synthesis of contrast-weighted images from MRF data is feasible and yields high-quality results.
• This approach holds promise for reducing MRI scan times and improving clinical workflow.
• Further research is necessary to validate the diagnostic accuracy of these synthesized images.