Synthesis of MR fingerprinting information from magnitude-only MR imaging data using a parallelized, multi network U-Net convolutional neural network

2024

Using Deep Learning to Create MR Fingerprinting Signals from Standard MRI Data

Sample size: 37 publication 10 minutes Evidence: moderate

Author Information

Author(s): Kiaran P. McGee, Yi Sui, Robert J. Witte, Ananya Panda, Norbert G. Campeau, Thomaz R. Mostardeiro, Nahil Sobh, Umberto Ravaioli, Shuyue (Lucia) Zhang, Kianoush Falahkheirkhah, Nicholas B. Larson, Christopher G. Schwarz, Jeffrey L. Gunter

Primary Institution: Mayo Clinic, Rochester, MN, United States

Hypothesis

Can a deep learning network synthesize MR fingerprinting signals from conventional magnitude-only MR imaging data?

Conclusion

The study demonstrates that MR fingerprinting signals can be synthesized from standard MRI data using a deep learning network, potentially allowing for quantitative relaxometry assessments without dedicated MR fingerprinting sequences.

Supporting Evidence

The concordance correlation coefficient for T1 and T2 MRF data pairs were 0.8793 and 0.9078, respectively.
The mean difference between true and DL relaxometry values was 48.23 ms for T1 and 2.02 ms for T2.
95% confidence limits for T1 were 0.8136–0.9383 and for T2 were 0.8981–0.9145.

Takeaway

Researchers created a computer program that can make special MRI images using regular MRI scans, which could help doctors get important information without needing extra scans.

Methodology

A U-Net deep learning network was developed to synthesize MR fingerprinting signals from magnitude-only 3D T1-weighted brain MRI data from 37 volunteers, and performance was evaluated using concordance correlation coefficients.