NNFit: A Self-Supervised Deep Learning Method for Accelerated Quantification of High-Resolution Short-Echo-Time MR
Alexander S Giuffrida1, Sulaiman Sheriff1, Vicki Huang1
1From the Department of Radiation Oncology (A.S.G., V.H., H.S.) and Department of Radiology and Imaging Sciences (B.D.W.), Emory University School of Medicine, 1701 Uppergate Dr, C5008 Winship Cancer Institute, Atlanta, GA 30322; Department of Radiology, University of Miami School of Medicine, Miami, Fla (S.S., A.A.M.); Department of Pathology, Northwestern University Feinberg School of Medicine, Chicago, Ill (L.A.D.C.); Department of Biostatistics and Bioinformatics, Emory University Rollins School of Public Health, Atlanta, Ga (Y.L.); Department of Psychology, Emory University, Atlanta, Ga (M.T.); and Department of Radiology, Duke University Medical Center, Durham, NC (B.J.S.).
NNFit, a deep learning method, quantifies echo-planar spectroscopic imaging (EPSI) data with performance comparable to traditional methods but significantly faster processing times. This advance addresses computational bottlenecks in clinical workflows for brain imaging.
Area of Science:
- Neuroimaging
- Artificial Intelligence
- Biomedical Engineering
Background:
- Conventional spectral quantification methods for echo-planar spectroscopic imaging (EPSI) present computational bottlenecks in clinical workflows.
- High-resolution short-echo-time (TE) EPSI datasets require efficient quantification for accurate analysis.
Purpose of the Study:
- To develop and evaluate NNFit, a self-supervised deep learning method for quantifying high-resolution short-TE EPSI data.
- To compare the performance of NNFit against conventional spectral quantification methods.
Main Methods:
- Retrospective analysis of 89 short-TE whole-brain EPSI scans from glioblastoma and major depressive disorder clinical trials.
- Development and application of a self-supervised deep learning model (NNFit) for spectral quantification.
- Comparative analysis with a parametric-modeling method (FITT) using structural similarity index measure (SSIM), linear correlation coefficient (R²), and Dice coefficient.
Main Results:
- NNFit demonstrated comparable performance to FITT across multiple metabolites, with high SSIM and R² values (e.g., 0.91/0.90 for choline, 0.98/0.98 for NAA in trial 2).
- Treatment volume delineation for glioblastoma using NNFit-generated metabolite maps yielded a mean Dice coefficient of 0.92.
- Mean processing time for NNFit was 90.1 seconds, compared to 52.9 minutes for FITT, representing a significant speed improvement.
Conclusions:
- Deep learning-based spectral quantification (NNFit) achieves performance similar to conventional methods for short-TE EPSI data.
- NNFit offers a substantial reduction in processing time, enhancing clinical workflow efficiency.
- The findings support the use of NNFit as a viable and accelerated alternative for spectral quantification in neuroimaging.
