Summary:
Researchers from UCLA’s Departments of Bioengineering and Radiological Sciences have developed a novel deep learning framework for accelerated quantitative magnetic resonance imaging.
Background:
Fatty liver disease, or hepatic steatosis, affects nearly 2 billion people globally and has pernicious implications for deadlier diseases including cancer, diabetes, and cardiovascular disease. Fortunately, steatosis can be reversible with intervention, and reduction in liver fat may diminish many of its associated risk. Thus, there is an urgent need for tools to more accurately detect its presence and assess its severity. Magnetic resonance imaging (MRI) has been used to quantify fat and diagnose hepatic steatosis. However, available MRI methods used today are either sensitive to motion, making acquisition difficult for some patients, or if robust to motion, are time-consuming to acquire and computationally expensive to process. With the prevalence of liver disease and mortality continuing to rise, the need for novel MRI methods that are cost-effective, accessible, and accurate is greater than ever.
Innovation:
UCLA researchers led by Dr. Holden Wu in the Departments of Bioengineering and Radiological Sciences have developed a deep-learning two-stage image-to-image-to-map (IIM) framework that is compatible with and accelerates free-breathing quantitative MRI. To test their framework, the researchers acquired axial free-breathing multi-echo stack-of-radial data from 68 adults and 22 children on 3T MRI scanners and retrospectively undersampled by one-third. Their IIM framework suppressed radial MRI undersampling artifacts and reconstructed accurate proton density fat fraction (PDFF) maps for 3-fold scan acceleration, corresponding to <1 minute free-breathing scan time. They demonstrated that with 20 hours of training on a NVIDIA v100 GPU, their framework had an average interference time of 65.79 milliseconds per slice for image enhancement and calculation of fat-water fraction and uncertainty maps, while conventional graph-cut (GC) algorithms took 28 seconds per slice. The researchers also showed that the built-in uncertainty estimation can identify regions with potential quantification errors. Using a novel deep-learning framework that combines image enhancement, parameter mapping, and uncertainty estimation, MRI techniques could be radically improved to make strides in diagnosing and reducing the prevalence of hepatic steatosis.
Potential Applications:
● Free-breathing and rapid liver fat mapping in stack-of-radial MRI that allows for more cost-effective, accessible, and accurate detection of hepatic steatosis
Advantages:
● Fast scan times (<1 min)
● Reduced computational time (<100 ms/slice) for signal fitting
● Removes artifacts and achieves high image quality from undersampled MRI data
● High accuracy for liver fat quantification
● Can detect uncertainty caused by noisy input data
● Robust to motion artifacts
Development Status:
Researchers have successfully trained and acquired liver fat maps using their deep-learning IIM framework in a prototype free-breathing multi-echo stack-of-radial MRI.
Related Papers:
S. Shih, S.G. Kafali, T. Armstrong, X. Zhong, K.L. Calkins, H.H. Wu, “Deep Learning‐Based Parameter Mapping with Uncertainty Estimation for Fat Quantification using Accelerated Free‐Breathing Radial MRI,“ 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI 2021).
S. Shih, S.G. Kafali, T. Armstrong, X. Zhong, K.L. Calkins, H.H. Wu, “Deep Learning‐Based Liver Fat and R2* Mapping with Uncertainty Estimation using Self‐Gated Free‐Breathing Stack‐of‐Radial MRI,“ 2021 International Society for Magnetic Resonance in Medicine Annual Meeting (ISMRM 2021).
Reference:
UCLA Case No. 2021-268
Lead Inventor:
Dr. Holden H. Wu