Magnetic Resonance Imaging

We investigate MRI physics, signal formulation, image reconstruction, and tissue modeling. We develop original MRI methods to understand fundamental pathophysiological processes of diseases and offer tools to assess tissue biological structures and activities in the brain, joints, and body.

A full list of our publications in this category can be found here.

Research Highlight

Quantitative MRI for Tissue Characterization

MRI is a diverse and powerful imaging modality with many applications in clinical diagnosis and basic scientific research. MRI offers superior soft-tissue characterization and flexible contrast mechanisms without radiation exposure, allowing the acquisition of functional, hemodynamic, and metabolic information and anatomical images with high resolution for a comprehensive examination. MRI also allows quantitative measurement of tissue properties in-vivo. There has long been an interest in using quantitative MRI to gain deeper insights into the disease environment in a routine clinical setting. Compared to conventional contrast-weighted MRI, quantitative MRI provides increased sensitivity to different diseases and may enable early identification of pathologies. Meanwhile, quantitative measurements can deliver important information specific to tissue composition and microstructure. Our group has dedicated research in quantitative MRI, with a specific focus on multi-contrast spin-lattice relaxation time (T1), spin-spin relaxation time (T2), spin-lattice relaxation time in the rotating frame (T1ρ), magnetization transfer imaging, ultra-short echo time imaging and susceptibility imaging in the applications of brain, body and musculoskeletal system.

T1, T2, and T1ρ Relaxation Mapping

The T2 and T1ρ are the most studied relaxation parameters for characterizing tissue degeneration in knee and brain diseases. T2 relaxation time is the most widely used method with the largest body of literature. T2 relaxation time can identify subtle changes in the macromolecular and water content and ultrastructure of tissue extracellular matrix associated with early tissue degeneration, which occurs before the onset of morphologic tissue damage. Cartilage T1ρ can reflect the changes in the cartilage extracellular matrix with high sensitivity to the loss of important macromolecules, suggesting the T1ρ can be used to estimate early Osteoarthritis. T1 time is also actively studied. T1 is reported to correlate with mechanical property changes of cartilage different from T2 and T1ρ and also has a high sensitivity to brain tissue degeneration.

We invent imaging techniques to provide an in-depth assessment of tissue microstructure and composition using advanced quantitative MR relaxometry. We have developed an MR imaging protocol to perform bi-component T2 mapping of cartilage and brain using two steady-state imaging sequences. Spoiled gradient-echo (SPGR) scans are acquired over multiple flip angles to infer T1 information. Balanced steady-state free precession (bSSFP) scans are acquired over multiple flip angles to infer combined T2/T1 information. The SPGR and bSSFP images were input into a bi-component tissue model to decouple two water pools. In assessing cartilage matrix, bi-component T2 maps can be used to evaluate proteoglycan and collagen content. In assessing brain tissue, this technique can assess myelin integrity through myelin water characterization.

Bi-component analysis of water fraction (FPG) and T2 relaxation time (T2PG) of water bound to the proteoglycan component is sensitive to the proteoglycan content of cartilage. FPG shows a positive correlation with proteoglycan content, while T2PG shows a positive correlation with proteoglycan denaturation. Trypsin is an enzyme to specifically degrades proteoglycan in the cartilage specimen. The second harmonic generation image shows clumped collagen fibers (d) in the degraded cartilage due to proteoglycan loss caused by trypsin.
FPG and T2PG are sensitive biomarkers to detect early cartilage degeneration. Arrows indicate a cartilage lesion in the trochlea of the cartilage of a 52-year-old male knee Osteoarthritis patient. The FPG has higher diagnostic values than all other T2 parameters for differentiating normal versus abnormal cartilage.
Magnetization Transfer Mapping

Cross-relaxation imaging is a promising technique that can probe the tissue extracellular matrix using the magnetization transfer (MT) effect between the water protons and the macromolecular protons of tissue. Cross-relaxation imaging can provide voxel-based measurements of the MT parameters, including the fraction of the macromolecular bound protons (f) and the T2 relaxation time of macromolecular bound protons. We showed the sensitivity of MT parameters to the cartilage extracellular matrix changes in both ex-vivo and in-vivo studies. We also demonstrated the capability of imaging the whole brain multi-component T2 and MT parameters using an efficient imaging acquisition protocol. MT parameters from cross-relaxation imaging may provide valuable information regarding the collagen fiber network of cartilage, the myelin content of the brain tissue, and various macromolecule content in other body parts. Combining advanced relaxation imaging and MT imaging can provide sensitive and specific imaging biomarkers for comprehensively assessing the contents of macromolecules in disease formation and progression.

Active development is ongoing in the lab to simultaneously acquire multi-contrast MT and relaxation parameters at an integrated acquisition framework compensating for various system imperfections and motion and at a super high spatial resolution, low signal-to-noise ratio, and short scan time.

The myelin water fraction (MWF) map obtained from mcDESPOT and mcRISE (proposed) and f map obtained from mcRISE for one slice of healthy volunteer is shown in a. The corresponding histograms in b show the overestimated MWF in mcDESPOT compared to mcRISE.
(a) Macromolecular fraction f, (b) fundamental rate constant k and (c) T1 maps obtained from MT phantom. (d) Plotting f as a function of Agar percentage shows a linear relationship, as expected. (e) The fundamental rate constant is independent of Agar percentage, corroborating previous studies. (f) Comparing T1 estimated using previous vFA and our proposed BTS shows that vFA consistently underestimates due to overlooking MT effects, and the degree of bias increases with macromolecular fraction. Representative fits of (g) 2%, (h) 4%, (i) 8% Agar and (j) egg white taken from pixels indicated by circles in (a). The difference between BL (blue) and BTS (orange) curves increases with increasing f.
Ultra-short Echo Time Imaging

Musculoskeletal tissues with an abundant content of highly organized collagen fibers, such as tendons, ligaments, and meniscus, appear dark when using conventional MRI methods due to their extremely rapid signal decay. Ultrashort echo time (UTE) techniques have recently been developed to capture the rapidly decaying signal within musculoskeletal tissues. By acquiring multiple echoes as short as 0.008 msec, these techniques can be used to calculate the ultra-short echo time T2* (UTE-T2*) relaxation time of tendon, ligament, meniscus, and cortical bone. We invent new multi-component UTE-T2* acquisition and image processing methods to further advance the imaging capability of UTE towards better characterizing the tissue properties against inherent low signal and various system confounding factors.

Sagittal images through the patellar tendon in a 27-year-old healthy male volunteer at all 16 echoes acquired using the 3D-Cones UTE-T2* mapping sequence. There is little MRI signal in the patellar tendon on images with TEs of 4.3 msec and longer. b,c: Signal intensity curves for a homogenous region of interest placed in a sagittal image through the central patellar tendon in a 27-year-old healthy volunteer and a 21-year-old patient with patellar tendinopathy, respectively. Note that there is a visibly improved curve fit of the signal intensity values when using a bicomponent exponential signal model.
UTE-T2* parameter maps in a 27-year-old healthy male volunteer and 32-year-old high-level recreational male runner and volleyball player with grade 2 patellar tendinopathy (ie, the increased signal intensity between 25% and 50% of the axial cross-sectional tendon width). Note the large focal area of increased T2F and decreased FF in the proximal patellar tendon in the patient with patellar tendinopathy with no visible change in T2S (arrows).
Rapid MRI using Compressed Sensing, Parallel Imaging and Non-Cartesian Acqusition

We investigate various methods for accelerating MRI, including compressed sensing, parallel imaging, and combined CS/PI. We also design new sequences to implement and optimize non-Cartesian k-space sampling to achieve high scanning efficiency, reduced imaging artifacts, three-dimensional volumetric imaging, and effective reconstruction in the brain, knee, and body applications.

Rapid 3D Knee MRI using Compressed Sensing

Three-dimensional fast spin-echo (3D-FSE-CUBE) sequences are commercially available on most MRI vendor platforms. 3D-FSE-CUBE sequences can acquire thin continuous slices through joints which can be reformatted in any orientation, thereby eliminating the need to repeat sequences with identical tissue contrast in multiple planes. The use of 3D-FSE-CUBE sequences in clinical practice could significantly decrease MRI examination times, which would improve patient comfort and increase the clinical efficiency of the MRI scanner.

However, 3D-FSE-CUBE sequences are currently limited by their long scan times needed to achieve high isotropic resolution. Compressed sensing (CS) is a method that could reduce the scan time of 3D-FSE-CUBE sequences by acquiring fewer image data through k-space undersampling. We performed studies to investigate the feasibility of using CS to accelerate 3D-FSE-CUBE imaging of the knee and determine the optimal imaging parameters of CS applied to achieve improved image quality for assessing different joint structures.

(a) Estimated low-frequency noise standard deviation maps for two healthy volunteers (#1 a central sagittal slice through the middle of the knee joint in a 30-year-old male and #2 a sagittal slice through the lateral femoral condyle in a 29-year-old female) for CUBE and CUBE-CS and corresponding (b) Color-coded noise amplification NA factor maps superimposed on top of the CUBE source image. (c) NA factor histograms show different noise amplification distributions for these two subjects, but both histograms have a mean value close to 1, indicating no noise amplification.
(a,b) CUBE and CUBE-CS images in a 26-year-old male show a similar appearance of an anterior cruciate ligament tear (arrows). (c,d) CUBE and CUBE-CS images in a 46-year-old female show a similar appearance of a posterior horn medial meniscus tear (arrows).
Novel Non-Cartesian MRI

In close collaboration with Dr. Li Feng from the Icahn School of Medicine at Mount Sinai, our group is also actively conducting technical development, optimization, and evaluation of non-Cartesian imaging in several clinical applications. In particular, we are interested in the stack-of-star imaging technique named GRASP, originally developed by Dr. Feng and his colleagues at New York University. We investigate the integration of GRASP and its variants with our AI-based reconstruction techniques, including SANTIS, MANTIS, and RELAX, to achieve unprecedented image speed, quality, quantification accuracy, and spatial-temporal resolution with a guaranteed algorithm convergence by theory.

The GRASP project represents a decade of innovation by our team consisting of imaging scientists, clinicians, and our industry partners. The GRASP paper was announced as the third most-cited MRM paper at the 2017 ISMRM annual meeting, and the XD-GRASP paper was announced as the top most-cited MRM paper at the 2019 ISMRM annual meeting. With the rise of Artificial Intelligence in recent years, we are now aiming to integrate GRASP MRI with deep learning approaches to enable further improvement in reconstruction quality and speed, as well as new uses of this imaging framework. The initial feasibility of deep-learning-enabled golden-angle radial MRI has been demonstrated by us with a technique called SANTIS (see Deep Learning for Rapid MRI), and we are in the process of developing new quantitative imaging methods based on a combination of GRASP with deep learning.


by Dr. Li Feng
IR-prepared stack-of-stars acquisitions. The imaging sequence was developed based on a stack-of-stars 3D GRE sequence (RAVE). A, An adiabatic non-selective 180° IR pulse is periodically played-out to achieve magnetization preparation. After each IR pulse, a series of radial stacks rotated by a pre-defined rotation scheme is acquired until the magnetization reaches a steady state. B, After synchronizing all the acquired repetitions, a composite IR-prepared dynamic image series can be generated where N consecutive golden-angle rotations form k-space at each time point to ensure uniform coverage. C, The IR-prepared stack-of-stars sequence can also be performed for multi-echo acquisitions, where the rotating angles for different echoes are the same, and the user can select the number of echoes.
Comparison of brain T1 maps obtained from MP2RAGE and MP-GRASP (proposed ) in one volunteer. The T1 maps are visually comparable except for the CSF and the skull region. The linear regression shows that mean T1 values across all the subjects exhibit a good correlation (R2 = 0.955). The Bland-Altman plot suggests that MP2RAGE yielded lower T1 values than MP-GRASP.