Updates on Compositional MRI Mapping of the Cartilage: Emerging Techniques and Applications

Abstract

Osteoarthritis (OA) is a widely occurring degenerative joint disease that is severely debilitating and causes significant socioeconomic burdens to society. Magnetic resonance imaging (MRI) is the preferred imaging modality for the morphological evaluation of cartilage due to its excellent soft tissue contrast and high spatial resolution. However, its utilization typically involves subjective qualitative assessment of cartilage. Compositional MRI, which refers to the quantitative characterization of cartilage using a variety of MRI methods, can provide important information regarding underlying compositional and ultrastructural changes that occur during early OA. Cartilage compositional MRI could serve as early imaging biomarkers for the objective evaluation of cartilage and help drive diagnostics, disease characterization, and response to novel therapies. This review will summarize current and ongoing state-of-the-art cartilage compositional MRI techniques and highlight emerging methods for cartilage compositional MRI including MR fingerprinting, compressed sensing, multi-exponential relaxometry, improved and robust radio-frequency pulse sequences, and deep learning-based acquisition, reconstruction, and segmentation. The review will also briefly discuss the current challenges and future directions for adopting these emerging cartilage compositional MRI techniques for use in clinical practice and translational OA research studies.

Quantitative cartilage compositional mapping is an important emerging topic in musculoskeletal imaging, particularly due to its potential for early diagnosis of osteoarthritis (OA) and subsequent monitoring of disease progression.¹ OA is a chronic and debilitating musculoskeletal disease causing progressive and irreversible degeneration of cartilage and other joint structures. OA currently affects over 27 million people in the United States alone and has a worse outlook in the years ahead due to the aging population and the current obesity epidemic.^1,2 The diagnosis of OA is made based on clinical symptoms such as joint pain and stiffness and the presence of definitive osteophytes on radiographs, with secondary findings of joint space narrowing (JSN) considered a surrogate marker of disease progression.³

Articular cartilage is mainly composed of water, which comprises 65%–85% of its net weight, and an abundant extracellular matrix (ECM) consisting of predominantly type II collagen fibers and large proteoglycan (PG) macromolecules.⁴ PG contains long glycosaminoglycan (GAG) side chains with negatively charged sulfate and hydroxyl groups that create a fixed charge density (FCD). The negative FCD attracts cations such as sodium that result in osmotic pressure that draws water into the cartilage.⁵ This produces a swelling pressure built up by the electrostatic repulsion between the GAG side chains in the cartilage ECM, which provides the underlying viscoelastic properties necessary for load distribution. The swelling pressure is constrained by the collagen fiber network, which provides the tensile force opposing the expansion of cartilage.⁴ Early OA is primarily characterized by a loss of PG, followed by the subsequent decrease in water content due to the reduction of osmotic pressure.⁵

Magnetic resonance imaging (MRI) is uniquely suited for characterizing human articular cartilage, providing both morphological and compositional evaluation.⁶ Compositional MRI is especially useful as it can detect changes in the composition and ultrastructure of cartilage during the earliest stages of cartilage degeneration before the onset of morphological cartilage loss. Compositional MRI can also monitor disease- and treatment-related changes in cartilage composition and ultrastructure over time in clinical practice and longitudinal OA research studies. Compositional MRI techniques are effective in evaluating cartilage in a variety of joints including the knee, hip, and ankle.^7–9

In this review article, section “Summary of current methods used for compositional MRI of cartilage” summarizes the current state-of-the-art cartilage compositional MRI techniques as well as some recent methods for cartilage morphometry that have been used in clinical practice and OA research studies. Methods described include T₂ and ${\text{T}}_2^{*}$ mapping, T_1ρ mapping, T₁ mapping with gadolinium contrast, ultrashort echo time (UTE) imaging, chemical exchange saturation transfer (gagCEST) imaging, sodium imaging, diffusion imaging, and morphometry (Table 1). Section “Emerging methods for quantitative MRI for cartilage” highlights emerging methods for cartilage compositional MRI including MR fingerprinting (MRF), compressed sensing (CS), multiexponential relaxometry, improved and robust radio-frequency (RF) pulse sequences, and deep learning (DL)-based acquisition, reconstruction, segmentation methods and MRI based biomechanical assessment. Finally, section “Summary of discussion and future directions” will briefly discuss the current challenges and future directions for adopting these emerging cartilage compositional MRI techniques for use in clinical practice and translational OA research studies.

TABLE 1.

Summary of Current State-of-The-Art Techniques Used for Cartilage Compositional MRI

MRI Method	Measurement	Interpretation
T2 Mapping	Water, collagen structure	T2 increase signifies the breakdown of the organized collagen structure
${\text{T}}_2{}^{*}$ Mapping	Water, collagen structure and local magnetic field inhomogeneity	${\text{T}}_2{}^{*}$ increase signifies the breakdown of the organized collagen structure and local magnetic field inhomogeneity
T1ρ Mapping	GAG	T1ρ increase signifies GAG depletion
T1 Mapping/dGEMRIC	GAG	T1 decrease signifies GAG depletion
UTE MRI	Short T2 components	Provides signal from tightly bound water in collagen-rich tissues
gagCEST MRI	Water and GAG	gagCEST decrease signifies GAG depletion and it is specific to GAG
Sodium MRI	GAG	Sodium signal decrease signifies GAG depletion and it is specific to FCD
Diffusion MRI	Water and PG/collagen	ADC signifies PG depletion FA correlates with collagen structure
Morphometry	Cartilage thickness, volume, and area	Reduction of these morphological measurements corresponds to loss of cartilage

Summary of Current Methods Used for Compositional MRI of Cartilage

T₂ Mapping

T₂ mapping is a quantitative MRI method that serves as a surrogate marker for water content and collagen structure. The T₂ relaxation time of cartilage primarily provides information regarding collagen fiber orientation and the integrity of the collagen fiber network.¹⁰ While increased hydration of cartilage should theoretically lead to an increase in T₂ relaxation time, not all studies have shown a direct correlation between T₂ relaxation time and the water content of ex vivo cartilage samples.¹¹ In healthy cartilage, the water content is high, but water is tightly bound to the hydrophilic GAG side chains of PG. With cartilage degeneration, there is a disorganization of the collagen fiber network and a decrease in proteoglycan content resulting in the presence of more unbound, freely mobile water. T₂ relaxation time measurements increase with the unbound water content, which enables T₂ mapping to pick up early signs of cartilage degeneration in OA that may not be apparent on morphological imaging sequences.¹²

T₂ mapping is typically performed with a range of echo times using 2D or 3D multiecho spin echo sequences or T₂ preparation modules, combined with later data fitting using mono-exponential models. More recently, 3D double-echo steady-state (DESS) sequences are used, which allows the joining quantification of T₁ relaxation and apparent diffusion coefficient (ADC) in addition to T₂ relaxation time.¹³ The main advantage of T₂ mapping is the commercial availability of T₂ pulse sequences on most MRI vendor platforms. Acknowledged drawbacks of T₂ mapping include long acquisition times, stimulated echoes, susceptibility to magic angle effects, imperfect RF pulses, reduced sensitivity to PG, magnetization transfer effects, and offline reconstruction methods.¹²

${\text{T}}_2^{*}$ Mapping

Similar to T₂, the ${\text{T}}_2^{*}$ mapping is also a surrogate marker for water content and collagen structure.¹⁴ However, ${\text{T}}_2^{*}$ is also influenced by coherent dephasing effects due to local magnetic field in-homogeneities. ${\text{T}}_2^{*}$ mapping is usually performed with fast gradient recalled echo (GRE) sequences, and, due to de-phasing, ${\text{T}}_2^{*}$ values are usually shorter than T₂. ${\text{T}}_2^{*}$ mapping offers the advantages of fast imaging, 3D acquisition, and high spatial resolution, with no additional hardware or special sequences required. ${\text{T}}_2^{*}$ studies have shown good reproducibility.^6,14,15

${\text{T}}_2^{*}$ mapping offers unique information useful for the characterization of cartilage. ${\text{T}}_2^{*}$ contrast captures shorter decay times in the cartilage and is less sensitive to bulk water. Decreases in ${\text{T}}_2^{*}$ values are reported in deeper layers of normal cartilage and increased ${\text{T}}_2^{*}$ values in pathology.^16,17 ${\text{T}}_2^{*}$ mapping combined with UTE sequences have been useful to characterize calcified cartilage with decay times lower than 1 millisecond (msec).¹⁸

T_1ρ Mapping

T_1ρ mapping is an endogenous contrast technique that is sensitive to the GAG content of cartilage.^19,20 The exchange of protons between water and the hydroxyl and amine groups on the GAG side chains of PG is thought to be the primary mechanism for T_1ρ dispersion within cartilage.²¹ A strong correlation has been found between cartilage T_1ρ relaxation time and FCD.²² The reduction of GAG content within cartilage has also been shown to increase cartilage T_1ρ relaxation times in in vitro and ex vivo studies.^23,24 However, T_1ρ relaxation time is not a specific measure of the GAG content of cartilage and is also influenced by other biological changes that occur with cartilage degeneration.²⁵ Although T₂ and T_1ρ relaxation time values are correlated, it has been shown that T_1ρ relaxation time is more sensitive than T₂ relaxation time for detecting early cartilage degeneration in human subjects.²⁶

T_1ρ mapping is performed by applying a preparatory spin-lock pulse (SLP) after the magnetization is tipped onto the transverse plane. The signal is then allowed to decay under the presence of the SLP, with the T_1ρ. T_1ρ contrast is sensitive to the kilohertz range because of the reliance on the RF-generated B₁ field. This makes the T_1ρ signal sensitive to chemical exchange between macromolecules and bulk water.²⁵ The advantages of T_1ρ mapping are that T_1ρ dispersion is sensitive to PG loss in cartilage, provide cartilage-specific underlying mechanisms, and are not as strongly influenced by magic angle effects as T₂ mapping. Some of the drawbacks of T_1ρ mapping include its long scan times, relatively high specific absorption rate (SAR) during the T_1ρ preparation pulses and the need for customized RF pulse sequences.

T₁ Mapping and dGEMRIC

T₁ mapping for cartilage applications²⁷ usually uses a contrast-based method called delayed gadolinium-enhanced MRI of cartilage (dGEMRIC). dGEMRIC employs a negatively charged gadolinium contrast agent to adhere to the negative FCD of GAG side chains of PG. GAGs repel the contrast agent and its distribution is mapped in inverse proportion to local PG content, with a higher concentration of contrast agent where GAG concentration is low. The higher concentration of contrast agent will reduce T₁ relaxation times in areas of low GAG concentration. The T₁ relaxation time of native and enzymatically degraded ex vivo cartilage samples in the presence of gadolinium contrast has been found to strongly correlate with the PG content²⁸ and compressive stiffness of cartilage.²⁹

The dGEMRIC technique has been shown to detect surgically and histologically confirmed early cartilage degeneration in human subjects due to its sensitivity and specificity to PG content and FCD.³⁰ However, challenges in the clinical adoption of dGEMRIC have limited its widespread use. Some of the challenges are that the technique uses exogenous contrast agent injections that may cause nephrogenic systemic fibrosis and gadolinium deposition in tissues, lack of standardization of dGEMRIC protocols, and the long wait time following contrast administration depending on the cartilage anatomy and thickness. The use of ionic vs. nonionic contrast agents has also been debated with varying results.²⁷

UTE MRI

UTE imaging allows the characterization of semi-solid tissues with highly organized collagen fibers and very short relaxation times including the menisci, tendons, ligaments, and deep calcified layers of cartilage, which appear dark on typical MR images due to signal loss. UTE techniques can be used to measure T₁, T₂, ${\text{T}}_2^{*}$, and T_1ρ relaxation times of the tissue of interest. However, UTE imaging assesses effective relaxation times and is also influenced by tissue susceptibility and magnetic field inhomogeneity.³¹

To achieve UTE imaging to characterize rapidly decaying signals, gradient echo sequences are used with half-excitation RF pulses in the 0.05–0.2 millisecond range.³¹ More recent work described a pointwise encoding time reduction with radial acquisition (PETRA) sequence for UTE imaging, which used combined Cartesian and radial imaging to achieve short TE and longer TE, to visualize all fast and slow decaying structures together.³² With UTE imaging, there are challenges with off-resonance, radial trajectory errors, and distortion of slice profile, but these can be addressed with off-resonance correction, gradient trajectory calibration, and efficient fat suppression.

gagCEST MRI

GAG chemical exchange saturation transfer (gagCEST) is a magnetization transfer contrast technique that uses off-resonance pulses to saturate exchangeable labile protons in the GAG side chains of PG.³³ A gagCEST technique at 3 T is comparable to dGEMRIC and T₂ mapping for characterizing degenerative cartilage.³⁴ A strong correlation between gagCEST and sodium values within both healthy cartilage and cartilage repair tissue has also been reported, which suggests that gagCEST can be used to assess the FCD of cartilage.³³ While gagCEST imaging is a potentially powerful technique for evaluating PG within cartilage, clinical studies exploring the feasibility of early OA detection are still limited due to the small contrast changes between normal and degenerative cartilage that occur on clinical 3 T scanners.

To perform gagCEST imaging, a spoiled gradient-echo sequence is used with a range of off-resonance saturation pulses applied to directly saturate water protons associated with the macromolecules. Although gagCEST provides sensitive and specific information regarding PG in cartilage, the technique has several drawbacks. These include low SNR on systems below 7 T, thus requiring ultra-high field systems for scanning, the need for customized RF pulse sequences, long RF pulses for off-resonance saturation leading to high RF energy deposition, and complex postprocessing techniques.^33,34

Sodium MRI

The sodium ion (²³Na) possesses a quadrupolar moment that interacts with electric field gradients. Sodium reacts to the FCD in cartilage and can be used as a biomarker for GAG content. Clinical validation is limited in the literature, in part due to the challenges associated with sodium imaging. Sodium imaging has been used to compare cartilage in healthy subjects and subjects with early OA, reporting higher FCD in healthy individuals.³⁵ High reproducibility and repeatability of sodium quantification in the cartilage at 3 T and 7 T field strengths have been reported.³⁶ Sodium imaging has been used to evaluate cartilage repair following surgery.³⁷ Multinuclear, dual-tuned sodium and proton RF coils for specific use in the knee have also been developed, which allow combined sodium and morphological imaging of cartilage.³⁸

The sodium concentration of healthy cartilage is 200–300 mmol/L, which is about 260–400 times lower than the proton concentration. This dramatically reduces the signal available and requires longer scan times or the use of higher magnetic fields to improve SNR.^35,36 Sodium imaging also requires specialized RF coils and customized UTE sequences to detect signals.³⁸ Despite these challenges, sodium MRI has the potential to be used as a sensitive and specific biomarker for GAG to quantitatively characterize the FCD of cartilage.

Diffusion MRI

Diffusion imaging has long been used for brain imaging in focal ischemia, as a marker for restricted diffusion. More recently, diffusion imaging has been applied to cartilage imaging applications. The apparent diffusion coefficient (ADC) can be used as a measure of the interaction of water with the surrounding macromolecular matrix. Fractional anisotropy (FA) allows the determination of the main direction of local diffusion of water protons. Studies using enzymatic degradation of ex vivo cartilage specimens have shown changes in ADC with PG depletion³⁹ and changes in both ADC and FA with collagen depletion.⁴⁰ FA has been found to correlate strongly with the orientation of the collagen fiber network of cartilage assessed using polarized light microscopy.⁴¹ ADC and FA have also been shown to be highly sensitive for detecting early cartilage degeneration in both ex vivo⁴² and in vivo⁴³ studies.

To perform diffusion-weighted imaging (DWI) in musculoskeletal applications, pulse sequences with steady precession (DW-FISP) are most commonly used.⁴⁴ The diffusion-weighted double echo steady-state free precession (SSFP) has also been used for quantitative in-vivo diffusion imaging.⁴⁵ Although it is a potentially powerful technique for evaluating both PG and collagen within cartilage, significant challenges remain in using diffusion imaging including sensitivity to motion, long scan times, and limited SNR.

Quantitative Morphometry

Morphological changes in cartilage in OA can also be quantitatively assessed by segmenting and measuring the segmented cartilage, in a process known as cartilage morphometry.⁴⁶ It usually requires high-resolution 3D sequences with sufficient contrast between cartilage and surrounding tissues, assessing morphological features such as cartilage area, volume, and thickness. Cartilage morphometry has shown great repeatability.⁴⁶ Also, morphometry with DL is an interesting approach to evaluating the loss of cartilage in longitudinal studies.⁴⁷ Despite these advantages, morphometry cannot assess premorphological changes in the cartilage ECM compared to most quantitative compositional MRI methods.

Emerging Methods for Quantitative MRI for Cartilage

Search Methodology

This section will discuss the emerging methods for quantitative MRI for cartilage. Reviewed papers were selected from Google Scholar and PubMed, searching for the words: “quantitative,” “MRI,” and “cartilage,” between 2020 and 2023. To reduce overlap with recent review articles on rapid knee MRI⁴⁸ and DL methods for fast imaging and relaxometry,^49,50 some references already in these reviews may not be cited here. This section provides an update on the newly emerging techniques described after the ones cited in these previously published review articles.

Updates on Accelerated MRI Using CS and DL for Reconstruction and Direct Quantitative Mapping

Rapid acquisition for quantitative MRI mapping techniques is fundamentally important to obtain information regarding cartilage composition and ultrastructure in clinically feasible scan times. Shortening scan time for quantitative MRI mapping techniques while maintaining high-spatial resolution has several advantages including reduced healthcare costs, increased patient comfort, and decreased motion artifacts. Undersampling k-space acquisitions, followed by advanced reconstructions such as CS or DL to remove aliasing artifacts, can achieve reduced scan time, which is known as accelerated MRI.

Model-based and CS reconstructions are still active research topics for accelerating cartilage compositional MRI techniques. In the literature,⁵¹ a CS reconstruction using robust tensor principal component analysis (PCA) models was proposed to accelerate cartilage T_1ρ mapping, showing good agreement of T_1ρ values between CS and fully sampled (FS) acquisitions. A model-based acceleration for diffusion imaging was proposed for musculoskeletal applications, and the feasibility of acceleration factors of 3.65 was shown without compromising diffusion data fidelity.⁵²

DL-based image reconstruction is currently one of the most active topics in medical imaging research. New studies have combined CS iterative methods with DL models, such as in the literature⁵³ where the authors proposed a multilayer basis pursuit framework to improve knee images, showing a better peak signal-to-noise ratio than previous approaches. CS and DL reconstructions were compared for T_1ρ mapping using mono- and bi-exponential models, where spatial and spatiotemporal priors were compared, as shown in Fig. 1.⁵⁴ Radmanesh et al⁵⁵ explored the limits of acceleration using the variational network. Different networks were trained using acceleration rates of up to 100 times. The assessments showed diagnostic quality for 2D acquisitions was obtained using an acceleration factor of R = 4, as shown in Fig. 2. New rapid MRI reconstruction methods have been proposed by Kim et al⁵⁶ for parallel imaging and simultaneous multislice acceleration with DL for evaluating the knee joint. The quality of the images of the proposed DL approach is equivalent to parallel MRI, but with a reduction of 71% of the acquisition time.

FIGURE 1:

Zibetti et al⁵⁴ compared compressed sensing (CS) with a spatiotemporal finite difference (CS-ST) against variational networks (VN) using independent spatial reconstructions (VN-S) and joint spatiotemporal reconstructions (VN-ST) with an acceleration factor of R = 6 on mono- and bi-exponential T_1ρ relaxation models.

FIGURE 2:

Radmanesh et al⁵⁵ explored the limits of acceleration using a network based on the variational network (VN) for 2D acquisitions. Source: Images courtesy of Dr. Matthew Muckley, PhD, Meta AI, USA.

One of the main limitations of DL-based reconstructions is the failure to recover enough image quality to identify subtle but clinically important abnormalities, such as the meniscal tear.⁵⁷ This failure could result from the training datasets that underrepresent this subtle yet salient feature. A generative adversarial network (GAN) was proposed to improve the quality of images reconstructed by DL methods, particularly edge details.⁵⁸ It was shown that training DL methods with loss functions for a specific region of interest (ROI) could improve the performance of the DL models for that particular ROI, as shown in Fig. 3 where the ROI is the cartilage in the knee and hip.⁵⁹

FIGURE 3:

The loss function used for deep-learning (DL) training contains a specific region of interest loss.⁵⁹ The neural network was trained to improve the accelerated MRI specifically in cartilage regions. Source: Images courtesy of Mr. Aniket Topaldi, University of California at San Francisco, USA.

In addition to DL-based reconstruction, MRI can be accelerated in the slice direction with image super resolution. The impact of DL super resolution on MRI biomarkers of knee OA was investigated. It was found that DL super resolution could improve the quality of morphological images for detecting knee joint pathology without biasing quantitative biomarkers estimation.⁶⁰

Large datasets may be needed for DL training. A recent public release of a knee dataset for quantitative T₂ mapping that also includes cartilage segmentation and pathology annotations could be found in.⁶¹ Also, augmentation schemes and mixtures of FS and undersampled data for training to better utilize data in a dataset were discussed.^62,63

Joint learning of sampling pattern (SP) and DL reconstruction was proposed in the literature.⁶⁴ The results showed that the quality of spatiotemporal undersampling could be improved by 20% or more if learned together with the DL reconstruction parameters. In Fig. 4, some results of the literature⁶⁴ illustrate how the learned SP may differ depending on the choice of DL architecture for reconstruction. Improvements can be easily seen in the joint learning approach when compared against the DL reconstruction with fixed SP, such as variable density with Poisson disc (VD + PD).

FIGURE 4:

Joint learning of the sampling pattern (SP) and deep-learning (DL) reconstruction was proposed, showing improvements over only learning of a DL reconstruction and a fixed SP.⁶⁴ Note that the variable density with Poisson disc (VD + PD) SP is one of the best human-made SPs for accelerated MRI. (a) VD + PD SP and UNET recon. (b) VD + PD SP and VN recon. (c) Joint learning SP and UNET recon. (d) Joint learning SP and VN recon.

While quantitative MRI maps are usually computed from the reconstructed relaxation-weighted images, several studies have shown that it may be advantageous to compute the quantitative maps directly from k-space data or undersampled images, particularly when DL methods are used. A deep neural network for cartilage T_1ρ and T₂ mapping was used together with spatiotemporal undersampling.⁶⁵ The results showed that it was possible to reduce the number of relaxation-weighted images from 8 to 2 (a temporal acceleration factor of R = 4), while also using k-space undersampling with acceleration factors of 2–6, obtaining a maximum combined spatiotemporal acceleration of R = 24. Similar spatiotemporal undersampling levels with neural network recovery were also demonstrated by⁶⁶ for cartilage T_1ρ mapping.

Improved training with GANs was demonstrated in the study by Liu et al,⁶⁷ where GANs produced sharper cartilage T₂ maps. Self-supervised learning for quantitative cartilage mapping with neural networks was also proposed in the literature,⁶⁸ improving training when FS data are not available. Figure 5 shows a diagram block of the proposed method, called reference-free latent map extraction (RELAX), with some illustrative results of cartilage T₁ and T₂ maps compared to zero-fill. The results showed that no reference maps were needed for model training.

FIGURE 5:

Self-supervised learning for quantitative mapping with neural networks was proposed, where no reference images were needed for the learning process.⁶⁷ The method relays on k-space consistency and image priors. Source: Images courtesy of Dr. Fang Liu, PhD, Harvard Medical School, USA.

Updates on Improved RF Pulse Sequences and Sequence Optimization

Improved RF pulse sequences are essential for acquiring data efficiently. While advanced reconstruction methods can remove undersampling artifacts and other imperfections, it is ultimately the data acquisition pulse sequence that defines how long the acquisition takes and how much SNR can be obtained from a particular scan. In this sense, the accuracy and precision of the measured k-space data are extremely dependent on the chosen pulse sequence and its parameters.

Data-driven approaches based on Cramer-Rao lower bounds (CRLB) and matched sampling-fitting methods were used to optimize the spin-lock times (TSLs) in mono-exponential cartilage T_1ρ mapping.⁶⁹ While it is known that two T_1ρ-weighted images are the minimum necessary for T_1ρ mapping, this study showed that the right choice of the TSLs could improve image quality and parameter estimation, obtaining results better than if three or more nonoptimized TSLs were chosen. The CRLB approach to choosing the TSLs was extended to bi-exponential and stretched-exponential T_1ρ mapping, showing that nearly the same quality can be obtained using half of the number of TSLs if they were optimized.⁷⁰

Efficient use of phase-cycling acquisitions in magnetization-prepared angle-modulated partitioned k-space spoiled GRE snapshots (MAPSS) sequences was proposed.⁷¹ To reduce the T₁ contamination that affects the precision of cartilage T_1ρ mapping, MAPSS required essentially two acquisitions with the same TSLs, one regular and the other with phase cycling. Peng et al⁷¹ demonstrated that a new complex-valued curve-fitting model could solve the issue of T₁ contamination and acquisition time could be reduced. In the study by Han et al,⁷² it was shown that MAPSS could be used efficiently to directly measure R₂–R_1ρ maps of knee cartilage without having to acquire data to produce T₂ and T_1ρ maps. However, the right preparation time (TSL and TE) must be chosen, as discussed in the literature.⁷²

Data-driven optimized variable flip-angles (OVFA) were used to improve the accuracy and precision of magnetization-prepared spoiled gradient echo (MP-GRE) and MAPSS sequences.⁷³ In this study, it was demonstrated that VFA could be optimized to reduce undesired effects in the signal evolution, improving accuracy by reducing T₁ contamination and k-space filtering effects in cartilage T_1ρ mapping, while increasing signal strength and consequently SNR. The optimized VFA could also reduce the dead time of the sequence, thereby reducing acquisition time. This is particularly important to improve the resolution of cartilage T_1ρ maps, as shown in Fig. 6.

FIGURE 6:

Examples of the use of data-driven optimal variable flip-angles (OVFA) in T_1ρ mapping.⁷³ The OVFA methods can be used to improve the SNR of magnetization-prepared angle-modulated partitioned k-space spoiled gradient echo snapshots (MAPSS), as in (a); to reduce the scan time of magnetization-prepared spoiled gradient echo (MP-GRE), as in (b); or even for other objectives, such as better resolution with less scan time, allowing for some decrease in SNR, as in (c). All images are produced from acquisitions using fully sampled data.

New spin-lock pulses for T_1ρ mapping have been proposed in the literature⁷⁴ using adiabatic pulses, and in the studies^75,76 using self-compensated and balanced spin-lock schemes. These new preparation pulses could be used in sequences such as MAPSS and MP-GRE to improve imaging robustness due to B0 and B1 inhomogeneities.

New extensions of DESS sequences have appeared in the study by Zijlstra et al,⁷⁷ with the proposal of a new sequence called multiple-echo steady-state (MESS) to jointly produce T₂ mapping and water–fat separation with chemical-shift correction. It was shown that quantitative DESS (qDESS) could be used to measure B₀ field inhomogeneity.⁷⁸ The diagnostic quality of a 5-minute (min) sequence based on qDESS was evaluated to detect knee joint pathology, showing high interreader agreement between the new 5-minute sequence and a conventional 20-minute MRI.⁶⁰ Figure 7 illustrates some of the images obtained in the study.

FIGURE 7:

(a) Sagittal T₂-weighted MR image shows an intact anterior cruciate ligament (ACL) graft with conspicuity (arrow), (b) quantitative double-echo steady-state (qDESS) MR image with mixed T₂- and T₁-weighted echo shows conspicuity of the anterior cruciate ligament graft (arrow) similar to (a), (c) qDESS T₂ map (values in milliseconds) shows focal increased T₂ relaxation time (arrow) of the trochlear cartilage despite morphologically normal cartilage identified in (a) and (b).⁶⁰ Source: Images courtesy of Dr. Akshay Chaudhari, PhD, Stanford University, USA.

Multiexponential Methods and UTE Sequences

Tissues may be composed of water in different compartments. In cartilage and other musculoskeletal tissues, there are two primary water components: fast-relaxing water tightly bound to the macromolecular matrix and slow-relaxing, freely mobile bulk water. The different water components of musculoskeletal tissues may lead to multiexponential relaxation within an image voxel. One common model to characterize multiexponential decay within a voxel is to use a two-compartment bi-exponential model. The bi-exponential models could provide better discrimination between healthy subjects and OA patients (KL-1,2) for early detection of OA, as illustrated in Fig. 8.

FIGURE 8:

Representative examples of multicomponent T_1ρ results of the knee cartilage, from age-matched (55 ± 2 years old) healthy control and early OA patients (KL-1,2). Maps refer to mono-exponential T_1ρ (T1ρ_m), bi-exponential long T_1ρ component (T1ρ_l), bi-exponential short T_1ρ component (T1ρ_s), and bi-exponential fraction of the short component (f_s). Left panel: medial cartilage and Right panel: lateral cartilage. Source: Unpublished data, courtesy of Dr. Ravinder Regatte, PhD, New York University, USA.

Most RF pulse sequences for quantitative mapping can measure tissues with relaxation times greater than 5 msec. However, some tissues in the knee joint including the deep and calcified cartilage, ligaments, tendons, and menisci have shorter T₂ relaxation times where the use of UTE or zero echo time (ZTE) sequences is needed to capture a signal.⁷⁹ This is important in clinical applications, such as OA,⁸⁰ where UTE and ZTE can better quantify the short T₂/T_1ρ tissues such as calcified cartilage, menisci, ligaments and tendons. UTE cones and adiabatic T_1ρ were used to assess cartilage degeneration,⁸¹ while the study by Wu et al⁸² also shows that adiabatic T_1ρ was more robust to magic angle effects.

A stretched-exponential fitting method was described to measure T₂ and T_1ρ relaxation of the intervertebral discs (IVD) of the spine.⁸³ The results demonstrated that the extra parameter of the stretched-exponential had a monotonic relationship consistent with reported variations in PG content within the IVD.

A method for rapid bi-exponential ${\text{T}}_2^{*}$ analysis was proposed in the litertature⁸⁴ using a single scan ramped hybrid-encoding (RHE) method. The approach produced bi-exponential maps of the knee joint with a 9-minute scan time. A bi-exponential analysis of ${\text{T}}_2^{*}$ relaxation time in the cartilage is shown in Fig. 9. A tricomponent model for ${\text{T}}_2^{*}$ analysis of cortical bone was described, which significantly improved the estimation of bound-water and pore-water in cortical bone.⁸⁵

FIGURE 9:

Mono and bi-exponential ${\text{T}}_2{}^{*}$ analysis of the knee cartilage. (a) For the more superficial cartilage region, the ${\text{T}}_2{}^{*}$ values of short and long components as well as the short fraction were 0.55 msec, 18.0 msec, and 11.0%, respectively. (c) For the deeper cartilage region, the values were 0.38 msec, 6.5 msec, and 27.0%, respectively. Regions marked in (b). Source: Images courtesy of Dr. Jiang Du, PhD, University of California at San Diego, USA.

MR Fingerprinting and Multicontrast Methods

Quantitative MRI mapping can reveal important information about knee joint tissues, particularly cartilage. In this sense, acquiring additional relevant parameters can provide more complementary and corroborative information. Acquiring several quantitative maps (eg T₁, T₂, and T_1ρ) separately can be very slow and inefficient. MRF⁸⁶ and multicontrast methods are efficient techniques to simultaneously estimate multiple MRI relaxation parameters. In general, they combine an improved pulse sequence with undersampling for an accelerated comprehensive multiparametric quantitative MRI mapping.

A novel MRF sequence was implemented for simultaneous T₁, T₂, and T_1ρ relaxation mapping of the human knee joint at 3 T.⁸⁷ With the use of T_1ρ preparation modules, the MRF sequence showed the feasibility of simultaneously estimating proton density (PD), T₁, T₂, T_1ρ, and B₁. The in vivo experiments also showed that multi-parameter cartilage relaxation time measurements from the novel MRF sequence could distinguish early OA patients from healthy subjects. The feasibility of the MRF sequence for simultaneous PD, T₁, T₂, T_1ρ, and B₁ mapping of the hip joint was further demonstrated in the literature.⁸⁸ The experiments showed excellent repeatability of PD, T₁, T₂, T_1ρ, and B₁ parameter estimations. The same group investigated the age-dependent changes of knee cartilage using a 3D MRF sequence for T₁, T₂, and T_1ρ maps in healthy subjects.⁸⁹ Higher global and regional cartilage T₂ and T_1ρ measurements were found for higher age groups as shown in Fig. 10.

FIGURE 10:

(a) Boxplot of cartilage T₁, T₂, and T_1ρ values for the different age groups, (b) illustrative cartilage proton density (PD), T₁, T₂, and T_1ρ maps of the lateral cartilage.⁸⁹ LFC-lateral femoral cartilage, LTC lateral tibial cartilage, MFC-medial femoral cartilage, MTC-medial tibial cartilage, PC-patellar cartilage.

A 3D MRF method was proposed for high-resolution (0.5 mm²) knee cartilage PD, T₁, and T₂ mapping.⁹⁰ Different trajectories were implemented in the k_xy and k_z directions followed by the singular value decomposition (SVD) compression reconstruction method to reduce reconstruction time. An algorithm was proposed to optimize the reconstruction of 3D MRF data.⁹¹ The low-rank property of the data was exploited and a subspace projection scheme was adopted to improve the accuracy of parameter estimation. Significant improvements in MRF time-series images were observed based on numerical experiments, and the technique was able to provide more accurate parameter maps.

Machine learning models have been investigated with quantitative MRI maps derived from MRF cartilage data in recent years. The accuracy and feasibility of different machine learning models using quantitative MRI parameters in predicting cartilage matrix components were compared.⁹² In this study, the ground truth was obtained from digital densitometry and polarized light microscopy. The Gaussian process regression (GPR) showed better performance than random forest, support vector regression, gradient boosting, and multilayer perceptron. A prospective, intraindividual ex vivo study was conducted to evaluate whether multiparametric quantitative MRI techniques could predict cartilage composition.⁹³ In the study, T₁, T₂, T_1ρ, and ${\text{T}}_2^{*}$ parameter maps of the knee joints were used to train DL models with micro-spectroscopically determined local proteoglycan and collagen fibers contents as ground truth. The results showed that advanced machine learning techniques could be used to determine compositional features of cartilage based on quantitative parameters with potential implications for the diagnosis of early cartilage degeneration and the monitoring of therapeutic outcomes.

DL-Based Segmentation

Cartilage segmentation for morphometry is still one of the most challenging tasks for quantitative assessment of OA as described in a recent review.⁴⁸ Currently, DL approaches have been among the most commonly used and efficient methods for segmenting cartilage. Cartilage segmentation and quantification are important for understanding disease-related structural changes in subjects with or at risk for OA⁴⁶ and serve as a pre-requisite to extracting relevant cartilage compositional MRI biomarkers such as T₂ relaxation time.⁹⁴

Manual segmentation, though usually used as a gold standard, is labor-intensive and subjective to operator bias and error. DL-based segmentation is a good alternative since it can be very fast. The recent development of DL methods for segmenting cartilage on knee MRI has been described in.^95–97 In the study,⁹⁵ a proposed 2D U-Net-based model showed high agreement with manual segmentation while preserving longitudinal reproducibility of quantitative cartilage morphometry from MRI knee datasets. In the study,⁹⁶ 3D convolutional neural networks (CNN) were used to investigate the 3D bone morphology in knee OA by segmenting the patella and distal femur cortex. The technique showed high accuracy in delineating the indistinct interfaces between bone and other joint structures. Transfer learning, also known as domain adaption, is a technique to transfer previously acquired knowledge to new problems (i.e. new datasets or tasks). In the study,⁹⁷ a GAN was combined with transfer learning using a small training dataset and showed comparable performance to humans in knee cartilage segmentation on heterogeneous knee MR images.

MRI-Based Biomechanical Assessment

Cartilage is a load-bearing structure and is subject to changes in biomechanical function with several factors including aging, trauma, pathology, and overuse. Recent techniques have shown the feasibility of compositional MRI in assessing cartilage overuse. The feasibility of ${\text{T}}_2^{*}$ mapping of the knee cartilage in assessing the response to mechanical loading in skiing was demonstrated.⁹⁸ More specific biomechanical measures, such as strain mapping, were evaluated in the knee cartilage,⁹⁹ as shown in Fig. 11, and intervertebral discs.¹⁰⁰ Displacement encoding with stimulated echoes (DENSE) MRI was used on human cartilage during cyclic varus loading to assess displacement and strain.¹⁰¹ Also in the study by,¹⁰¹ T_1ρ maps were acquired before and after varus loading to investigate changes with this kind of loading. This emerging topic shows the utility of assessing biomechanical function using noninvasive MRI methods under mechanical loading.

FIGURE 11:

(a) A photograph of the experimental setup used to asses cartilage strain maps under mechanical loading during MRI scan.⁹⁹ (b) Illustration of the process, consisting of an 18 minutes continuous acquisition using golden-angle radial acquisitions while the knee joint at rest, during mechanical loading, and during recovery. The image reconstruction is performed with compressed sensing methods, followed by motion correction using optical flow methods, and finally, Lagrangian strain is calculated to produce cartilage strain maps.

Summary of Discussion and Future Directions

Current Challenges of Emerging Methods

Although recent emerging quantitative MRI techniques have many advantages for accelerating cartilage compositional MRI, several challenges need to be addressed before their widespread application in musculoskeletal imaging. For example, the quality of many CS reconstructions depends on the choice of regularization parameters. In retrospective CS studies, the parameter is usually optimized to provide an image as similar as possible to the FS reference image. However, this is not possible in prospective studies leading to the need for methods for automatically choosing the regularization parameters.

Recently, DL methods showed promising initial results for image reconstruction, fully automated cartilage segmentation, and quantitative compositional mapping. One challenge when using DL approaches is the difficulty in obtaining large training datasets to optimize model performance. Simulated datasets have been recently used to generate large training datasets with good results. Alternatively, studies have utilized GANs to synthesize MR images for augmenting the training datasets. A more comprehensive investigation of the robustness and generalizability of DL-based cartilage compositional mapping techniques is also needed.

The MRF framework is highly flexible for multi-parametric cartilage mapping but still poses a significant computational burden and scan time as the dimensionality increases. Rigorous technical improvements and robust validations are necessary for pulse sequence optimization, faster image reconstructions, clinical pathology assessment, and quantitative metrics for error analysis before the widespread use of MRF for musculoskeletal imaging.

Standardization of Compositional MRI Techniques across Different Scanners, Vendors, and Institutions

T₂ and T_1ρ mapping are by far the most common cartilage compositional MRI techniques used in clinical practice and OA research studies.⁶ However, cartilage T₂ and T_1ρ relaxation time are influenced by multiple factors including the type of scanner, coil, sequence, and imaging parameters used to obtain the measurement.^102,103 Coefficients of variation for scan–rescan reproducibility for the six articular surfaces of the knee joint using the same scanner, coil, and imaging protocol have ranged between 2.3% and 6.5% for T₂ and between 4.2% and 7.4% for T_1ρ with much higher coefficients of variance for sub-regional and laminar analysis.^104,105 This reproducibility is adequate for cross-sectional and longitudinal OA research studies performed at a single institution. However, additional standardization of cartilage T₂ and T_1ρ measurements is needed for multicenter OA research studies and clinical trials.

Recent studies have investigated the scan–rescan reproducibility of cartilage compositional MRI techniques across different scanners, vendors, and institutions. Coefficients of variation for reproducibility of cartilage T₂ and T_1ρ measurements of the knee joint for subjects evaluated at three separate institutions using the same 3 T scanner model from the same manufacturer with the same imaging protocols were 4.4% for T₂ and 4.9% for T_1ρ.¹⁰⁶ However, coefficients of variance increased to 10.1% for T₂ and 8.1% for T_1ρ when the same imaging protocols were performed on three different 3 T scanner models from three different manufacturers.¹⁰⁷ Coefficients of variance were even higher when different compositional MRI sequences were used to measure cartilage T₂ and T_1ρ on different scanner models from different manufacturers.^103,106,107 These inherent variations in cartilage T₂ and T_1ρ relaxation time may serve as confounding factors that may make it challenging to detect changes in cartilage MRI biomarkers due to degeneration.

Current studies suggest that with careful quality control and cross-calibration, compositional MRI can be applied in multicenter research studies and clinical trials for evaluating cartilage degeneration if the same imaging protocol is used on the same scanner model from the same manufacturer.¹⁰⁶ However, additional future collaborative efforts between researchers and MRI vendors will be needed to accomplish the ultimate goal of better standardization of cartilage MRI biomarkers across different scanner models from different manufacturers at different institutions.¹⁰⁸ This will require all MRI vendors to agree on a single sequence for measuring cartilage T₂ and T_1ρ relaxation time and then work together to limit variability in factors such as MRI system and coil design, B₀ homogeneity profiles, shimming algorithms, and reconstruction and postprocessing methods to better standardize quantitative measurements.¹⁰⁷

Clinical Translational Impact on MSK Diseases

Quantitative MRI techniques are highly reliable for detecting early OA and assessing disease- and treatment-related changes in cartilage composition and ultrastructure.⁶ However, the current widespread translation of cartilage compositional MRI techniques into clinical practice and multicenter research studies has been limited by multiple factors including long acquisition times, inefficient image analysis methods, and lack of standardization of quantitative MRI measurements. There have been many recent advances in compositional MRI pulse sequence acceleration which have greatly reduced acquisition times. New DL methods have also been described for rapid and fully automated cartilage segmentation, which is the first and most time-consuming step in the image analysis process. However, much additional work is needed to improve the standardization of cartilage compositional MRI techniques.

Standardization of rapidly acquired and efficiently analyzed cartilage MRI biomarkers would allow for large multicenter OA research studies and a better comparison of quantitative imaging data acquired at different institutions. Cartilage T₂ measurements obtained on older adults in the Osteoarthritis Initiative (OAI) have led to a wealth of important information on knee OA.¹⁰⁹ Imagine the potential knowledge obtained if rapidly acquired measurements of cartilage T₂ and T_1ρ could be standardized and analyzed from larger and more diverse patient populations as part of routine clinical care at hundreds of institutions worldwide. Larger patient populations evaluated during a clinical knee MRI examination would facilitate cost-effective research for a better understanding of disease mechanisms and a more rapid evaluation of new surgical and pharmaceutical therapies. In addition, standardization of cartilage MRI biomarkers would allow for the development of normative data for cartilage T₂ and T_1ρ relaxation times on different articular surfaces for individuals of different ages, gender, and ethnicity. The development of normative data and threshold values could be used to define abnormal cartilage compositional values and values associated with different disease burdens and different risks for disease onset and progression.

Summary

This review of compositional cartilage MRI has systematically introduced and discussed the current state-of-the-art methods as well as recent emerging techniques for the quantitative assessment of cartilage. The potential challenges and opportunities for emerging techniques were also discussed including standardization across different scanners, vendors, and institutions. It is expected that some of these emerging accelerated cartilage compositional MRI techniques will be translated to widespread clinical use in the near future to improve the detection of disease onset and progression, enhance treatment monitoring, and ultimately aid in the management of patient care.