Current status of functional MRI of osteoarthritis for diagnosis and prognosis

Abstract

Purpose of review

Osteoarthritis is a major source of disability, pain and socioeconomic cost worldwide. The epidemiology of the disorder is multifactorial including genetic, biological and biomechanical components, some of them detectable by MRI. This review provides the most recent update on MRI biomarkers which can provide functional information of the joint structures for diagnosis, prognosis and treatment response monitoring in osteoarthritis trials.

Recent findings

Compositional or functional MRI can provide clinicians with valuable information on glycosaminoglycan content (chemical exchange saturation transfer, sodium MRI, T_1ρ) and collagen organization (T₂, T₂*, apparent diffusion coefficient, magnetization transfer) in joint structures. Other parameters may also provide useful information, such as volumetric measurements of joint structures or advanced image data postprocessing and analysis. Automated tools seem to have a great potential to be included in these efforts providing standardization and acceleration of the image data analysis process.

Summary

Functional or compositional MRI has great potential to provide noninvasive imaging biomarkers for osteoarthritis. Osteoarthritis as a whole joint condition needs to be diagnosed in early stages to facilitate selection of patients into clinical trials and/or to measure treatment effectiveness. Advanced evaluation including machine learning, neural networks and multidimensional data analysis allow for wall-to-wall understanding of parameter interactions and their role in clinical evaluation of osteoarthritis.

INTRODUCTION

Osteoarthritis is a debilitating disease that causes pain and affects mobility; degeneration of joint structures is a major manifestation of the disease process. Over more than a decade a number of MRI-based biomarkers for collagen and proteoglycan content in articular cartilage affected by osteoarthritis were developed. These compositional (often referred to as functional or biochemical) markers are related to collagen content and organization [1,2], proteoglycan content [3–5] and mechanical properties [6,7] and they can also predict focal cartilage and meniscus disruption [8]. T₂ relaxation constant mapping is a well investigated technique reflecting the combination of water content and collagen matrix organization [1,9,10]. Proteoglycans with covalently attached glycosaminoglycan (GAGs) can be potentially assessed with T_1ρ (relaxation in rotating frame) [4], glycosaminoglycan chemical exchange saturation transfer (gagCEST) [3] and sodium MRI [11]. All these GAG-specific contrast-free techniques are preferred over conventional delayed Gadolinium Enhanced MRI of Cartilage (dGEMRIC) [12] as the Federal Food and Drug Administration recently warned about gadolinium-based contrast agents which can be potentially harmful for patients due to gadolinium deposition in central nervous system structures [13]. Some joint structures such as ligament, tendons and subchondral bone tendons are characterized by ultrashort relaxation rates resulting in loss of magnetic resonance (MR) signal in conventional MRI images. Ultra-short echo time techniques are able to either acquire signal directly from these structures [14] of even calculate T₂* and T_1ρ relaxation to describe the collagen fibres organization [15,16]. Diffusion weighted imaging and diffusion tensor imaging also provide an important information about cartilage status by calculating apparent diffusion coefficient [17] and diffusion tensor map [18], respectively. Modern computer science approaches are pushing the field further by implementing machine learning algorithms for image postprocessing and multidimensional data analysis for searching the relations between various biomarkers. Osteoarthritis as a complex multifactorial joint disease which is characterized by structural alteration including GAG loss and collagen remodelling, but is also accompanied by volumetric changes. Here, the deep learning approach helps to automatically segment joint structures such as cartilage and meniscus which is a typically extremely demanding task in terms of man-power [19^■]. All these modern methods have a common goal: to establish quantitative MRI biomarkers as reproducible and reliable indicators of joint structures suffering from osteoarthritis. This review summarizes the most recent findings and improvements in the field of MRI-based biomarkers for osteoarthritis assessment and their usage for diagnosis, prognosis and treatment monitoring.

T₂ MAPPING

The transversal relaxation constant, T₂, in cartilage reflects the interaction of water molecules and the extracellular matrix on a molecular level. Anisotropy of tissue results in different T₂ values from deep to superficial cartilage layers depending on the collagen fibre orientation. Previously, T₂ was validated as a marker for osteoarthritic alterations in cartilage [20], as a physiological indicator of loading induces changes in cartilage [21] and also as a biomarker for monitoring the patients after surgical treatment of cartilage lesions [22]. Advances in MR hardware, mostly in increasing main magnetic field strength, accompanied with the challenges such specific-absorption limitations and special radiofrequency coils need, initiated using novel MR sequences for T₂ mapping, such as triple-echo and double-echo steady-state-based approaches [22,23]. T₂ can be also exploited as a predictive marker for focal lesion development which may result in osteoarthritis. Kretzschmar et al. [24] investigated T₂ mapping for early osteoarthritis changes in cartilage of 57 patients and they showed that both diffuse and focal changes of cartilage composition precede the development of cartilage lesions. Evaluation of T₂ in cartilage can be further enhanced and standardized using human-independent machine learning techniques. Pedoia et al. demonstrated the machine learning capabilities using whole Osteoarthritis Initiative (OAI) Dataset (>4000 individuals). Their results showed that feature learning from T₂ maps has a great potential in uncovering information that can prospectively help to better diagnose osteoarthritis [25^■■]. This concept can be extended and further improved by combining other features with quantitative MR, such as biomechanical parameters. In another study of Pedoia et al. [26^■■], it was shown that the integrated parameter set of demographic, clinical information, gait kinematics and kinetics, cartilage compositional T_1ρ and T₂ provides an information about their cross-interactions and may serve as a robust early predictor of cartilage lesion progression. A combined network of morphological MRI, biomechanics and quantitative MRI is depicted in Fig. 1.

FIGURE 1.

Combined network of morphological MRI, biomechanics and compositional MRI. The network shows a gradient with severe patients appearing in the lower right (marked with dashed black circles) and less severe in the upper left based on Kellgren–Lawrence grading. Reproduced with permission [26^■■].

T_1ρ MAPPING

T_1ρ is sensitive to low-frequency interactions of macromolecular protons and bulk water as well as early biochemical deterioration of cellular matrix [16]. T_1ρ imaging of connective tissues affected by osteoarthritis may provide a more systematic assessment of this disease. In the previous studies the correlation was found between both proteoglycan-specific methods like gagCEST [27] and collagen-specific methods like T₂ [28]. Atkinson et al. [29] demonstrated using a complex meta-analysis that increased T_1ρ and T₂ values in cartilage may serve as a risk factors for osteoarthritis. Higher signal-to-noise ratio increases the accuracy of T_1ρ calculation, using higher field strengths, for instance; Wyatt et al. [30] found T_1ρ at 3 and 7 T significantly higher in the lateral femoral condyle and patella in patients with osteoarthritis; however, more regions were significant at 7 T compared with 3 T. The recent meta-analysis demonstrates that T_1ρ provides more discrimination than T₂ for osteoarthritis [31]. Advances in MRI techniques and image postprocessing allow for detecting more than a single component of T_1ρ in connective tissues which may help to better understand the osteoarthritis pathogenesis in the future [32,33^■].

GLYCOSAMINOGLYCAN CHEMICAL EXCHANGE SATURATION TRANSFER

Glycosaminoglycan-specific chemical exchange saturation transfer (gagCEST) is a recently developed technique, which provides information about the amount of GAG in articular cartilage. It is based on selective saturation of the exchangeable hydroxyl protons of GAG, which exchange with water protons [3]. The feasibility of using gagCEST for GAG content determination has been previously validated by number of studies [27,34,35]. Further development aims towards faster clinical MRI acquisition methods and more quantitative acquisition and analysis routines [36]. Stability and feasibility of gagCEST imaging to detection of early cartilage damage was investigated by Brinkhof et al. [37^■■]; they developed a 7-min gagCEST protocol and used it to calculate intraclass correlation coefficient (ICC) to show high reproducibility in the medial femoral condyle (ICC = 0.87) and in the lateral femoral condyle (ICC = 0.97) and they also found statistically significant change in GAG between damaged and healthy cartilage. In terms of technical development, Windschuh et al. [38] investigated the effects of a frequency drift on gagCEST contrast in the human knee at 7 T and suggested a retrospective correction method that can eliminate errors. This technique is illustrated in Fig. 2. The water T₂, and thus the frequency width of the water peak, may impact the extent of the so-called spillover effect in CEST. Peterson et al. [39^■■] demonstrated that T₂ substantially influences gagCEST which is obvious especially in femoral cartilage with a wide range of T₂ and showed this effect to be more pronounced at 3 T compared with 7 T as the gagCEST effect is weaker at lower field strength.

FIGURE 2.

Cropped M₀ images superimposed with glycosaminoglycan chemical exchange saturation transfer_M0 contrast in patellar and femoral cartilage from identical measurements from a 32-year-old healthy volunteer, acquired at t = 5 min (a, d, g, j), t = 17 min (b, e, h, k), and t = 29 min (c, f, i, l). The glycosaminoglycan chemical exchange saturation transfer_M0 contrast of data using the standard B₀ correction (a, b, c) shows an increase over time, whereas the dynamic WASABI (simultaneous mapping of water shift and B₁) correction (d, e, f) and the dynamic WASABI/phase correction method (g, h, i) provide a stable contrast. The last row represents data from the same individual acquired 1 month later (j, k, l). Reproduced with permission [38].

SODIUM MRI AND PET/MAGNETIC RESONANCE

Sodium MRI is another technique which has potential to noninvasively determine GAG content in human articular cartilage; however, it requires special radiofrequency hardware (dedicated coils, ultra-high field scanners) and complicated postprocessing [40]. Despite that, the number of studies was performed demonstrating a high sensitivity and reproducibility of sodium MRI in detecting GAG content levels in cartilage either in osteoarthritis [5,41] or in cartilage repair [42,43]. In a recent study, Madelin et al. [44^■■] used two sodium MR sequences (radial three-dimensional and fluid suppression by adiabatic inversion recovery) to investigate whether they can detect the sodium volume changes in patients with osteoarthritis after 16 months. A significant decrease of measured apparent sodium concentration in cartilage using IR in both femoral condyles and patella proved the sensitivity of sodium MRI to detect subtle changes of GAG concentration. Fluid suppression in sodium MRI was also investigated by Xia et al. [45] they developed quadrupolar jump-and-return pulse sequence to suppress the fluid signal from the artery and enhance the contrast of knee cartilage in vivo. Another modern approach towards the advanced evaluation of osteoarthritic alterations in joints is the hybrid PET/MRI scanning. Wandler et al. [46] investigated the potential of ¹⁸F-fluoro-deoxyglucose uptake pattern within the shoulder cartilage and they found out it can be is associated with signs and symptoms of osteoarthritis or bursitis. Kogan et al. [47] used PET/MR to assess the metabolic activity in osteoarthritis and concluded that PET/MR may detect metabolic abnormalities in subchondral bone, which appear normal on MRI. In another study by Savic et al. [48], significant cartilage and bone interactions were demonstrated in osteoarthritis of the knee joint using simultaneous [¹⁸F]-sodium fluoride PET/MR.

CARTILAGE VOLUMETRIC MEASUREMENTS

Cartilage tissue loss is a hallmark of osteoarthritis. Thinning and cartilage/bone deformation can be manually assessed from high-resolved morphological MR images, however, this approach requires enormous workload. There are many techniques for automated cartilage segmentation which were developed in recent years, including intensity-based [49] and edge-based [50] approaches, deformable models [51], clustering [52] and atlas/graph-based methods [53]. Machine learning has vast ranging applications including automated segmentation of cartilage allowing for reliable segmentation using convolutional neural networks (CNN) [19^■,54–56]. In addition to knee cartilage, other joints have been segmented using automated techniques such as hip cartilage [57], shoulder [58], bones [59] or even wrist cartilage [60]. The logical translation of automated cartilage segmentation is the use for quantitative MR assessment which is also often related to tedious manual segmentation. Hesper et al. [61^■■] found automated hip cartilage segmentation for dGEMRIC assessment to be reliable and reader-independent approach for assessment of biochemical cartilage status. Norman et al. used a CNN-based approach for automated T_1ρ assessment demonstrating the ability to quantify, in a longitudinally repeatable way, relaxometry and morphology in a single session [55].

ADVANCES IN IMAGE POSTPROCESSING

Machine learning is more and more employed in extracting the predictive or diagnostic value from the existing multidimensional patient data. Up-to-now, there were many attempts to developed patient-specific prediction models for knee osteoarthritis by analyzing multivariable ‘big data’ using either conventional statistical approach [62] or machine learning [63]. To classify early stages of osteoarthritis is also of interest for machine learning; Ashinsky et al. [64] used such an approach to predict early symptomatic osteoarthritis by combining T₂ mapping and clinical outcome variables and they found the classification accuracy to be 0.75 ± 0.9%. Lazzarini et al. [65] proposed five highly predictive small models that can be possibly adopted for an early prediction of knee osteoarthritis where all the models showed high performance (area under curve >0.7). Sodium imaging was also tested with various machine learning to classify osteoarthritis patients with accuracy ranging from 63 to 78% [66]. Du et al. [67^■■] suggested a novel approach to predict knee osteoarthritis combining four machine learning methods and the clinical data. In general, selecting the best variables to incorporate into prediction models can be a complex task owing to the great number of variables to choose from [68].

Quantitative MR analysis can be extended using texture analysis techniques, typically feature extraction from grey-scale co-occurrence matrix (GLCM). Texture analysis provides a reliable tool for assessing knee osteoarthritis with more sensitive detection of cartilage degeneration compared with the simple mean T₂ or T_1ρ value in an identical region of interest. As cartilage is a complex-shaped complex organ, original texture analysis suggested by Haralick [69] needs to be modified. Peuna et al. [70^■■] developed a cartilage-specific GLCM analysis algorithm taking into account-specific cartilage features such as curvature and angulation. Joseph et al. [71] studied osteoarthritis patients using conventional T₂ maps and texture analysis; they found out individuals at risk for osteoarthritis have both higher and more heterogeneous T₂ values than controls and that individuals with cartilage abnormalities have elevated cartilage T₂ parameters compared with individuals without abnormalities (Fig. 3) [71]. Heilmeier et al. [72^■■] analyzed more than 300 individuals from the OAI and according to their findings combined clinical score, quantitative MR and texture can help predict the patient’s individual risk for an incident total knee replacement 4–7 years later.

FIGURE 3.

Representative T₂ maps from an individual from the control cohort (left) and an individual from the incidence cohort (right). Cartilage T₂ maps are median-filtered with a 3 × 3 kernel for visualization. Both individuals have no cartilage abnormalities and no pain; however, the individual from the incidence cohort has elevated mean T₂ (39.12 versus 33.39 ms), elevated grey level co-occurrence matrix variance (311.63 versus 190.50), elevated grey-scale co-occurrence matrix contrast (466.16 versus 266.82), and elevated grey-scale co-occurrence matrix entropy (7.17 versus 6.80). Reproduced with permission [71].

CONCLUSION

Functional compositional MR techniques provide an important extension to clinical, morphological and biochemical assessment of osteoarthritis diagnosis and progression staging. There are still remaining issues to be addressed such as lack of standardization and tedious image postprocessing; however, the clinical value of quantitative magnetic resonance imaging is undeniable already in the presence. Future challenges include advancements in imaging techniques, searching relationships between various markers by machine learning and multidimensional data analysis, and parameter validation using large freely available database such as the OAI.

KEY POINTS.

• Advances in image-processing, image reconstruction and data analysis yield better understanding of MRI-based markers for joint structure status.

• Quantitative MRI provides noninvasive markers for osteoarthritis diagnosis and prognosis.

• Recent developments allow for faster magnetic resonance scanning, reliable cartilage and meniscus segmentation and complex multidimensional data analysis.