Adjacent cartilage tissue structure after successful transplantation: a quantitative MRI study using T₂ mapping and texture analysis

The articular surfaces of the knee joint are covered by hyaline cartilage, which can withstand high repeated pressure and provides low-friction surfaces for joint motion [1, 2]. The prevalence of cartilage defects in the knee is estimated to be approximately 12% of the overall population [3]. In young adults, these defects are mostly the result of sport injuries, other trauma, or mechanical malalignment of the joint [4]. Subsequent damage to the chondrocytes leads to degeneration and death of cells, cartilage matrix degradation, decreased production of proteoglycans, increased water content, and loss of zonal architecture [2]. In cases where conservative treatment is not effective, bone marrow stimulating techniques (MFx) and cartilage restoration techniques, such as matrix-associated autologous chondrocyte transplantation (MACT), are often used [5‐8]. Previous studies have shown that the outcome of cartilage repair surgery can vary from hyaline-like tissue to mixed fibrous-hyaline tissue or fibrous tissue with inferior biomechanical properties often seen after MFx [9‐11]. Furthermore, the tissue adjacent to the cartilage repair site might be affected by degenerative changes in the perilesional zone caused by insufficient defect preparation [12] or suturing of articular cartilage [13].

Early stages of these changes in adjacent cartilage tissue are often undetectable on routine morphological scans. Quantitative MRI, especially T₂ mapping, which is sensitive to collagen matrix organization and water content [14, 15], has been proven to be a reliable non-invasive diagnostic technique [16, 17], as degradation of cartilage matrix leads to increased water mobility and subsequent T₂ elevation [15]. Differences in T₂ distributions are often visible on T₂ maps, which suggests that the analysis of spatial T₂ distribution might be a valuable tool for assessing changes of repair site and the surrounding tissue during standard post procedure follow-ups. Gray-level co-occurrence matrix (GLCM) texture analysis [18] has been explored in recent years in the context of detecting early osteoarthritic (OA) changes [19‐24]. The GLCM determines the co-occurrence of different signal intensities in a specific offset in the image, creating a co-occurrence matrix from which different quantitative features can be extracted. GLCM features (second-order image statistics) provide information about linear and quadratic relationships between pixel pairs. There are various properties of the GLCM calculation. The displacement distance (d) determines the distance between neighboring pixels [18, 25], the angle (direction) at which GLCM is calculated, and the number of gray levels. However, in most cases, the quantization into 16 gray levels is sufficient [25]. The angle (direction) at which GLCM is calculated is an important property to consider, because the structure of cartilage is not homogenous in all directions. Since the region of interest (ROI) might not be rectangular, ROI flattening was implemented in various studies to improve the reproducibility of GLCM calculation [21, 26]. The aim of this study was to assess the properties of the repair site after MACT and its adjacent cartilage tissue and compare these to healthy cartilage using quantitative MRI techniques (T₂ mapping) and advanced image texture analysis techniques, such as GLCM.

The mean MOCART score increased from 74.6 ± 12.2 at 12 months to 88.7 ± 8.8 at 24 months (for details, see Table 3). A significant decrease in mean T₂ values between the two post-surgical time points (12 and 24 months) was observed only for repair cartilage (p < 0.001). In terms of GLCM, significant changes (p < 0.05) over time were observed for repair cartilage in the following features: contrast, correlation, difference entropy, difference variance, dissimilarity, homogeneity, information measure, information measure of correlation 2, inverse difference moment normalized, and inverse difference normalized. Similarly, significant changes (p < 0.05) from 12 to 24 months were found in the following GLCM features in tissue adjacent to the repair site: contrast, correlation, difference entropy, difference variance, dissimilarity, homogeneity, information measure, inverse difference moment normalized, and inverse difference normalized. The changes in adjacent cartilage were similar to those for repair cartilage. In reference to cartilage, autocorrelation, sum average, sum of squares, and sum variance were significantly different between 12 and 24 months (p < 0.05). Mean absolute values with standard deviations and mean differences between time points are specified in Table 4. We found a moderate correlation between MOCART and autocorrelation (0.52), cluster shade (−0.53), sum of squares (0.53), sum average (0.53), and sum variance (0.51).

The following variables were modelled by means of linear-mixed effects models: mean T₂, autocorrelation, correlation, homogeneity, contrast, and difference entropy. Variables were chosen based on results from paired test results and correlations between features (Fig. 2). Because many GLCM features are highly correlated with each other, we analyzed only those that were easily interpretable in the context of cartilage texture and that satisfied model assumptions.

Our model confirmed a significant (p < 0.001) estimated decrease of T₂ of −5.82 (−8.23 to −3.41) ms from 12 to 24 months after surgery. The rate of decrease was different for each ROI (significant interaction between adjacent tissue and time, and reference tissue and time), showing a more stable mean T₂ value compared to repair cartilage (Fig. 3a).

The effect of time on autocorrelation was not significant (p = 0.08). Adjacent tissue (p < 0.001) and reference tissue (p < 0.001) had significantly higher mean values compared to repair tissue.

Homogeneity decrease in time was significant (p = 0.001). Adjacent tissue (p = 0.027) and reference tissue (p = 0.002) had significantly lower mean values compared to repair tissue. (Fig. 3c).

Models of the cube of correlation, cube root of contrast, and difference entropy showed a significant effect (p < 0.05) of time and a significant difference (p < 0.05) between repair and reference tissue. However, the difference between repair and adjacent tissue was not significant.

Summary tables of all models and estimated values can be found in the Supplementary material.

The purpose of this study was to evaluate the T₂ relaxation time and GLCM texture features in knee cartilage treated with MACT. We were specifically interested in MACT transplant maturation and changes in cartilage tissue adjacent to the repair site (Table 5).

Elevated T₂ relaxation time in cartilage is associated with degeneration or trauma and its decrease over time signals changes in cartilage structure [19, 33]. It has been shown that variation of T₂ in healthy hyaline cartilage was highly correlated with collagen anisotropy fibril angle [34]. Randomly oriented collagen fibers (e.g., in the superficial cartilage zone) allow more mobility of protons, and thus, higher T₂ relaxation times were found [35, 36]. Early in the maturation process, the MACT graft might have a fluid-like appearance, because the repair tissue is initially poorly organized and highly water-permeable [35‐37]. It has been hypothesized that a decrease in T₂ is connected to the reorganization of the matrix structure and a decrease in free water [35].

In our study, the analysis showed a significant decrease in mean T₂ between 12 and 24 months only in repair cartilage (∆ = 5.82 ± 6.48, p < 0.001). Decreased T₂ values in combination with a high morphological score (MOCART) indicated a favorable clinical outcome, which is in correspondence with the clinical results of this study published by Niemeyer et al. [27]. A decrease in the mean T₂ value of a selected ROI is very general and represents only one mean quantitative value for the whole area without providing information about changes in cartilage texture; therefore, texture analysis might be a useful add-on. GLCM features are statistical texture features that describe relationships between individual pixels [18]—information that is lost when comparing mean T₂ values. These features are highly correlated; hence, the selection of the most fitting features must be considered. This can be done either automatically using an algorithm, or manually, based on correlation and cluster analysis and an understanding of the image intended for analysis [38].

In our study, we used quantization, a process in which a range of values is compressed to a single quantum [39]. If the variance of T₂ values is large, similar values fall into one quantization bin, which results in larger areas of one gray level. In tissues with a smaller variance of T₂ values, quantization results in greater disparity in gray levels. This is the reason the texture of reference cartilage is richer. Moreover, the variability of T₂ values at 12 months among patients is high, which leads to high variability of features. Thus, the specific feature values might not be as important as the direction of differences between ROIs and the two time points. Since our cohort consisted of comparatively young individuals with focal cartilage defects, we can consider our reference to be healthy cartilage tissue.

We also found significant differences in autocorrelation, sum average, sum of squares, and sum variance in reference tissue between 12 and 24 months (Table 2). These features are linked to gray-level dispersion, which is higher in reference tissue. However, the features that are linked to the presence of edges or disorderliness were not significantly different compared to the other two areas. A possible explanation could be a subsidiary improvement of healthy cartilage, with the direction of change in these features the same as that in repair and adjacent tissue, although not significant.

Our analysis of GLCM features showed that autocorrelation [39] does not change between 12 and 24 months, but is significantly higher in adjacent and reference tissue compared to repair tissue, suggesting more prominent patterns and possibly the normal collagen fiber network in healthy reference cartilage. Concerning correlation, adjacent tissue was more similar to repair cartilage, but values approached those of the reference tissue between 12 and 24 months. The linear relationship between pixels is more predictable in repair cartilage and reveals a more uniform texture compared to healthy cartilage.

Images with more uniform gray levels result in a higher overall value of homogeneity [18]. Homogeneity was lower in adjacent (p = 0.027) and reference tissue (p = 0.002) compared to the repair site. In combination with contrast and difference entropy [18, 40], which both showed significant differences only between repair and reference cartilage, we see adjacent tissue might have been more similar to repair tissue than to reference cartilage, or is between those two, texture-wise.

After transforming contrast to −1/3, we expected lower values in reference tissue (higher values before transformation). Because feature values in repair and adjacent tissue approached the reference at 24 months, we assume that GLCM analysis of T₂ maps can show a positive effect of MACT on surrounding tissue.

In recent years, several articles about GLCM analysis of knee cartilage have been published. Although these articles concern osteoarthritis (OA), there may be some similarities between OA lesions and MACT grafts that have not fully maturated, namely, in elevated T₂.

Carballido-Gamio [22] reported elevated correlation, homogeneity, and entropy in osteoarthritic cartilage in the direction parallel to the cartilage surface between patient follow-ups. Joseph et al [19] reported that contrast was elevated in individuals with risk factors for OA. Schooler et al [21] performed a study with similar GLCM parameters (offset = 1, angle = 0° and 90°, flattened ROI) in patients with OA and healthy volunteers and they found elevated contrast 1, 2, and 3 years after baseline measurement and elevated entropy in the direction parallel to the cartilage. Heilmeier et al [41] reported an up to 40% higher risk for total knee replacement with a one-SD increase in contrast. Blumenkrantz et al reported elevated entropy in OA patients [24].

There are also few limitations that have to be considered. The number of cases was relatively low (n = 23). The first examination was done 12 months after MACT. At that time, the MACT transplant is expected to be fully matured and should not be as highly hydrated as it is less than six months after surgery. Therefore, a more rapid change in texture that could have occurred prior to the measurements in this study would have been missed. Due to the relatively large size of the repair tissue, reference cartilage was often chosen near the magic angle, but the overall increase of signal should not have affected the texture. The texture analysis was performed on three consecutive slices and not in a 3D manner, because of the non-zero slice distance used in multi-echo spin-echo T₂ mapping.

In conclusion, GLCM texture analysis is a useful add-on to T₂ mapping in the evaluation of knee cartilage after MACT by increasing the sensitivity to changes in cartilage structure. Some texture features, namely, correlation, homogeneity, contrast, and difference entropy, changed significantly in repair tissue and adjacent tissue between 12 and 24 months after surgery, indicating a sizable change in texture, and likely, collagen structure. The results of this study, in accordance with previous research, suggest that the use of texture analysis provides additional information not only about the state and maturation of the repair tissue but also about the texture and composition of the repair cartilage and adjacent morphologically normal-appearing cartilage tissue.