Optimization of knee cartilage texture analysis of quantitative MRI T2 maps

Abstract

Purpose: T2 mapping in magnetic resonance imaging (MRI) is commonly used technique in assessment of osteoarthritis (OA) in clinical research. Recently, gray-level co-occurrence matrix (GLCM) texture analysis is drawing more and more attention as it provides add-on evaluation of collagen specific T2 mapping. Up to date, however, this approach raises some questions concerning clinical interpretation of GLCM features, which can be potentially used as non-invasive biomarkers of degenerative and regenerative processes. Objective of our study was to compare the GLCM features of knee cartilage lesions treated with ACI (NOVOCART® Inject, TETEC AG, Reutlingen, Germany) at 12 months after surgery and healthy reference regions. We assessed the effect of GLCM calculation offset (a direction) and number of parameters used for T2 mapping on GLCM features considering microanatomy of the cartilage. Furthermore, based on GLCM mathematical definitions and cartilage structure, we selected various features and created their visual representation in both lesions and healthy reference cartilage. Methods: This study was approved by the appropriate ethics committees and regulatory authorities. Twenty-five patients underwent a knee MR examination in a multi-center study (with 3 radiology sites). MRI scanner, coil and T2 mapping sequence information is listed in Table 1. T2 mapping was performed using mono-exponential decay fitting with 2-parameters (M0: ‘zero magnetization‘ and T2: ‘transversal relaxation constant‘) and with 3-parameters (including ‘offset’, i.e. noise estimation). Regions-of-interest (ROI) were defined by an experienced radiologist on two or three consecutive slices on morphological images using JiveX (Visus, Bochum, Germany) based on pathophysiological appearance of cartilage and subchondral bone. Number of slices depended on the size of the lesion. The ROIs were then transferred onto T2 maps using the script written in MatLab. In each subject, two locations were selected: 1) cartilage lesion; 2) healthy reference cartilage. Inclusion criterion for GLCM analysis was the size of ROI greater than 15 pixels, therefore only 20 lesions and their corresponding references were analyzed. ROIs were rotated and flattened (Figure 1) and consecutive GLCM analysis was computed with two offsets: 0° (parallel to cartilage surface) and 90° (perpendicular to cartilage surface). Selected GLCM features were: 1) autocorrelation; 2) correlation (in visualization); 3) contrast; 4) homogeneity. GLCM analysis was performed using custom written script in MatLab environment. Mean T2 time was calculated for each ROI. Shapiro-Wilk normality test was used to assess normality of examined variables. Wilcoxon signed-rank test was used to determine differences in mean T2 time and GLCM features values between chondral lesion and reference cartilage. Wilcoxon signed-rank test was also used to determine differences in GLCM features values, when different offsets and number of T2 mapping parameters were used. We present results as mean value ± standard deviation and/or p-value. All statistical analysis and visual representations of GLCM features were performed using Rstudio version 1.3.1093 (Rstudio, PBC, USA).Tabled 1Table 1Scanner and sequence parameters.Site123ScannerAchieva (Philips, Netherlands)3 TAchieva (Philips, Netherlands)3 TMAGNETOM Skyra (Siemens, Germany)3 TCoil8-channel knee16-channel knee15-channel kneeSequenceMultiecho spin echoMultiecho spin echoMultiecho spin echoOrientation planeSagittalSagittalSagittalSlice thickness (mm)333Slice spacing (mm)3.33.63.3Repetition time (ms)200020002000Echo time (ms)number112.512.512.52252525337.537.537.54505050562.562.562.56757575787.587.587.58100100100Averages111Acquisition matrix268 x 320268 x 320320 x 256Field-of-view (cm)16 x 1616 x 1616 x 16Total acquisition time7min 52s9min 52s8min 4s Open table in a new tab Results: In our cohort, mean T2 times were 56.7 ± 8.8, 52.9 ± 8.6, 64.3 ± 14.3 and 62.6 ± 17.8 ms in lesion and reference 2-parameter mapping and lesion and reference 3-parameter mapping respectively. There was statistically significant difference between lesion and reference mean T2 times in 2-parameter mapping (p = 0.03) but difference was not significant in 3-parameter mapping (p > 0.05), although apparent differences between T2 times were visible on T2 maps.Table 2GLCM features values.Number of parametersROIAngle (°)AutocorrelationContrastHomogeneity2Lesion073.9 ± 35.61.5 ± 1.90.68 ± 0.102Lesion9073.4 ± 36.13.3 ± 2.50.54 ± 0.122Reference0112.7 ± 34.31.8 ± 1.10.61 ± 0.122Reference90112.3 ± 34.83.2 ± 2.60.51 ± 0.133Lesion039.4 ± 20.94.4 ± 2.10.57 ± 0.083Lesion9038.4 ± 20.66.9 ± 4.70.47 ± 0.113Reference053.0 ± 24.45.4 ± 2.90.49 ± 0.123Reference9051.7 ± 22.77.3 ± 5.40.44 ± 0.13 Open table in a new tab When comparing GLCM features of lesion and reference, we found significant difference in autocorrelation, contrast and homogeneity. Difference in autocorrelation was significant for 2-parameter mapping and GLCM offset 0° (p < 0.001) and 90° (p < 0.001) and also for 3-parameter mapping with offset 0° (p < 0.001) and 90° (p < 0.001). Difference in contrast was only significant with offset 0° for both 2-parameter (p = 0.02) and 3-parameter mapping (p = 0.03). Similarly, difference in homogeneity was significant with offset 0° for 2-parameter (p ≪ 0.05) and 3-parameter mapping (p < 0.001). Contrast was significantly different when using different offsets (0° and 90°) for both 2-parameter (p < 0.001) and 3-parameter (p < 0.001) mapping of lesion and for 2-parameter (p < 0.001) and 3-parameter (p ≪ 0.05) mapping of reference. Similarly to contrast, there was significant difference in homogeneity for 2-parameter (p < 0.001) and 3-parameter (p < 0.001) mapping of lesion and also for 2-parameter (p < 0.001) and 3-parameter (p ≪ 0.05) mapping of reference. Autocorrelation was not significantly different using different offsets. We also found significant difference in GLCM features values between 2- and 3-parameter mapping. Autocorrelation, contrast and homogeneity were significantly different for lesion and reference texture analysis with both offsets (0° and 90°). All p-values were < 0.001. In Figure 2 and Figure 3 we present visualization of GLCM features calculated with offset 0° from lesion and reference ROIs.Figure 3GLCM parameters visualization for flattened lesion and reference cartilage ROIs computed from 3-parametric T2 maps with offset 0°.View Large Image Figure ViewerDownload Hi-res image Download (PPT) Conclusions: Although significant difference in mean T2 in 3-parameter mapping was not shown, different T2 distribution in lesion and reference cartilage is visible in Figure 2 and Figure 3. Considering zonal architecture of healthy cartilage, its loss in OA chondral lesions and resultant T2 distribution, autocorrelation, contrast and homogeneity might be reliable indicators of cartilage damage. Autocorrelation represents the extent of pattern repetition. Higher autocorrelation is expected in healthy cartilage, which was proven by our data. Correlation provides a measure similar to autocorrelation and low values in lesion can be seen in Figure 2 and Figure 3. Contrast represents local gray level variation and high values suggest presence of edges, which is typical for healthy cartilage. Homogeneity measures the smoothness of gray level distribution and therefore might serve as a tool for zonal architecture loss detection. Because texture of healthy cartilage is smooth (homogeneous) in horizontal direction when ROI is rotated and flattened, offset 0° has been proven to be a key parameter in distinguishing lesion and reference, with contrast and homogeneity being significantly different only in this direction. Number of T2 mapping parameters has an effect on absolute values of mean T2 times, but this difference does not affect the significance of difference in parameters values between lesion and reference. Texture analysis with GLCM seems to be an useful add-on to T2 mapping of articular cartilage lesions which could potentially serve as a robust marker for monitoring patients in interventional studies.