Introduction

Quantitative MRI (qMRI) methods have been investigated extensively to detect early changes in articular cartilage¹. Established qMRI parameters, such as T₁, T₂ and T_1rho, have been shown to reflect, each to a different extent, the macromolecular content and structural organization of cartilage^2-4. However, accurate tissue characterization remains challenging due to limited specificity of qMRI parameters to single cartilage matrix components. Multiparametric approaches have shown improved performances in estimating cartilage degradation^5,6. The aim of this study was to use different ensemble regression models to predict compositional and structural properties of articular cartilage, i.e. proteoglycan (PG) content and collagen fiber orientation from combined use of multiple qMRI parameters (T₁, T₂, continuous wave T_1rho, adiabatic T_1rho, adiabatic T_2rho and T_RAFF).

Methods

The study had been approved by the local ethical committee of Utrecht University in compliance with the Dutch Act on Animal Experimentation. Mildly osteoarthritic and control samples (n=20) were obtained from the stifle joints of Shetland ponies (N=10). The imaging was performed at 9.4 T using a 19-mm quadrature RF transceiver (RAPID Biomedical GmbH, Rimpar, Germany). qMRI relaxation time measurements for T₁, T₂, CWT_1rho at varying spin-lock amplitudes (γB₁ = 100, 200, 300, 400, 500, 600, 800, 1000 and 2000 Hz), adiabatic T_1rho, adiabatic T_2rho and T_RAFF were measured by coupling a preparation block to a fast spin-echo readout (TR = 5 s, ETL = 8, TE_eff = 4.2 ms, matrix size = 192×192, slice = 1 mm, FOV = 19.2×19.2 mm²).
For the analysis, four full-thickness cartilage regions of interest (ROIs) were defined for each sample (Figure 1), and then the qMRI relaxation time maps in the ROIs were calculated using MATLAB script. The relaxation time values across cartilage thickness (20 depth-wise points) obtained from the ROIs, were concatenated throughout all samples, creating a vector of 1600 points (20 depth-wise points x 4 ROIs x 20 samples = 1600 data points) for each qMRI parameter. For each ROI, corresponding regions from digital densitometry (DD), indicative of PG content, and polarized light microscopy (PLM), indicative of collagen fiber orientation, were utilized.
Sklearn package in Python and SPSS (version 25) were used to build the machine learning pipeline and conduct statistical analysis. Three ensemble-based regressors were developed: Gradient Boosting Regressor (GBR), Random Forest Regressor (RFR) and Extra Trees Regressor (ETR). To test the regressors, 20 percent of the independent variables (qMRI parameters) and dependent variables (PLM and DD) were extracted randomly. To avoid over-fitting and optimize hyper-parameters, cross-validated grid-search object was implemented in the training phase. To evaluate the performance of the models, regression metrics in testing phase were used: mean absolute error, median absolute error and coefficient of determination (r²). In addition, Pearson correlation coefficients (r) were calculated for all regressors. To compare contribution of each qMRI parameter, feature importance attribute, based on mean decrease in impurity, was used. In addition, ANOVA filter procedure, calculation of F-score between each independent and dependent variable, was performed as a baseline feature selection method. Feature-importance values of the regressors and F-scores were normalized.

Results

ETR outperformed GBR and RFR regressors in terms of lower mean absolute error and median absolute error, and higher r². ETR was found to have the highest prediction accuracy for both PLM (r²=0.77, p<0.01) and DD (r²=0.62, p<0.01), as compared to GBR and RFR (Figure 2). Strong correlations were observed between the predicted and actual PLM, and between the predicted and actual DD with all regressors (p<0.01) (Figure 2). ANOVA filter, ETR, GBR and RFR indicated that CWT_1rho at (γB₁=100 Hz) and (γB₁=200 Hz) had the greatest variable importance to PLM and to DD prediction (Figures 3-4).

Discussion

In the current study, multiparametric MRI was able to explain 77% of the variance embedded into PLM and 62% in DD. Previously, models based on the combination of 2 to 4 qMRI parameters measured at 9.4 T have shown the ability to distinguish between immature and mature engineered cartilage⁷. Furthermore, different combinations of the same qMRI parameters have demonstrated high accuracy to classify degenerated tissue and strong correlations with tissue hydration and sGAG (PG) content in bovine nasal cartilage^5,6. Similarly at 3 T, combinations of 2 to 4 qMRI parameters have shown high accuracy in detecting histological degeneration in human articular cartilage⁸. Significant correlations have been determined with PLM (r = 0.68) and DD (r =0.67) using combined T₂* and quantitative susceptibility maps from the same specimens as those used in this study⁹. In the present study, 14 combined qMRI parameters were strongly correlated with collagen fiber orientation (r = 0.86-0.88) and PG content (r = 0.77-0.79). Analyzing the variable importance of each parameter, CWT_1rho at low spin-lock amplitudes (γB₁=100, 200 Hz) was found to be the most important predictor for collagen fiber orientation and PG content.

Conclusion

These findings show that multiparametric MRI can be used for non-invasive prediction of quantitative histology of cartilage tissue. CWT_1rho at low spin-lock amplitudes (γB₁=100-200 Hz) appeared as the most important qMRI parameter to predict collagen orientation and proteoglycan content in articular cartilage.

Acknowledgements

Support from Jane and Aatos Erkko Foundation, and the Academy of Finland (grants #285909, #293970, #319440 and #297033) is gratefully acknowledged.