An Artificial Neural Network to Evaluate Cartilage Collagen and Proteoglycan Fractions Based on Multiparametric Quantitative MRI Techniques

Objectives: Cartilage composition, that is, proteoglycan (PG) and, to a lesser degree, collagen (CO) fractions, is altered in cartilage degeneration. This study applied an artificial neuronal network (ANN) to assess CO and PG fractions in intact cartilage noninvasively based on quantitative magnetic resonance imaging (qMRI).

Methods: A total of 11 macroscopically intact human osteochondral samples (source: medial femoral condyle) were harvested from total knee arthroplasties. Spatially resolved T1, T1ρ, T2, and T2* maps were obtained using inversion recovery, spin-lock multiple gradient-echo, spin-echo, and gradient-echo sequences (3T, Achieva, Philips). For reference, Fourier transform infrared microspectrometry was applied to generate spatial CO and PG maps along the same plane. An ANN was implemented to predict these fractions in a pixel-wise manner (training stage) followed by appropriate 11-fold cross-validation. Correlations of qMRI parameters and fractions were determined using Spearman’s ρ and the predictive power of the ANN by mean squared errors (MSEs).

Results: Highly significant correlations (p < 0.001) between qMRI parameters and fractions were found: T1 (ρCO: − 0.65; ρPG: − 0.63); T1ρ (ρCO: − 0.34; ρPG: − 0.33); T2 (ρCO: − 0.59; ρPG: − 0.66); and T2* (ρCO: − 0.53; ρPG: − 0.56). ANN-based predictions of CO and PG fractions were excellent (MSECO = MSEPG = 0.005).

Conclusion: Although no qMRI parameter is specific to any particular cartilage component, these parameters give information on compositional features in cartilage. The implemented ANN allows prediction of composition based on multiparametric qMRI techniques with minimal error, thereby making feasible the noninvasive determination of local CO and PG fractions.

Conflict of Interest: None declared.