Conventional arthroscopic evaluation of articular cartilage is subjective and insufficient for assessing early compositional and structural changes during the progression of post-traumatic osteoarthritis. Therefore, in this study, arthroscopic near-infrared (NIR) spectroscopy is introduced, for the first time, for in vivo evaluation of articular cartilage thickness, proteoglycan (PG) content, and collagen orientation angle. NIR spectra were acquired in vivo and in vitro from equine cartilage adjacent to experimental cartilage repair sites. As reference, digital densitometry and polarized light microscopy were used to evaluate superficial and full-thickness PG content and collagen orientation angle. To relate NIR spectra and cartilage properties, ensemble neural networks, each with two different architectures, were trained and evaluated by using Spearman's correlation analysis (ρ). The ensemble networks enabled accurate predictions for full-thickness reference properties (PG content: ρin vitro, Val= 0.691, ρin vivo= 0.676; collagen orientation angle: ρin vitro, Val= 0.626, ρin vivo= 0.574) from NIR spectral data. In addition, the networks enabled reliable prediction of PG content in superficial (25%) cartilage (ρin vitro, Val= 0.650, ρin vivo= 0.613) and cartilage thickness (ρin vitro, Val= 0.797, ρin vivo= 0.596). To conclude, NIR spectroscopy could enhance the detection of initial cartilage degeneration and thus enable demarcation of the boundary between healthy and compromised cartilage tissue during arthroscopic surgery.
Keywords: Arthroscopy; Deep learning; Equine; Mosaicplasty; Neural networks; Osteoarthritis; Post-traumatic osteoarthritis.