The relationship between articular cartilage reflectance spectra and its functional properties in various spectral NIR regions has been shown in several in vitro stud-ies [18, 30, 31, 33, 34, 41, 43, 121]. In this thesis, spectral regions 0.7–1.05µm (studies II–III) and 0.4–1.9µm (studyIV) were utilized for evaluation of cartilage thickness, composition, and biomechanical properties. In studyII, the prediction error when
mapping equine cartilage thickness (0.09 mm) was smaller than the typical pixel size (0.5× 0.5 mm2) of clinical MRI [33, 142], thus demonstrating the feasibility of the technique to accurately measure and map cartilage thickness on large areas of interest. Similar findings have been presented by Afaraet alwith rat, bovine, and human samples [31, 33, 41], and by Oberg et al with bovine samples [121]. These studies utilized varying spectral regions (between 0.4–2.5 µm) with coefficient of determination varying between 0.48–0.97. Furthermore, similarly to studyII, Afara et al utilized an independent test set validation approach, which provides an un-biased evaluation on model generalization. Therefore, studyIItogether with these previous studies, highlights the potential of NIRS for accurate estimation of cartilage thickness.
Biomechanical parameters are good proxies for estimating cartilage health as cartilage has a mechanical function crucial for locomotion [52, 143]. In studies II andIV, biomechanical parameters correlated (r= 0.77–0.98) significantly with data from spectral regions of 0.7–1.05 µm and 0.75–1.9 µm, respectively. Martickeet al utilized the spectral region of 1.1–1.7 µm and calculated the ratio of two spectral peaks, which showed an adequate correlation (ρ = 0.54) with equilibrium modu-lus [18]. Furthermore, Afara et al utilized data in the spectral region 0.4–1.1µm and demonstrated high correlations with equilibrium (r= 0.89) and dynamic (r = 0.90) moduli [41]. These studies [18, 41] provided valuable information since hu-man articular cartilage samples were used; however, in studyII, model validation and mapping was achieved using an independent test set while Afaraet alapplied only cross-validation [41]. Furthermore, mapping enabled visualization of the ex-tent of degeneration in damaged regions. Additionally, in studyIV, a wider spectral region (0.40–1.90µm) and more sophisticated mathematical approaches (i.e., ANN and variable selection) were used compared to the aforementioned studies.
Spectral analysis is conventionally used to determine sample (e.g., tissue) com-positional and even structural changes [96]. Therefore, several studies have utilized NIRS for evaluation of cartilage PG [40,41] and water contents [35,41]. Nevertheless, no study has shown the relationship between NIR spectra and collagen content or orientation in native articular cartilage. Although, Spahn et al concluded that no relationship exists between NIR spectra and PG or collagen contents [30]. However, in studyIII, contrary to findings by Spahnet al, NIR spectra were observed to cor-relate with the PG and collagen contents. In addition, a relationship between colla-gen network orientation and NIR spectra was observed, making NIRS a promising optical technique for detecting superficial collagen disruption, which is generally considered the first visual cue of cartilage degeneration. These correlations between NIR spectra and cartilage PG and collagen contents arise from overtone vibrations of these cartilage constituents. For superficial collagen orientation, the correlation with NIR spectra is expected to arise from the birefringence property of cartilage. Fur-thermore, in studyIII, the average absorbance of visually healthy cartilage (ICRS0) was lower compared to damaged cartilage, thus indicating more light to be reflected from intact superficial collagen network.
Although a significant body of knowledge has demonstrated the role of cartilage damage in the development of OA and PTOA [3, 5], substantially less research have focused on the role of the underlying subchondral bone [60,144]. In studyIV, ANN models reliably predicted subchondral bone plate and subchondral trabecular bone properties from tissue spectral response (0.4–1.9µm). In a previous study [32], NIR spectral region (0.8–2.5µm) was successfully utilized for prediction of bone param-eters in a rat model by utilizing cross-validation. However, cartilage thickness in rat
knee is substantially thinner compared to that in equine and human; thus, this thesis presents for the first time the potential of NIRS technique for evaluation of subchon-dral bone in larger animals (i.e., equine) and humans [50,51]. In addition, the ability of light in the VIS region to penetrate deeper into soft tissues [42, 121] enhanced the reliability of predicting subchondral trabecular bone properties. Articular cartilage overlays the subchondral bone and thus the spectral response (0.4–1.9µm) includes contributions from both tissues [39]. In study IV, a variable selection technique was applied to determine the most relevant wavelengths for each cartilage and bone parameter, thus optimizing model performance. For cartilage properties, the NIR spectral region alone had better prediction performance without the VIS region. To accurately account for contributions from different tissues, the wavelength-specific absorption and scattering coefficients of osteochondral tissues should be determined to build a physical model of light propagation; however, these tasks are beyond the scope of this thesis.
The correlations of VIS and NIR spectra with articular cartilage and bone proper-ties arise from combination and overtone bands of main constituents of tissue: water, collagen (types I and II), PGs, and hydroxyapatite [52]. The main bonds associated with spectral changes are OH, NH, CH, SH, and PO4. As the water is the most abundant in cartilage, the effect of OH combination and overtone bands are most distinguishable in the absorbance spectrum at 2.0–2.2µm, and 1.40–1.55µm (first overtone), and 0.95–1.10 µm (second overtone), respectively [95, 96]. Additionally, NH and CH overtones are easily distinguishable in the NIR spectral region [95, 96].
In studyIV, a wider spectral region (0.4–1.9µm) including stronger overtones was utilized; nevertheless, no substantial improvement compared to study II was ob-served in prediction of biomechanical parameters. Arguably, models incorporating the wider spectral region should generalize better; however, for reliable comparison of the results, the role of analysis techniques and different hardware should be con-sidered. In studyIV, the masking effect of water was evident in wavelength region between 1.9 µm and 2.5µm; therefore, this spectral region was not utilized in the analysis.
PLSR and ANN techniques were employed in studyII, and studiesIII andIV, respectively. These techniques enabled the evaluation of complex data and out-performed univariate analysis, which is not suitable for analysis of spectra with overlapping features [95, 96]. Although ANN has outperformed PLSR in multiple occasions [97, 124, 125], no significant difference was reported between the two tech-niques in a conference proceeding utilizing cartilage NIR spectra and biomechanical properties [145]. In studies III andIV, the forward variable selection method en-hanced prediction performance of the models. Similar findings were observed in the aforementioned conference proceeding by applying the genetic algorithm ap-proach [145]. These variable selection methods had a substantial effect on model performance when compared to the preprocessing techniques in studies IIto IV.
Scatter correction techniques (i.e., SNV and MSC) had no positive effect on model performance arguably due to the stable and controlled measurement environment.
With ANN, derivative preprocessing slightly decreased model performance, which somewhat contradicts the findings of Rieppo et al [146]. The MIR spectral region utilized by Rieppo et al includes stronger and less overlapped contributions than NIR spectral region used in this thesis; therefore, derivative preprocessing may not be optimal for the VIS and NIR spectral regions. Thus, for optimal results, the fol-lowing steps are proposed: spectral smoothing without loss of spectral fidelity and information, discarding the regions with poor SNR, and applying variable selection
techniques to identify the most contributing wavelengths.