1.

Introduction

Collagen and proteoglycan (PG), two main macromolecules in articular cartilage, play important roles in maintaining cartilage’s biophysical and mechanical properties. A disruption of the collagen network or a reduction in PG content would lead to cartilage diseases, mainly marked by the functional incapacitation of articular cartilage accompanied by joint pain and loss of mobility for the patients.¹ Articular cartilage can be conceptually subdivided into three structural zones from the surface to subchondral bone superficial zone (SZ), transitional zone (TZ), and radial zone (RZ).^2,3 Both collagen and PG are inhomogeneously distributed in these zones.⁴ In addition, the PG content, especially in SZ,^5,6 shows clear differences between healthy and osteoarthritic (OA) cartilages.

Fourier transform infrared imaging (FTIRI), which is a newly developed spectroscopic imaging technique with high resolution and sensitivity, can quantify the collagen and PG contents in articular cartilage. FTIRI provides the possibilities to study the concentration variations of collagen and PG with microscopic resolution. Most FTIRI studies in the literature are limited to the qualitative analysis,⁷ which is due to the significant overlapping among the IR characteristic bands of the principal macromolecules. Recently, the absorption bands of the collagen and PG in cartilage were analyzed by principal component regression (PCR), which has produced the distribution profiles of both principal macromolecules.⁴

Partial least squares (PLS) regression method combines the features of principal component analysis and multiple regression algorithm,⁸ which uses the entire range of the spectrum rather than the individual peak absorption bands. PLS-discriminant analysis (PLS-DA), a classification method based on PLS, possesses a robust and objective ability in classification since it extracts information from both the predictor matrix (spectral matrix) and response matrix (categorical matrix).⁹ By replacing the concentration matrix in PLS with a categorical matrix in PLS-DA, the PLS-DA approach has been successfully used for discriminating analysis in biomedical fields.^10,11

Following our recent work that successfully quantified the percentage concentrations of collagen and PG in both healthy and OA cartilages by FTIRI combined with PCR⁴ or PLS,¹² the main goal in this research was to classify the healthy and OA cartilages. This would be the first report that the hyphenated methods of FTIRI with PLS-DA were employed to differentiate healthy and OA cartilage samples.

2.

Materials and Methods

3.

Results and Discussion

Figure 1 shows the visible and total absorption IR images of healthy and an OA cartilage sections, respectively. Some chondrocytes are interspersed in cartilage sections and shown as irregular circles or shapes in the visible images; the quantity of the cells becomes less in SZ, TZ, even the upper part of RZ for OA cartilage [Fig. 1(b)]. In addition, the cells in the subchondral areas of OA cartilage were bigger than these in healthy cartilage [Fig. 1(a)].¹⁴ When comparing the total absorption images of healthy and OA cartilages, the absorption intensity was significantly reduced by 10% in SZ and TZ of OA cartilage ($100\text{-}\mu \mathrm{m}$ depth beneath the cartilage surface). The main differences between OA and healthy cartilages, therefore, appear to be in SZ and TZ.⁵ In the processing of the PLS modeling, IR spectra would be extracted from SZ and TZ. The PLS-DA calculated results are shown in Fig. 2.

Fig. 1

Visible and total absorption infrared images of (a) a healthy and (b) an osteoarthritic (OA) whole-thickness cartilage from the surface to subchondral bone ($\text{size}=900\mu \mathrm{m}\times 160\mu \mathrm{m}$). The absorption limits of healthy and OA cartilages were 0 to 1. The color bar shows the relatively absorbance intensities of Fourier transform infrared images.

Fig. 2

Partial least squares-discriminant analysis (PLS-DA) calculated category values of the spectral matrix, including 100 spectra sequenced from 1 to 100. Healthy (open square) and OA (open circle) cartilages spectra of calibration matrix were predicted in cross-validation; healthy (filled square) and OA (filled circle) cartilages spectra of prediction matrix were predicted by PLS-DA.

In Fig. 2, spectra #1 to 50 come from the healthy cartilage, which are shown as the rectangular points in this plot; spectra #51 to 100 originate from the OA cartilage, shown as the round points. There are 100 calculated values in the plot, including the calibration and prediction matrices. Prediction accuracies of the calibration and prediction matrices by the PLS-DA model are 100% and 90.24%, respectively.

The discriminant accuracy of healthy cartilages was 100%, which means that no healthy cartilages spectra were misjudged by this model. By comparison, there were eight misjudged spectra in the prediction matrix from the OA cartilage. The predicted category values of these eight misjudged OA spectra were over the cut-off value of 0.5, which were shown as the solid circles above the dashed line in Fig. 2. According to the spectral number in the prediction matrix, these eight misjudged spectra appear in the surface part of OA cartilage (0 to 20 and 80 to $100\mu \mathrm{m}$ positions).

In addition, one-way ANOVA tests were carried out for the cartilage sample. The statistical analyses based on the calculated values from PLS-DA showed that differences among the healthy cartilage were unremarkable but became clear for the OA cartilage. There were also significant differences between the healthy and OA cartilage groups, which confirms the significant decrease in the PG content in OA cartilage when compared with the healthy cartilage.⁵

The procedure of ACL transection is known to induce a mechanically instability in knee and change in the joint reactive force, which causes cartilage damage.⁵ Although all OA cartilages were obtained from the 2-year ACL transection knee joints, differences among the individual specimens might contribute to the misjudgment. For example, the OA degree of some misjudged OA samples might be weaker than those of other OA samples, where the OA degeneration did not developed deeper to TZ for these sections. PG loss could consequently become unremarkable in this area.⁵ In addition, the SZ of some individual OA samples may be seriously damaged, which causes difficulties in defining the actual surface of the OA cartilage section in imaging. Since the PG loss in RZ occurs later in the degradation process, an inaccurate identification of the surface location may cause the last 2 to 3 spectra to be extracted from deeper locations than 80 to $100\mu \mathrm{m}$,^4,5 which results in misjudging in the spectra from the 80- to $100\text{-}\mu \mathrm{m}$ location.

For the surface misjudgment (the 0- to $20\text{-}\mu \mathrm{m}$ region), the peripheral regions of cartilage sections may curl up slightly during the air-drying process. In addition, the changes of chondrocyte in the OA cartilages¹⁵ might cause stronger diffusion reflections and scattering effects in SZ than those in the intermediate layer. These practical factors might contribute to the misjudged cases in the surface position for OA cartilage.

4.

Conclusions

FTIRI combined with PLS-DA was used successfully to classify healthy and OA cartilages, with accuracies of 100% and 90.24% for calibration and prediction matrices, respectively, which closely approached the Fisher’s discrimination results obtained in our lab (unpublished). To the best of our knowledge, this hyphenated technique of FTIRI and chemometrics discriminance had not yet been applied in research of articular cartilage and OA. This approach could become a promising classification-discriminant technique for discriminating between healthy and OA specimens, though it is currently limited in section research in lab.