Fig 1.
The hierarchical structure of ligaments, and organization of the tropocollagen from triple-helical polypeptide chains.
Fig 2.
Sample processing and measurement techniques.
a: Four knee ligaments and the patellar tendon were extracted from 10 bovine stifle joints (right human knee used here for visualization purposes). b: Full ligaments were cut into smaller sample pieces which were then measured using NIRS from five equispaced locations along the longitudinal axis. A duplicate piece of tissue adjacent to the sample piece was extracted for histological analysis which determined the length and angle of the collagen crimp. c: The ligament samples were subjected to destructive mechanical testing protocol (see [32] for details) which ruptured the samples. Pieces of the samples were then used for hydroxyproline, elastin, and uronic acid quantification.
Fig 3.
NIRS measurements and reference variables.
a: Full wavelength range of the collected NIR spectra from detector 1 (blue) and detector 2 (orange). Thick lines represent the average spectra. Absorption values from both detectors have been rescaled to range from 1 to 0 to enable easier comparison. b: Normalized distribution of the reference properties. c: Correlation between the reference properties. Color of the boxes represents the sign and strength of the correlation.
Fig 4.
An example optical retardance image of ACL sample used to determine the crimp parameters. The values for crimp length and angle were calculated as an average from manually selected regions of interest.
Table 1.
Distributions of the biochemical and structural reference variables.
Fig 5.
Accuracy of the predictive models.
Median prediction performance (in terms of explained variance) for nine different biochemical and structural properties. Median performance was computed using a 10-fold repeated (N = 10) cross-validation. Only the prediction for water and hydroxyproline content (of wet weight) produced models with over 50% of the variance explained.
Fig 6.
Feature importance of the predictive models.
a: Coefficients of the PLS models for water and hydroxyproline content. The five most influential wavelengths (in terms of absolute magnitude) have been highlighted in red. b: Spectrum used as input for the water and hydroxyproline models. Spectra from detector 1 was smoothed with a 3rd order polynomial Savitzky-Golay filter using a 18 nm filtering window and the second spectral derivative. Similar treatment was applied to spectra from detector 2 with the exception of filtering window being set to 58 nm. Spectra from detector 1 was limited to the wavelength range of 370–980 nm and the spectra from detector 2 to 1000–1900 nm.
Table 2.
Median cross-validated correlation (rCV), R2CV, and root mean squared error (RMSECV) of different reference properties.