Polarimetric data-based model for tissue recognition

Abstract

We highlight the potential of a predictive optical model method for tissue recognition, based on the statistical analysis of different polarimetric indicators that retrieve complete polarimetric information (selective absorption, retardance and depolarization) of samples. The study is conducted on the experimental Mueller matrices of four biological tissues (bone, tendon, muscle and myotendinous junction) measured from a collection of 157 ex-vivo chicken samples. Moreover, we perform several non-parametric data distribution analyses to build a logistic regression-based algorithm capable to recognize, in a single and dynamic measurement, whether a sample corresponds (or not) to one of the four different tissue categories.

Fig. 1.

Intensity image measured at 625 nm of a) muscle, b) tendon, c) myotendinous junction and d) bone tissues of a given chicken thigh. Images correspond to an area of 1.1 × 1.1 cm².

Fig. 2.

Boxplot of P₂ index for all tissues measured at 625 nm illumination channel. Red-dashed line visually represents the potential of the metric to discriminate muscle among remaining tissue types: the median of the muscle box does not fit within the other tissues’ boxes. Boxplot points out the low quantity of outliers on data distributions (which can be extrapolated to the remaining metrics): mild and extreme values are represented by circles and stars, respectively.

Fig. 3.

Scree plot of the principal component analysis.

Fig. 4.

The plot showing principal components C₁ against C₂ represents the correlation coefficients between the 27 polarimetric indicators and the two first extracted principal components. The notation of the polarimetric indicators names is composed of the word that represents the measured parameter: P1, P2, P3 and PA (IPPs and P_Δ, respectively), D and P (diattenuation and polarizance), R and Delta (global and linear δ retardance) and Phi (optical rotation Ψ), followed by R, G or B (corresponding to red, green and blue measured wavelength, respectively).

Fig. 5.

ROC curve of the principal component C₁ for (a) muscle, (b) tendon, (c) myotendinous junction and (d) bone.

Fig. 6.

ROC curve of the probabilistic model for (a) tendon, (b) muscle, (c) myotendinous junction and (d) bone.

Fig. 7.

Intensity image M00 (a) and probability image outcome when applying the Muscle-model (b), Tendon-model (c), Myotendinous junction-model (d) and Bone-model (e) on an arbitrary chicken tendon sample. The gray level bars, placed to the right of the corresponding probability function images, defines the probability of the pixel to be recognized as a particular tissue, in a range between one (white) or zero (black).

Fig. 8.

Intensity image M00 (a) and probability function of muscle (b), tendon (c), myotendinous junction (d) and bone (e) for chicken muscle measurements. The gray level bars, placed to the right of the corresponding probability function images, defines the probability of the pixel to be recognized as a particular tissue, in a range between one (white) or zero (black).