Characterization and classification of ductal carcinoma tissue using four channel based stokes-mueller polarimetry and machine learning

Abstract

Interaction of polarized light with healthy and abnormal regions of tissue reveals structural information associated with its pathological condition. Even a slight variation in structural alignment can induce a change in polarization property, which can play a crucial role in the early detection of abnormal tissue morphology. We propose a transmission-based Stokes-Mueller microscope for quantitative analysis of the microstructural properties of the tissue specimen. The Stokes-Mueller based polarization microscopy provides significant structural information of tissue through various polarization parameters such as degree of polarization (DOP), degree of linear polarization (DOLP), and degree of circular polarization (DOCP), anisotropy (r) and Mueller decomposition parameters such as diattenuation, retardance and depolarization. Further, by applying a suitable image processing technique such as Machine learning (ML) output images were analysed effectively. The support vector machine image classification model achieved 95.78% validation accuracy and 94.81% testing accuracy with polarization parameter dataset. The study's findings demonstrate the potential of Stokes-Mueller polarimetry in tissue characterization and diagnosis, providing a valuable tool for biomedical applications.

Keywords: Machine learning; Mueller matrix; Polar decomposition; Polarization; Stokes vector; Tissue.

Fig. 1

Polarization-resolved four-channel Stokes-Mueller polarimeter setup: HWP: half wave-plate, QWP: quarter wave-plate, D1, D2, D3, D4: CMOS camera

Fig. 2

Reconstructed Stokes vector 2D images from (a) normal regions of ductal carcinoma samples, and (b) tumour regions of ductal carcinoma samples with the input polarization of 0⁰ and 90⁰. The color bar has a value ranging from 0 to 1 for S₀ and -1 to 1 for the remaining parameters

Fig. 3

Polarization parameter images such as DOP, DOLP, DOCP and anisotropy. (a, b) For the normal regions and (c, d) for tumour regions ductal carcinoma sample, incident with the light of 0⁰ and 90⁰ polarization state respectively. The colour bar for DOP, DOLP, and DOCP has a value ranging from 0 to 1. Anisotropy has a colour bar ranging from -0.5 to 1

Fig. 4

Shows the validation and testing confusion matrices for SVM classifiers trained following polarization parameter datasets with (a) DOP, (b) DOLP, (c) DOCP and (d) anisotropy

Fig. 5

Shows 2D reconstructed Mueller matrix images for (a) normal region and (b) tumour regions of ductal carcinoma sample. The colour bar shows the value ranging from -1 to 1