A Non-Invasive Polarized Light System To Detect Breast Cancer On Mice Based On The Application Of Machine Learning To Classify Collected Images Utilizing Mueller Matrix Transformation

Abstract

On the basis of the variation in microstructural properties of tissue caused by changes in the degree of polarized light, it could be conceivable to create an automated optical polarimetry system capable of enhancing the distinction between normal and malignant samples. This thesis aims to establish a framework for a non-invasive Mueller matrix polarimetry approach to discriminate sensitive optical parameters for a better descriptions of cancerous specific features, particularly breast cancer. Therefore, sample preparation, data collection, and extraction of optical properties were performed on a mouse model of breast cancer. The testing group consists of 10 Swiss albino females with estrogen- and DMBA-induced breast cancer, while the control group consists of 10 healthy Swiss albino females. A non-contact optical technique was utilized to measure the Mueller matrices of normal and malignant mice in order to obtain 30 best datasets for each mouse type. In order to evaluate the two samples, the parameters derived from the Mueller matrix were statistically processed and analyzed. The proposed theory was canceled, however, due to the uncertainty in the testing mice, who were exposed to numerous cancer lesions, such as lymphoma and intraductal papilloma. Only the two sets of in vivo measurements were further processed in the classification stages using machine learning algorithms with the five common classifiers Support Vector Machine, K-Nearest Neighbors, Gradient Boosting, and the Random Forest. By using the method of defining correlation matrix and two-way ANOVA test to extract the significant sensitive parameters, the training dataset comprising 295 data for each feature D, m12, m13, m23, m34, m42, and m44 of both healthy and abnormal mice was trained by five classifiers. Five trained models have demonstrated promising outcomes, with the lowest accuracy reaching 88.14% and the greatest reaching 94.92%. All of the aforementioned findings convince it as a remarkable non-invasive, automated, and sensitive polarimetry optical system utilizing Mueller matrix transformation for early characterization of the differences between the two testing and control mouse models. This approach could pave the way for the noncontact and computer-aided early diagnosis of breast cancer and other types of cancer in humans.