Multivariate Analysis of Fourier Transform Infrared Spectroscopy Data for Characterization of Breast Cancer

Abstract

This thesis work presents a method for characterization of breast cancer using Fourier Transform Infrared (FTIR) spectroscopy. FTIR spectroscopy is a tool used to analyse the structure and chemical composition of both organic and inorganic materials. Recently, it has become a key technique for biomedical applications and achieved considerable advancements in the field of cancer diagnosis. Analysing typical group frequencies in FTIR spectrum enables qualitative chemical composition estimations for the materials detected. In the current thesis work, FTIR spectroscopy was applied with multivariate explorative tools and machine learning classification models to characterize breast cancer tissues, extract important chemical components from the tissues, and automatically subtype and grade the breast cancer spectral data. A total of 462 and 126 FTIR spectra were used for breast cancer subtype and grade analysis, respectively. The obtained results showed that as breast cancer progresses, changes were visually differentiated on the spectra by analysing the peak in the lipid, nucleic acid, protein and carbohydrate regions of the normal, malignant, and benign samples. Then using principal component analysis (PCA) score plots and loading plots, and based on the respective sample variance, wavenumbers holding important biochemical components were extracted for each breast cancer subtype and each grade. Finally, Radial basis function (RBF) Kernel support vector machine (SVM) was used to classify breast cancer tissue spectra subtypes into adenosis, fibroadenoma, hyperplasia, fibrocystic change, normal breast, lobular carcinoma and ductal carcinoma and that resulted in a training accuracy of 91.0% and testing accuracy of 90.7%. In terms of grading, the method was also able to classify the dataset into Grade I, Grade II and Grade III with a training accuracy is 84.0% and testing accuracy of 83.7%. This work is significant as it makes the work of pathologists easier to find different breast cancer biomarkers and easily classify the breast types, subtypes and grades automatically. It comes with great promises for use in early detection of breast cancer providing accurate diagnoses and cut down on the time-consuming labour effort.