Date of Award

16-9-2023

Document Type

Thesis

School

School of Electrical & Electroncis Engineering

First Advisor

Dr.K.Narasimhan

Keywords

Machine Learning, Deep Learning, Artificial Intelligence

Abstract

The prevalence of breast cancer in women worldwide is far higher than that of cancers of the lungs, brain, or liver. Increasing ageing populations and poor lifestyle habits among the general public, primarily in industrialized nations, are significant factors contributing to the rise in cancer-related mortality rates worldwide. Approximately one woman in every three will develop breast cancer. This research proposes several advanced computer methods for analyzing breast cancer images. This work analyses breast cancer in four imaging modalities: mammography, thermography, ultrasonography and histopathology.

Each modality has some limitations in diagnosing tumors in the breast region. Heavy dose in mammogram imaging leads to cancer risk among patients. The drawback of using thermal and ultrasonography imaging is that hazy imaging with speckle noise. The advantages of using these three imaging techniques (mammography, thermography, ultrasonography) are invasive and economical. Which are suitable for diagnosing cancer early on. The biopsy is a diagnostic method that is used to identify the stages of cancer with sub-types.

An examination of histopathology can deliver a more robust and reliable analysis. Using this approach, tissue samples are taken from the affected breast area and investigated under a microscope. The limitation of this histopathology diagnosis method is it is an invasive, expensive technique. The performance of various texture analysis methods was analysed in this research so that the number of false positives in breast cancer detection could be reduced. This study includes various pre-processing techniques such as image resizing, filtering, segmentation and ROI (Region of Interest) extraction. The four categories of texture analysis methods have been utilised in this work: statistical, structural, model-based and transform-based models.

Some well-known texture analysis methods are Grey level co-occurrence matrix features (GLCM), Curvelet features and Gabor filter. This research mainly focuses on the advanced classification techniques of various machine learning and deep learning classification to categorize breast tumors. The different machine learning classifiers are studied in support vector machine (SVM), k-nearest neighbor (KNN), trees, regressions, and linear discriminant analysis. The advantage of the machine learning algorithm is that it performs intelligent prediction without human intervention. The construction of the machine learning algorithm based on statistical analysis using training data.

However, the machine learning-based classifiers require a manual feature extraction process. The optimal feature selection and several trials are essential for convincing, precise results. For this reason, CNN has been utilised to achieve promising results. CNN is one of the deep learning approaches developed in recent years. In CNN, the image features are extracted directly from the raw input data without using feature extractors. It provides results in each intermediate stage. CNN is utilised adequately in the field of breast cancer classification research. The proposed method quantifies and visualizes breast tumor changes and may help physicians plan treatment. Overall, the strategies presented in this thesis improve the performance of state-of-the-art approaches and may help improve breast cancer diagnosis.

Download

Share

COinS