On the Development of Improved Mammogram Detection System using Machine Learning Approaches

Abstract

In recent years, breast cancer has become one of the most prevalent causes of death among women. Once the malignant cells are developed in the breast, it spreads to different body organs very quickly. Detection at the early stages and diagnosis is the only way to prevent mortality. Mammography, a noninvasive, non-radioactive imaging technique, has been widely used in diagnosing breast tissue abnormalities. Manual diagnosis based on visual inspection of mammograms is time-consuming, inconvenient, and necessitates skilled supervision. Thus, automated detection using modern imaging, machine learning, and deep learning approaches has become vital for quick, reliable, and correct conclusions. In the last decade, the development of automated computer-aided diagnosis/detection (CAD) models has progressed remarkably. However, there is still an opportunity for improvement in terms of automation, usability, and accuracy. This dissertation is aimed at designing automated CAD frameworks that will help radiologists validate their clinical diagnoses. This research primarily proposes various feature extraction techniques and classifiers for detecting breast tumors in mammography images. The first contribution consists of three frameworks with various feature extraction techniques like discrete wavelet transform (DWT), lifting wavelet transform (LWT), and fast curvelet transforms (FCT). For all frameworks, a combined feature reduction technique such as principal component analysis (PCA) and linear discriminant analysis (LDA) has been employed for feature vector computation. Finally, a simple and flexible learning scheme called the extreme learning algorithm (ELM), back-propagation neural network, k-nearest neighbors, and support vector machine have been used separately to obtain the classification accuracy. This contribution describes an empirical analysis of ELM with other classifiers. In the second contribution, a set of innovative hybrid classification systems proposes to reduce the bottleneck caused by extreme learning machines and contemporary meta-heuristic optimization techniques to classify mammogram images. The optimization techniques have been utilized to obtain the hidden node parameters of the ELM. Here, the same feature reduction technique is used as in the previous contribution. Different hybrid classification systems have examined the three handcrafted feature extraction techniques: DWT, LWT, and FCT. The third contribution is about designing a framework based on non-handcrafted features. Here, deep learning algorithms are used to solve the challenge of manually selecting appropriate features for mammogram classification. The different deep CNN models such as VGG-16, ResNet-50, and Inception-V3 have been utilized for feature extraction, and the same feature reduction method is applied as previous frameworks. Finally, various hybrid classifiers are used for the classification task. The final contribution involves designing a customized CNN model for multiclass mammogram images. This model is designed to extract high-level features from mammogram images automatically. End-to-end learning is facilitated by the proposed deep architectures, which aid in generating promising results. To evaluate the efficiency of each suggested CAD framework, a significant number of experiments have been conducted individually utilizing binary and multiclass mammogram classification. Various performance measurements have been used to compare the suggested CAD frameworks with existing standard techniques. Experimental results demonstrate that the proposed methodologies are superior to existing binary and multiclass breast cancer detection models. The customized CNN model removes manual handcrafted feature extraction issues and avoids feature reduction tasks. As a result, the proposed CAD frameworks are faster and can be used as an enhanced tool by clinicians to validate their diagnoses.