Development of a discrete wavelet transform and artificial neural network based classification system for mammogram images
Loading...
Date
2016-09-01
Authors
Luqman Mahmood Mina
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Nowadays, numerous computer-aided diagnosis (CAD) systems have been developed to assist radiologists in the recognition of mammographic lesions that may indicate the presence of breast cancer. However, the performance of CAD is limited by two main issues; (i) unwanted regions (i.e. high-intensity rectangular label, tape, artefact, skin-air interface, etc.) could disturb the detection of breast cancer and reduce the accuracy rate of CAD, (ii) the irregularity of mammograms’ texture in which features such as entropy, energy, skewness, kurtosis, mean, and standard deviation are correlated in the spatial domain and insignificant for classification. Therefore, to address the aforementioned problems, an improved CAD system for the mammogram image is proposed. The proposed CAD consists of three main stages, namely pre-processing, feature extraction, and classification of mammogram images. In pre-processing step, Adaptive Multilevel Threshold (AMLT) is proposed, which successfully removes the above-mentioned unwanted regions. It gives the advantage to the system where it allows the search for abnormalities to be constrained to the region of the breast tissue without the effect of the unwanted regions in the image background. In feature extraction stage, two new features, namely medians of maximum and minimum of high-frequency subbands have been proposed to classify the mammogram images into normal, benign and malignant. Box plot analysis has proven that both new features are uncorrelated and significant for classification of mammogram images as compared to the conventional features. In the classification stage, multilayer perceptron (MLP) network is employed to classify normal and abnormal mammograms in the first phase and benign and malignant in the second phase. The average results produced from 322 mammogram images in the first phase concluded that the proposed approach attained reliable results with an accuracy of 96.27%, sensitivity of 94.78% and specificity of 96.60%. In addition, the average results produced from 115 abnormal images for accuracy, sensitivity, and specificity are 95.65%, 96.18%, and 95.38% respectively. The final experimental results show that the developed mammogram classification system is able to achieve the highest classification as compared to the other state-of-the-art systems. These promising classification performances show that the proposed system could probably be used to assist pathologists in their diagnosis process.