Asian Research Thesis Index

Abstract

Breast cancer is considered to be one of the most fatal types of cancer among women. The early detection of breast cancer increases the survival probability. Computer Aided Diagnosis (CAD) system for detection and classifying of masses in mammograms is essential. In this thesis, new pre-processing techniques for mammograms have been proposed. First, an improved segmentation technique using image pre-processing and modified reaction diffusion based level set method is developed. The mammogram is passed through morphological operations, contrast enhancement and interpolation algorithm. The pre-processed image is then passed through a modified reaction diffusion based level set algorithm for accurate breast boundary segmentation. In Second pre-processing technique, pectoral muscle is suppressed by the proposed technique based on geometry and contrast variance between the breast tissues and pectoral muscles. Simulation results verify the significance of proposed schemes visually and quantitatively as compared to state of art existing schemes. Three class classification based on deep learning techniques i.e. CNN and RBMs is proposed. In the CNN-DW method, enhanced mammogram images are decomposed as its four subbands by means of two-dimensional discrete wavelet transform (2D-DWT), while in the second method discrete curvelet transform (DCT) is used. Proposed methods have been compared with existing methods in terms of accuracy rate, error rate, and various validation assessment measures. Simulation results clearly validate the significance and impact of our proposed model as compared to other well-known existing techniques. Last but not the least, a novel three-class classification technique for a large dataset of mammograms using a deep-learning method of restricted Boltzmann machine (RBM) is proposed. The augmented dataset is generated using mammogram patches. The dataset is filtered using a non-local means (NLM) filter to enhance the contrast of patches. These patches are decomposed using (2D-DWT) and (CT). The proposed method is compared with existing methods in terms of ROC curve, accuracy rate and various validation assessment measures. The simulation results clearly demonstrate the significance and impact of our proposed model compared to other well-known existing techniques.

Add/Update Thesis