Segmentation, Super-resolution and Fusion for Digital Mammogram Classification

Majeed, Taban Fouad (2016) Segmentation, Super-resolution and Fusion for Digital Mammogram Classification. Doctoral thesis, University of Buckingham.

Preview

Text Majeed_2016_Segmentation.pdf - Submitted Version Download (3MB) | Preview

Abstract

Mammography is one of the most common and effective techniques used by radiologists for the early detection of breast cancer. Recently, computer-aided detection/diagnosis (CAD) has become a major research topic in medical imaging and has been widely applied in clinical situations. According to statics, early detection of cancer can reduce the mortality rates by 30% to 70%, therefore detection and diagnosis in the early stage are very important. CAD systems are designed primarily to assist radiologists in detecting and classifying abnormalities in medical scan images, but the main challenges hindering their wider deployment is the difficulty in achieving accuracy rates that help improve radiologists’ performance. The detection and diagnosis of breast cancer face two main issues: the accuracy of the CAD system, and the radiologists’ performance in reading and diagnosing mammograms. This thesis focused on the accuracy of CAD systems. In particular, we investigated two main steps of CAD systems; pre-processing (enhancement and segmentation), feature extraction and classification. Through this investigation, we make five main contributions to the field of automatic mammogram analysis. In automated mammogram analysis, image segmentation techniques are employed in breast boundary or region-of-interest (ROI) extraction. In most Medio-Lateral Oblique (MLO) views of mammograms, the pectoral muscle represents a predominant density region and it is important to detect and segment out this muscle region during pre-processing because it could be bias to the detection of breast cancer. An important reason for the breast border extraction is that it will limit the search-zone for abnormalities in the region of the breast without undue influence from the background of the mammogram. Therefore, we propose a new scheme for breast border extraction, artifact removal and removal of annotations, which are found in the background of mammograms. This was achieved using an local adaptive threshold that creates a binary mask for the images, followed by the use of morphological operations. Furthermore, an adaptive algorithm is proposed to detect and remove the pectoral muscle automatically. Feature extraction is another important step of any image-based pattern classification system. The performance of the corresponding classification depends very much on how well the extracted features represent the object of interest. We investigated a range of different texture feature sets such as Local Binary Pattern Histogram (LBPH), Histogram of Oriented Gradients (HOG) descriptor, and Gray Level Co-occurrence Matrix (GLCM). We propose the use of multi-scale features based on wavelet and local binary patterns for mammogram classification. We extract histograms of LBP codes from the original image as well as the wavelet sub-bands. Extracted features are combined into a single feature set. Experimental results show that our proposed method of combining LBPH features obtained from the original image and with LBPH features obtained from the wavelet domain increase the classification accuracy (sensitivity and specificity) when compared with LBPH extracted from the original image. The feature vector size could be large for some types of feature extraction schemes and they may contain redundant features that could have a negative effect on the performance of classification accuracy. Therefore, feature vector size reduction is needed to achieve higher accuracy as well as efficiency (processing and storage). We reduced the size of the features by applying principle component analysis (PCA) on the feature set and only chose a small number of eigen components to represent the features. Experimental results showed enhancement in the mammogram classification accuracy with a small set of features when compared with using original feature vector. Then we investigated and propose the use of the feature and decision fusion in mammogram classification. In feature-level fusion, two or more extracted feature sets of the same mammogram are concatenated into a single larger fused feature vector to represent the mammogram. Whereas in decision-level fusion, the results of individual classifiers based on distinct features extracted from the same mammogram are combined into a single decision. In this case the final decision is made by majority voting among the results of individual classifiers. Finally, we investigated the use of super resolution as a pre-processing step to enhance the mammograms prior to extracting features. From the preliminary experimental results we conclude that using enhanced mammograms have a positive effect on the performance of the system. Overall, our combination of proposals outperforms several existing schemes published in the literature.

Actions (login required)

View Item