Fusion scheme of segmentation and classification for breast cancer static ultrasound images

Abstract

Breast Cancer (BC) is defined as cancer that forms in the ducts of the breast (tubes that convey milk to the nipple) and lobules of the breast tissue. This study aims to develop a Computer-Aided Diagnosis (CAD) approach that provides a multidisciplinary skill for breast ultrasound images that could aid specialists in improving accuracy in disease identification, thus reducing the rate of false-positive and falsenegative results. To achieve this goal and build a fully automatic solution, the main limitations faced with the breast ultrasound image will be highlighted. First, ultrasound images suffer from speckle noise and artefacts. Second, the similarity between the textures inside the Region of Interest (ROI) and the background region, and that will end up with overlapping between the ROI and the backgrounds. Third, the similarity between the texture of the benign images and the malignant images, and this challenge will reduce the accuracy of the diagnosis by decreasing the sensitivity and the specificity of the proposed solution. Fourth, the borders of the ROI are not clear. Finally, applying a traditional segmentation method, i.e., the threshold method, will end up with a number of false-positive cases and false-negative cases, and both will affect the result of the automatic solution. In the segmentation stage, we have proposed a trainable schema based on multi-texture features to avoid problems related to the similarity between the texture of the ROI and the background. It also used to avoid the noise and the artifact by training the schema on good samples including regions with noise and artifacts. The trainable schema has solved the poor border problems by training the schema on blocks with poor borders. Forth, feature extraction stage (for the segmentation stage), an existing schema, a single feature that is Local Binary Pattern (LBP), was employed to describe the cancer region. This study has developed a hybrid model based on a multi descriptor (texture feature) to enable the effective extraction of the ROI. Furthermore, this thesis focuses on proposing a new describer that can help to identify the breast abnormality by enhancing the LBP texture features and the LBP descriptor using a new threshold that can help to identify the important information required for the identification of abnormal cases. Eventually, multi-level fusion for automatic classification of static ultrasound images of breast cancer is a method that makes it possible to diagnose breast diseases quickly and accurately compared to a manual approach. This study has used median and Wiener filters to reduce the speckle noise to enhance the ultra sound image texture. This process has helped to extract a powerful feature that can help to reduce the overlapping between the benign and malignant class. This process, followed by the fusion process, has helped to produce a significant decision based on different features produced from different filtered images. The experimental results show the proposed method can apply LBP based texture feature for categorizing ultrasound images, which registered a higher accuracy of 98.8%, the sensitivity of 98.01%, and specificity of 99.3%.