Classification of histopathological breast cancer images using convolutional neural networks

Breast cancer is one of the leading cancer types among women worldwide. Every year, many breast cancer patients die due to late diagnosis and treatment. Therefore, early diagnosis of breast cancer is of vital importance. Although X-ray mammography is the most preferred imaging modality, it is not sufficient for cancer diagnosis alone. Computer aided detection and diagnosis (CAD) methods are developed to assist the radiologist in the diagnosis process and to free the patient from unnecessary pain. In this context, there are many CAD methods in the literature. In this thesis, detailed information was given about convolutional neural networks from deep learning architectures, which are high performance in image recognition. In addition, classification was carried out with low-level texture features preferred primarily for texture based image recognition. Convolutional neural networks consist of feature extraction and transfer learning. Feature extraction was performed on the fc6 and fc7 layers of AlexNet and Vgg16 models. Pre-trained AlexNet model was used for Transfer learning. The low-level texture features methods are respectively Local Binary Pattern (LBP), Histogram of Oriented Gradients (HOG), Scale-Invariant Feature Transform (SIFT) and Grey Level Co-occurrence Matrix (GLCM). In both studies, Support Vector Machines (SVM) method was used in the classification stage. The highest accuracy value for feature extraction in convolutional neural networks was obtained from the combined properties of AlexNet-fc7 and Vgg16-fc7 with 93.78%. Transfer learning is the highest accuracy value of 93.57%. The highest results in low-level texture features methods were obtained as 75.48% by combining LBP-SIFT methods.