Classification of breast cancer histopatological images using deep learning methods

Breast cancer is a dangerous type of cancer that is usually found in women, and this is an important research topic in medical science. Breast cancer is a disease that occurs as a result of tumor formation by the uncontrolled proliferation of some of the breast cell structures. Early diagnosis makes it difficult to treat cancer by spreading to other organs in undetected patients. In the diagnosis of breast cancer, the accuracy of the pathological diagnosis is of great importance in order to shorten the decision-making process, minimize the undetected cancer cells and get a faster diagnosis. However, similarity of images in histopathological breast cancer image analysis, finding different levels of healthy or tumor tissues in different regions is a sensitive and difficult process that requires high proficiency for field experts. In recent years, researchers have been looking for a solution to this process with machine learning and deep learning methods, which have contributed to important developments in medical diagnosis and image analysis. In this thesis, two basic studies on computer-assisted cancer diagnosis were carried out on breast cancer pathological images. In the first study, a hybrid method DESA+ReliefF is proposed with the size reduction-based ReliefF feature selection algorithm, utilizing the enabling features of pre-trained deep convolutional neural network (DESA) models for the classification of breast cancer histopathological images. The model is based on a fine-tuned transfer learning technique in fully connected layers. In addition, models were compared with k-nearest neighbors (kNN), navie bayes (NB), and support vector machine (DVM) machine learning approaches. The well-known VggNet-19 model was also used for comparison with the proposed model. The performance of each combination of feature extractor and classifier is analyzed using the precision, precision, F1-score and ROC curves. The proposed hybrid model was individually trained at 40X, 100X, 200X, 400X magnifications using the BreakHis dataset. The results show that the model is an effective classification model with high performance up to 97.8% diagnostic accuracy. In the second study carried out within the scope of the thesis, an optimized histopathological convolutional neural network (HCNN) model was proposed by utilizing the activation features of current DESA networks for the dual classification of breast cancer histopathological images as benign and malignant tumors. The model basically uses a fine-tuned transfer learning technique. In the training of the model, the optimization of the model was achieved by using stochastic gradient descent (SGD), Nesterov accelerated gradient (Nag), adaptive gradient (AdaGrad), RMSprop, AdaDelta and Adam solvers. These solvers calculate the initial values of the proposed network and determine the parameters of the network; updates according to learning speed, history and method. It provides faster back propagation learning, optimization and optimum updating of parameters. The proposed HCNN model is individually trained at 40X, 100X, 200X, 400X magnifications using the BreakHis dataset. The results show that the model is an effective classification model with high performance up to 99.05% accuracy.