Classification of invasive ductal carcinoma grades from histopathological images using deep learning

Breast cancer is one of the most common types of cancer worldwide and many women are diagnosed with this disease every year. The lethality rate of breast cancer can be reduced by appropriate treatment and detection methods. It is crucial to develop the methods and diagnostic tools used in the diagnosis of breast cancer and automate them with computer-aided methods. Histopathological images are used in the definitive diagnosis of breast cancer. Histopathological images are obtained by examining biopsied breast tissue under a microscope and are used in the diagnosis and grading of breast cancer. In this thesis, it is aimed to classify the degrees of invasive ductal carcinoma from histopathological images by creating a classification model using deep learning methods. Since invasive ductal carcinoma accounts for a large proportion of breast cancer cases, accurate grading is of great importance in treatment planning. In this thesis, classification models were created using ResNet50, DenseNet121 and Xception deep learning architectures. Histopathological images were used for training and evaluation of these models and the images were classified into three grades. Balanced accuracy and area under the receiver operating characteristic curve metrics were used to evaluate the performance of the classification model. Balanced accuracy rates of 58,97% were obtained for the validation data at 4× magnification and 47,81% for the test data at 40× magnification. This thesis demonstrates that convolutional neural network-based transfer learning methods can be used with reported performance values in classifying invasive ductal carcinoma grades from histopathological images.