A new decision support system design used in the evaluation of mammography images

Breast cancer is the leading cause of deaths related with cancer among women. Breast cancer makes up approximately 24,5% of cancer diseases in the world, and 23,9% in Türkiye. Early diagnosis of the disease is the most effective step for treatment. Mammography is one of the most used method in the early diagnosis of breast cancer. However, misdiagnosis can be made due to some reasons such as used technologies, experience of the radiologists, breast density, breast structure and shooting errors. Therefore, a second radiologist's diagnosis is needed to reduce ambiguity in the diagnosis of the mammography image. Recently, many studies have been carried out on the Computerized Decision Support System (CDSS) which help for the most accurate diagnosis with mammography images. In this thesis, a CDSS is designed in order to contribute the radiologist by providing a second opinion for the diagnosis of breast cancer. The main aim of this system, in which the final decision will be made by the radiologist, is to reduce the rate of human error. Todays, deep neural networks are preferred more than traditional methods in image processing studies. Although deep learning methods are successful in especially image classification problems, it need a lot of labeled data for training. Limitations in accessing correctly labeled mammography data makes these types of studies difficult. Because of this, most of the former studies in this area were carried out with too limited mammography images. The mammography images used in this study were got from 4 different sources: Elazig Fethi Sekin Research Hospital, INBREAST, DDSM and MIAS. In the study, first, the images were preprocessed by using morphological image processing methods to eliminate labels and noises in the images. Then, 224x224 and 227x227 pixels of sections from the breast tissue areas were cropped and 3 different classes (cyst, calcification and normal) consisting of relatively simple and small images were created. MKESAS model which was developed for this study, AlexNet and GoogLeNet models were trained with these images. The results show that the MKESAS is the most accurate model with 98,71%. GoogLeNet network is more accurate (96%) than AlexNet network (95,55%). Keywords: Breast cancer, convolutional neural networks, biomedical image processing, image classification