A Deep Learning Approach for Categorizing Breast Carcinoma Histopathology Images

The prominence of breast cancer (BCa) as the most prevalent type of cancer worldwide, especially among women, underscores the critical importance of its early diagnosis. However, the insufficient number of experienced physicians in the field of pathology and the intense work of existing experts make digital pathology a more important need day by day. In this context, in this thesis study, a deep learning approach is proposed for the systematic categorization of histopathology images of breast cancer patients. The images utilized in the study were obtained from ICIAR 2018, and a region-based convolutional neural network (R-CNN) integrated with ResNet-18 was employed as the method. Customizations were implemented in the image augmentation, patch size determination, and training parameters adjustment stages, and for the effective operation of the proposed model, color distribution variances in the images were reduced and the transfer learning technique was applied. The network achieved an accuracy rate of 93.75% in categorizing the training dataset into four groups at the patch level and 97.06% in classifying it into two groups. When the effectiveness of the method was tested on a blind test set, it achieved accuracy rates of 73.44% for quadruple grouping at the patch level and 87.24% for binary grouping. Image-based categorization yielded accuracy rates of 69.79% for four groups and 86.46% for two groups. These results indicate that the proposed model can compete with studies incorporating the latest technologies in the literature and can be utilized as an assistive tool for pathologists during the diagnosis process.