Categorization of Breast Carcinoma Histopathology Images by Utilizing Region-Based Convolutional Neural Networks

Abstract

The inadequacy of experienced pathologists worldwide, combined with the workload of current specialists, has increased the need for digital pathology. Accordingly, in this study, the categorization of breast cancer histopathology images supplied by ICIAR 2018 was carried out using region-based convolutional neural networks (R-CNN) based on transfer learning with custom augmentation, training parameters and patch size selection. The images were normalized using a stain normalization method to reduce inequalities in color distribution. Image patches were extracted, and a transfer learning technique was performed to solve the lack of data. ResNet-18 was utilized for transfer learning. Image augmentation was also performed to increase the training data. The network achieved a test accuracy of 93.75% and 97.06% for four classes and two classes, respectively, on the training dataset. The success of our method was also examined on the blind test set, and it had 73.44% accuracy for four classes and 87.24% accuracy for two classes in patch-wise classification, while it obtained 69.79% accuracy for four classes and 86.46% accuracy for two classes in image-wise classification. Our model achieved a score very close to the highest result in the literature, with a difference of 1.51% for two classes and 4.36% for four classes in patch-wise classification. The outcomes show that R-CNN with transfer learning gives competitive results with state-of-the-art studies in the literature in this dataset and can be used as a tool to aid pathologists.