Automated classification of histopathology images using transfer learning

Early and accurate diagnosis of diseases can often save lives. Diagnosis of diseases from tissue samples is done manually by pathologists. Diagnostics process is usually time consuming and expensive. Hence, automated analysis of tissue samples from histopathology images has critical importance for early diagnosis and treatment. The computer aided systems can improve the quality of diagnoses and give pathologists a second opinion for critical cases. In this study, a deep learning based transfer learning approach has been proposed to classify histopathology images automatically. Two well-known and current pre-trained convolutional neural network (CNN) models, ResNet-50 and DenseNet-161, have been trained and tested using color and grayscale images. The DenseNet-161 tested on grayscale images and obtained the best classification accuracy of 97.89%. Additionally, ResNet-50 pre-trained model was tested on the color images of the Kimia Path24 dataset and achieved the highest classification accuracy of 98.87%. According to the obtained results, it may be said that the proposed pre-trained models can be used for fast and accurate classification of histopathology images and assist pathologists in their daily clinical tasks.