Computer-aided automatic detection of breast cancer using deep neural networks on histopathological images

Today, breast cancer is the most common type of cancer among women and ranks second after lung cancer with a very high mortality rate. In case of late detection, breast cancer treatment becomes very difficult. Although there are various methods for the detection of breast cancer, the need for additional detection and treatment methods/tools is still important. In this study, a model using deep neural networks is proposed for the detection of breast cancer on histopathological images. The used dataset is BreakHis, which has 40X, 100X, 200X and 400X magnification ratios and contains 7909 histopathological images in total. In the proposed deep neural network model, successful results were obtained using four different pre-trained networks such as DenseNet201, Inception V3, ResNet50 and Xception. The architectures of the used pre-trained networks were examined and it was determined why successful networks were better. Dropout, data augmentation, test time data augmentation and batch normalization methods were used to further increase the performance of the models. The performances of each pre-trained network at different magnification ratios were examined by different performance criteria (accuracy, F1 score, area under curve) and graphical accuracy value. The confusion matrices of the pre-trained networks were achieved and the predictions realized on the images in the dataset as true or false. As a result of experimental results, it was seen that the results obtained with the Xception network are more successful than the other networks. In the experimental studies performed with the Xception network at 200X magnification ratio, the most successful results were obtained compared to other networks and other magnification ratios, and an accuracy score of 98.01%, a sensitivity value of 98.21% and a recall value of 98.92% were obtained. The area value under the ROC (Receiver Operating Characteristic) curve at 200X magnification ratio of the Xception network was achieved as 0.975. The predictions and actual results of the Xception network on histopathological images with a randomly selected 200X magnification ratio were compared. As a result, it was denoted that more successful results were obtained in DenseNet201 at 100X magnification ratio, InceptionV3 at 200X magnification ratio, Xception network at 200X and 400X magnification ratios for this dataset. Keywords:Breast cancer, Histopathological image, Classification, Deep neural network, Xception, InceptionV3, ResNet, DenseNet, Pre-trained networks