This paper considers the modern classification methods of breast cancer histopathology. The main purpose of the study is to conduct an extended test of the trained model on data that fundamentally differs from the training dataset. We chose a large Russian dataset with different types of classification as the training dataset. The dataset contains images with different resolutions and magnifications. As testing data, the same dataset was used, but the resolution, color balance, brightness, and contrast of the images were changed. The classes in the dataset were unbalanced, so we applied augmentation methods (flipping and rotation). The models ResNet 152, DenseNet 121, Inception_resnet_v2 were selected for training. The transfer learning approach was used for training. The preprocessing of images consisted of normalizing the values of all image channels in the range from 0 to 1. The models had good results with standard testing methods. The resolution change slightly reduced metrics. The change in color balance, brightness, and contrast significantly reduced all metrics. The test results show that elementary normalization is not enough for high-quality training of models resistant to changes in input data.
Keywords: neural network, model, machine learning, breast cancer, cancer classification, artificial intelligence, transfer learning, histopathology