Multiple instance learning for classifying histopathological images of the breast cancer using residual neural network

Digital histopathological images have complex textures and high variability. Thus, classifying histopathological images requires an accurate classification and recognition of the tissue components in these images. In this article, we propose a novel classification layer based on multiple instances learning (MIL). In regular convolutional neural network (CNN) a flatten or a global pooling layer is used before the fully connected layers. However, in our proposed layer, we consider each last feature map in the network as an instance that will be classified by the output layer. Then, an aggregation function will be applied to get the class of the image (bag). This mapping helps the model to classify each feature independently to catch the micro-objects of the complex tissue images. Also, our method succeeded in achieving high accuracy without the preprocessing of the images with color normalization, stain normalization, or any other techniques. Additionally, we trained our models in two different strategies. The first one is by combining the images from all the magnification factors, and the second is by training a model for each magnification factor. We show in this work that our model outperforms several previous works on breast cancer classification.