Deep multiple instance learning for digital histopathology

Multiple instance learning (MIL) is a variation of supervised learning where a single class label is assigned to a set of instances, e.g., image patches. After providing a comprehensive introduction, we give a probabilistic definition of MIL. We move on introducing three different MIL approaches and show how they dictate the design of deep neural networks for MIL. Consequently a variety of MIL pooling functions is presented. We compare those pooling functions regarding their interpretability and flexibility. Finally, we evaluate the different MIL approaches and pooling functions on two histopathology datasets. Here, we put great emphasis on the details of the experiment design, including histopathology-specific augmentation techniques and MIL-specific evaluation metrics.