GCN-based MIL: multi-instance learning utilizing structural relationships among instances

Multi-instance learning (MIL) aims to learn mappings between bags of instances and bag-level labels. Therefore, the relationships among instances are very important for learning mappings. In this study, we propose an MIL algorithm based on a graph built by the structural relationships among instances within a bag. Then, we use a graph convolutional network (GCN) and a new graph-attention mechanism to learn the bag-embedding. In the task of high-resolution medical image classification, the GCN-based MIL algorithm makes full use of structural relationships among patches (instances) in an original image spatial domain. Experimental results verify that the proposed method is more suitable for handling high-resolution medical images. We also experimentally verify that the proposed method achieves better results than previous methods on five benchmark MIL datasets and five medical image datasets. Moreover, we verify that GCN and the proposed graph-attention mechanism outperform GNN and attention-MIL by ablation experiments on MIL benchmark datasets.