Instance Explainable Multi-instance Learning for ROI of Various Data

Estimating the Region of Interest (ROI) for images is a classic problem in the field of computer vision. In a broader sense, the object of ROI estimation can be generalized to the bag containing multiple data instances, i.e., identify the instances that probably arouse our interest. Under the circumstance without instance labels, generalized ROI estimation problem can be addressed in the framework of Multi-Instance Learning (MIL). MIL is a variation of supervised learning where a bag containing multiple instances is assigned a single class label. Though the success in bag-level classification, when bags contain a large number of instances, existing works ignore instance-level interpretation which is the key to ROI estimation. In this paper we propose an instance explainable MIL method to solve the problem. We devise a generalized permutation-invariant operator with the idea of utility and show that the interpretation issues can be addressed by including a family of utility functions in the space of instance embedding. Following this route, we propose a novel Permutation-Invariant Operator to improve the instance-level interpretability of MIL as well as the overall performance. We also point out that existing approaches can be regarded as a special case of our framework and qualitatively analyze the superiority of our work. Furthermore we give a criterion to measure the linear separability in the instance embedding space. We conduct extensive evaluations on both classic MIL benchmarks and a real-life histopathology dataset. Experimental results show that our method achieves a significant improvement in the performance of both instance-level ROI estimation and bag-level classification compared to state-of-the-art methods.