Dual Space Multiple Instance Representative Learning for Medical Image Classification

Medical image classification plays a vital role in AI-aided medical diagnosis and is often addressed as a Multiple Instance Learning (MIL) issue (i.e., each sample is a bag of instances). For medical images, the disease area or the discriminative area is usually smaller than the whole tissue. In other words, most of the instances in a bag are irrelevant and could interfere with the bag label inference. To address this issue, we add an instance representative selection process before MIL and propose a novel MIL approach named Dual Space Multiple Instance Representative Learning (DSMIRL). DSMIRL consists of two core steps, Adaptive Instance Representative Selection (AIRS) and Multiple Instance Representative Learning (MIRL). In AIRS, the instances in the same bag are grouped into different sub-bags via clustering, and only one sub-bag is selected as the final collection of instance representatives by ranking the maximum instance predictions of sub-bags, thus adaptively filtering out the irrelevant instances. In MIRL, we perform aggregations on the selected instance representatives in label and feature spaces to further exploit the complementary information of the two spaces. Finally, these two steps are iteratively conducted in each iteration to optimize all modules of DSMIRL progressively. Extensive experiments on five standard MIL benchmarks and two medical image datasets demonstrate the promising performance of DSMIRL over the state-of-the-art MIL approaches.