MULTIPLE INSTANCE LEARNING WITH CRITICAL INSTANCE FOR WHOLE SLIDE IMAGE CLASSIFICATION

Histopathological whole slide image (WSI) classification is a fundamental task in clinical diagnosis. The extremely high resolution of WSIs makes it laborious to obtain localized annotations. Hence, WSI classification is often cast as a multiple instance learning (MIL) problem when only slide-level labels are available. A major challenge for MIL approaches is that the proportion of malignant regions in a malignant WSI could be very small, resulting in the misclassification of these slides. In this work, we propose to find the critical instance in the projection space of instance representations. A novel critical instance pooling is proposed to aggregate instance representations with normalized distances between instances and the critical instance. Additionally, a feature magnitude loss is utilized to enforce the discrepancy between positive and negative instances in the projection space. Experimental results on existing benchmark datasets showed the effectiveness of our proposed method.