Ensemble Distillation of Divergent Opinions for Robust Pathological Image Classification

The construction of highly accurate deep neural networks (DNNs) requires consistent labeled data. However, there are numerous cases wherein the ground truth is not uniquely determined, even for the same data, owing to different interpretations depending on observers' decision criteria. Studies on the definition of labels and the building of a DNN model under such circumstances are scarce. Thus, this study addresses this issue in the field of pathological image diagnosis, where opinions occasionally vary, even among medical experts. We propose a method for constructing DNN models that are more robust to inter-observer variability by exploiting the knowledge about the relationships among the data learned by multiple DNN models. Comparison experiments were conducted using multiple pathology datasets of independently labeled images by different pathologists for the same set of images. The proposed method exhibited good generalization capability and outperformed the classification accuracy of baseline models specific to certain decision criteria.