Abstract
Constructing highly accurate artificial intelligence (AI) models requires a large amount of consistently correct data (teacher labels). However, there are many cases where the correct answer is not uniquely specified, even for the same data, due to different interpretations depending on observers’ decision criteria. Under such circumstances, how to define the correct answer and build an AI model has not been sufficiently discussed. In this study, we address this issue in the field of pathological image diagnosis, where opinions are occasionally varied even among medical experts. In this paper, we propose a method for constructing AI models that are more robust to inter-observer variability by training several AI models based on teacher labels with the variability and exploiting the relationships among the data predicted by these models. We conduct comparison experiments using multiple datasets of independently labeled images by different pathologists for the same set of images. The proposed method show higher classification accuracy than baselines for most datasets, outperforming MacroF1 by up to 0.1.
