Open-Set Semi-supervised Medical Image Classification with Learnable Prototypes and Outlier Filter

Semi-supervised learning (SSL) is an effective way to utilize unlabeled data due to the high annotation cost. Unfortunately, the collected unlabeled data will inevitably contain outliers not belonging to the labeled classes in many clinical practices, which is named open-set Semi-supervised learning (OSSL). Although existing methods have achieved decent performance on natural images, they ignore the fine-grained characteristics of medical images, and thus they are not suitable for medical image. In this paper, we propose a framework for the challenging Open-set Semi-Supervised Classification for medical image, named OpenSSC and it consists of three components. First, the learnable prototypes is proposed to learn the compact representation of the fine-grained seen classes. Then, we propose a multi-binary discriminator that integrate the closed-set output to distinguish seen and unseen classes. Based on the two components, a joint outlier filter is proposed to classify seen classes and identify unseen classes from unlabeled data. Our proposed method can handle well both the seen and unseen classes. We conduct extensive experiments to demonstrate the superiority of our method, and it outperforms other state-of-the-art OSSL methods on two medical image classification tasks.