Reciprocal Collaboration for Semi-supervised Medical Image Classification

To acquire information from unlabeled data, current semisupervised methods are mainly developed based on the mean-teacher or co-training paradigm, with non-controversial optimization objectives so as to regularize the discrepancy in learning towards consistency. However, these methods suffer from the consensus issue, where the learning process might devolve into vanilla self-training due to identical learning targets. To address this issue, we propose a novel Reciprocal Collaboration model (ReCo) for semi-supervised medical image classification. ReCo is composed of a main network and an auxiliary network, which are constrained by distinct while latently consistent objectives. On labeled data, the main network learns from the ground truth acquiescently, while simultaneously generating auxiliary labels utilized as the supervision for the auxiliary network. Specifically, given a labeled image, the auxiliary label is defined as the category with the second-highest classification score predicted by the main network, thus symbolizing the most likely mistaken classification. Hence, the auxiliary network is specifically designed to discern which category the image should NOT belong to. On unlabeled data, cross pseudo supervision is applied using reversed predictions. Furthermore, feature embeddings are purposefully regularized under the guidance of contrary predictions, with the aim of differentiating between categories susceptible to misclassification. We evaluate our approach on two public benchmarks. Our results demonstrate the superiority of ReCo, which consistently outperforms popular competitors and sets a new state of the art.