Federated Semi-Supervised Learning Through a Combination of Self and Cross Model Ensembling

Abstract

Medical image segmentation (MIS) plays a vital role in modern computer-aided diagnosis systems. Deep learning technology has achieved promising results in MIS in recent years. However, deep learning (DL) is a data-hungry technology and in the domain of healthcare, a single hospital usually cannot afford to collect adequate medical images to train a robust DL model. Moreover, a medical image usually contains sensitive data related to patients' privacy, making it's infeasible to build a larger dataset by collecting images from different hospitals. Federated learning (FL), a recently proposed privacy-protecting collaborative paradigm aiming at allowing different data owners to collaboratively train a model without exposing raw data, seems to be a proper solution to these problems. Some recent works have verified the feasibility of applying FL to MIS but most of these works only confine to fully supervised scenarios. Unfortunately, most hospitals in realistic usually cannot provide fully labeled data due to lack of labor. In this paper, we study a challenging but more practical problem in which each hospital can only provide a few labeled data combined with some other unlabeled data. To effectively handle such a problem, we propose a novel and robust federated semi-supervised learning (FSSL) framework, which improves over the mean teacher mechanism with a cross-clients ensemble module and a model-wise self-ensembling module. We evaluate our method on two public medical image datasets and the results show that, in the challenging FSSL scenario, our method can effectively leverage unlabeled data to boost the model performance by a considerable margin. Notably, our method also outperforms other existing FSSL approaches designed for MIS.