Bilateral Supervision Network for Semi-Supervised Medical Image Segmentation

Massive high-quality annotated data is required by fully-supervised learning, which is difficult to obtain for image segmentation since the pixel-level annotation is expensive, especially for medical image segmentation tasks that need domain knowledge. As an alternative solution, semi-supervised learning (SSL) can effectively alleviate the dependence on the annotated samples by leveraging abundant unlabeled samples. Among the SSL methods, mean-teacher (MT) is the most popular one. However, in MT, teacher model’s weights are completely determined by student model’s weights, which will lead to the training bottleneck at the late training stages. Besides, only pixel-wise consistency is applied for unlabeled data, which ignores the category information and is susceptible to noise. In this paper, we propose a bilateral supervision network with bilateral exponential moving average (bilateral-EMA), named BSNet to overcome these issues. On the one hand, both the student and teacher models are trained on labeled data, and then their weights are updated with the bilateral-EMA, and thus the two models can learn from each other. On the other hand, pseudo labels are used to perform bilateral supervision for unlabeled data. Moreover, for enhancing the supervision, we adopt adversarial learning to enforce the network generate more reliable pseudo labels for unlabeled data. We conduct extensive experiments on three datasets to evaluate the proposed BSNet, and results show that BSNet can improve the semi-supervised segmentation performance by a large margin and surpass other state-of-the-art SSL methods.

I. Introduction

Medical image segmentation of anatomical structures and lesion regions is vital for the medical image analysis and clinical diagnosis. Convolutional neural networks (CNN) have achieved great progress on medical image segmentation [1], [2], [3], [4], [5]. However, these fully-supervised methods depend on a large amount of labeled data. For medical image segmentation, pixel-wise annotation requires domain knowledge, and obtaining massive amounts of labeled images is expensive and time-consuming. Three medical image segmentation tasks are shown in Fig.1, and labeling the lesion or vessel masks is very time-consuming, especially for thin vessels of fundus images. Therefore, training deep models in the low data regime is a practical and challenging task for medical image segmentation. Fig. 1.

Examples of three medical image segmentation tasks. The first row shows gastrointestinal polyp image, fundus image and dermoscopy image, respectively. The second row shows the corresponding ground truth masks, respectively. It can be seen from the fundus image that some thin vessels are very small in size, and the surrounding tissue in the polyp image is very similar in texture, and the pixel-level annotation is time-consuming.