Self-Paced Co-Training and Foundation Model for Semi-Supervised Medical Image Segmentation

Semi-supervised learning is an effective approach for image segmentation, especially in medical images where segmentation labels are scarce and require expertise. Currently, there is still much room for improvement in these existing semi-supervised methods. In this paper, we propose a semi-supervised learning method for image segmentation with self-paced learning and foundation model. We first synthesize virtual unlabeled images by mixing the original unlabeled images in training set with normal images without lesions through mixup operation. Then we design a novel self-paced co-training strategy that constrains the different networks to learn consistent prediction distributions for the paired virtual unlabeled images that are synthesized from different normal images but the same unlabeled images. In addition, by incorporating both pixel-level and image-level importance into the self-paced learning, the designed approach allows the network to learn both pixel-level and image-level features from easy to hard. We perform the experiment using the pre-training foundation model as the network architecture. The experimental results show the effectiveness of our method in semi-supervised segmentation and can further improve the performance of the foundation model.

I. Introduction

Medical image segmentation plays a crucial role in assisting clinicians with the quantitative analysis of pathological features. While deep learning has demonstrated effectiveness in image segmentation, its application in the medical field presents many challenges. The annotation in medical images typically requires specialized knowledge, making it difficult to obtain a large amount of labeled data. Many efforts have been made to reduce the dependence on labeled data. Techniques like semi-supervised learning [1], [2] and foundation models [3], [4] have attracted much attention and can improve segmentation performance when annotated data is scarce.