Better Pseudo-Labeling for Semi-Supervised Domain Generalization in Medical Magnetic Resonance Image Segmentation

Abstract

Abstract Magnetic resonance image (MRI) is the primary diagnostic test used clinically for the diagnosis and assessment of a wide range of diseases. In recent years, many studies have employed artificial intelligence techniques for MRI segmentation. Deep learning methods have demonstrated potential to enhance segmentation performance. However, they still face two challenges: annotation scarcity and domain shift. The annotation of MRI is both challenging and costly, and well-annotated datasets are scarce and valuable. Moreover, due to variations in MRI machines, ensuring the independence and identical distribution between model training data and real-world data is difficult, which may lead to noisy model predictions and weak generalization ability. We aim to address the challenges through a multi-pronged approach. First, we propose a method that integrates confidence and uncertainty for generating reliable pseudo-labels. Second, we introduce a consistency learning method that employs self-perturbation at both the image and feature levels to encourage the learning of more generalized feature representations. Finally, we optimize pseudo-labels end-to-end with the teacher–student framework. To evaluate the effectiveness of our method, we conduct experiments on six different MRI segmentation datasets. The results showed that our method was superior to the existing methods in DSC, ASD and HD95 metrics. In addition, we evaluated the quality and quantity of the generated pseudo-labels, and the results showed that our method generated better pseudo-labels than other methods. Overall, our proposed method shows promising potential in assisting clinicians in practical applications.

Keywords

• Semi-supervised learning
• Domain generalization
• Medical image segmentation
• Pseudo-labels