Source-free domain adaptation framework based on confidence constrained mean teacher for fundus image segmentation

Unsupervised domain adaptation (UDA) has been gradually applied in fundus image segmentation, mitigating the challenge of insufficient data annotation by transferring pre-trained knowledge from source domain to target domain. However, UDA requires access to source data, which is usually restricted due to privacy and security concerns. Consequently, Source-Free Domain Adaptation (SFDA), particularly mean teacher-based methods, has garnered widespread attention. The mean teacher primarily includes two components: student model (f_s), which updates parameters through backpropagation, and teacher model (f_t), which is the exponential moving average (EMA) of f_s. During this process, the EMA parameter heavily influences the efficiency and performance of f_t. However, current methods employ fixed EMA parameter, i.e., f_t can only learn knowledge from f_s based on a fixed proportion, which affects the effectiveness of feature extraction. To address this issue, this study proposes a Confidence Constrained Mean Teacher (CCMT) framework, which dynamically adjusts the exponential moving parameter by measuring the outputs difference between f_t and f_s, enables the model to actively adjust the learning progress and direction, thereby enhancing the update efficiency of the model and improving its performance. Additionally, we introduce JS divergence to constrain the output distributions of f_t and f_s, ensuring their consistency, since f_s relies on the outputs of f_t as pseudo-labels. Experimental on two publicly available benchmark datasets demonstrate the advantages of CCMT in achieving accurate and robust segmentation of the fundus optic disc and optic cup in source-free settings, outperforming the state-of-the-art result.