Source-free domain adaptive fundus image segmentation with entropy-minimized mean teacher

Source-free domain adaptive fundus image segmentation offers a novel means of segmenting retinal images without access to source-domain data. By leveraging a pretrained model alongside unlabeled target-domain images, this approach circumvents the privacy and data-access challenges inherent to conventional domain adaptation. However, the resulting domain shift often injects noise that degrades segmentation performance. To address this, we introduce the EntropyMinimized Mean Teacher (EMMT) framework. In EMMT, a teacher network generates pseudo-labels from weakly augmented target images, which supervise a student network trained on their strongly augmented counterparts, thereby promoting model stability. We further enhance prediction confidence by integrating an entropy-minimization term into the student’s loss function. The teacher’s weights are updated via an exponential moving average of the student’s parameters, reducing the influence of incorrect pseudo-labels and improving adaptation to the target domain. Experimental results on benchmark fundus segmentation datasets demonstrate that EMMT outperforms state-of-the-art methods.