Local-global consistency guided test-time domain adaptation fundus image segmentation

Fundus image analysis often suffers from distribution shifts across imaging devices, leading to degraded segmentation performance when directly applying source-pretrained models. Test-time adaptation (TTA) offers a solution by enhancing model adaptability to unseen target samples, but existing methods are hindered by distribution biases and yield unstable results. To address this, we propose LGCTTA, a local-global consistency guided test-time adaptation method. It enforces local boundary consistency to exploit complementary knowledge from segmentation and boundary branches, and global prototypical feature consistency to improve robustness across varying test distributions. Experiments on benchmark fundus datasets show that LGCTTA consistently outperforms state-of-the-art TTA methods.