Glaucoma is a leading cause of irreversible blindness worldwide. While deep learning approaches using fundus images have largely improved early diagnosis of glaucoma, variations in images from different devices and locations (known as domain shifts) challenge the use of pre-trained models in real-world settings. To address this, we propose a novel Graph-guided Test-Time Adaptation (GTTA) framework to generalize glaucoma diagnosis models to unseen test environments. GTTA integrates the topological information of fundus images into the model training, enhancing the model's transferability and reducing the risk of learning spurious correlation. During inference, GTTA introduces a novel test-time training objective to make the source-trained classifier progressively adapt to target patterns with reliable class conditional estimation and consistency regularization. Experiments on cross-domain glaucoma diagnosis benchmarks demonstrate the superiority of the overall framework and individual components under different backbone networks.

翻译：青光眼是全球范围内不可逆性失明的主要原因。尽管利用眼底图像的深度学习方法已显著改进了青光眼的早期诊断，但不同设备和地点采集的图像差异（即领域偏移）对预训练模型在实际场景中的应用构成了挑战。为解决此问题，我们提出了一种新颖的图引导测试时自适应（GTTA）框架，旨在将青光眼诊断模型泛化至未见过的测试环境。GTTA将眼底图像的拓扑信息整合到模型训练中，从而增强模型的可迁移性并降低学习伪相关性的风险。在推理阶段，GTTA引入了一种创新的测试时训练目标，通过可靠的类别条件估计与一致性正则化，使源域训练的分类器逐步适应目标域模式。跨领域青光眼诊断基准测试的实验结果表明，该整体框架及其各组成模块在不同骨干网络下均表现出优越性能。