Pseudo-Label Clustering-Driven Dual-Level Contrast Learning Based Source-Free Domain Adaptation for Fundus Image Segmentation

Source-Free Domain Adaptation (SFDA) has gained attention as a promising solution to address the domain shift issue, eliminating the requirement for labeled data from the source domain. However, current SFDA methods heavily rely on self-training, which are confronted with two main challenges: inevitable occurrence of noisy pseudolabels and insufficient adaptation across a single scale or level. To overcome these limitations, a novel SFDA method is developed for fundus image segmentation across different datasets. Our method encompasses two essential phases: the generation phase and the adaptation phase. In the generation phase, we introduce clustering to SFDA segmentation and propose a feature-enhanced clustering method to generate robust pseudo-labels. This process improves adaptation quality particularly when the source model's feature learning capability is limited in the target domain. In the adaptation phase, we develop a dual-level contrast learning method aimed at mitigating domain shift through self-supervision. First, we present a full-scale feature-level contrast loss that utilizes low-level and high-level features from both the target domain data and its augmented version. This enables the model to acquire discriminative characteristics while minimizing disparities between the original and augmented data. Second, we design a clinical prior-guided label-level contrast loss to filter out low-quality pseudo-labels, providing favorable guidance for the segmentation model. Extensive experiments on cross-domain datasets of fundus images demonstrate its superiority over mainstream SFDA methods. In the challenging Drishti-GS target domain, our method surpasses SOTA models by 3.14% and 2.18% in optic disc and optic cup Dice scores, respectively. Codes are available at https://github.com/M4cheal/PCDCL- SFDA.