Grad-CL: Source Free Domain Adaptation with Gradient Guided Feature Disalignment

2025-09-12

AI Summary

To address the significant performance degradation of segmentation models on fundus images across domains—primarily caused by discrepancies in imaging protocols—this paper proposes a source-free unsupervised domain adaptation framework. Methodologically, it innovatively integrates gradient-guided pseudo-label refinement, cosine-similarity-based contrastive learning, uncertainty-aware prototype feature modeling, and a corresponding contrastive loss to enhance class discriminability and cross-domain feature alignment. Extensive experiments on multiple cross-domain fundus image benchmarks (e.g., REFUGE→RIM-ONE, Drishti-GS→REFUGE) demonstrate that our approach consistently outperforms state-of-the-art methods: average Dice scores for optic disc and cup segmentation improve by 2.1–3.8%, boundary localization accuracy is markedly enhanced, and the model exhibits superior generalizability and robustness under domain shift.

Technology Category

• Computer Vision: Segmentation
• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Application Domains: Security

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Security and Privacy: Data transparency and provenance
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation

Abstract

Accurate segmentation of the optic disc and cup is critical for the early diagnosis and management of ocular diseases such as glaucoma. However, segmentation models trained on one dataset often suffer significant performance degradation when applied to target data acquired under different imaging protocols or conditions. To address this challenge, we propose extbf{Grad-CL}, a novel source-free domain adaptation framework that leverages a pre-trained source model and unlabeled target data to robustly adapt segmentation performance without requiring access to the original source data. Grad-CL combines a gradient-guided pseudolabel refinement module with a cosine similarity-based contrastive learning strategy. In the first stage, salient class-specific features are extracted via a gradient-based mechanism, enabling more accurate uncertainty quantification and robust prototype estimation for refining noisy pseudolabels. In the second stage, a contrastive loss based on cosine similarity is employed to explicitly enforce inter-class separability between the gradient-informed features of the optic cup and disc. Extensive experiments on challenging cross-domain fundus imaging datasets demonstrate that Grad-CL outperforms state-of-the-art unsupervised and source-free domain adaptation methods, achieving superior segmentation accuracy and improved boundary delineation. Project and code are available at https://visdomlab.github.io/GCL/.

Problem

• Adapting optic disc segmentation across domains without source data
• Reducing performance degradation from imaging protocol differences
• Refining pseudolabels and enhancing inter-class separability features

Innovation

• Gradient-guided pseudolabel refinement module
• Cosine similarity-based contrastive learning strategy
• Source-free domain adaptation framework