Abstract / Summary
Accurate identification of retinal diseases often requires multi-label annotation from domain specialists, yet high-quality labels remain expensive and inconsistent across experts. To address this challenge, we propose a knowledge-driven multi-expert active learning framework that integrates graph-based reasoning and adaptive annotation strategies for multi-label retinal disease diagnosis. The framework models both expert knowledge and pathology correlations within a heterogeneous pathology information network (HPIN), where graph convolutional propagation captures semantic dependencies between diseases and images. A loss prediction network is then employed to estimate the informativeness of unlabeled samples. Furthermore, a multi-expert collaboration mechanism is introduced to emulate real-world ophthalmic diagnostic workflows, where each expert contributes partial labels within their domain of expertise. This design not only reduces redundant annotation effort but also enhances label completeness and consistency. Experiments conducted on two public retinal datasets demonstrate that the proposed framework significantly improves annotation efficiency and diagnostic accuracy compared with state-of-the-art multi-label active learning methods. The results highlight the framework’s potential as an intelligent knowledge-based system for scalable medical image understanding.