Foundation Model With Uncertainty Estimation-Based Active Learning for Retinal Image Classification

Retinal disease diagnosis faces critical annotation bottlenecks due to extensive expert-labeling requirements, which limit the deployment of automated diagnostic systems. We developed an FMUE (Foundation Model with Uncertainty Estimation)-based active learning framework that integrates evidential uncertainty estimation to enable efficient retinal diagnosis across optical coherence tomography (OCT) and color fundus photography (CFP) modalities. Our approach employs FMUE's uncertainty-aware classifier to guide sample selection through evidential deep learning principles. Evaluation across four retinal datasets demonstrates consistent superiority over traditional active learning methods, including entropy-based selection and Bayesian Active Learning by Disagreement (BALD), with pronounced improvements at low-annotation scenarios. The framework achieves accuracy improvements of 0.249 for CFP and 0.194 for OCT from 2% to 4% annotation data, substantially outperforming competing methods (0.014-0.059 and 0.047-0.058 ranges, respectively). Additionally, it delivers a sample selection speed that is 9 times faster compared to BALD in a specific environment. Sample selection analysis reveals that evidential uncertainty guidance results in a more balanced category distribution and increased attention to underrepresented diseases. These findings establish that combining foundation models with evidential uncertainty estimation effectively addresses retinal imaging annotation challenges, providing practical clinical advantages through improved selection mechanisms and computational efficiency.