Category‐Specific Uncertainty Learning for Retinal Disease Diagnosis

ABSTRACT Integrating artificial intelligence (AI) in retinal disease diagnosis faces challenges in effectively communicating prediction confidence and managing data uncertainty, which is critical for clinical acceptance. To address these limitations, we propose Category‐Specific Uncertainty Learning (CSUL), an advanced AI framework for diagnosing eight retinopathies using optical coherence tomography (OCT) images. CSUL leverages a fine‐tuned ResNet‐50 architecture and an uncertainty‐based classification mechanism to quantify prediction confidence. It further establishes category‐specific thresholds based on uncertainty distributions from the validation set, optimizing the retention of reliable predictions. For the unlabeled test data, CSUL introduces a pseudo‐labeling strategy using K‐means++ clustering, enabling the assignment of appropriate category‐specific thresholds. The proposed framework achieves exceptional performance, with a thresholded diagnostic accuracy of 99.58% and a retention rate of 85.18% on deterministic samples, outperforming conventional single‐threshold models. Notably, for the challenging DRUSEN category, it attains a 97.10% accuracy and a 59.14% retention rate, marking a significant improvement over the existing state‐of‐the‐art methods. In conclusion, the innovative integration of category‐specific thresholding and pseudo‐labeling in CSUL enhances diagnostic accuracy and reliability, addressing the key challenges in AI‐driven retinal disease screening and advancing its clinical applicability.