Revolutionizing Corneal Staining Assessment: Advanced Evaluation Through Lesion-aware Fine-Grained Knowledge Distillation

Abstract Corneal staining is crucial for evaluating ocular surface diseases, yet existing AI models for CSS (Corneal Staining Score) assessments struggle with detailed lesion identification and lack applicability in real-world clinical settings. Moreover, the output of current AI-assist staining evaluation system only provides categories of grades, leading to potential “plateau” effect, which could misrepresent treatment response in clinical practices. Addressing these gaps, we developed the Fine-grained Knowledge Distillation Corneal Staining Score (FKD-CSS) model, which effectively distills fine-grained features into the CSS grading process and outputs continuous, nuanced scores for thorough assessments. Trained on 1471 images from 14 centers of heterogenous sources, FKD-CSS demonstrates robust accuracy with a Pearson's r of 0.898 against ground-truth and an area under the curve (AUC) of 0.881 in internal validation, rivaling senior ophthalmologists. Additionally, the model achieved expert performance with considerable Pearson's r (0.844–0.899) and AUCs (0.804–0.883) in external tests in six regions of China using 2376 corneal staining images of dry eye across 23 hospitals, and generalizes to multi-ocular-surface-disease test (Pearson's r: 0.816, AUC: 0.807), underscore its efficiency and explainability for CSS assessment. These results highlight FKD-CSS's potential as a precise, valuable tool for staging and outcome measurement of ocular surface diseases.