CoNV-AutoGrader: A Deep Learning System for Automated Severity Grading of Corneal Neovascularization From Slit-Lamp Images

Purpose: The purpose of this study was to develop and validate a fully automated framework (CoNV-AutoGrader) for clinically interpretable severity grading of corneal neovascularization (CoNV) from slit-lamp photographs. Methods: A retrospective dataset of 200 slit-lamp images with CoNV was collected from a tertiary eye center. Corneal and CoNV regions were manually annotated and quality controlled by ophthalmologists. CoNV-AutoGrader integrates two U-Net-based segmentation models for corneal region and neovascularization detection with a deterministic geometric scoring module based on a standardized clinical grading system incorporating quadrant involvement and centripetal extent. Model performance was evaluated using five-fold cross-validation and quantified with Dice coefficient, F1 score, intersection over union (IoU), continuous linear Dice (clDice), accuracy, sensitivity, specificity, precision, and area under the receiver operating characteristic curve (AUC). Agreement between automated and expert grading was assessed using intraclass correlation coefficient (ICC) and Bland-Altman analysis. Results: For corneal region segmentation, the model achieved a mean Dice coefficient of 0.971 ± 0.003 and AUC of 0.992 ± 0.003. For CoNV segmentation, the mean Dice coefficient was 0.405 ± 0.011, sensitivity was 0.522 ± 0.037, and AUC was 0.872 ± 0.013. The automated severity score demonstrated good agreement with expert ratings (overall ICC = 0.918, P < 0.001). Bland-Altman analysis demonstrated good agreement between automated predictions and ophthalmologist-assigned scores, with negligible systematic bias and limits of agreement within clinically acceptable ranges. Conclusions: CoNV-AutoGrader enables automated, objective, and reproducible CoNV severity grading directly from slit-lamp images, achieving strong agreement with expert grading and consistent performance across diverse clinical scenarios. Translational Relevance: By transforming U-Net-based vessel segmentation into standardized clinical severity metrics, this framework bridges algorithmic image analysis and real-world CoNV management.