Deep Learning-Based Assessment for Media Haze and Retinal Vascular Leakage of Uveitis

To apply a deep learning-based approach for the automated assessment of media haze and vascular leakage in uveitis using CFP and FFA, and to evaluate its performance against conventional methods.

A total of 756 CFP images and 740 FFA images from 213 uveitis patients were collected. EfficientNetV2-L, InceptionV3, and MobileNetV3 models were developed for media haze assessment using annotations from intermediate ophthalmologists. LadderNet was used for segmenting vascular and leakage areas. Correlation analyses were conducted between media haze, inflammatory factors, and vascular leakage. K-means clustering was applied to identify leakage patterns, and follow-up validations were performed to evaluate treatment efficacy.

In the 9-level media haze classification, EfficientNetV2-L achieved the highest performance with an average Micro-AUC of 0.933, outperforming InceptionV3 (0.893) and MobileNetV3 (0.683). Under a simplified 6-level scoring system, EfficientNetV2-L maintained its superiority with an average Micro-AUC of 0.906. LadderNet demonstrated high accuracy in vascular and leakage segmentation, with Dice similarity coefficients (DSC) of 0.95 and 0.89, respectively. Significant positive associations were found between media haze and leakage area, as well as between the neutrophil-to-lymphocyte ratio (NLR) and leakage area, relative leakage area, and leakage rate. K-means clustering identified distinct leakage patterns, and follow-up validations indicated reductions in leakage severity and NLR post-treatment.

This study underscores the potential of deep learning in automating uveitis diagnosis, improving accuracy, and offering novel indicators for disease activity and treatment outcomes.