Automated detection of vitritis using ultrawide-field fundus photographs and deep learning

Purpose: Evaluate the performance of a deep learning (DL) algorithm for the automated detection and grading of vitritis on ultra-wide field (UWF) imaging. Design: Cross-sectional non-interventional study. Method: UWF fundus retinophotographs of uveitis patients were used. Vitreous haze was defined according to the 6 steps of the SUN classification. The DL framework TensorFlow and the DenseNet121 convolutional neural network were used to perform the classification task. The best fitted model was tested in a validation study. Results: 1181 images were included. The performance of the model for the detection of vitritis was good with a sensitivity of 91%, a specificity of 89%, an accuracy of 0.90 and an area under the ROC curve of 0.97. When used on an external set of images, the accuracy for the detection of vitritis was 0.78. The accuracy to classify vitritis in one of the 6 SUN grades was limited (0.61), but improved to 0.75 when the grades were grouped in three categories. When accepting an error of one grade, the accuracy for the 6-class classification increased to 0.90, suggesting the need for a larger sample to improve the model performances. Conclusion: We describe a new DL model based on UWF fundus imaging that produces an efficient tool for the detection of vitritis. The performance of the model for the grading into 3 categories of increasing vitritis severity was acceptable. The performance for the 6-class grading of vitritis was limited but can probably be improved with a larger set of images.