Fig. 2.

Validation set performance for glaucoma diagnosis. Performance of CNN, ophthalmologists and traditional algorithms are presented. There were 9 ophthalmologists participating in evaluation of VFs. On the validation set of 300 VFs, CNN achieved an accuracy of 0.876, while the specificity and sensitivity was 0.826 and 0.932, respectively. The average accuracies are 0.607, 0.585 and 0.626 for resident ophthalmologists, attending ophthalmologists and glaucoma experts, respectively. Both AGIS and GSS2 are not able to achieve satisfactory results. Three traditional machine learning algorithms were also included in the experiments. SVM performed best among these machine learning methods, but still much worse than CNN. We also examined the receiver operating characteristic curve (ROC) of CNN and the compared methods. CNN achieved an AUC of 0.966 (95%CI, 0.948–0.985), which outperformed all the ophthalmologists, rule based methods and traditional machine learning methods by a large margin