Table 1. A summary of studies depicting automated diagnosis of glaucoma using fundal images.
| Paper | Classifier | Number, age | Training and testing | Results | Glaucoma diagnosis | Database |
| Nayak et al 2009[42] | ANN | 61, 37 G, 24 H, 25 to 60 | 46 images used for training, 15 images used for testing | AUROC 0.984 (sensitivity 100%, specificity 80%), no CI | Ill-defined but by an ophthalmologist | Kasturba Medical College, Manipal, India |
| Bock et al 2010[51] | SVM | 575, 239 G, 336 N, 56.1±11.4 | 5 fold cross validation | AUROC 0.88 P<0.07, sensitivity 73%, specificity 85% | Ill defined, stated gold standard | Erlangen Glaucoma Registry, Germany |
| Acharya et al2011[43] | SVM | 60, 30 G, 30 N, 20-70 | 5 fold cross validation | 91% accuracy, no CI, stated P significant is <0.05 | Ill defined | Kasturba Medical College, Manipal, India |
| Mookiah et al 2012[44] | SVM | 60, 30 G, 30 N, 20-70 | 3 fold stratified cross validation | Accuracy 93.33%, sensitivity 86.67%, specificity 93.33%, AUROC 0.984, no CI, stated P significant is <0.05 | Ill-defined but by an ophthalmologist | Kasturba Medical College, Manipal, India |
| Chakrabarty et al 2016[41] | CNN | 314, 169 G, 145 N | 1926 to train, 314 to test | AUROC 0.792 | Gold standard. Diagnosed by 4 glaucoma specialist | Aravind Eye Hospital, Madurai and Coimbatore, India |
| Issac et al 2015[45] | SVM | 67, 32 G, 35 N, 18-75 | Leave one out cross validation | Accuracy 94.11%, sensitivity 100%, specificity 90%, no CI, P significant if less than 0.05 | Ill-defined but by an ophthalmologist | Venu Eye Research Centre, New Delhi, India |
| Maheshwari et al 2017[46] | SVM | Two databases, 60, 30 G, 30 N, 505, 250 G, 255 N, no age range | Three fold and tenfold cross validation | Accuracy 98.33%, sensitivity 100%, specificity 96.67%, no CI, P significant if less than 0.05 | Ill-defined but by an ophthalmologist | Medical Images analysis Group Kasturba Medical College, Manipal, India |
| Singh et al 2016[47] | SVM | 63, 33 G, 30 N, 18-75 | Leave one out cross validation, 44 to train 19 to check | Accuracy 95.24%, sensitivity 96.97%, specificity 93.33%, no CI, P significant if less than 0.05 | Ill-defined but by an ophthalmologist | Venu Eye Research Centre, New Delhi, India |
| Maheshwari et al 2017[48] | LS-SVM | 488, 244 G, 244 N, no age range | Three fold and tenfold, cross validation | Accuracy 94.79%, sensitivity 93.62%, specificity 95.88% | Ill-defined but by an ophthalmologist | Kasturba Medical College, Manipal, India |
| Raghavendra et al 2018[49] | SVM | 1426, 837 G, 589 N | 70% raining, 30% testing, repeated 50 times, random training and testing partitions | Accuracy 98.13%, sensitivity 98%, specificity 98.3%, no CI, P significant if less than 0.05 | Ill-defined but by an ophthalmologist | Kasturba Medical College, Manipal, India |
| Ahn et al 2018[63] | CNN | 1542, 756 G, 786 N, no age range | Randomly partitioned into 754 training, 324 validation and 464 test datasets | AUROC 0.94, accuracy 87.9%, no CI | Ill-defined but likely Anderson Patella Criteria | Kim's Eye Hospital, Seoul, South Korea |
| Christopher et al 2018[52] | CNN | 14822, 5633 G, 9189 N | 10 fold cross validation | AUROC 0.91 (0.9-0.91 CI) | Independent masked graders | The ADAGES study, New and Alabama DIGS Study, California |
| Li et al 2018[53] | CNN | 39745, 9279 G, 30466 N | 8000 images as the validation set, and 31745 images as training set | AUROC 0.986 (95%CI, 0.984-0.988) | Grading by trained ophthalmologists | Label me Data Set |
G: Glaucoma; N: Normal; AUROC: Area under the receiver operating characteristics curve; CI: Confidence interval; CNN: Convolutional neural networks; ANN: Artificial neural network; SVM: Support vector machine; LS-SVM: Least squares support vectors machine; ADAGES: African descent and glaucoma evaluation study; DIGS: Diagnostic innovations in glaucoma study.