Table 5 Summary of DL methods using FP and OCT to detect eye disease
From: Application of machine learning in ophthalmic imaging modalities
Authors | Year | Imaging Modalities | Aim | Data sets | DL techniques | Performance |
|---|---|---|---|---|---|---|
Arcadu F et al. [111] | 2019 | FP | Diabetic macular thickening detection | Local:17,997 FPs | Inception-v3 | AUC:0.97 (central subfield thickness ≥ 250 μm)0.91 (central foveal thickness ≥ 250 μm)0.94 (central subfield thickness ≥ 400 μm)0.96 (central foveal thickness ≥ 400 μm) |
Nagasawa T et al. [112] | 2019 | FP | Treatment-naïve proliferative diabetic retinopathy detection | Local:132 FPs | VGG-16 | Sensitivity: 94.7%Specificity: 97.2%AUC: 0.969 |
Phan S et al. [113] | 2019 | FP | Glaucoma detection | Local:3312 FPs | VGG-19ResNet-152DenseNet-201 | AUCs of 0.9 or more (3 DCNNs) |
Nagasato D et al. [114] | 2019 | FP | Branch retinal vein occlusion detection | Local:466 FPs | VGG-16SVM | Sensitivity: 94.0%Specificity: 97.0%positive predictive value (PPV): 96.5%negative predictive value (NPV): 93.2%AUC: 97.6% |
Burlina PM et al. [115] | 2019 | FP | To develop DL techniques for synthesizing high-resolution realistic fundus images | Local:133,821 FPs | GAN | AUC:0.9706 (model trained on real data) 0.9235 (model trained on synthetic data) |
Girard F et al. [116] | 2019 | FP | Joint segmentation and classification of retinal arteries and veins | Public:DRIVE, 40 FPsMESSIDOR, 1200 FPs | CNN | Accuracy: 94.8% Sensitivity: 93.7% Specificity: 92.9% |
Coyner AS et al. [117] | 2018 | FP | Image quality assessment of fundus images in ROP | Local: 6043 FPs | VGG-19 DCNN | Accuracy: 89.1% AUC: 0.964 |
Keel S et al. [118] | 2018 | FP | Detection of referable diabetic retinopathy and glaucoma | Public:LabelMe, 114,906 FPs (referable DR) | Â | Sensitivity:90% (glaucomatous optic neuropathy) 96% (referable DR) |
Sayres R et al. [119] | 2018 | FP | Assist grading for DR | Public: EyePACS, 1796 FPs | Inception v-4 | Sensitivity:79.4% (unassisted) 87.5% (grades only) 88.7% (grades plus heatmap) |
Peng Y et al. [120] | 2018 | FP | Automated classification of AMD severity | Public: AREDS, 59302 FPs | DeepSeeNet (Inception v-3) | Accuracy: 0.671 AUC: 0.94 (large drusen) 0.93 (pigmentary abnormalities) 0.97 (late AMD) |
Guo Y et al. [121] | 2018 | FP | Retinal vessel detection | Public: DRIVE, 20 FPs STARE, 20 FPs | Multiple DCNNs | Accuracy: 95.97% (DRIVE training dataset) 96.13% (DRIVE testing dataset) 95.39% (STARE dataset) AUC: 0,9726 (DRIVE training dataset) 0.9737 (DRIVE testing dataset) 0.9539 (STARE dataset) |
Khojasteh P et al. [122] | 2018 | FP | Detection of exudates, microaneurysms and hemorrhages | Public: DIARETDB1, 75 FPs e-Ophtha, 209 FPs | CNN | Accuracy: 97.3% (DIARETDB1 dataset) 86.6% (e-Ophtha) Sensitivity: 0.96 (exudates) 0.84 (hemorrhages) 0.85 (microaneurysms) |
Gargeya R et al. [123] | 2017 | FP | Automated identification of DR | Public: EyePACS, 75,137 FPs MESSIDOR 2, 1748 E-Ophtha, 463 FPs | DCNN | Sensitivity: 94% Specificity: 98% AUC: 0.97 |
Burlina PM et al. [63] | 2017 | FP | Automated grading of AMD | Public: AREDS, more than 130,000 FPs | DCNN | Accuracy: 88.4% (SD, 0.5%)-91.6% (SD, 0.1%) AUC: 0.94 (SD, 0.5%)-0.96 (SD, 0.1%) |
Ordóñez PF et al. [124] | 2017 | FP | To improve the accuracy of microaneurysms detection | Public: Kaggle, 88,702 FPs Messidor, 1200 FPs DiaRerDB1, 89 FPs | Standard CNNVGG CNN | Sensitivity > 91% Specificity > 93% AUC > 93% |
Takahashi H et al. [58] | 2017 | FP | Improving staging of DR | Local: 9939 FPs | GoogleNet DCNN | Prevalence and bias-adjusted Fleiss’kappa (PABAK): 0.64 (modified Davis grading) 0.37 (real prognosis grading) |
Abbas Q et al. [125] | 2017 | FP | Automatic recognition of severity level of DR | Local: 750 FPs | DCNN | Sensitivity: 92.18% Specificity: 94.50% AUC: 0.924 |
Pfister M et al. [126] | 2019 | OCT | Automated segmentation of dermal fillers in OCT images | Local: 100 OCT volume data sets | CNN (U-net-like architecture) | Accuracy: 0.9938 |
Fu H et al. [127] | 2019 | OCT | Automated angle-closure detection | Local: 4135 anterior segment OCT images | CNN | Sensitivity: 0.79 ± 0.037 Specificity: 0.87 ± 0.009 AUC: 0.90 |
Masood S et al. [128] | 2019 | OCT | Automatic choroid layer segmentation from OCT images | Local: 525 OCT images | CNN (Cifar-10 model) | Accuracy: 97% |
Dos Santos VA et al. [129] | 2019 | OCT | Segmentation of cornea OCT scans | Local: 20,160 OCT images | CNN | Accuracy: 99.56% |
Asaoka R et al. [130] | 2019 | OCT | Diagnosis early-onset glaucoma from OCT images | Local: 4316 OCT images | CNN | AUC: 93.7% |
Lu W et al. [131] | 2018 | OCT | Classification of multi-categorical abnormalities from OCT images | Local: 60,407 OCT images | ResNet | Accuracy: 0.959 AUC: 0.984 |
Schlegl T et al. [132] | 2018 | OCT | Detection of macular fluid in OCT images | Local: 1200 OCT scans | CNN | Intraretinal cystoid fluid detection: Accuracy: 0.91 AUC: 0.94 Subretinal fluid detection: Accuracy: 0.61 AUC: 0.92 |
Prahs P et al. [133] | 2018 | OCT | Evaluation of treatment indication with anti-vascular endothelial growth factor medications | Local: 183,402 OCT scans | GoogleNet inception DCNN | Accuracy: 95.5% Sensitivity: 90.1% Specificity: 96.2% AUC: 0.968 |
Shah A et al. [134] | 2018 | OCT | Retinal layer segmentation in OCT images | Local: 3000 OCT scans | CNN | Average computation time: 12.3 s |
Chan GCY et al. [135] | 2018 | OCT | Automated diabetic macular edema classification | Public: Singapore Eye Research Institute, 14,720 OCT scans | AlexNet, VGG, GoogleNet | Accuracy: 93.75% |
Muhammad H et al. [136] | 2017 | OCT | Classification of glaucoma suspects | Local:102 OCT scans | CNN, Random forest | Accuracy: 93.1% (retinal nerve fiber layer) |
Lee CS et al. [81] | 2017 | OCT | Segmentation of macular edema in OCT | Local:1289 OCT images | U-Net CNN | cross-validated Dice coefficient: 0.911 |
Lee CS et al. [137] | 2017 | OCT | Classification of normal and AMD OCT images | Public:Electronic medical records, 101,002 OCT images | VGG-16 | Accuracy: 87.63% AUC: 92.78% |