Table 4.
Test Dataset | Model | AUC (%) | F1max | Recall (%) | Precision (%) | TP | TN | FP | FN | Threshold∗ |
---|---|---|---|---|---|---|---|---|---|---|
Kaggle (internal) | VGG19 | 86.1 | 0.68 | 63.8 | 72.3 | 102 | 846 | 39 | 58 | 0.651 |
ResNet50 | 89.4 | 0.69 | 66.9 | 71.3 | 107 | 842 | 43 | 53 | 0.434 | |
DenseNet201 | 88.7 | 0.70 | 61.9 | 81.2 | 99 | 862 | 23 | 61 | 0.706 | |
InceptionV3 | 88.3 | 0.68 | 71.9 | 63.9 | 115 | 820 | 65 | 45 | 0.203 | |
EfficientNetV25 | 89.2 | 0.69 | 60.6 | 78.9 | 97 | 859 | 26 | 63 | 0.775 | |
VAN_Small | 90.7 | 0.68 | 68.1 | 68.6 | 109 | 835 | 50 | 51 | 0.114 | |
SWIN_Tiny | 95.7 | 0.80 | 83.1 | 76.4 | 133 | 844 | 41 | 27 | 0.690 | |
CrossViT_Small | 93.8 | 0.74 | 74.4 | 73.0 | 119 | 841 | 41 | 41 | 0.704 | |
ViT_Small | 94.5 | 0.76 | 72.5 | 80.0 | 116 | 856 | 29 | 44 | 0.840 | |
SEED (External) | VGG19 | 94.0 | 0.69 | 67.5 | 70.1 | 368 | 4753 | 157 | 177 | 0.992 |
ResNet50 | 92.4 | 0.63 | 62.9 | 62.1 | 343 | 4701 | 209 | 202 | 0.986 | |
DenseNet201 | 92.9 | 0.63 | 60.7 | 66.1 | 331 | 4740 | 170 | 214 | 0.984 | |
InceptionV3 | 91.7 | 0.65 | 61.1 | 70.1 | 333 | 4768 | 142 | 212 | 0.974 | |
EfficientNetV25 | 94.1 | 0.67 | 66.7 | 68.4 | 363 | 4742 | 168 | 182 | 0.991 | |
VAN_Small | 95.2 | 0.71 | 71.4 | 69.7 | 389 | 4741 | 169 | 156 | 0.946 | |
SWIN_Tiny | 97.3 | 0.79 | 79.5 | 78.2 | 433 | 4789 | 121 | 112 | 0.990 | |
CrossViT_Small | 94.5 | 0.70 | 69.2 | 71.7 | 377 | 4761 | 149 | 168 | 0.976 | |
ViT_Small | 95.5 | 0.74 | 73.4 | 74.1 | 400 | 4770 | 140 | 145 | 0.926 | |
Messidor (External) | VGG19 | 86.5 | 0.77 | 71.5 | 82.7 | 358 | 624 | 75 | 143 | 0.547 |
ResNet50 | 89.7 | 0.80 | 75.1 | 85.1 | 376 | 633 | 66 | 125 | 0.244 | |
DenseNet201 | 87.8 | 0.77 | 78.8 | 75.5 | 395 | 571 | 128 | 106 | 0.143 | |
InceptionV3 | 89.7 | 0.80 | 73.9 | 87.1 | 370 | 644 | 55 | 131 | 0.646 | |
EfficientNetV25 | 87.9 | 0.78 | 76.5 | 80.0 | 383 | 603 | 96 | 118 | 0.198 | |
VAN_Small | 90.1 | 0.79 | 80.0 | 78.8 | 401 | 591 | 108 | 100 | 0.084 | |
SWIN_Tiny | 96.3 | 0.90 | 87.8 | 91.5 | 440 | 658 | 41 | 61 | 0.810 | |
CrossViT_Small | 90.5 | 0.81 | 81.0 | 81.2 | 406 | 605 | 94 | 95 | 0.623 | |
ViT_Small | 91.4 | 0.82 | 79.6 | 84.9 | 399 | 628 | 71 | 102 | 0.691 |
AUC = area under curve; F1max = maximum F1 score; FN = false negative; FP = false positive; SEED = Singapore Epidemiology of Eye Diseases; TN = true negative; TP = true positive; SWIN = Hierarchical Vision transformer using Shifted Windows; VAN = Visual Attention Network; VGG = Visual Geometry Group; ViT = vision transformer.
Bold values represent the model with the highest F1max value in the respective test dataset.
Threshold based on maximum F1 score.