Table 1. The quantitative comparisons of our approach and other state-of-the-art methods for binary-class and multi-class DR grading tasks on Messidor and EyePACS datasets. “-” denotes no results reported in corresponding works. The best results were shown in bold. a .
| Methods | Parameters | Messidor |
EyePACS |
||||||
|---|---|---|---|---|---|---|---|---|---|
| AUC | Acc | Pre | Recall | F1-score | Kappa | Kappa | Acc | ||
| Our DACNN | 157.86M | 0 . 985 | 0.938 | 0.946 | 0.907 | 0.926 | 0.879 | 0.886 | 0.880 |
| CABNet [12] | 25.19M | 0.969 | 0.931 | 0.929 | 0.902 | 0.915 | - | 0.868 | 0.862 |
| LAT [9] | 75.5M | 0.979 | - | - | - | - | 0.851 | 0.884 | - |
| CANet [11] | 29.03M | 0.963 | 0.926 | 0.906 | 0.920 | 0.913 | - | - | - |
| DR|GRADUATE [58] | 7.82M | 0.910 | 0.912 | 0.933 | 0.614 | 0.741 | 0.710 | 0.740 | 0.536 |
| ResNet-50 [30] | 25.54M | 0.880 | 0.929 | 0.857 | 0.796 | 0.826 | 0.781 | 0.653 | 0.815 |
| MobileNet-1.0 [59] | 3.2M | 0.867 | 0.927 | 0.872 | 0.764 | 0.815 | 0.769 | 0.513 | 0.798 |
| Inception-v3 [60] | 21.77M | 0.876 | 0.932 | 0.883 | 0.780 | 0.828 | 0.786 | 0.657 | 0.824 |
a
AUC: the area under the receiver operating characteristic (ROC) curve; Acc: accuracy; Pre: precision