Table 6 Performance of the ResNet50 enhanced model used different numbers of attention heads on the test dataset.
From: Diabetic retinopathy classification using a multi-attention residual refinement architecture
Number of attention heads | Sensitivity | Specificity | Kappa | AUC |
|---|---|---|---|---|
\(H=2\) | 0.78 | 0.80 | 0.80 | 0.87 |
\(H=3\) | 0.79 | 0.80 | 0.81 | 0.88 |
\(H=4\) | 0.80 | 0.81 | 0.82 | 0.89 |
\(H=5\) | 0.79 | 0.81 | 0.82 | 0.88 |
\(H=6\) | 0.78 | 0.80 | 0.80 | 0.87 |
\(H=7\) | 0.77 | 0.80 | 0.80 | 0.86 |
\(H=8\) | 0.77 | 0.79 | 0.80 | 0.85 |
