Table 6.
Classification performance of the CSDB CNN and DFL over a baseline CNN with 7 convolutional layers connected in a feedforward manner. Class-wise F1-scores and AUC values are reported for the ablation experiments. It is assumed that ‘n’ denotes the number of convolutional layers on CNN.
| Experiments | F1-score |
AUC |
Accuracy | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| COVID-19 | Normal | Bacterial Pneumonia | Viral Pneumonia | Macro average | COVID-19 | Normal | Bacterial Pneumonia | Viral Pneumonia | Macro AUC | ||
| Baseline 7-layer feedforward CNN | 73.60 | 84.48 | 69.26 | 64.78 | 73.56 | 90.29 | 84.49 | 84.75 | 83.62 | 85.78 | 79.96 |
| Baseline CNN with DFL at n-1 conv layer | 81.78 | 89.90 | 76.02 | 72.89 | 80.14 | 94.51 | 88.16 | 88.45 | 88.24 | 89.84 | 84.87 |
| CSDB CNN | 94.26 | 95.62 | 90.15 | 83.18 | 76.85 | 95.21 | 94.88 | 95.60 | 94.78 | 87.81 | 93.42 |
| CSDB CNN with DFL at n-2, n-1 conv layers | 96.86 | 98.28 | 95.50 | 94.74 | 96.34 | 98.16 | 97.99 | 97.93 | 99.10 | 98.23 | 97.35 |
| CSDB CNN with DFL at n-1 conv layer | 97.20 | 98.71 | 97.05 | 94.66 | 96.90 | 98.01 | 98.46 | 98.84 | 98.27 | 98.39 | 97.94 |