Table 2.
Metrics-based classifier confusion matrices. the models were evaluated with 100 covid-19, 33 ILD, 33 other pneumonia, and 34 no pathologies CT scans. The operating point was chosen as the closest point to the top left corner on the ROC computed over the test dataset (without bootstrapping). Note: the table shows the prediction vs ground truth for each of the negative class categories (ILD, other pneumonia, no pathology). M1, metrics-based random forest classifier; M2, metrics-based logistic regression classifier; M3, Deep learning–based classifier; CO-RADS, SCORING system [16]
| Ground truth | |||||
|---|---|---|---|---|---|
| Positive | Negative | ||||
| COVID-19 | ILD | Pneumonia (non-COVID-19) | No pathology | ||
| Predicted (M1) | Positive | 86 | 21 | 19 | 0 |
| Negative | 14 | 12 | 14 | 34 | |
| Predicted (M2) | Positive | 74 | 11 | 10 | 0 |
| Negative | 26 | 22 | 23 | 34 | |
| Predicted (M3) | Positive | 90 | 3 | 12 | 2 |
| Negative | 10 | 30 | 21 | 32 | |
| Predicted (CO-RADS) | Positive | 74 | 8 | 15 | 0 |
| Negative | 26 | 19 | 18 | 34 | |