Previous COVID-19 research has observed lung lesions of COVID-19 are generally peripherally distributed and bilateral, but other viral pneumonia lesions are not subject to those standards. Since these factors distinguishing the pneumonias are location based, we propose a novel method of COVID-19 classification by using segmented overlays as part of the classification process.
The performance of our proposed method on Santa Barbara Cottage Hospital dataset is showin the following table:
| Model | Train | Val | Test | Avg Precision | Avg Recall | Subject-wise Accuracy |
|---|---|---|---|---|---|---|
| ResNet152 | 99% | 0.6888 | 0.8420 | 0.8473 | 0.8444 | 0.880 |