Table 5.
Performance of CoSev vs. other models.
| Paper | Brief Description | Metrics |
|---|---|---|
| Multi-Dataset Papers | ||
| Jin et al. [25] | External validation on MosMet; trained on other datasets. | 93.25 |
| Yousefzadeh et al. [26] | External validation on MosMet; trained on other datasets. | 95.4 |
| Meng et al. [28] | Trained on several datasets. Segmentation and classification data used. | 94.9 |
| Methodological Deviations | ||
| Dara et al. [29] | Multiple federated learning models used; potential performance inflation. | 94.00 |
| Goncharov et al. [36] | Trained on classification and segmentation data from MosMed dataset. | 93.00 |
| Mittal et al. [37] | Uses upsampled 2d slices. | 94.12 |
| Kollias, Arsenos [39] | CNN+RNN model on 2D slices. | 89.87 |
| State-Of-The-Art CNN Architectures | ||
| DenseNet169 [38] | SOTA 2D CNN. | 61.66 |
| VGG16 [38] | SOTA 2D CNN. | 65.18 |
| VGG19 [38] | SOTA 2D CNN. | 65.62 |
| ResNet-50 [38] | SOTA 2D CNN. | 58.80 |
| Inceptionv3 [38] | SOTA 2D CNN. | 64.48 |
| Comparable 3D Networks | ||
| DenseNet3D121 [39] | SOTA 3D CNN | 79.95 |
| ResNet3D [39] | SOTA 3D CNN | 79.95 |
| MC3 18 [39] | SOTA 3D CNN | 80.24 |
| CovidNet3D [40] | DNAS on 3D CNN Architectures. | 82.29 |
| DecovNet [12] | Custom 3D CNN w/Residual Blocks. | 82.43 |
| CoSev (Ours) | Sequential data-driven training | 81.57 |