TABLE III. Semantic Segmentation Results: Segmentation Performance Measured by Accuracy Across All Categories (Acc.), the Dice Coefficient for the Union of COVID-19 Related Scored (Dice), and the Mean Dice Across Scores 0, 2 and 3 (Cat. Dice) as in [23]. Accuracy and Dice Scores Are Heavily Biased Towards the Dominant Background Class, While Cat. Dice Reflects Better the Performance on the Relevant Annotated Pixels.
| Model | Input | Acc. | Dice | Cat. Dice | |||
DeepLabV
 (Royetal.) |
✓ | 0.95 | 0.71 | 0.62 | |||
Ensemble
 (Roy etal.) |
✓ | 0.96 | 0.75 | 0.65 | |||
DeepLabV
 (ours) |
✓ | ✓ | ✓ | 0.93 | 0.76 | 0.70 | |
| Ablation | DeepLabV
|
✓ | 0.92 | 0.72 | 0.64 | ||
DeepLabV
|
✓ | ✓ | 0.93 | 0.74 | 0.70 | ||
DeepLabV
|
✓ | ✓ | 0.93 | 0.74 | 0.68 | ||
| * Note that Roy et al. [23] trained and evaluated only on convex frames, while our method used both linear and convex. | |||||||