Table 7.
Classwise segmentation accuracy of models in comparison.
| Model | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| (Pre-trained and data augmented) | BKG | BCC | SCC | IEC | EPI | GLD | INF | RET | FOL | PAP | HYP | KER |
| Thomas et al. (32) | 0.95 | 0.865 | 0.857 | 0.707 | 0.831 | 0.873 | 0.574 | 0.702 | 0.615 | 0.808 | 0.962 | 0.846 |
| MiT-B0 (default) | 0.983 | 0.905 | 0.707 | 0.787 | 0.734 | 0.873 | 0.692 | 0.909 | 0.558 | 0.646 | 0.869 | 0.757 |
| MiT-B0 (proposed) | 0.987 | 0.915 | 0.786 | 0.814 | 0.791 | 0.897 | 0.715 | 0.912 | 0.658 | 0.748 | 0.935 | 0.813 |
| Percentage difference | 4% | 6% | -8% | 15% | -5% | 3% | 25% | 30% | 7% | -7% | -3% | -4% |
The highest accuracy for each class is highlighted as bold. The percentage difference in fourth row has been calculated considering accuracy achieved by Thomas et al. (32) as original and by proposed model as the new one therefore, all the positive values which are highlighted as bold as well, indicating classes for which proposed model has performed better than others.