Table 6.
The following table compares how accurate the extracted features are and how computationally complex they are. Computing times are an average of three training runs using various TL models.
| TL network | Dataset | Accuracy | Runtime-training | Runtime-test (s) | Computational complexity |
|---|---|---|---|---|---|
| Inception v4 | Kather texture 2016 | 97.31 | 20 min, 58 s | 0.36 | Low |
| NCT-CRC-HE-100K | 98.47 | 7 h, 16 min, 29 s | 0.38 | ||
| VGG-16 | Kather texture 2016 | 97.93 | 23 min, 44 s | 0.33 | Medium |
| NCT-CRC-HE-100K | 98.86 | 8 h, 09 min, 27 s | 0.34 | ||
| VGG-19 | Kather texture 2016 | 98.08 | 27 min, 51 s | 0.46 | High |
| NCT-CRC-HE-100K | 99.17 | 9 h, 33 min, 12 s | 0.43 | ||
| DenseNet-169 | Kather texture 2016 | 98.48 | 50 min, 38 s | 0.76 | High |
| NCT-CRC-HE-100K | 99.57 | 19 h, 40 min, 30 s | 0.72 | ||
| DenseNet-201 | Kather texture 2016 | 98.54 | 1 h min, 16 s | 1.12 | High |
| NCT-CRC-HE-100K | 99.68 | 22 h, 06 min, 13 s | 1.43 | ||
| ResNet-101 | Kather texture 2016 | 98.30 | 49 min, 26 s | 0.61 | Low |
| NCT-CRC-HE-100K | 99.29 | 18 h, 36 min, 44 s | 0.64 | ||
| ResNet-110 | Kather texture 2016 | 98.43 | 54 min, 12 s | 0.67 | Low |
| NCT-CRC-HE-100K | 99.35 | 20 h, 54 min, 21 s | 0.72 | ||
| ResNet-152 | Kather texture 2016 | 98.64 | 1 h, 17 min, 28 s | 0.75 | Medium |
| NCT-CRC-HE-100K | 99.42 | 23 h, 35 min, 37 s | 0.78 | ||
| ResNet-164 | Kather texture 2016 | 98.73 | 1 h, 32 min, 20 s | 0.83 | High |
| NCT-CRC-HE-100K | 99.63 | 26 h, 03 min, 41 s | 0.89 | ||
| dResNet-101 | Kather texture 2016 | 98.80 | 1 h, 07 min, 04 s | 0.66 | Medium |
| NCT-CRC-HE-100K | 99.79 | 21 h, 23 min, 31 s | 0.69 |