Table 2.
Average patch and WSI prediction time, slowdown and number of trainable parameters for each of the CNN architectures considered in this work.
| Model | Avg. Prediction Time (patch) | Avg. Prediction Time (WSI) | Slowdown * | Trainable Parameters |
|---|---|---|---|---|
| PROMETEO | 3.054 ± 4.845 ms | 4.201 ± 1.739 s | 1× | 1,107,010 |
| ResNet34 | 8.982 ± 10.086 ms | 10.712 ± 3.134 s | 2.55× | 21,800,107 |
| InceptionV3 | 41.301 ± 44.282 ms | 49.076 ± 14.353 s | 11.68× | 23,851,784 |
| VGG16 | 28.664 ± 9.241 ms | 34.921 ± 10.160 s | 8.31× | 138,357,544 |
| VGG19 | 29.931 ± 9.305 ms | 36.250 ± 10.536 s | 8.63× | 143,667,240 |
| MobileNet | 25.689 ± 10.986 ms | 31.110 ± 9.030 s | 7.41× | 4,253,864 |
| DenseNet121 | 42.489 ± 16.859 ms | 51.483 ± 14.945 s | 12.25× | 8,062,504 |
| Xception | 34.050 ± 11.789 ms | 41.764 ± 12.175 s | 9.94× | 22,910,480 |
| ResNet101 | 43.287 ± 14.679 ms | 52.517 ± 15.266 s | 12.50× | 44,707,176 |
* Calculated by using the average prediction time per WSI and taking the PROMETEO architecture as reference. A slowdown of A× means that model B is A times slower than PROMETEO.