Pretrained Models
January 11, 2021 ยท View on GitHub
The naming convention for models uses the following format:
COVID-Net CT-<dataset version> <architecture version> (<minor dataset version if applicable>)
For example, a COVID-Net CT model using the Large (L) architecture which was trained on the COVIDx CT-2 dataset's "A" variant would be called "COVID-Net CT-2 L (2A)".
COVID-Net CT-1 Models
These models are trained and tested on the COVID CT-1 dataset.
| Model | Type | Input Resolution | COVID-19 Sensitivity (%) | Accuracy (%) | # Params (K) | FLOPs (G) |
|---|---|---|---|---|---|---|
| COVID-Net CT-1 L | ckpt | 512 x 512 | 97.3 | 99.1 | 1399.38 | 4.18 |
| COVID-Net CT-1 S | ckpt | 512 x 512 | 94.7 | 98.5 | 447.57 | 1.94 |
COVID-Net CT-2 Models
These models are trained and tested on the COVIDx CT-2 dataset. Notably, COVID-Net CT-2 L (2A RAD) is a special version of the model trained exclusively on cases where slice selection or segmentation was performed manually by a radiologist.
| Model | Type | Input Resolution | COVID-19 Sensitivity (%) | Accuracy (%) | # Params (K) | FLOPs (G) |
|---|---|---|---|---|---|---|
| COVID-Net CT-2 L (2A) | ckpt | 512 x 512 | 96.2 | 98.1 | 1399.38 | 4.18 |
| COVID-Net CT-2 S (2A) | ckpt | 512 x 512 | 95.7 | 97.9 | 447.57 | 1.94 |
| COVID-Net CT-2 L (2A RAD) | ckpt | 512 x 512 | 96.4 | 98.3 | 1399.38 | 4.18 |