Two different dataset has been tried.
- X-ray dataset
- CT-scan dataset
There are three classes in this dataset
- covid
- normal
- pneumonia
No seprate test holdout dataset is used. Instead 5 fold cross validation technique is used. The dataset is taken from tawsifur/COVID-19-Chest-X-ray-Detection github repo. The author achieved an F1-score of 0.983, where as we got an F1-score of 0.981.
| Precision | Recall | F1-score | Accuracy | |
|---|---|---|---|---|
| Fold 1 | 0.99 | 0.98 | 0.98 | 0.98 |
| Fold 2 | 0.99 | 0.98 | 0.99 | 0.98 |
| Fold 3 | 0.97 | 0.98 | 0.98 | 0.97 |
| Fold 4 | 0.98 | 0.98 | 0.98 | 0.97 |
| Fold 5 | 0.98 | 0.98 | 0.98 | 0.97 |
We achieved an F1 score of 0.98. Result can be reproduce through this notebook
In order to download data in notebook you need to provide it a kaggle API. Go to kaggle.com, click on my account and then go to API section. Click on create new API token. It will download a file , kaggle.json . upload this kaggle.json to content folder of colab. And run the whole notebook.
Dataset is taken from github.com/UCSD-AI4H/COVID-CT repo, where author achieved an F1-score of 0.85.
All CT-scan related files are in CT-scan folder
Here i study effect of learning rate on model performance.
| Plateau LR | Cyclic LR | Constant LR | |
|---|---|---|---|
| Macro Average Precision | 0.90 | 0.89 | 0.85 |
| Macro Average Recall | 0.89 | 0.86 | 0.82 |
| Macro Average F1 Score | 0.90 | 0.86 | 0.82 |
| Accuracy | 0.90 | 0.86 | 0.86 |
| AUC score | 0.90 | 0.86 | 0.82 |
Jeremy Kohn has compiled all image based diagnostic of coronavirus at one place. Do check his repo github.com/jeremykohn/rid-covid
@INPROCEEDINGS{9318212,
author={T. {Anwar} and S. {Zakir}},
booktitle={2020 IEEE 23rd International Multitopic Conference (INMIC)},
title={Deep learning based diagnosis of COVID-19 using chest CT-scan images},
year={2020}, volume={}, number={}, pages={1-5},
doi={10.1109/INMIC50486.2020.9318212}}