COVID-19 detection is only for educational purposes that don't mean highly accurate COVID-19 diagnosis system.

Positive samples come from this repo. And I used Chest X-Ray Images (Pneumonia) dataset for the samples of negative COVID-19.

Train

• Positives: 67 images
• Negatives: 67 images

20% of training samples were used for the validation set.
Test

• Positives: 6 images
• Negatives: 6 images
  Positive samples
  Negative samples

As we can see here, the resolution and quality of images of each category are a bit different, but I just ignore that for the educational purpose.

I grabbed VGG16 and added some layers, Avg_Pool2D, 2 Dense layers, and Dropout between the dense layers. Then trained only layers added.

Added layers

As we can see here, it is underfitting on training data but the results on test data I got were 1.0 for Recall, Precision, and F1 score.

I can definitely say that the number of samples for test data is not enough, however, I could get a good result even small samples of training data with finetuning.