The Dataset used for training this model was made public by Dr. Joseph Paul Cohen, a Post-Doctoral Fellow at University of Montréal, through an open-source database of COVID-19 chest X-Ray and CT Images on this GitHub Repo.
Dr. Cohen's Repo contains COVID-19, MERS, SARS, and ARDS infected patients' X-Ray Scans. I have extracted the COVID-19 Positive Samples from his Repo.
For the Healthy Patient Samples, I extracted images from the Kaggle Chest X-Ray Images Dataset.
I have saved both the COVID-19 Positive and Healthy X-Ray Samples in the dataset folder in this GitHub Repo. The size of the dataset is very limited. The community is still working to enhance the dataset.
Here are samples from our COVID-19 patient Chest X-Ray Images Dataset.
| COVID-19 Positive Samples | COVID-19 Negative Samples |
|---|---|
 |
 |
The Data Prepare file contains the code to rename and preprocess the images in the dataset folder. Execute this code this before heading for the training code.
The Train_Covid19 file contains well documented code to build and train the deep learning model from scratch.
Model Performance
| Accuracy Graph | Loss Graph |
|---|---|
 |
 |
From the above plots we can observe that the Model does not Overfit even though our dataset is having limited training data.
Classification Report
| Classficiation Report |
|---|
 |
From the above results, we can see that the Model has obtained 93% Accuracy on our dataset based only on X-Ray Images and no other data.
The Model has also obtained 100% sensitivity and 88% specificity which implies:
-
Patients that are infected with COVID-19 (True Positives), have been accurately identified by the model as “COVID-19 Positive” 100% of the time.
-
Patients that are NOT infected with COVID-19 (True Negatives), have been accurately identified by the model as “COVID-19 Negative” 88% of the time.
NOTE:
We are able to accurately detect COVID-19 with 100% Accuracy which is amazing, however, our true negative rate is a not convincing enough, for instance, we don’t want to classify someone as “COVID-19 negative” when they are “COVID-19 positive”.
DISCLAIMER
The above project does not intend to 'solve' the COVID-19 Virus detection problem, nor does it claim to be a certified/authorized medical solution. The results obtained are shared just for educational purposes.