COVID-19 Prediction through Convolutional Neural Network Project Manual

The project consists of two major segments, which can be found in the folder named “Project Files”.

First One is the Jupyter Notebook (a .ipynb file), which consists of the actual work. As it was created in google colab thats why to run this notebook we don’t need to worry about any dependencies or package installations, what we do need is a stable internet connection to open this notebook in google colab. After that we won't face any problems running each cell.

Fig-1: Google Colab Jupyter Notebook

For more convenient here is the Google Colab Jupyter Notebook Link:
https://colab.research.google.com/drive/1IqxPYdg94sMKmbeMTkfCOPhvLBeuwauc?usp=sharing

Just click on the link, and we are good to go to run each & every cell.

Second One is the web app (a covid.py file)

Location: 201714039_Tasfik_SEC-A_Pattern_Project\Project Files\Covid19-model\covid.py, which is the user interface to interact with the research work. To run this project you need to install below python packages & libraries:

1. Flask
2. Numpy
3. Tensorflow
4. Base64
5. PIL
6. io

After successfully installing all the above mentioned packages, we have to open the windows terminal and go to the location where the covid.py file is. Then we have to type the following commands in the terminal-

1. Set FLASK_APP=covid
2. flask run

Fig-2.1: Running flask app from the terminal

We can ignore the warnings and if it runs successfully then our web app will be available at this address http://localhost:5000/static/predict.html .

We can goto the above address and there we will find an option to choose an image button, which will lead us to our gallery where we can choose any x-ray image that we want to predict if it is covid positive or not.

Fig-2.2: COVID19 Prediction Web APP

After uploading the image, we can see the uploaded image in the browser, now the only thing that we need to do is to push the predict button and within a few seconds we will be able to see the results.

Fig-2.3: COVID19 Prediction Result

**To exit from the web app, we must close the running terminal and then close the browser

Title:

“COVID19 Prediction Using Convolutional Neural Network”

Introduction:

Right now we are going through a very tough time, this COVID19 pandemic has hit almost all the industry and damaged the health industry severely and the most alarming situation is that we have not yet been able to invent any medications that can prevent or cure this disease even after 6 months, more than 250 countries are its victims already with 9,210,543 total cases, 474,818 total deaths, 3,778,373 active cases out of which 57,909 are in critical condition.

As of till now to test if a patient is covid +ve or not he or she has to do blood/saliva test. We all know these testing procedures can take hours for each patient. With this deep learning approach this can be reduced to minutes and its accuracy also is quite high.

Methodology:

1. Dataset Acquisition:

   For the training and testing purpose we collected our data from two different sources. One we collected from a github repo, consisting 142 covid19 positive x-ray images and the other one we collected from kaggle for normal x-ray images, we picked 142 such images from this kaggle dataset also to keep the dataset balanced.

2. Splitting the data:

   After we acquired our data we split the whole data into two sets one for training and another for testing purposes.

   So, these two sets consist of 71 Covid +ve and 71 Covid -ve images each.

3. Designing the CNN model architecture:

   To solve this problem using a deep learning approach, we considered Convolutional Neural Network.

   Which consists of 5 convolutional layers, 3 max pooling layers to reduce the dimensionality, 4 dropout layers to prevent the network to be biased on any particular perceptron.

4. Training our data with the designed model:

   Before feeding data into our model we normalized all pixel values in 0 to 1 range. With the batch size 32 we feed our data to the model in 7 steps per epoch. Which took a total of 10 epochs to train the model.

Result Discussion:

After the training finished the result that we got:

Loss	0.1138
Accuracy	0.9598
Validation Loss	0.0694
Validation Accuracy	0.9667

Fig-1: Ground Truth vs Prediction Result

As we can see from the above diagram our model only predicted two incorrect predictions, which are two covid -ve as covid +ve which is less severe, in false positive perspective as we can not afford false negative states.

Fig-2: Confusion Matrix

We can also observe the same result from the figure 02 confusion matrix, we don’t have any false negatives. We have 30 true negatives, 28 true positives and 2 false positives which is affordable in a medical sense.