A critical component in the fight in the Covid-19 pandemic and to defeat the coronavirus is large scale testing of people and isolating the ones that need intensive care. To make this large scale testing accessible and feasible for the governments across the world, we thought of leveraging deep learning for an effective and reliable screening of people for Covid-19. Deep Learning, especially neural networks are phenomenally good at finding patterns in the data and learning the correlations between data points in the datasets so it is plausible to think that in theory Neural networks can be used to detect Covid-19 infections from the chest radiography images of the patients. It is important to note that we do not mean to replace thr RT-PCR method of testing but provide and additional tool which may help in increasing the testing percentages.
What it does
We developed a web app where anyone can upload a recent xray scan of their chest and get the result within seconds whether they are tested positive or negative for Covid-19. Currently, the model works with 87% accuracy.
How we built it
Dataset: We combined two different open-sourced datasets 1. RSNA dataset for Pneumonia patients and 2. Covid-19 Image Dataset Collection. That gave us a total of 18,707 images. The dataset is biased towards normal and pneumonia images as we did not have sufficient chest scans of COVID patients.
Model Architecture: The model consists of 5 convolutional layers, 3 Dense layers and on final output layer with 3 neurons and Softmax activation function. Convolutional layers: • The output size of the convolutional layers double after every layer ie from 32—>64, 64—>128 etc • The activation function used in the convolutional layers is rectified linear unit (ReLU) and convolved with a kernel size of 3x3 which provided the best balance between accuracy and faster convergence rate. • MaxPooling layer with a pool size of 2x2 is applied at the end of each convolutional layer to effectively reduce the spatial size of the image thereby reducing the number of parameters to back-propagate and hence reducing the computational cost of the architecture. The Dense layers are regularised with dropout regularisation to prevent overfitting. We had also thought of implementing Image augmentation techniques for better results but were limited by the computational limitations of our devices.
The inspiration behind this was from a Canadian startup Darwin.ai
Challenges we ran into
The main challenge that we ran into was the time constraints of the Hackathon. To build a functional model within 2 days which provides acceptable accuracy scores was a challenge for us but somwhow we managed to build the model.
Accomplishments that we're proud of
To build a fairly accurate ML model and deploying it on the web using GCP within a limited amount of time.
What we learned
This helped us in implementing the theories of deep learning and the web development concepts that we learned to solve a real world problem.
What's next for DTechT
The main thing that we are focusing on right now is to reduce the data imbalance by add more Covid-19 data points to the dataset and increasing the acccuracy of the model to about 95%. We also plan to improve the scalability of the website if it id deployed in various hospitals around the country.
Try It out
css, flask, html, keras, tensorflow