What it does
SC2-Net uses a Convolutional Neural Network to predict a possible disease like Covid-19 from a chest x-ray image. It was fed 2k+ images to train on and is able to make predictions with 70-80% accuracy on most new samples. I also developed a web user interface that allows you to upload an image and requests a prediction from the model. The prediction result is then visualized.
Challenges I ran into
It is really hard to build a usable chest x-ray dataset with Covid-19 samples to train the model and get good generalized results. The main problem is the distinction between normal pneumonia and Covid-19. My first model was really bad at predicting with approx. 20% accuracy, but I was able to improve it over time.
What I learned
This was my first time working with Convolutional Neural Networks, so I had to research and learn a lot about them. That helped me learn something new and interesting about this type of network and how to use them efficiently. I also learned a lot about creating large image datasets for the network.
My project is similar to another existing solution: https://github.com/lindawangg/COVID-Net
What's next for SC2-Net / Covid-19 X-Ray Detector
The model is still overfitting and needs more data for training and validation. It can also be improved by tweaking some parameters to make predictions more accurate.