cxd

A deep learning based chest x-ray diagnosis web application

Comment

Inspiration

In 2020, when COVID was wreaking havoc, I was suffering from a disease that was later diagnosed to be tuberculosis. Hospitals were not safe, especially for me. Local doctors were not able to diagnose my condition. Long before the successful diagnosis, due to my symptoms, I got my chest x-rayed. But being an engineer I for the love of the world could not understand it. So, one day while holding my x-ray it strikes me. I am an engineer, software developer no less. So, if I can't do the work, I can make my computer do it for me! Hence, began a long rigorous journey of me building this application.

What it does

After long research (the research paper is in the process of begin published), I was able to not only replicate chexnet but also create my own dense convolutional neural net models that helped diagnose chest x-rays. Later I built a web application using which one can diagnose his/her chest x-ray.

How I built it

With the inspiration in my mind, I started searching on the internet about any previous researches that would help me diagnose chest x-ray using convolutional neural networks. After going through a few papers, I found the ChextNet paper by Stanford and really liked it. I spent some time reading and understanding it. After I understood it, I tried to replicate it, there were some challenges but finally I was able to replicate it. Then I tried training a few more models and compared them to the chexnet replica. Finally, I used Django to build the backend APIs, React Js to build the frontend and TensorFlow Js to integrate deep learning models to react js.

Challenges we ran into

Training models were computationally expensive, I did not have any GPU and the dataset was also large. So, I used TPU on Kaggle to train the model.

Accomplishments that we're proud of

Successfully build the application. I am most proud that I learnt how to integrate Deep Learning models in the client end.

What we learned

I learned a lot:

  1. How to use TPUs to train DL models
  2. Dockerizing and inter-container communication
  3. Building frontend application
  4. Using TensorFlow Js

What's next for cxd

Well medicine is a very sensitive field. It would not only require more data but also a lot more research to make it live.

Built With

Share this project:

Updates