Seeing doctors and nurses working days and nights tirelessly on the front-line of this pandemic, we thought to ourselves "surely, there has to be something we can do to help them". To help combat this pandemic, we believe it is crucial to be able to test more people for COVID-19 without overloading medical staff. Thus, we set out to develop an application to help medical staff diagnose COVID19 faster with higher precision using deep learning based ML models

What it does

Our web application takes X-ray or CT images of a patient's lung (PA), and predicts whether or not a patient has been infected by COVID-19. We create two classifiers — the first one is to differentiate Pneumonia and healthy lungs. Then, with a second classifier, we differentiate COVID-19 patients from Pneumonia patients to be more effective in diagnosing COVID-19 patients. Both classifiers achieve accuracy of around 90%. Moreover, we use a centralized database to collect every single submission for farther improvement of our ML models. Also, we design a function to help healthcare workers record all the patients' information automatically in Excel to save a significant amount of their time.

How we built it

We implement a multi-layer prediction (chain prediction) to improve the reliability of our deep learning-based ML models using tensorflow framework. Our back-end is developed with Python and Express.JS by utilizing microservice architecture. React was used for front-end and our web application is fully deployed on AWS

Challenges we ran into

It is to be expected that validated data for training purposes is hard to come by for a relatively new disease. Moreover, our model cannot predict the early stage of COVID-19 while the abnormalities of the patient's lungs are still invisible. We lack some knowledge about COVID-19, and we could not consult experts either due to the lockdown policy.

Accomplishments that we're proud of

Even though we did not have as much training data as we hope for, our models are still able to achieve high accuracy. For such a short period of time and lockdown in place, we successfully deploy our web application, ready to help stop this pandemic.

What we learned

As a team, we learned many new skills from each other such as training ML, Web development, web hosting, and reading scientific papers. And this is our first experience working remotely without a single in-person meeting.

What's next for Detective COVID-19

As mention previously, we want to continually improve our models' accuracy when more validated data becomes available. We are exploring the idea of cooperating with new data types for better diagnosis of COVID-19, such as temperature, voice analysis for detecting dry cough, and other COVID-19 symptoms. Our ultimate goal is for our app to be able to diagnose with minimum supervising from medical staff to reduce their workload. Please check out our application by clicking on the link below:

Covid Detection App

Built With

Share this project: