Inspiration
Medical Imaging using computational methods is a very promising area of work right now. The current computer vision models have the potential to drastically transform the landscape right now. This will not only help in providing quality health care to more people but also will reduce the burnout of people in the medical industry.
What it does
CorDet is an online tool for detection of COVID-19 from chest xray radiographs
How I built it
CorDet main CNN classifier architecture follows a MobileNet V1 CNN architecture. This model is then trained by transfer learning on dataset of 250+ chest xrays with both the positive and negative classes equally distributed. The dataset was subjected to image augmentations etc. to prevent overfitting.
Challenges I ran into
Managing the strict memory limits on heroku dynos was not trivial. Had to change the network architecture many times to come up with a memory efficient CNN.
The Dataset
The dataset used here consisted of around 125 COVID-19 positive and around the same number of COVID-19 negative CXRs. COVID-19 positive images came from dataset published on this Github repo maintained by University of Montreal. COVID-19 negative CXRs came from a kaggle challenge There is clearly lesser datat than what you would expect but with time better quality datasets will be available!
What I learned
Taking a trained CNN and integrating it with web applications is not as easy as it sounds. It involves plenty of work and sweat!
What's next for CorDet : An Online tool for detection of COVID-19 from CXRs
Improving the accuracy of classisifers, adding the support for CT-Scans. There is also a possibility of CNN able to diagnose people on the basis of cough sounds. Once there is enough data for the classfier available in the public this feature could be integrated.
Log in or sign up for Devpost to join the conversation.