Tricorder

Inspiration

Coming from computer science background, we had no idea about medical image analysis, so we started exploring on available resources on internet, and tried building model for lung cancer classifier. And we came know that there are a million different parameters come into the picture in medical image analysis, and developers require a lot of help from radiology experts. On the other side, radiologists require faster way to analyse the medical images (like CT scan ). So, we thought of a platform which can connect both.

What it does

Tricorder is a platform where you can upload an image for given disease, i.e., lung cancer, covid19 and get test results from the classifier.

How we built it

As a newbie in medical image analysis, we searched and learned from different types of models. We collected data from multiple sources and used different types of CNN architectures to make classifiers for different tasks. To make these classifiers accessible to end users, made a web page where user can upload image and get test result of particular disease. Technologies used: tensorflow, keras, fastai, flask, jinja templates

Challenges we ran into

Getting lots of errors on model training and over fitting. Less amount of training data, Architecture tuning.

Accomplishments that we're proud of

Troubleshooting problems overnight by researching about topics.

What we learned

We learned how to work with medical images and what kind of research goes into making such systems. How CNN's are built and how we can fine-tune it. Methods to preprocess medical images.

What's next for Tricorder

Our aim is to provide a platform that can help both developers to improve their models and radiologists to get faster test results. Radiology experts can help the developers to improve model accuracy by providing feedback on test result, this way model can be debugged and improved over time. New types of models can be uploaded on the platform and tested for wider audience.