Story

Healthcare check-ups are expensive, invasive, and time-consuming. What if there was a way to automate these procedures? A computer could take the role of a doctor in determining whether or not you have certain diseases. We know that there are a lot of models already exist for detection some specific diseases, given a photo or a scan. But how could to average person know how to deploy (much less train) a neural net according to a research paper?

We introduce HealthDe.tech to solve this problem, an centralized app which gives users the ability to detect a multitude of diseases, all a doctor's visit or a medical degree. We curated some of the most common diseases that could be detected accurately using machine learning and computer vision. We then trained state-of-art TensorFlow models according to medical research papers. Finally, we deployed these models on Google's Cloud Platform and verified them against a model trained on Google's AutoML Vision API to deliver fast, accurate results to anyone who decides to use our service.

Challenges

We were first challenged with curating a database of images and pre-processing it to get the best result in any transformation. Medical data from hospitals and universities often contained too many large images to create a model efficiently, even on Google's Cloud. So, we filtered out those that we felt were unfit for training while resizing and recoloring the rest. However, this cleaning process took many hours on Google's Cloud machines, so we turned to another approach: training models locally and deploying afterwards. We made custom tensorflow models with our own architectures tailored for classification of our images, which was challenging too! We also started our backend work in Node.js, so getting Tensorflow on Python interfacing with a React app hosted on an Express backend (which would make calls to some models we deployed using Google's Cloud Vision AutoML) was confusing, to say the least.

Conclusion

We know that doctor/medical resources are limited and are expensive. So by using our app people can do a basic analysis with satisfactory precision and decrease the doctor visits thereby increasing the efficiency of doctors and saving a lot of money!

All you need for our app to run a detection is an x-ray or blood sample diagram or in many cases just a photo!

We do not claim responsibility for any injury or death as a result of false predictions from our service.

Built With

Share this project:
×

Updates