Inspiration

Covid-19 is infecting and killing a lot of people, and the infecting curve is not flattening even after half a year, so we need to prepare for fighting with it for a long time. Hospitals would not be able to focus on this single disease, instead covid19 patients will be mixed with patients with all kinds of diseases. So it would be helpful to have a tool that distinguishes covid19 from other diseases. Given that one important step for covid19 diagnosis is x-ray scan, we developed a model that distinguishes different diseases including covid19, pneumonia, breast cancer, and brain tumor. (more can be integrated in future) We believe this tool can help doctors diagnose covid19 and other diseases from x-ray scans in seconds.

What it does

We developed a model that distinguishes different diseases including covid19, pneumonia, breast cancer, and brain tumour from x-ray images with deep learning techniques.

How we built it

Using the AWS Deep Learning AMI service, we built the model using a concatenation of four datasets, and each dataset is multiclass. From this dataset, we created a train and validation set that was used to train and simultaneously validate the model. We loaded a pre-trained (VGG16) model with our custom parameters and hyperparameters to train the model. Next, the model was deployed using Amazon Sagemaker. From Sagemaker, we got our endpoint which was used by our frontend. Finally, the web application was deployed using AWS Amplify.

Challenges I ran into

Using the AWS Deep Learning AMI service, we built the model using a concatenation of four datasets, and each dataset is multiclass. From this dataset, we created a train and validation set that was used to train and simultaneously validate the model. We loaded a pre-trained (VGG16) model with our custom parameters and hyperparameters to train the model. Next, the model was deployed using Amazon Sagemaker. From Sagemaker, we got our endpoint which was used by our frontend. Finally, the web application was deployed using AWS Amplify. Model Training: Training our model on a concatenated dataset, each with multi classes, required lots and lots of experiments. It was challenging to decide on a specific architecture as well as hyperparameters for tuning the model to achieve a fair loss and accuracy.

Data Source: It’s not easy to find enough data for training. What makes it worse is that the hackathon requires us to use some data from AWS data exchange, in which most of the datasets are not free. The free data sets are usually too small, and some of them took a long time to get permission.

App Responsiveness: We also faced some challenges in making the app responsive to multiple screens.

Accomplishments that I'm proud of

Achievements we are proud of includes running many experiments on a concatenated dataset instead of the popular single dataset. Based On our experiments and tuning, seeing we were able to successfully train and improve the model accuracy was a great accomplishment. We are proud that we are able to develop and deploy the entire web application in the given time frame. As of now, there is no application that can do what our application can so it gives us pleasure to share and demonstrate with others.

What I learned

What we learned that we would always remember and remain happy about was how to work with some AWS services. AWS Sagemaker and Amplify to be specific. For us, we could regard this as our first deployed/production model. We were able to turn something outside the classroom or general learning into a real-world solution.

Also as a team, we learned to work very well with each other considering the differences in cultural and technical background. It posed little challenges but overall was a smooth run.

What's next for A.I. Radiologist

Well, our model is multi-class so we can add more classes and make it a full-fledged AI Radiologist with many other organs. With a better dataset, the accuracy will be increased and our model will give better results.

In the near future, we will be able to predict with any X-ray images for medical use, including all kinds of cancer and even fractures. The accuracy is not too high for now, but it’s purely restricted by the training data size, as long as enough data is available, the model will be more accurate than any human doctors!

Share this project:

Updates