Hack4Health AI for Alzheimer's Challenge

Inspiration

Dementia is a neurodegenerative condition that currently has no cure, and we have personally seen the devastating effects it has on families. Early detection is crucial to the treatment of dementia, and we wanted to focus on the image-based machine learning classification as much as possible.

What it does

This model uses convolutional neural networks to classify images of MRI scans of the brain according to severity of dementia

How we built it

Firstly we built, the basic CNN model, then we used the Densenet model to improve the model. After that w e used VGG19 to compare the models. We also used Resetnet and GRAD CAM fr explainability of the models.

Challenges we ran into

As we were collaborating online, the greatest challenge we ran into was communicating with each other throughout the whole process, especially during the holidays. We also had to troubleshoot models and come to consensus on which models and metrics we wanted to use.

Accomplishments that we're proud of

We are firstly proud of completing and submitting our project which we all contributed to. In addition, we’re proud of creating a model with a high degree of accuracy and F1 score, and implementing a Grad CAM based approach for the explainability of the model as well.

What we learned

We learned how important the brainstorming phase is, and how important it is to discuss ideas and have a clear vision before beginning any project. We also learned a lot about neural networks and how each layers connects to each other in the process of implementing Grad CAM.

What's next for Hack4Health AI for Alzheimer's Challenge

We would like to continue working on image classification using larger datasets, and taking advantage of better compute resources to attempt better methods for classification, including upscaling images and testing different activation functions. We would like to have the opportunity in the future to carry out live demos and include submissions which are not necessarily in notebook format. eg: apps

Built With

• https://github.com/srarnab7/hakathon-al