Always searched for an answer to eliminate the hassle of asking Medical Professionals to collect appropriate data for Deep Learning Models. It can be very time consuming and difficult task to get a regulated and reliable dataset.
Being a researcher, I was interested in implementing the new paper and also inspired by the structured format of a dataset recently released by NIH for chest X-rays. Gave a deep thought and considered implementing it in the Hackathon.
What it does
Automating the system of Dataset generation as well as Sharing for guided assistance to the Medical Professionals by the Medical Professionals.
How I built it
Trained a DenseNet121 Deep learning architecture on a Dataset of 112,120 frontal-view X-rays. The network was configured in PyTorch Python and trained on a Rutgers University GPU Cluster . The Web Dev is done in Django python to create a platform for professionals to share it with others.
Challenges I ran into
Challenged severely by Web Dev part of the project. Reason - The Fine Tuning of the Model consumed a lot of time thus leaving very less time for Web Dev and UI subsection of the project.
Accomplishments that I'm proud of
An amazing average accuracy of 85.0846 on the test set trained on a 121 layer Dense CNN.
What I learned
Implementation of Class Activation Maps to localize the disease visualizations on XRays. Specifically 3 different approaches and significance of each.
What's next for DocBase
Possible Deployment as a Web App and Improvements like addition of more deep learning models to diagnose more diseases as well as create a structured dataset on the go.