Arbitrium

• 

  X-Ray upload Option

• 

  Reporting Abnormalities

Inspiration

Amount of Data generated in radiology is much higher than that can be consumed by available radiologists for diagnosis. Availability of radiologists is growing at a very slow rate, while the data goes on growing at faster rate. The advent in Deep Learning has provided us tools to solve this problem in a way that wasn't possible earlier. And thus lead us to affordable healthcare.

What it does

Asks user to upload an X-Ray, It runs an innovative Deep Learning Algorithm over it, It then reports the findings with their respective probabilities & areas on the X-rays suggestive occurrence. All this happens in real-time.

How we built it

Based on the Data Set CXR8, Chest X-Rays dataset, released by NIH a couple of weeks back containing 100,000+ Chest X-Rays of 30,000+ patients with 14 different abnormalities and as well included cohort of normal patients. Next we normalized the data to be consumed for training a novel deep learning algorithm over it which does predicts the occurrences of abnormalities and goes beyond that to also infers the region of importance for the abnormalities present. The validation of the regions was done against the bounding boxes released in the same data set.

Challenges we ran into

Biggest challenge was to train an algorithm with end to end network within 30 hours, availability of GPUs. As opposed to larger networks, we give real time solution back within 1-2 seconds which further constraint us to look for further research to boost the performance of network.

Accomplishments that we're proud of

We have having a perfectly fine working platform where anyone upload an XRay & get probable findings and areas of concern. Though trained within a short time, algorithm has accuracy of more that 80% for all the tags which is bound to increase if spent more time with.

What we learnt

Taking the research idea into real world and doing that in a short time is difficult but possible. Integrating Deep learning with web based application comes with its own set of exciting opportunities and challenges. We finally feel upon further work on our Hack, we can deliver a very promising solution for TB screenings & a complete Chest X-ray diagnostics Solution.

What's next for Arbitrium

Currently we've worked only on 3 tags (3 parameters to report findings on chest x-rays). Next stage is to include other tags, improve accuracies and working further to research on training with small amounts of data by combining domain knowledge.

Built With

• deep-learning
• python
• tornado