Inspiration

• 2015: About 415 million adults with diabetes worldwide - projected to grow by 50% to 642 million adult diabetics by 2040. • About 35% - 50% of all diabetics may have retinopathy. • Of those, 10% may be at risk, i.e., 10 million+ people may become totally blind in 12 years. • Only 55% of diabetics obtain an eye-screening exam. This number is even lower in under-developed countries. • Global need for improved access to enhanced screening and rapid detection is immense. • Currently, detecting DR is a time-consuming and manual process that requires a trained clinician to examine and evaluate digital color fundus photographs of the retina. • By the time human readers submit their reviews, often a day or two later, the delayed results lead to lost follow up, miscommunication, and delayed treatment.

What it does

• Automated classification/evaluation of digital color fundus photographs using models with realistic clinical potential in classifying diabetic retinopathy symptoms

How we built it

• Transfer learning by retraining Google’s InceptionV4 model on AWS P2 instance NVIDIA K80 GPU. • Front end using node.js and react.js and tensorflow models being served from python flask.

Challenges we ran into

• Determining models and initial hyper-parameter tuning. • Networking and infra issues. • Developing web interface to serve tensorflow trained models.

Accomplishments that we're proud of

• Given the limited number of training sets, we achieved the accuracy of 83% and validation accuracy of 78%. • We overcame the challenge of employing google inception v4 model.

What we learned

• Building E2E product using diverse technologies mentioned above and the troubleshooting skills.

What's next for AIconic

• Improving the models accuracy by training with a much larger data set and with training image augmentation.

Share this project:
×

Updates