Combating Preventable Vision Loss in Kids with AI-powered Eye Screening

Machine Learning Can Scale Doctors’ Expertise for More Targeted Diagnosis and Treatment

Pr3vent sought to utilize the power of AI to combat preventable vision loss in infants. Because trained doctors who can diagnose eye diseases by a newborn’s retina are rare, the team’s vision was to marry Deep Learning and data to scale the expertise of ophthalmologists who can, to cut per-screen cost, increase accuracy, and improve screening availability along the way.

• Reduce cost per screening through automation
• Scale doctors’ expertise through image analysis, anomaly detection, and data
• Ensure diagnosis accuracy of no less than 97%

The solution as such should be designed as highly accurate in detecting pathology in a newborn’s retina, to receive FDA approval. Resolving this required Pr3vent to accurately label a database of 350K fundus and retina images by a team of experienced ophthalmologists, build an AI-driven image analysis and anomaly detection engine, and develop an application for ophthalmologists to handle retina images.

Pr3vent teamed up with Provectus to deliver an FDA-compliant eye screening solution in a bid to improve lives for over four million infants in the United States.

Embracing Computer Vision with Provectus to Build a Disease Screening Solution for Doctors

To deliver an advanced and user-friendly eye screening solution to Pr3vent, Provectus designed and built an ecosystem consisting of three essential parts that are bundled in a single application for ophthalmologists.

#1 Image Labeling Infrastructure

Pr3vent has exclusive access to a database of more than 350K retina and fundus images of newborns. Pr3vent had the good fortune to attract some of the best ophthalmologists in the US to label the training dataset, but they needed a solution to speed up labeling and make it more efficient than allowed for by standard image-labeling tools.

Provectus moved forward to build an image labeling infrastructure as the first step. Designed for ophthalmologists, it facilitated and accelerated labeling of medical images, enabling doctors to process up to 72 eye screens per minute (12 screens per patient). Faster labeling saved Pr3vent dozens of hours of ophthalmologists’ precious time.

Since data quality is paramount in machine learning, the image labeling infrastructure was designed to eliminate any signs of bias and to ensure objectivity of the training dataset, thereby increasing accuracy of the final machine learning model.

#2 ML Model Building and Training Infrastructure

In addition to labeling quality and data quality, underlying infrastructure impacts the accuracy of models in machine learning. The FDA also requires that any ML infrastructure used to build and train ML models for AI in healthcare must be auditable, explainable, and well-monitored.

As step two, Provectus built an FDA-ready AI infrastructure with fully auditable labeling, dataset management, model training, model evaluation, model release management, model inferencing, prediction explainability, and model monitoring components integrated into an end-to-end AI platform for healthcare.

To build the infrastructure, a variety of Amazon services was used, including but not limited to Amazon S3, AWS Glue, Amazon EKS, and Amazon SageMaker. Additionally, PyTorch, Petastorm, and Provectus Machine Learning Infrastructure were utilized.

#3 Disease Detection and Diagnosis Application

Disease screening is performed in the main application. Ophthalmologists can access the app to check how accurately pathology has been detected and classified by a given ML model. To display screening results, the application features the following fields: Label — normal or abnormal; Title and Version of the applied ML model; Prediction result — disease detected or not; Prediction accuracy in %; and Explainability — to highlight pathology areas in the image. Ophthalmologists can quickly sort images to prioritize the ones that have been labeled as abnormal, to scrutinize and analyze them for a specific diagnosis. The application helps them use their precious time more efficiently and focus on patients who require immediate assistance, instead of looking through dozens of normal images of healthy individuals.

Scalable Eye Screening Can Mean Earlier Access to Treatment, Saving Eyesight for More Babies

Pr3vent has received an ML-powered eye screening solution capable of detecting pathology and screening diseases in a newborn’s retina and fundus images with high accuracy. Demonstrating 85% of recall and 95.7% of precision, the solution is ready for FDA approval.

If approved, Pr3vent’s solution could scale newborn screening in the United States and globally. By reducing cost per screening by 10x (and by scaling ophthalmologists’ expertise through automation), it makes timely eye screening accessible to no less than 4 million babies in the US alone. No longer do ophthalmologists need to look through dozens of thousands of eye screens of healthy patients, but focus on those with signs and symptoms of possible vision problems.

Given that up to 97% of eye conditions can be treated in the first few months of life, Pr3vent may significantly reduce the number of infants and adults suffering from vision impairment and vision loss, thereby saving the society up to $3 billion per year in associated costs.

Moving Forward

1. Learn more about the Provectus AI Solutions and ML Infrastructure
2. Watch the webinar on MLOps and reproducible ML on AWS
3. Apply for Machine Learning Infrastructure Acceleration Program to get started