Pr3vent: ML-powered Disease Screening Platform
Pr3vent scales newborn screening and combats preventable vision loss with Machine Learning
10x reduction in eye screening cost
99% reduction in a doctor’s manual work
$3bn saved for society in vision loss costs
4 mln. newborns screened per year
Pr3vent’s eye screening solution had to demonstrate accuracy of no less than 97% in detecting pathology of a newborn’s retina to receive FDA approval. Resolving this task required:
- Accurate and proficient labeling of a database of 350K fundus and retina images by experienced ophthalmologists to enable ML model training
- Building and training a binary image classification model capable of identifying pathology in an image
- Developing a user-friendly UI for ophthalmologists to sift out normal retina images, scrutinize abnormal ones, as well as label images to train the model
In an effort to build a solution that could nail all of these boxes, Pr3vent approached Provectus’ AI/ML team to deliver it from scratch.
Provectus’ development efforts were focused on three constituent parts of Pr3vent’s eye screening solution:
Image labeling tool
ML model building and training tool
UI for disease detection and diagnosis application
The image labeling tool is built for trained ophthalmologists to process and label fundus and retina images stored in Amazon S3. Using Amazon SNS and SQS, the tool gets notified about newly uploaded, unlabeled images and moves them to a separate S3 bucket to initiate image pre-processing. The tool’s user interface is web-based, and it is easily accessible to medical professionals. They can look through, sort, and label medical images as normal or abnormal. They can also delete low-quality images to increase an ML model’s accuracy and performance.
The ML model building and training tool accesses the labeled dataset to feed it to Amazon SageMaker to build and train the binary image classification model. The training logs, as well as the model itself, are stored in RDS PostgreSQL. MLflow is applied for versioning. The ML models demonstrating the highest accuracy in detecting pathology, their pre- and post-processing scripts are containerized using Docker, and they are moved to Amazon ECR. Amazon ECR is pinged to initiate the relevant model in Amazon EKS to start using the main application in production.
Disease screening is performed in the main application. An ophthalmologist can access the app’s UI to check how accurately pathology has been detected and classified by applied ML models. To assess screening results, the application displays: Label — either normal or abnormal; Title and Version of the applied ML model(s); Prediction; Prediction accuracy; and Explainability section highlighting pathology areas in a given image. The solution’s logs are collected and stored in Amazon CloudWatch.
Pr3vent 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 95% of recall (ML success metric), the solution is ready for FDA approval.
If approved, Pr3vent’s solution has the potential to scale newborn screening. By reducing cost per screening by 10x (and by scaling ophthalmologists’ expertise through automation), it makes timely eye screening accessible to no less than four million babies in the US alone.
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 $3bn per year in related costs.
Looking to explore the solution? Contact Us!