One of our first projects to tackle global challenges in our portfolio is for the early detection of Diabetic Retinopathy (DR) at scale using AI. As described in an earlier post, DR is today the leading cause of blindness worldwide and a global healthcare challenge requiring innovative solutions.
In designing a system to handle this challenge we recognized that while an AI model could tackle many routine screening tasks for DR, trained human experts are an essential part of the solution. We need specialists and doctors in the loop for model validation and for providing advice, counseling, referral, follow-up and treatment options to patients. Our solution is therefore designed to use AI to augment and assist rather than replace specialists and experts.
The ideal system should combine AI, technology and human expertise in ways that they can complement each other to address the global DR challenge. The system should satisfy the following objectives:
- Provide the capability to be deployed in remote and under-served areas of the world as this is where we have the most acute shortages of healthcare professionals for screening.
- Provide the capability to capture retinal images with low-cost and portable devices or smartphones that can be easily deployed in the field and remote areas.
- Provide the capability of using AI in an automated way for easily and quickly screening for DR into referable and non-referable cases. A good analogy is that it should be as simple and quick as checking for blood-pressure using an automated blood pressure monitor.
- Provide the capability of remote diagnosis by professionals for cases identified by the AI as referable or at high risk, or that are difficult to diagnose with AI, and to allow them to validate the AI diagnosis, and to interact and counsel patients.
Our solution for DR Diagnosis meets all of these objectives – an overview of the solution and its major components is as follows:
The various system components and process flow is as follows:
- Patient’s retinal images are captured via fundus cameras at local screening centers or clinics, and are uploaded via the web to a cloud-based server for further processing. There also exist several low-cost devices on the market that can be attached to a smart-phones for capturing .the image which can be used as well.
- A cloud-based web application for patient registration and data entry, image capture and uploading, integration with the AI model, remote diagnosis by trained specialists, as well as patient reporting, messaging and notification.
- Automated AI Diagnosis: The AI model runs on a remote server and automatically diagnoses and classifies the image nearly instantly as either normal or showing evidence of DR, the stage of the DR, along with the probability of the model’s predicted diagnosis.
- Remote Human Diagnosis: Eye-care professionals can login remotely to the web application to review and validate the AI diagnosis, add notes, provide referral to a specialist, follow-up and treatment options. The system design also incorporates the ability to fine-tune the AI algorithm based on corrections of diagnosis errors by the specialists.
- Integrated administrative, reporting, and messaging for patient communication, system performance reports, and overall statistics.
A beta version of our solution was launched recently and is currently undergoing field trials. Click here to learn more.
The solution is cloud-based and uses the following technologies: PyTorch, Keras. Flask, Nginx, and runs on Amazon AWS and the Google Cloud infrastructure and is built to be fast, reliable and scalable to handle millions of images.