Screening for Diabetic Retinopathy by an Artificial Intelligence

There are 425 million people with diabetes worldwide (1 in 11 adults), with a projected 600 million by 2035. It is estimated that 30 to 50% of people with diabetes have diabetic retinopathy.

The WHO recommends an annual screening, which means 2.3 million eyes to be analyzed every day!

To face this challenge, a fundscopic dialated eye exam performed by an ophthalmologist cannot be the only solution, which is why the screening technique is quickly gaining ground.

In terms of organization, delegating tasks has already proven its effectiveness. However, today, we see that:
- there is a lack of qualified professionals to do the telescreening
- long and tedious traditional methods are widely used   
- there is a lack of regulation to allow telescreening in some countries

All these points need to be resolved to provide early and effective care for a mass volume of patients (mass screening).

We have developed a decision support algorithm that allows all screening professionals to carry out their activity on par with the best retina experts. It was developed by analyzing a database of several hundred thousand retinas, a database which was created by the largest French telescreening network within the Parisian hospital system (Hôpitaux de Paris (AP-HP)).

Our solution is a certified medical device for:

- Screening organization, thanks to a platform hosted by geographical area which is accessible anywhere by Internet and which allows for:

• organization into sub-networks, the networking of stakeholders
• possible integration with hospital IT systems
• the creation of a patient file (profile, clinical context)
• automatic photo integration (DICOM)
• picture archiving (PACS) and patient follow-up over time
• automatic sending of reports
• a patient file to be accessible at any time by Internet

- Assistance with screening through Artificial Intelligence that works with any type of fundus camera and that indicates:

• (to the operator / screener) in real time, if the retina image will be able to be interpreted
• (to the person reading the results) within a few seconds: the severity of the Diabetic Retinopathy, the location of the lesions, the actions to be taken (degree of urgency), the associated pathologies, and the confidence interval.

Performance: 99% Sensitivity (number of true positives) / 87% Specificity (number of true negatives)