Grupo ULMA – UMI.dr - AI-based Software for Automatic DR Screening Through Automatic Readings of Retinal Images
Grupo ULMA
Sector: Health
Business Case
37 million adults (20-79), or 1 in 10, live with diabetes. This number is expected to increase to 643 million by 2030 and 784 million by 2045. Diabetic Retinopathy (DR) is a diabetes complication and a significant cause of blindness worldwide. 1 in 3 diabetics suffer some degree of DR. More than 90% of cases are preventable with early detection and proper treatment. 50% of people with diagnosed diabetes do not get regular eye exams.
Objectives
• Increase coverage of the DR screening programme. • Streamline and improve the efficiency of the DR diagnostic process. • Gain greater reproducibility in diagnosis independently of the health professional. • Improve resource management by freeing up specialists for higher value-added tasks. • Reduce the costs associated with treatments related to patients suffering from the most advanced stages of the disease. • Improve patient experience.
Use case
Screening for DR is a process for early detection and diagnosis of this eye disease. The “Gold Standard” for the diagnosis of DR is retinal or fundus imaging. Diabetic patients visit their health centre annually/biennially to have them done. After that, ophthalmology is in charge of making the diagnosis. At ULMA we have developed software based on AI for automatic DR screening through automatic reading of retinal images.
Infrastructure
On Premise Cloud Edge
Technology
Automatic or Deep Learning
Data
Retinal images and their respective diagnoses. Each image has been dually diagnosed by two ophthalmology professionals independently. In case of discordant diagnoses, a third professional was the person in charge of making the final diagnosis.
Resources
Multidisciplinary team of: ophthalmology and primary care medical professionals, health research and engineering technicians with specific training in medical imaging, deep learning, clinical validations, MLOps, interoperability standards, medical device regulation, frontend & backend development, etc. Infrastructure: Linux server with 512 GB of RAM and dedicated 40 GB nvidia A100 graphics card.
Difficulties and learning
1. Access to quality data; retinal images from DR screening programmes with their respective diagnoses, GDPR requirements and limited availability of specialised professionals for the diagnostic process of such images. 2. Knowledge management of the multidisciplinary team. 3. New medical device regulation (MDR) and integration of AI models in hospital information systems based on different interoperability standards.
KPIs (business impact and metrics of the model)
Results of the last clinical validation performed in the SESPA (Servicio de Salud del Principado de Asturias) with 3,458 images of the Asturias DR screening programme. Sensitivity: 95.7% / Specificity: 94.5% / Accuracy: 94.5%.
Funding
Own Funds SME Instrument Phase 2 of Horizonte 2020 2020 call for proposals on development based on artificial intelligence (red.es) Applied Artificial Intelligence Programme (Basque Government)
Collaborators, Partners
Osakidetza, SESPA, IDIAP Jordi Gol, Vicomtech, etc.
