Artificial intelligence for the automatic detection of diabetic retinopathy with feedback from key areas

  1. David Carmona-Ballester 1
  2. Sergio Bonaque-González 2
  3. Jose Manuel Rodriguez-Ramos 2
  4. José G. Marichal-Hernández 1
  5. Ricardo Oliva 1
  6. Sabato Ceruso 1
  7. Alicia Pareja-Ríos 1
  1. 1 Universidad de La Laguna, San Cristobal De La Laguna, Spain
  2. 2 Wooptix, Spain
ARVO annual meeting. Publicado en "Investigative Ophthalmology & Visual Science"

Publisher: The Association for Research in Vision and Ophthalmology

ISSN: 1552-5783

Year of publication: 2019

Volume: 60

Issue: 9

Pages: 1435-1435

Congress: ARVO annual meeting

Type: Conference paper


Purpose : Diabetic retinopathy (DR) is one of the main causes of blindness in developed countries. However, an ocular fundus examination at least once a year in diabetics could reduce severe visual impairment risk in over 90%. Nevertheless, the population of diabetics is very large and growing, so it is a real challenge to identify DR signs at an early stage in already saturated health systems. Consequently, significant efforts have been made in applying artificial intelligence (AI) to classify ocular fundus images as pathological or not in an automated way.One of the problem of current AI for DR diagnosis is that it is not possible to find out why and how the AI arrived to a certain conclusion. This knowledge would allow introducing a feedback that progressively improves the accuracy of the AI in a real implementation. We developed an AI based algorithm to, not only classify an image as pathological or not, but also to detect and highlight those signs that enabled the identification of the DR.Methods : From 2007 to the present day, the Retisalud project is being carried out in the Canary Islands, Spain. It is a telemedicine project where diabetics have an eye fundus photograph that is later analysed by a trained family doctor and, in the case of suspected pathology, later by a specialist in ophthalmology. This has allowed to have a database of nearly 800,000 classified images according almost exclusively to its degree of DR.We used this set of images to develop a deep convolutional neural network to be used to classify images as with DR or not. To verify the accuracy obtained, it was tested with a validation set extracted from our dataset. Features learned by the AI were then used to elaborate an additional detection algorithm capable to highlight the key areas.Results : The final accuracy achieved by the classification algorithm was 92%. The detection algorithm successfully detected DR signs, having a sensitivity of 32 pixels per feature. Discrepancies between AI and original labels occurred mainly in doubtful cases with very mild DR, and were later re-evaluated by an ophthalmologist showing a higher objectively of the AI.Conclusions : It was created a high accuracy algorithm, capable of detecting DR signs over real data showing the reasons that have led AI to arrive that decision.