Session Title: Free Paper Session 7: Vascular Diseases & Diabetic Retinopathy II
Session Date/Time: Thursday 07/09/2017 | 14:30-16:00
Paper Time: 15:42
Venue: Room 117
First Author: : P.De Boever BELGIUM
Co Author(s): : B. Elen
Screening for Diabetic Retinopathy (DR) can be done more cost efficient and on a larger scale with the help of Automated Retinal Image Analysis Systems (ARIASs). Doubt about the reliability of automated DR screening software based on deep learning models can still be present among eye healthcare professionals, and might slow down its adoption. We have evaluated the Iflexis automated DR screening software on a public available data set to mitigate this doubt.
Artificial intelligence algorithm for diabetic retinopathy classification
The Iflexis DR screening machine learning algorithm uses an ensemble of two large deep convolutional neural networks to detect signs of the presence of DR on fundus images. The software has been used to score all 1748 fundus images from the public available Messidor-2 database. It are fovea-centreed fundus images taken from both eyes of 874 diabetes patients. A consensus reference standard is set by three US board certified retinal specialists for comparison. No images from the Messidor-2 database have been used for the training of the deep learning models or for the tuning of its parameters.
All 190 patients in the Messidor-2 database with referable DR (moderate nonproliferative DR or worse) have been successfully detected by the Iflexis DR screening software (sensitivity 100%, 95% CI: 98,1%-100%). The software labeled 444 of the 684 diabetes patients without referable DR accordingly (specificity 64,9%, 95% CI: 61,2%-68,5%). The automated DR screening software demonstrates to be quite conservative while screening for the presence of referable DR on fundus images, allowing eye health care professionals to safely focus their attention on patients in which signs of referable DR has been detected by the automated screening.
Automated screening software for the detection of referable DR based on deep learning models can obtain a very high sensitivity combined with a good specificity. The automated screening is done in a conservative way, allowing eye healthcare professionals to spend significantly less time on manually scoring fundus images for the presence of DR. The application of ARIA’s enables more cost efficient, and on larger scale screening for referable DR.