HONOLULU — Deep learning methods have great potential to extend the expertise of physicians who diagnose and treat diabetic retinopathy, Rory Sayres, PhD, a researcher for Google, said at the Association for Research in Vision and Ophthalmology meeting here.
Sayres and colleagues looked at the impact on accuracy, speed and confidence of physicians when algorithms developed from machine learning methods were applied to assist in grading retinal fundus images for diabetic retinopathy.
Using more than 1,800 images from EyePACS, Sayres and colleagues recruited nine ophthalmologists to read the images with one of three retina grading models applied: unassisted, grades only or grades plus masks, in which explanatory maps were applied.
Most accurate readings were achieved when the physicians were assisted by the algorithms.
The assisted reading model prevents underdiagnosis, improves accuracy and sensitivity of the diagnosis and improves confidence of the human readers, Sayres said. Grading time was increased by 10 to 30 seconds per image, which is not unreasonable because it is part of a mechanism by which better human decisions are made, Sayres said. – by Patricia Nale, ELS
Reference: Sayres R. Assisted reads for diabetic retinopathy using a deep learning algorithm and integrated gradient explanation. Presented at: Association for Research in Vision and Ophthalmology annual meeting; April 28-May 3, 2018; Honolulu.
Disclosure: Sayres reports he is an employee of Google.