[Skip to Navigation]
Invited Commentary
June 13, 2019

Observations and Lessons Learned From the Artificial Intelligence Studies for Diabetic Retinopathy Screening

Author Affiliations
  • 1Singapore National Eye Center, Singapore Eye Research Institute, Singapore
  • 2Duke-NUS Medical School, National University of Singapore, Singapore
  • 3Duke University, Durham, North Carolina
  • 4Department of Ophthalmology and Visual Sciences, University of Iowa, Iowa City
JAMA Ophthalmol. 2019;137(9):994-995. doi:10.1001/jamaophthalmol.2019.1997

Deep learning (DL) is a form of artificial intelligence (AI) with a long history that has recently taken off given improved computational power (eg, graphic processing units), the availability of objective data at scale, and the availability of many publicly available deep neural networks, such as convolutional neural networks (eg, AlexNet, VGGNet, and GoogleNet) and coding software (eg, Caffe and Tensorflow).1,2 Deep learning has shown excellent performance across a range of data modalities, such as images, video, text, and audio. In this issue of JAMA Ophthalmology, Gulshan et al3 focus on applying DL to image analyses, particularly in ophthalmology where it has shown promising application in detecting various retinal diseases,4 including diabetic retinopathy (DR).5-7 Most of these DL systems were trained to detect referable DR, often defined as moderate nonproliferative DR or worse (Early Treatment Diabetic Retinopathy Treatment Study [ETDRS] level 35 or higher) or diabetic macular edema (DME) on either 1-field or 2-field retinal photographs.5,6

Add or change institution