Purchase this article with an account.
Laura C. Huang, Caroline Yu, Robert A. Kleinman, Ryan A. Shields, Ryan G. Smith, Carson Lam, Darvin Yi, Daniel Rubin; Opening the Black Box: Visualization of Deep Neural Network for Detection of Disease in Retinal Fundus Photographs. Invest. Ophthalmol. Vis. Sci. 2017;58(8):94.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Deep learning, a set of algorithms inspired by our understanding of how neurons are connected in the brain, has shown strong results in image interpretation, particularly in detection of diabetic retinopathy within retinal fundus photos. Retinopathy is estimated to affect 40% of diabetics over 40 years old. Screening adherence rates are 35 to 60%, making streamlining the process an important goal. Several limitations must be addressed before such algorithms are used towards clinical applications. Neural networks incorporate millions of parameters in its interpretation but are unable to communicate with users how they reach their conclusions, usually in the form of a number representing the probability of disease in an image. This has left much of the medical community skeptical of its clinical applications. Elucidating the components that generate a neural network’s output allows clinicians and patients to make more informed decisions and may reveal important new features with high predictive value for diabetic retinopathy
The neural network used in this study was a CNN that uses the GoogLeNet architecture. For training and testing purposes, a retrospective set of 35,126 retinal images from the publicly available Kaggle diabetic retinopathy database were used. Images were labeled 5 times each by a clinician from EyePacs, a US residency trained ophthalmologist, and 3 ophthalmology trainees. Interrater reliability was calculated using pairwise comparisons. A test set of 5000 images was randomly selected from the database to assess for algorithm sensitivity and specificity
The development set included 30,126 images, with (73%) normal, (7%) mild nonproliferative diabetic retinopathy (NPDR), (15%) moderate NPDR, (2.5%) severe NPDR, and (2%) proliferative diabetic retinopathy. The algorithm achieved 96% sensitivity and 92% specificity on 5000 images set aside for testing. Visualization of the algorithm revealed both intuitive (exudates) and non-intuitive features used to predict disease
Deep learning is a powerful tool that can achieve high sensitivity and specificity for detecting diabetic retinopathy. However, the way the algorithms achieve results must be presented to the public in a human interpretable format. Visualization and further study of the features used by these algorithms are needed to safely introduce them into the clinical workflow
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.
This PDF is available to Subscribers Only