NEW YORK (Reuters Health) – Automated assessment of age-related macular degeneration (AMD) from fundus images could play a role in managing the disease, according to a new study.
Dr. Neil M. Bressler from the Wilmer Eye Institute at Johns Hopkins University in Baltimore and colleagues compared automated grading of AMD with that of expert human assessment in distinguishing disease-free/early-stage AMD from intermediate/advanced-stage AMD.
The 12-year data set, comprising more than 13,000 digitized color fundus images from 4,613 patients, came from the National Institutes of Health (NIH) Age-Related Eye Disease Study (AREDS).
Two automated approaches were tested. Each involved deep convolutional neural network (DCNN) analysis of fundus images features “without relying on manual feature selection,” according to the September 28 online report in JAMA Ophthalmology.
The two DCNN approaches (sometimes called “deep learning methods”) were compared with human graders for their accuracy in assessing the two broad categories of AMD severity. Performance was measured against gold-standard assessments from the NIH’s AREDS data set.
The authors report accuracy of the DCNN methods ranging from 88.4% to 91.6% with kappa scores nearing or exceeding 0.8, “comparable with human expert performance levels.” Area under the receiver operating characteristic curve for the DCNN methods was 0.94 to 0.96.
“With eye care specialists a scarce resource globally and the increasing use of digital photography to evaluate patients, processes like deep neural networks and machine learning methods are being expanded to help grade images in fields like diabetic retinopathy and macular degeneration,” said Dr. Rishi Singh, an ophthalmologist at the Cleveland Clinic Cole Eye Institute in Ohio.
Dr. Singh, who was not involved in the study, told Reuters Health by email that although these algorithms are still in the research stage, they have significant importance for detecting disease and risk-stratifying patients to determine who needs care.
“The study is an interesting application of deep learning-based assessment of age-related macular degeneration from color photographs of the macula,” said Dr. Johanna Seddon, director of the Ophthalmic Epidemiology and Genetics Service at Tufts University School of Medicine in Boston.
Dr. Seddon, who was not involved in the study, also cautioned that distinguishing one category of no or early disease from intermediate or advanced AMD is not a big challenge. “The study does not apply to more-refined differences and classifications in the phenotypes,” she said in an email to Reuters Health.
Dr. Manju Subramanian, a specialist in vitreoretinal diseases, surgery, and vice chairman of operations for eye clinics at Boston Medical Center, said the study tests a novel approach to using artificial intelligence in AMD assessment.
“The sensitivity and specificity of this new automated method are high, indicating its potential for clinical utility in diagnosing and staging AMD,” said Dr. Subramanian, who was not involved in the study.
“If properly implemented,” he said by email, “such advances in artificial intelligence could aid in early identification and treatment, providing improved access to eye care in large populations that are currently underserved in the United States and worldwide.”
Dr. Subramanian said the next step is to test the technology prospectively in a more diverse population than the one represented in the bank of photos used in this study.
The study’s corresponding author did not respond to requests for comment.
Three of the six authors hold a patent on a system and method for the automated detection of AMD and other retinal abnormalities.
SOURCE: http://bit.ly/2fPCkwQ
JAMA Ophthalmol 2017.
Tidak ada komentar:
Posting Komentar