Abstract
Purpose :
To determine the concordance between an Unsupervised Learning Algorithm and eye fundus color photos interpretation by a specialist for the identification of the optic disc damage.
Methods :
Cross-sectional study of diagnostic concordance for an Unsupervised Learning Algorithm was made. The Cohen's kappa coefficient was calculated for identification of the optic disc damage in eye fundus color photos and were assessed according to Armaly's cup/disc ratio classification.
Results :
The Unsupervised Learning Algorithm evaluated 689 color optic disc images of subjects classified as: healthy (no damage), mild, moderate and severe damage. The second stage is composed of a k-means classifier using the first dense layer with 4096 units to cluster the extracted features in four groups. Obtained a performance measured of Cohen's kappa coefficient of 0,47, 0,82, 0,71 and 0,74 respectively. While classifying the images in two groups: Healthy and with damage, we found a Cohen's kappa coefficient of 0,70.
Conclusions :
The Unsupervised Learning Algorithm for the classification of optic disc damage on color fundus photos showed a good concordance with the one done by the glaucoma specialist, using Armaly's cup/disc ratio classification. The concordance was better when the classification only included optic discs with damage (mild, moderate and severe).
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.