Abstract
Purpose :
Deep Learning Methods (DLM) based on neural networks are systems that use processing cores (using specific software and hardware) to simulate neural connections. The use of these methods on automatic analysis of eye color fundus images as a tool to support medical diagnosis/screening has been a challenge in terms of achieving the best accuracy, the lowest computational cost and lowest runtime. The purpose of this study was to determine the accuracy of a Deep Learning Method to classify color eye fundus images from diabetic patients as referable or non-referable
Methods :
A cross-sectional study was designed for this research. We used a pre-trained method VGG-16 with non-medical images from Imagenet with a support vector machine classifier with Gaussian kernel using first dense layer with 4096 units. The proposed method was evaluated in 700 color eye fundus images (from diabetic patients) acquired by hand-held dispositive (Volk Pictor® camera) and a tabletop design dispositive (Zeiss VISUPAC FF450plus fundoscopic camera). The methodology was carried out to improve the method performance for classify the images as referable (with any of diabetic retinopathy or diabetic macular edema findings) or non-referable (without abnormal findings). Finally, the method performance was compared to 2 Retina especialists
Results :
The method performance had a sensitivity of 0.97, a specificity of 0.99 and an accuracy of 0.98 in the detection of images for referral
Conclusions :
The present method had high sensitivity and specificity for detecting referable images. Further research is necessary to determine the feasibility of applying this algorithm in the clinical screening setting, as well as in cases of poor quality images
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.