GoogLeNet-based Diabetic-retinopathy-detection
Bojia Shi, Xiaoya Zhang, Zhuoyang Wang, Jiawei Song, Jiaxuan Han, Zaiye Zhang, Teoh Teik Toe
Abstract
This paper is about a research in applying different neural networks for diabetic-retinopathy-detection. Respectively using basic CNNs, VGG16 and GoogLeNet trained on datasets from Aravind Eye Hospital in India including 8929 photos and validated on other 1114 photos. Experiment showed that GoogLeNet model could better identify diabetic retinopathy with a higher train accuracy around 97%, compared to the CNN model’s performance of 84% and VGG16’s 94%. Meanwhile, the test accuracy of GoogLeNet is 85%, relatively higher than other proposed models. The excellent performance of the GoogLeNet model shows its great potential and promises to be extended to replace ophthalmologists in the screening of patients in the future.