Application of Convolutional Neural Network in Classification of Autofluorescence Image of Diabetic Retina Fundus

摘要

Diabetic retinopathy is one of thecomplications of diabetes. The common medical diagnosismethod is fluorescein fundus angiography, but thediagnosis process requires fluorescein sodium injection.The fundus autofluorescence technology used in this studycan be harmless and has a better application prospect forpatients who cannot be angiographic examinations.However, the naked eye cannot recognize the early fundusimages and need to introduce computer-aided diagnosis.This paper's research object is 190 fundusautofluorescence images, and the accuracy of the 10-foldcross-check is used as the evaluation index. Compare theeffects of convolutional neural network algorithms onclassification performance under different imageresolutions and image enhancement operations. Theoptimal image resolution is 64*64, the image enhancementoperation is horizontal flip, and the optimal accuracy rateis 0.92105. After exploring the network structure, it isfound that there is a better result without modifying thenetwork. This article summarizes the following trainingsteps: first, use the basic model to select the appropriateimage resolution and image enhancement operation, andsecondly, modify the network layer and explore thenetwork through trial and error.