Automatic diagnosis of multiple lesions in fundus images based on dual attention mechanism

Abstract

Glaucomatous optic neuropathy (GON), retinal exudates and retinal hemorrhage are the main basis for the diagnosis of fundus diseases. Traditional methods can diagnose fundus diseases and their severity, but there are few studies on the characteristics of fundus diseases, which cannot give a reasonable explanation for the diagnosis of fundus diseases. Therefore, a convolutional neural network based on dual attention mechanism was proposed to realize automatic diagnosis of multiple fundus lesions with high accuracy. Convolutional neural network uses a residual structure with jumping connections, and channels and spatial attention mechanisms are embedded after each group of convolution to improve the accuracy of fundus lesions diagnosis. The model was tested on the clinical data of Ningbo Eye Hospital Affiliated to Wenzhou Medical University. The diagnostic accuracy of GON, retinal exudates and retinal hemorrhage were 98.17%, 97.49% and 97.15%, respectively. The experimental results showed that: the model showed good feature extraction ability and diagnostic performance in multi-lesion diagnosis of fundus, which provided reference value for subsequent medical artificial intelligence diagnosis research.