An Automatic Glaucoma Grading Method Based on Attention Mechanism and EfficientNet-B3 Network

Abstract

Deep learning has received considerable attention in the computer vision field and has been widely studied, especially in recognizing and diagnosing ophthalmic diseases. Currently, glaucoma recognition algorithms are mostly based on unimodal OCT, the visual field for glaucoma auxiliary diagnosis. Such algorithms have poor robustness and limited help for glaucoma auxiliary diagnosis; therefore, this experiment is proposed to use a 2D fundus image and 3D-OCT scanner two modal data as the experimental dataset and use the EfficientNet-B3 network and ResNet34 network models for feature extraction and fusion to improve automatic glaucoma grading accuracy. Since fundus images usually contain a large number of meaningless black background regions, this may lead to feature redundancy. Therefore, this experiment employs an attention mechanism that focuses the attention of the convolutional neural network on eye subject features to improve the performance of the glaucoma autoclassification model.