Automatic grading of Diabetic macular edema based on end-to-end network

Abstract

Diabetic macular edema (DME) is one of the main causes of permanent blindness. Automatic grading is helpful to quick diagnosis of DME and reduction of vision loss. Most of the existing methods grade DME according to the shortest distance from the hard exudates(HEs) to the fovea, which is required to detect the fovea and HEs beforehand. The end-to-end network takes the whole fundus image as input to obtain the grading result, but the accuracy is often low because of the complexity of DME. Some end-to-end architectures simplify the grading to two grades instead of three grades or use other disease information to improve the accuracy. In order to improve the efficiency and accuracy of grading without introducing the lesion segmentation or simplifying the grading task, this paper proposes a new end-to-end architecture with ResNet50 combined with channel attention (SENet) to extract features and introduces the disease attention module to supplement the disease-specific information of DME, which can achieve higher grading results and reduce the grading cost without additional acquisition of fovea and HEs. In order to solve the problem of class imbalance and insufficient training samples in grading, this paper adds class weight to the crossentropy function in the design of the loss function and augmentation techniques to expand data. The proposed network is verified on the MESSIDOR dataset and has a good grading result. The experimental results show that the proposed method has high accuracy, specificity, sensitivity and F1 score, which are 97.06%, 98.97%, 88.64% and 0.9177, respectively.