Red lesion detection in fundus images based on convolution neural network

Early diagnosis of Diabetic Retinopathy (DR) can effectively control and delay the process of deterioration. The traditional detection method uses the manual design features to classify the lesions after the candidate regions are extracted, and the feature extraction effect is poor. This paper mainly studies the method of lesion detection based on convolutional neural network. We propose a lesion detection method based on improved LeNet convolutional neural network. Three innovations are proposed: (1) The network inherits the size of LeNet's convolution kernel and deepens the depth of the network, and can learn more features; (2) The weighted cross entropy loss, which is used for the sample class imbalance problem; (3) a dynamic learning rate based on the rate of change of loss, which is used to deal with overfitting and underfitting. Finally, the experiments in the DIARETDB1 and E-ophtha databases show that the improved LeNet convolutional neural network has achieved good results in the detection of background retinal image lesions.