Research On Detection Algorithm Of Diabetic Retinal Red Lesions Based On Convolution Neural Network And Deep Feature Fusion

Diabetic Retinopathy(DR)is a popular research in present medical image processing field.Detection and treatment of retinopathy in its early stages is the key to controlling disease progression,so this paper mainly studies DR auto-detection based on red lesions.Red lesions are difficult to detect due to their similar color to the background,low contrast,tissue structure such as blood vessels,and disturbing imaging by fundus images.The traditional detection methods have a problem that its effect of feature extraction is poor due to using the hand-crafted features to train classifier after the candidate detection.Relatively this paper mainly propose a novel method for lesion detection based on convolutional neural network.The main contents and contributions are as follows:(1)Retinopathy candidate detection:the traditional Bot-Hat transformation extraction algorithm can only extract lesion with mono-scale and uniform morphology.To solve the problem,this paper gives a method for red lesion detection based on combining both multiscale Bot-Hat transformation and threshold segmentation in connective region.Firstly,the lesions under this scale were extracted by multi-scale Bot-Hat transformation,which contained too many false positive samples,and then the connected domain threshold segmentation and area filter was used to reduce the number of false positive samples,which was finally used for classification detection.(2)Lesions detection based on the improved LeNet convolutional neural network:An improved LeNet convolutional neural network is proposed for the problem of relatively small number of fundus image samples and low resolution.The network inherits the receptive field size of LeNet’s convolution kernel and deepens the depth of the network,and improves the ability to learn features;it uses a class balanced cross-entropy loss for the imbalance problem of data set;in the case of overfitting and underfitting which are easy to appear in the training process,a dynamic learning rate based on loss rate is proposed.(3)Lesions detection based on deep feature fusion:A method for detecting lesions by combining traditional hand craft features with depth features is proposed.Compared with the commonly used manual features,the hand craft features this paper proposes incorporate shape features from the blood vessel based on U-Net segmentation and pixel features based on multichannel.The complementarity of deep features and manual features is utilized to improve the ability to describe features.Such ensemble vector of descriptors is used to train a random forest classifier which has had a good performence.The experimental results using DIARETDB1 and E-ophtha indicate that the proposed method for lesion detection based on the improved LeNet convolutional neural network in the complex background has a good performance;by extracting the deep features of the convolutional neuron network and the hand-crafted features,the lesion characteristics can be futher enhanced.The ability to describe lesion characteristics proves the complementarity between deep features and manual features.