Fundus Image Classification Algorithm Based On Metric Learning

In recent years,diabetic retinal fundus disease has gradually become one of the main factors causing visual impairment and blindness.Timely detection can enable patients with fundus disease to detect symptoms as soon as possible,slow down the rate of exacerbation of patients’ symptoms,thereby reducing the probability of serious visual impairment and blindness.With the rapid development of artificial intelligence,computer-aided diagnostic systems have been widely used in various fields.Computer-aided diagnostic systems can increase the chance that doctors will get the diagnosis correct,reduce patient consultation time,and reduce the cost of consultations.The research on image classification methods for fundus lesions will make great contributions to the further development of the medical field.In the previous fundus image classification work,the fine-grained characteristics in diabetic retinal fundus lesion images were ignored,and it was difficult for traditional convolutional neural network models to fully obtain the key features with discrimination in the dataset.In real life,datasets related to the medical field often have different degrees of uneven sample distribution.The samples of patients with severe disease usually occupy only a small part of the dataset,and too few samples of patients with severe illness will directly lead to the lack of information of this category sample in the model training process,which will limit the classification ability of the network.In addition,due to the memory capacity of the graphics card,it is difficult for the network model in the deep learning method to load all the sample data at once during the training process,and the model can only be trained by loading the data in batches.However,this training method will make it difficult to update the network model according to the overall distribution of all training datasets during the training process.Training only from the samples in the Mini-batch during the training process will limit the update of the network model to a certain extent,or even go in the wrong direction.For the problems in the above classification tasks of fundus images,this paper proposes the following methods to address these challenges:This paper proposes a virtual global proxy method,so that the network model can be updated with reference to the distribution of the overall dataset without increasing the memory occupation during the training process.The introduction of the virtual global proxy method can enable the network model to have an overall view of the global sample distribution during the update process,so that the model can avoid the drawback that it can only be optimized based on the local information in the current training batch.At the same time,we also refer to the design idea based on sample pair loss in the virtual global proxy method to enable the model to fully explore the relationship between samples in the training process.In this paper,a dynamic linear mirror feature synthesis method is proposed,which enables the model to dynamically synthesize features on a small number of class samples during the training process.Therefore,the weight of minority class samples in the training process is increased,so that the model can give more attention to minority class samples,thereby reducing the interference caused by the uneven sample distribution problem on the model.Different from the traditional method of broadening or increasing the weight of minority samples,the dynamic linear mirror feature synthesis method proposed in this paper can dynamically control the synthesis of minority class samples according to the average loss ratio difference of minority class samples in the training process,so that a dynamic balance can be achieved between each class sample in each batch training process,so as to enhance the generalization ability of the model.