Research On Multi-label Classification Based On Medical Fundus Images

Medical fundus pictures are a kind of special natural pictures generated by professional cameras for fundus information shooting.These pictures are mainly used in medical eye diagnosis,and provide a strong scientific basis for ophthalmologists to diagnose fundus diseases.However,each fundus picture generated by the camera requires a detailed analysis by a professional doctor.Due to the large number of fundus pictures,it will undoubtedly waste a lot of time and effort to analyze the types of fundus diseases.Today,deep learning technology has made breakthrough progress in the classification of fundus pictures,and its accuracy has been greatly improved.Most of these algorithms are for single-label classification algorithms whether they have a certain type of disease.For a fundus picture,they may often be affected.There are many diseases.In this paper,we study the multi-classification algorithm for fundus pictures,and jointly consider eliminating category imbalance and known category number to guide classification optimization.The picture multi-label classification task is a more complex problem.Unlike single-label multi-category classification,only one category needs to be predicted.It needs to predict all the existing categories in the picture.This is usually because the number of categories is uncertain and Its combination is diverse.This topic uses the data-driven approach to predict the number of categories of fundus pictures,and extracts the patient type information contained in each picture from the medical fundus picture data set.In the case of a priori number of categories,use the output category to determine the difficulty of simplifying multi-label classification and improve the accuracy of multi-label classification,specifically:First of all,this paper improves the classification feature extraction network,designing a new network with only 10 layers in VGG structure,and then using a metalearning model similar to the MAML learning process to construct a similar perception of the loss value generated by the classification network The weighted mapping is performed by the filter network,and finally the new loss weighting is used to eliminate the impact of the imbalance of the category on the training data set on the final classification accuracy.This will eventually lead to more accurate category classification predictions.Then use the obtained category information to design a multi-label classification network based on the number of categories a priori.In this paper,for the feature extraction of multi-label classification network,the residual network Res Net50 is used,and the final output layer is slightly adjusted so that the multi-label classification task is converted into a multi-valued binary classification task.Then,a Topk module is designed using the condition that the number of types is known,where Topk is used to filter out the most likely combination of categories,and the network parameters are updated by using the predicted loss value between the combination of types and labels.And the original loss function is improved to make it fit the Topk process.Through comparative experimental analysis,the method in this paper is superior to other methods in accuracy,and achieves the best results on the ODIR-5K fundus image data set.