Research On Multi-label Fundus Image Classification Algorithm Based On Deep Learning

The eyes are one of the most important organs of the human senses,and eye diseases can lead to vision loss and even blindness in severe cases.Currently,common fundus diseases include glaucoma,cataracts,and myopia,etc.These diseases do not cause serious damage to vision in their early stages,so early detection and timely treatment are essential to reduce vision loss.However,due to the wide variety of fundus diseases and the large differences in the range of diseased areas,it brings certain difficulties to the identification work.In addition,there are problems such as poor quality and insufficient quantity of fundus disease data samples.Therefore,deep learning algorithms that fuse image information with patient meta-information are of great research significance.In view of the urgent need for diagnostic tools for fundus diseases in clinical practice,this thesis presents a deep learning method for aiding diagnosis of fundus diseases and proposes a deep learning method for fusing metadata.By introducing the idea of metadata information fusion and the Triple M-Tran model,the relationship between images,known labels,and additional information is used to improve the accuracy of multi-label image classification and recall rate.The main research works of this thesis are as follows:(1)A new fundus image classification method is proposed,which introduces patient meta data while adding multi-label attributes and performs multi-label multiclassification studies of fundus diseases by combining image data with patient meta data(e.g.,age,gender,previous medical conditions)information.(2)For the fundus image multi-label classification task,the features of fundus image data are first extracted using a convolutional neural network CNN,then fused with the processed meta data(meta data)with patient information,and finally fed into the designed classifier for multi-label multi-class classification of fundus diseases.This approach of information fusion makes full use of the complementary effects between multiple information sources,and the recognition accuracy on the dataset is improved by 0.77%and 1.84%,respectively.(3)In order to better explore the relationship between image and metadata information,this thesis integrates a new dataset MLFi D,which contains 22 different types of fundus diseases,by increasing the types and numbers of fundus diseases identified,to enrich the categories of fundus diseases and make the classification of diseases more accurate.(4)A Triple M-Tran model for extracting fundus image features and patient metainformation label features is designed.Through Transformer’s Encoder structure,the relationship feature maps between images and metadata are obtained,and then these feature maps are sent to the classification network for classification.Multi-label multiclass classification of fundus diseases.By introducing the Transformer Encoder mechanism and using the Embedding method to train the labels,the Triple M-Tran model can better utilize the dependencies between the input data.The experimental results show that the Triple M-Tran model proposed in this thesis is compared with other methods.When identifying and classifying 22 different fundus diseases,the ML AUC and Modal Score scores increased by 0.3% and 0.28%,respectively,proving its superiority in processing and fusing multi-category data.