Research On Chest DR Image Disease Classification Model And Lightweight Method Based On Label Graph Embeddin

Hundreds of millions people around the world suffer from chest diseases every year.These diseases are various types,such as Corona Virus Disease 2019(COVID-19)since2019,according to the statistics of the World Health Organization(WHO),As of February2021,420 million people have been infected.If not treated in time,it will have a great impact on patients and even endanger their lives.Another example is lung cancer,lung cancer is a cancer with high morbidity and mortality all over the world.In China,lung cancer has become the main disease of cancer death.For the pre-diagnosis of chest diseases,medical imaging examinations are widely used,such as the detection of chest masses,pleural effusion,cardiac hypertrophy,pulmonary edema,pleural thickening and other lesions,as well as assessment of disease severity,early screening of lung cancer,etc.The current chest imaging examinations are mainly chest X-rays and chest CT images.Compared with chest CT examination,X-ray examination has the advantages of low price,fast speed,easy popularization,etc,and the radiation dose of X-ray examination is small,and the radiation dose of one CT examination is several hundred times that of X-ray examination.Medical X-rays,as a routine inspection method for chest diseases,can diagnose early non-obvious chest diseases and observe the lesions.However,the presence of multiple disease features on the same radiological image presents a challenge for classification tasks.In addition,there are different correspondences between disease labels,further causing difficulties in classification tasks.In view of the above problems,this paper takes chest DR radiographic images as the main research object,deeply studies the relationship between disease labels in chest disease classification task,and performs lightweight processing.The main research contents are as follows:(1)In this paper,a graph convolutional neural network is combined with a traditional convolutional neural network to propose a multi-label disease classification method for chest radiography that fuses label features with image features.This method takes into account the correspondence between different disease labels,and uses a graph convolutional neural network to model the global correlation of labels,that is,constructs a directed relation graph on the disease labels,and each node in the directed graph represents A label category,which is then fed into a graph convolutional neural network to extract label features,and finally fused with image features for classification.The experimental results of the method proposed in this paper on the Chest X-ray14 dataset show that the average AUC for 14 chest diseases reaches 0.843.Compared with the current three classic methods and advanced methods,the method in this paper can effectively improve the classification performance.(2)When deep learning algorithms are applied to medical image detection and classification tasks,there are often problems such as large number of parameters and large memory usage,which are difficult to run on mobile terminals or embedded devices.In this case,this paper designs a light-weight model for chest radiographic disease classification based on knowledge distillation.The model uses the joint model completed in the previous step as the teacher model,and the light-weight model is the student model.Knowledge transfer is performed to make the light-weight models reduce overall runtime memory and greatly improve model speed with minimal loss of accuracy.The average AUC of the model proposed in this paper is 0.817,which still exceeds the current three classical methods.Compared with the joint model in(1),the speed is increased by 34%,and the memory usage rate is reduced by 35%,which proves that Knowledge distillation on this model is effective.