Research On Medical Image Recognition Based On Transformer

With the development of medical technology,medical image plays a very important role in the diagnosis and treatment of more and more diseases.Although medical images can assist medical staff to make more accurate diagnosis and treatment,they also increase the burden of medical staff in the process of continuous application.Therefore,more and more attention has been paid to the research of medical image analysis using computer vision.The current mainstream medical image classification model based on convolutional neural network has the problems of low classification accuracy and insufficient generalization.It is difficult to find a model suitable for different medical image classification tasks.At the same time,due to the difficulty of collecting medical images and the cost of manual labeling,there are very few medical image data samples with labels,and general deep learning models cannot accurately achieve classification.In order to meet the challenges in the task of medical image classification,this thesis proposes two medical image classification models based on Transformer.By designing the appropriate structure of medical image classification network,the proposed two models combine the powerful modeling ability of Transformer and the advantages of attention mechanism,show excellent results in feature learning and classification prediction,and adapt to the task of medical image classification in different scenes.Aiming at the problem that the model based on convolutional neural network does not perform well in the task of medical image classification,a feature pyramid Transformer model is proposed in this thesis.The main idea of feature pyramid Transformer model is to enrich the feature information of image by multi-scale features and combine the long-range modeling ability of Transformer.The application of Transformer structure makes up for the lack of feature learning caused by the local convolution operation of convolutional neural network.Through comparative experiments and ablation experiments,this thesis proves that the feature pyramid transformer model has higher classification accuracy and generalization in the decathlon task of medical image classification.Aiming at the difficulty of obtaining labeled medical image data,this thesis proposes a few-shot medical image classification model based on Transformer.Using the idea of prototype network and Transformer,the model generates task-related slice prototypes.Then,the distance between query sample slice prototypes and support set slice prototypes is measured by hybrid metric,and the category of query samples is predicted based on the distance.The attention ideas of Transformer and the application of hybrid metrics enhance the capability of feature learning and prototype matching.In order to verify the classification effect of the model,two small sample learning datasets were constructed on the medical image classification decathlon task and a body part X-ray image dataset,and experiments were carried out.The proposed model is compared with some commonly used small sample image classification models,and the classification accuracy of the small sample medical image classification model based on Transformer can reach the highest.