Research On Small Data Classification And Generation Of Breast Ultrasound Images Based On Graph Neural Network

Breast cancer is one of the most common cancers.Early detection and timely treatment can greatly increase the survival rate of patients.Therefore,it is considerable for the early diagnosis of breast cancer.Compared with mammography,CT and other methods,ultrasound detection is extensively adopted because of its safety,convenience and other advantages.The auxiliary discrimination model for breast ultrasound images is gradually being valued.Unlike massive human face and natural image data,typical,accurate,and effective medical image gold standard data is extremely limited.At the same time,rare diseases data and early stage data of emerging diseases are also extremely lacking.Such conditions greatly limit the capabilities of powerful models such as deep learning.And the existing few-shot learning research does not solve the actual scene problems.Therefore,it is significant to study small data learning of medical images for limited data.From the perspective of breast ultrasound images and small data learning,combined with human cognition of such images and thinking mode of few-shot learning,this thesis devises a multi-type feature extraction module to obtain important identification features in the image.And on the basis of few-shot learning,the graph neural network is introduced to extract the features of sample classes,so as to guide the recognition and generation of breast ultrasound images with small data.The main contents are as follows:1.The graph neural networks model that can summarize class information and compute more efficiently is constructed.The message propagation ability of graph neural networks can better express the relationship information,but this kind of research is less.In graph neural networks based on the idea of metric-learning,vertices represent sample features,and edges represent similarity between samples.The update processing of multi-layer graph neural networks is optimized to highlight the class features contained in small data and improve the independence of the graph updating module.The introduction of the dropout-edge operation alleviates the over-smoothing problem,and the design of the graph-pooling operation reduces the time and space complexity of graph calculations.They make the calculation of multi-layer graph more effective.This graph neural networks model has the capability to learn and summarize class information from small data,and can be directly embedded into the classification,generation and other related methods.2.A small data classification method using multi-feature aggregation based on graph neural network is proposed.Although the research on few-shot classification is abundant,the research on small data in the actual scene is less.Based on this problem,the multi-type feature extraction module is used to mine the important recognition information from the edge details and tumor contour of ultrasonic images,and this thesis combine the original image to extract rich features to construct the graph neural networks.The classification information contained in small data is highlighted by updating the multi-layer graph neural networks.Taking the information contained in the graph into full consideration,the vertex and edge values of each layer graph are used,and the complete metric-features are obtained from the perspective of sample features and relationships between samples.The multi-dimensional metric-features are aggregated to obtain the final similarity matrix,and finally the tumor classification is completed according to the information.3.A small data generation method based on generative adversarial networks and introducing graph neural networks and real features is proposed.To generate data for downstream tasks based on small data,the structure of generator and discriminator is adopted,and graph neural networks is introduced to learn class features contained in small data to generate the corresponding class of image data.The feature introduction module is designed to supplement relevant details with real image data.The loss function is designed to ensure the effective training and improve the effect of generated images.In this thesis,four breast ultrasound images datasets are selected to carry out small data classification and generation experiments respectively,and the effect of each module is verified by relevant ablation experiments.The classification accuracy of the method in this thesis is higher than the second-best model about 2%.The generated images in this thesis perform better in FID(Fréchet Inception Distance)and IS(Inception Score)values,and the introduction of the generated images into the classification model increases the accuracy by about 0.5%.The results show that the small data classification and generation methods have achieved good results,which verifies the effectiveness of the model.In this thesis,the theoretical content and model framework of few-shot learning are applied to the actual scene,and the recognition information in breast ultrasound images is fully extracted.Based on the metric-learning method,combined with few-shot learning and graph neural networks,the classification and generation methods of breast ultrasound images suitable for the small data problem are proposed,which provides ideas and ways for the related research of practical small data problem.