Research On Few-shot Medical Image Classification With Transfer Learning And Meta Learning

In the medical field,judging the type of disease through medical images has become one of the methods in daily diagnosis.The existing solutions include extracting features such as edges,textures,and morphological filtering in the image for analysis.We can also use deep learning models to automatically learn hidden features from a large number of data samples.However,the data obtained in the medical field usually has the problem of uneven distribution of categories or too small amount.In response to this situation,this paper proposes to adopt a meta-learning method to increase the extraction of effective features and learn through training models.To counter the problem of too little amount of data,this paper chooses to use GAN to generate images.DCGAN has been proved to be able to train high-quality images even with a small sample size.This paper improves the DCGAN,changing the original batch normalization to spectral normalization,and changing optimization from the original Adam to the RMSprop.The quality of the model generation is greatly improved through these improvements,and the training is more stable.This paper chooses a prototypical network for training.Prototypical network is a metric-based learning method.The model trained by this method can help train the new model suitable for solving the target task well,but it uses the ordinary convolution block to extract information and the extracted information has no obvious effect on the subsequent classification,so in this paper,the attention mechanism and1×1 convolution are added to the original convolution block to enhance the model's ability to express in space and channels.1×1 convolution can also be used to increase the nonlinear characteristics of the model while reducing the number of parameters without loss of resolution.Experiments have proved that after the above modifications to the model,the classification accuracy can reach 92.44%.The above method is to add part of the structure to the original network.This paper also uses transfer learning to transfer the Res Net network to extract features.Res Net has a deep depth.Therefore,this paper first uses the improved DCGAN to generate a series of medical image-related data to pre-train Res Net,and then part of the underlying convolution parameters in Res Net are frozen.This structure is used as the feature extraction part of the prototypical network.Then we use the target dataset to train high-level convolution Finally,we use the classifier for classification.Experiments show that the combination of transfer learning and meta-learning can further improve the effectiveness of the model,and the accuracy of classification on the brain tumor dataset can reach more than 93%.