Research On Brain Tumor Image Segmentation Algorithm Based On Few-shot Learning Of Dual-branch Network

With the continuous improvement of the level of medical imaging,brain tumor MR(Magnetic Resonance,MR)image segmentation technology based on artificial intelligence is of great significance for the diagnosis of brain tumor regions.The traditional brain tumor image segmentation technology has shortcomings such as time-consuming and resource-consuming for identifying data sets.In order to save resources and minimize the dependence on medical experts to manually label images,the brain tumor MRI image segmentation method based on the few-shot learning model is researched in the thesis.Based on the dual-branch U-Net network,the interaction module is optimized,and the image loss function and difficult sample segmentation are also optimized.The main work of the thesis is as follows:(1)Aiming at the problem of brain tumor image segmentation in the case of small data sets,the U-Net network is used as the backbone network to design and implement an dual-branch network few-shot learning model.Firstly,the U-Net network most commonly used in medical image segmentation is introduced.Afterwards,a multi-branch network is introduced,and the U-Net network is used as the backbone network of the dual-branch network.The theory of few-shot learning is researched,and the advantages and disadvantages of various network models for few-shot learning are analyzed.Finally a dual branch network model including conditional branch and segmentation branch is realized,which is used for brain tumor MRI data to perform 5-Shot few-shot image segmentation experiment.(2)Aiming at the problem of low performance of dual-branch network brain tumor MRI image segmentation under the condition of small samples,technologies as the data amplification,dual-branch interactive module and mask average pooling are researched and realized for the few-shot learning network.An improved few-shot learning model for dual-branch networks is proposed.Firstly,on the basis of the dual-branch network few-shot learning model,the Mixup data enhancement method is used to enhance the data and enrich the data set,and then a Dual-Branch Concurrent Spatial and Channel Squeeze and Channel Excitation module is proposed to be applied to the network interaction module,so that the segmentation branch can make full use of the performance of the conditional branch.The mask average pooling technology is used to extract the representational vectors of the target in the supporting set.The cosine similarity is used to measure the distance between the feature vector of the query set and the support set,which is used to guide the segmentation of the query set.The results of 5-shot brain tumor MRI image segmentation show that the loss function converges faster,and the Dice coefficients of the three regions have been significantly improved.(3)Aiming at the optimization problem of enhanced tumor region,which is a difficult sample region in brain tumor image segmentation,a dual-branch network model based on the improved few-shot learning model for dual-branch networks for enhanced tumor region is designed,and a loss function optimization strategy for difficult sample segmentation based on contrastive learning is proposed.The network makes full use of the feature interaction characteristics of the condition branch and the segmentation branch in the few-shot learning framework.On this basis,it focuses on the optimization of the loss function in the difficult sample area of the brain tumor image,and uses the training loss of the support set data to assist the query set.The segmentation of difficult samples has been improved.This model is a hard sample optimization form under the few-shot model.The experiment results show that the segmentation accuracy is more accurate.