Design And Implementation Of Brain Medical Image Segmentation Algorithm Based On Swin Transformer

MRI of brain tumors can provide a large amount of information about the lesions,which is an important basis for doctors to diagnose and treat brain tumors.MRI evaluation of brain tumors is a highly specialized task that requires experienced doctors to complete.Due to the shortage of medical resources,the subjectivity and insufficient efficiency and accuracy caused by manual segmentation,an automated segmentation technology is urgently needed to assist medical diagnosis.With the rapid development of artificial intelligence,big data technology and computer technology,the medical assistant diagnosis system based on machine learning technology has gradually come into reality.The existing medical image segmentation methods are mostly based on convolution neural network,which ignore the relationship between different parts of the image due to the inability to balance the receptive field and the model size,leading to low accuracy and efficiency.Swin Transform is a network model based on the self-attention mechanism that has emerged in the past two years.It has stronger modeling and representation capabilities,and can extract image features from a larger receptive field,achieving state-of-the-art performance in many computer vision subdivision fields.This paper focuses on brain tumor MRI data and segmentation models,explores effective data preprocessing mechanisms,and uses the current advanced Transformer structure and multi-scale fusion techniques to build a network model to achieve accurate and efficient segmentation of brain tumor images.The main work of this paper is as follows:(1)Through the experiments and analysis of the data format of brain tumor MRI image,the appropriate data preprocessing process is explored.In view of the difference in the distribution of pixel values and the imbalance of categories in the data set,data clipping,loss function optimization and z-score normalization strategies are adopted,and the preprocessed data is stored in TIFF format to facilitate image visualization and model migration.(2)In view of the bottleneck of the current medical image segmentation model,this paper introduces the current advanced Swin Transformer into the field of medical image processing,extracts image features from a larger receptive field through the self-attention mechanism,and constructs a hierarchical feature pyramid structure.Then,the multiscale information of the feature pyramid is fused using the UPerNet,which makes full use of the low-level and high-level semantic information of the image to increase the feature representation capability of the model.Therefore,the UP-SwinT model is proposed in this paper,and the feasibility and effectiveness of the scheme are verified by experiments.(3)Aiming at the problem of detail information loss caused by the initial quadruple down-sampling of the Swin Transformer model,a new CSwinT backbone network integrating CNN and Swin Transformer structure is proposed.In order to further improve the accuracy of tumor segmentation for small lesion areas,this paper uses U-shaped network for multi-scale fusion and proposes CSwinT-Unet model.The model adopts a U-shaped structure of coder-decoder.The encoder is composed of a twolayer CNN module and Swin Transformer network.The image is extracted from the local low-level features of the CNN,and then input into the Swin Transformer network for the operation of the self-attention mechanism.The multi-scale feature map is obtained,and then the final segmentation map is generated by the multi-scale feature fusion of the decoder.Finally,this paper verifies the enhancement effect of the model through the ablation experiment,and the comparison experiment with other models verifies that the model has good segmentation accuracy for the point-like discontinuous tumor regions and the tumor edges.In summary,this paper proposes the corresponding data preprocessing strategy according to the characteristics of brain tumor images,and proposes an accurate and efficient CSwinT-Unet model.Through the experiments on the BRATS2020 and BRATS2021 data sets,the accuracy,efficiency and good generalization of the model are verified,which is of great significance for computer-aided diagnosis and medical treatment.