Research And Implementation Of Brain Lesion Segmentation And Classification Based On Multipath Features

In recent years,deep learning technology has been widely used in the medical field,and medical image segmentation and medical image classification have attracted the attention of researchers.However,it is still challenging to quickly segment lesion regions and classify tumor types due to the small size of sellar region lesions with multiple brain tumors,large differences in different modal features of brain MRI images,and scarce data sets.Deep learning technology still has important research value and application prospects in the segmentation and classification of MRI medical images of brain sellar region lesions.This paper solves the above problems from the perspective of multi-path feature fusion,and realizes efficient automatic segmentation and classification of brain sellar region lesions.The main research contents are as follows:1)Aiming at the scarcity of medical image datasets,this paper proposes a brain sellar region lesion segmentation method based on multi-path visual Transformer.This method uses the multi-path visual Transformer as the backbone of the model,constructs a U-Net network,and uses skip connections with residuals to replace the original skip connections to directly stitch features,which can reduce the semantic difference in the process of downsampling and upsampling,and further improve Accuracy of model brain sellar region lesion segmentation.In the T2-weighted imaging data set,the mean intersection over union and mean Dice coefficient of the proposed model are 61.08%and 74.42%,respectively,which are 1.45%and 0.8%higher than the mainstream methods.On the T1 enhanced weighted imaging dataset,the mean intersection over union and mean Dice coefficient of the proposed model are 64.1%and 76.37%,respectively,which are 1.64%and 0.54%higher than the mainstream methods.2)Aiming at the problem of large semantic differences in multimodal MRI image features,existing methods ignore the information between modalities,which affects the accuracy and effect of disease diagnosis and treatment.In this paper,a classification model of sellar region lesions based on multi-path feature pyramid is proposed.In order to apply the process of real physician diagnosis to model design,a multi-path structure is used to extract the features of different modalities of MRI images.Each path uses 3D convolution for dimensionality reduction and uses a 3D pyramid with skip connections.The module performs multi-scale feature extraction,and then uses the nested modality-aware feature aggregation module to fuse multi-modal features,thereby improving the classification accuracy of the model.The accuracy rate of the proposed model is 69.35%,the precision is 69.68%,the recall rate is 68.90%,and the F1 score is 69.12%.Compared with the mainstream method,the accuracy rate increased by 3.23%,the precision increased by 3.01%,the recall rate increased by 3.75%,and the F1 score increased by 3.57%.3)An automatic segmentation and classification system for lesion images in the sellar region of the brain was constructed,which can apply the content of this research to the actual medical environment.The system can read the MRI images uploaded by doctors.It can not only segment specific lesion areas,but also assist doctors in diagnosing lesion types based on model prediction results,making medical imaging diagnosis more scientific,accurate and efficient.Work provides important technical support.