Research On Segmentation Algorithm Of Multimodal Tumor Medical Image

Automatic segmentation of tumors can solve the problem of inefficient and errorprone manual segmentation.In recent years,there have been many works on the automatic segmentation of medical images of tumors generated by MRI.However,MRI is prone to noise and artifacts,and the current segmentation work is affected by noise and cannot fully utilize the MRI multimodal imaging features,resulting in poor segmentation results.To address these problems,this paper focuses on the denoising and segmentation algorithms of multimodal brain tumor medical images,proposes a denoising model based on multi-generator generative adversarial networks and a segmentation model that makes full use of multimodal contextual information,and develops a 3D visualization software.The main work and contribution of the paper are as follows:First,a multimodal brain tumor image denoising model based on a multi-generator generative adversarial network is proposed for the noise or artifact problem in MRI,using the multimodal features of MRI and incorporating the subsequent segmentation task.Using the multimodal features of tumor images,a multi-generator architecture network is built,and the loss of the subsequent segmentation task is incorporated into the denoising loss function to remove noise information and retain important features by establishing a connection with the subsequent segmentation task.The experimental results show that the model can significantly decrease the number of artifacts or noise that appear in MRI,thus improving the segmentation accuracy.Second,to address the problem that the existing segmentation work fails to fully utilize multimodal information,a segmentation model that can fully utilize multimodal contextual information is proposed based on deep residual learning and UNet.By introducing a multiscale feature extraction module,a coordinate attention module,and a dense atrous spatial pyramid pooling module,the complementarity between multimodalities is fully utilized,thus paying more attention to the representation of fine-grained features and achieving information extraction and integration of multimodal images at different scales.Experiments on the denoised multimodal brain tumor dataset show CA-Res2UNet++ performs well on several currently popular medical segmentation metrics.Finally,in order to visualize multimodal brain images and segmented brain tumor images in three dimensions from current MRI,a visualization software based on the VTK tool library is developed.With a clear interface and a variety of functions,the software can visualize brain images of different modalities and segmented brain tumor images in three dimensions,thus showing the characteristic information of brain tumors in three dimensions and better assisting clinical treatment work.