Design And Implementation Of MRI-Image Segmentation System For Rectal Tumors Based On Deep Learning

Accurate segmentation of rectal tumors can provide an important basis for clinical treatment and prognosis monitoring of tumors.However,there are many problems in the current rectal tumor segmentation task: the lack of high-quality datasets;The mainstream segmentation network cannot complete the high-precision segmentation task of rectal tumors.Lack of convenient and reliable automatic segmentation system.In order to solve the above problems,this thesis reseach the image enhancement and segmentation algorithm based on convolutional neural network to achieve data enhancement and automatic segmentation of Magnetic Resonance Imaging images of rectal tumors.The main work and innovation of this thesis can be summarized as follows:(1)To solve the problem of scarcity of high-quality MRI datasets of rectal tumors,this thesis improves the Cycle GAN network to realize modal transfer and reconstruction of CT images and MRI images of rectal tumors and uses the transferred and reconstructed images and the affine transformation method to construct a high-quality rectal tumor dataset.Firstly,the UNet++ network is used as the basic network structure of the Cycle GAN network to balance the proportion of shallow features and deep features in the UNet network.Secondly,the mixed attention mechanism of spatial domain and channel domain was introduced at the jump junction of the UNet++ network,and the weight of channel domain feature information was adaptively assigned.Finally,the SSIM loss function is introduced to improve the cyclic consensus loss and ontology mapping loss function in the Cycle GAN network to ensure the structural consistency between the reconstructed image and the original image.(2)To achieve high-precision segmentation of the lesion area in MRI images of rectal tumors,a novel segmentation network DCMSG-UNet is designed based on the U-Net network.Firstly,the hollow convolution and multi-head self-attention mechanisms were used to improve the basic feature extraction module of the segmentation network.Without changing the size of the convolution kernel,the feature information of a larger receptive field was extracted,and the multi-head self-attention mechanism was used to capture the long dependence of feature information in different receptive fields.Secondly,a decoder path was added to the network,which was used to fuse the down-sampling information of multi-scale features to realize the coarse segmentation of rectal tumors.Finally,the network uses GAM hybrid attention mechanism to amplify the dimensional interaction characteristics of the additional decoder path and then transmits the original decoder path to refine the segmentation results,to obtain more accurate segmentation results.(3)To realize the visualization and interactive segmentation of rectal tumors,this thesis developed a rectal tumor MRI image segmentation system based on Py Qt5 to assist doctors in the clinical medical diagnosis of rectal tumors.The system realizes the functions of automatic segmentation and visualization of rectal tumor images,digital quantification of tumor segmentation results,and true labels.In summary,the DCMSG-UNet network proposed in this thesis integrates the advantages of the UNet network,cavity convolution,multi-head self-attention mechanism,and channel attention mechanism,and is verified by experiments on the rectal tumor dataset constructed by Cycle GAN and affine transformation.The effectiveness and rationality of the rectal tumor image sample enhancement method and segmentation network designed in this thesis are proven.The segmentation system based on the DCMSG-UNet segmentation network can effectively realize the automatic segmentation task of rectal tumor images,which provides important reference information for clinical medical diagnosis and has certain research significance.