Research On Multi-Scale Organ Segmentation Of CT In Head And Neck Based On Deep Learning

Medical image segmentation is a necessary pre-element for many clinical disease treatments,because the precision of traditional segmentation method is difficult to meet the requirements.More segmentation transactions will be handed over to doctors for manual processing,which consumes a lot of unnecessary manpower.Convolutional neural network has adapted the filter to allow for data driven error feedback,making the algorithm model more accurate for a particular image segmentation model.Due to the characteristics of CT image characteristics and less morphology,voxel segmentation is required,each time the single organism segmentation is relatively low,and the direct mixing multi-organ label will allow the model to ignore a small volume of organs.The objective of this study is to study the Parotid,Mandible,Optical Chiasm and Parotid in head and neck CT images.Some segmentation methods based on deep learning and driven by the combination of reinforcement modules are explored.According to the characteristics of head and neck CT images,the method can be divided into two stages.The first stage is to get the 3D detection frame of organs by regression method and cut the images along the detection frame.The second stage is the pixel-level image segmentation of the model in the frame.In the second stage,we perform two-stage image preprocessing including image normalization and nonlinear numerical mapping and data expansion based on elastic deformation.Then the algorithm of multi-scale head and neck CT image is studied and improved.The following sampling-upper sampling structure is mainly focused on attention U-Net,which is improved in three aspects: symmetrical structure,classification assistant module and information exchange module: firstly,we add an up-sampling-downsampling branch to the U-Net part of attention to capture the small features in the data;secondly,the information exchange module is added between the main branch and the branches to exchange the information of the main branch and the branches to restrict each other,so that the two features complement each other instead of separating the different parts completely;thirdly,the spatial dimension of different organs is fixed by using classification assistant module,which can enhance the ability of feature extraction of different organs and a new classification model based on few-shot learning proposed for this module.According to each innovation point of the model,detailed experiments are carried out to verify the effect of each module.Finally,the method is applied to the head and neck CT images,and the performance improvement of the algorithm model is verified by the combination of auxiliary modules.Through the comparative experiment of head and neck CT image data set,the results show that the evaluation results of the algorithm model proposed in this paper are better than the mainstream u-net network model,which realizes the accurate segmentation of head and neck CT image,and provides a better solution for multiscale organ segmentation.