Lung Tumor Image Segmentation Based On Convolutional Neural Network And Transforme

Medical image segmentation is an auxiliary medical technology that classifies medical images at the pixel level,and the category of each pixel is obtained,and all the pixels finally summarized into the outline of the segmentation target in the medical image.The automatic segmentation of the tumor is mainly because its lesion location may locate anywhere in the organ,and the size of the tumor varies according to the type and time of onset.It is because of these uncertainties that clinicians and radiologists can make it challenging to judge the condition and locate the tumor.For a long time,tumors were judged by the clinical experience of clinicians and radiologists.With the continuous improvement of tumor incidence and the continuous upgrading of the accuracy of medical instruments,the workload of relevant doctors has gradually become very large,which affects the diagnosis and follow-up treatment of patients.Automatically segmenting CT images of lung patients based on informatics technology helps doctors to promote early diagnosis of lung cancer patients.We utilize the deep learning techniques to segment lung tumors from CT volumes is important for the diagnosis of lung cancer.(1)Lung tumor segmentation method based on multi-scale receptive fields.Extracting and integrating spatial dependence and semantic connections in lung images is very important for lung CT image segmentation.We propose a module based on multi-scale receptive fields to learn the spatial connections inside CT images,and finally segment lung tumors from CT images.We propose a new multi-scale learning strategy for3 D lung CT images to learn the multi-scale content representation of image nodes,and design an attention mechanism based on the scale level to achieve adaptive fusion of multiple content representations.(2)We design a lung tumor segmentation method based on 3D axial transformer,which is used to learn and integrate spatial associations and global information between image region nodes.An attention mechanism at the head level is designed to extract more representative information by automatically recognizing the different contributions of different head features in the attention of multiple heads,so as to achieve the purpose of adaptive fusion of the internal spatial connection of CT images.(3)Lung tumor segmentation method based on mixed attention transformer.We design a way to blend global and local information.First of all,the original multi-head self-attention of the Transformer is improved,and while retaining the original attention mechanism,we add branches at the spatial level and the channel level respectively to extract global information and local information.For different types of information,a category-level attention mechanism is designed,and more representative information is extracted by identifying different contributions of different information to tumor segmentation.The effectiveness of each innovation in our network is demonstrated by comparing experimental results with ablation experiments,especially when tumor boundaries are blurred,internal contrast changes are evident,and shapes are irregular.After the models are embedded in different segmentation trunks,the experimental results also prove the robustness of the methods.