Research On Lung Tumor Image Segmentation Method Based On Hybrid Convolutio

Lung tumor is a common malignant tumor,which usually refers to a tumor that grows in the lung and may invade surrounding tissues and organs.Lung tumors are a serious medical issue and receive significant attention due to their high incidence and mortality rates.An important research direction in the field of medical imaging is the segmentation of lung tumors,which involves separating lung tumors from normal lung tissue for more accurately diagnose and treat lung cancer,and improve treatment outcomes and patient survival rates.Lung tumor segmentation techniques are mainly achieved through computer vision technology and deep learning algorithms.Despite their high accuracy and stability in the field of lung tumor segmentation,these methods still face challenges such as differences in imaging under different scanning devices and conditions,the complexity of lung anatomy,and the diversity of tumor morphology and size.To overcome these challenges,further research is needed to explore more efficient,accurate,and stable segmentation methods for lung tumors to improve their clinical value.This article proposes three different lung tumor segmentation methods based on methods in deep learning,which address the local and spatial relationships of nodes in lung tumor CT images,the topological relationships between nodes,and the global information of all nodes.(1)Feature fusion lung tumor segmentation method with attention gating for 2D and3 D convolution(FALTS)FALTS balances the strengths and weaknesses of 2D and 3D convolutions by fusing them together and utilizes the advantages of 3D convolutions to improve the representation ability of features between slices in 2D convolutions.Additionally,it uses an attention gating mechanism to filter features in cross-layer connections,focusing more on features in the region of interest and reducing the impact of irrelevant background noise,thereby improving segmentation accuracy.(2)Lung tumor segmentation method based on graph convolutional autoencoder and convolutional neural network(GCALTS)Tumors in lung images usually have complex,irregular shapes and structures.GCALTS uses graph convolution for graph inference to learn the topological features of image regions,including the connectivity and relationships between nodes.However,graph convolution operations have limitations in that they can only consider nodes that are directly adjacent to the current node and cannot capture more extensive contextual information.Therefore,as the network layers increase,the features of nodes are continuously averaged with the features of their surrounding adjacent nodes,which may lead to feature loss and information loss,known as over-smoothing.To alleviate this problem,GCALTS uses a convolutional neural network to learn specific features of image region nodes and supplements them to the corresponding graph convolutional encoding layer.(3)Lung tumor segmentation method based on multilayer perceptron and maxpooling(MPLTS)The topological and specific features learned in the GCALTS method can both be regarded as local features.To better segment lung tumors,MPLTS models a global feature learning module,which selects a most representative semantic feature from each channel of the initial feature map to obtain a global feature representation vector.MPLTS uses multilayer perceptron to obtain feature maps of different dimensions,thus obtaining multiple global feature representation vectors.By combining GCALTS with MPLTS,a model with both local and global feature learning capabilities was obtained,greatly improving the segmentation accuracy of lung tumors and the generalization ability of the model to different datasets.To verify the effectiveness and reliability of the proposed method,five lung tumor segmentation methods were used for comparison,and the experimental results and segmentation cases demonstrated their accuracy.