Research On Lung Tumor CT Image Segmentation Method Based On Spatial Enhancemen

The lung cancer is a malignant tumor of the lung that is the most common and deadly worldwide.At present,the screening method based on computed tomography(CT)scanning images is the most widely used auxiliary means in the diagnosis and treatment of lung cancer.Accurately segmenting lung tumors from CT images can be challenging for physicians because the size,shape,and location of growth of tumors can vary greatly between patients.Therefore,it is very important for doctors to segment lung tumors from a large number of lung CT images with the help of automatic segmentation algorithm of lung tumors for diagnosis and treatment.In this paper,based on the artificial intelligence method,the research task of CT image segmentation of lung tumors is carried out.The main work completed is listed as follows.First,based on the research of encoding spatial information and mixed channel information,this project designs a lung tumor segmentation method based on Space-MLP and Channel-MLP(referred to as PCMLP).This topic discusses the shortcomings of the previous multi-layer perceptron network.The previous multi-layer perceptron network always encodes spatial information along the flattened spatial dimension,and the position information on different dimensions of the feature map and the connection between different dimensions have not been fully utilized.In order to encode spatial information of different dimensions,a Space-MLP is proposed to separately encode the features of different dimensions in the feature map through linear projection.Channel-MLP models the relationship between different channels through a non-linearly activated two-layer MLP.For the experimental results,this paper evaluates by using the Dice coefficient,Io U coefficient and HD distance,and achieves higher segmentation results on the public lung tumor dataset,which demonstrates that separately encoding the height,width,depth and channel information of a 3D CT volume can improve segmentation performance.Second,based on the research on the main factors of self-attention mechanism to improve image segmentation performance,this topic proposes a lung tumor segmentation method based on recursive gated convolution(referred to as RGC).Through the in-depth study of the self-attention mechanism,this topic finds that the self-attention mechanism can improve the image segmentation effect through the high-order interaction of spatial information and the capture of long-distance dependencies.However,convolution-based improved models can equally efficiently implement these factors that improve segmentation performance.And the higher-order spatial interaction mechanism has not been studied in depth.To achieve higher-order spatial interactions,recursive gated convolutions module are built to extend self-attentional two-order spatial interactions to arbitrary orders by recursive reason and gated convolutions.Experimental results demonstrate that recurrent gated convolution can be a promising alternative to selfattention mechanisms.The generality of the RGC model is verified by using multiple different 3D image segmentation backbone networks.Ablation studies and comparisons with other approaches demonstrate the effectiveness and improved segmentation performance of technical contribution of this method.Third,based on the study of multi-channels contextual relations,spatial and position dependencies across image regions,this topic establishes a segmentation model based on convolutional bi-directional learning and spatial enhanced attention(referred to as PRCS).This topic discusses and finds that the state-of-the-art tumor segmentation models do not fully consider the global relationship between image region locations and the contextual connections between multiple channels.Therefore,to extract contextual relationships between different deep image feature tensor channels,we propose a new convolutional bi-directional gated recurrent unit based module for forward and backward learning.In addition,a novel cross-channel region-level attention mechanism is designed to discriminate the contributions of different local regions and associated features to the global learning process.Finally,to formulate spatial and position dependencies,a new position enhanced self-attention mechanism is built.The new attention can measure the different contributions of other positions to the target position and obtain the enhanced adaptive feature vector of the target position.In terms of spatial overlap and shape similarity,PRCS model outperforms 7 state-of-the-art segmentation methods on public lung tumor datasets.Ablation studies and segmentation results on different 3D CNN segmentation backbones demonstrate the effectiveness and generality of the three technical innovations.