Research And Implementation Of Lung Nodule Segmentation Model Based On Improved V-Net Network

Lung cancer is a common malignant tumor and one of the most fatal cancers in the world.According to the survey,about 1,8 million people die from lung cancer each year worldwide,posing a serious threat to human health.Despite the continuous improvement and refinement of lung cancer treatment methods,the medical community still faces many challenges.Early symptoms of lung cancer manifest as the formation of lung nodules.Therefore,timely screening and diagnosis of lung nodules is an effective means to prevent and treat lung cancer at an early stage.However,lung nodules appear in different locations in the lung,and their tiny size,variety,and different shapes make it difficult to perform accurate lung nodule segmentation in clinical practice.Especially in the case of manual segmentation of pulmonary nodules,the surrounding light,the clinician’s personal experience,fatigue,and the influence of subjective consciousness all become factors that make it difficult to segment pulmonary nodules accurately.In recent years,with the development of computer technology,the use of image segmentation technology in computer image processing to assist physicians in clinical diagnosis can effectively reduce the workload of physicians,improve the accuracy and efficiency of lesion segmentation,provide the basis for lung cancer screening,and defend people’s life and health.In this paper,we present a review of current lung nodule segmentation schemes,and aim to implement a network model capable of segmenting lung nodules with high accuracy,using convolutional neural networks in deep learning as the method.This paper demonstrates the accuracy and effectiveness of the model with comparative segmentation experiments on the Luna16 dataset.The main research contents and results of this paper are as follows:1.To address the problem that the small sample size of medical image datasets affects the segmentation accuracy of the network model training results,we propose and use several enhancement methods specifically for 3D medical images,horizontal image transposition and horizontal multi-angle mirror flip combined with vertical image inverse order emission,which preserves the semantic features of the planar images while taking into account the orderliness of the image sequence in the vertical direction.It does not have a negative impact on the 3D segmentation network.The scheme is experimented on the Luna16 dataset,which is augmented and expanded,and the effectiveness of the data augmentation method is confirmed.2.To address the problems of weak feature extraction and inefficient semantic feature utilization in V-Net and its derivative networks,first,we design a portable Dig＿Sep Block,a feature separation extraction module based on pixel threshold,which we embed into the input layer of the V-Net network model while also retaining the original semantic information,so that the network model We embed this module into the input layer of the V-Net network model,while retaining the original semantic information,so that the network model can extract richer and more hierarchical original semantic information features in the initial stage.Second,due to the complex structure of the 3D semantic segmentation network model and the deeper network layers,we replace the PReLU activation function used in this network with the ELU activation function that is more suitable for deeper networks,so that the network can perform better during the training process.Finally,we constructed a new network architecture DigVNet that incorporates the Dig＿Sep module and uses the ELU activation function.Through a series of network model comparison and ablation experiments,we demonstrated that the DigV-Net model has a certain improvement in segmentation accuracy compared with some other classical models.In the ablation experimental stage,the improvement of the model in the prediction accuracy of positive samples is not obvious.3.Aiming at the problem that the ablation experiment effect in the DigV-Net model is not obvious,an improved network model Dig-CS-VNet is proposed.By changing the dual attention mechanism of CBAM spatial channel and integrating it into the end of each upsampling layer of the encoder part of the improved model,the model enhances the network’s perception ability of features and the extraction of specific feature information,so as to further improve the prediction ability of positive samples.At the same time,the effectiveness of the 3D-CBAM attention mechanism is confirmed by ablation experiments,which shows that the improved network Dig-CS-VNet has new improvements in edge contour segmentation and positive sample prediction accuracy.In summary,this paper improves the shortcomings of the existing 3D segmentation network model,such as insufficient sample data and insufficient ability to extract hierarchical and detailed features during model training.A network model for 3D medical image segmentation with higher segmentation performance than most existing network models is implemented.The segmentation results of the model have certain clinical medical reference value.