Research On Lung Tissue Segmentation Based On Convolutional Neural Network

In recent years,due to the rapid development of medical technology,various medical image imaging technologies have emerged and are widely used in major medical and health systems.Medical image analysis has also become an important basis for doctors to diagnose diseases,and medical image segmentation technology,as one of the three main tasks of medical image analysis at present,has been a key and difficult point of visual research in graphic imaging.The emerging field of artificial intelligence has allowed deep learning techniques to show great advantages in the field of image segmentation,attracting the attention of a wide range of scientific scholars.At present,deep learning-based methods have achieved more satisfactory results in the field of natural image segmentation,but the application of deep learning technology in the field of medical image segmentation is still more challenging due to the characteristics of medical images,such as modal diversity,small differences in gray values and high influence by noise.This thesis mainly focuses on the existing deep learning-based segmentation methods for lung CT images on the problems of a series of research work,the main work and innovation points of this thesis include the following three parts:(1)In response to the problems that existing segmentation algorithms are computationally complex and resource-consuming,as well as the small differences in grayscale distribution in lung CT images,which are seriously affected by noise and so on,this thesis designs a lightweight two-dimensional convolutional neural network-based model for segmenting lung parenchyma.The model uses multi-scale convolutional groups to extract feature information of different scales of CT images,which enriches the semantic features of lung parenchyma while reducing the model parameters.And a multi-scale hybrid attention module is designed to enhance the segmentation performance of the network by combining the attention weights jointly with the high-level semantic features in the low-resolution images to highlight the overall structure of the lung parenchyma while fully filtering the noise components in the current images to strengthen the boundary features of the lung parenchyma.In this thesis,the effectiveness of this algorithm is verified on three public datasets,and Dice achieves an average value of 93.25%,Io U achieves an average value of 91.37%,the number of model parameters is only 1.98 MB,and the computational volume is 18.06 GMac.experimental results fully demonstrate that this method has better segmentation performance on the lung parenchyma segmentation task compared with existing deep learning-based methods,and The experimental results fully demonstrate that the present method has better segmentation performance for the lung parenchyma segmentation task than existing deep learning-based methods,and has significant advantages in the number of model parameters and computation volume.(2)For the above proposed algorithm is only limited to two-dimensional basis,considering that CT images belong to three-dimensional imaging,based on the two-dimensional convolutional neural network model in feature extraction for the relationship between image and image feature information is missing,so the above proposed network model is extended to three-dimensional space,and a segmentation model based on three-dimensional convolutional neural network is designed,using three-dimensional multi-scale convolutional group can extract more spatial feature information of CT images,so as to achieve the true meaning of three-dimensional segmentation of lung parenchyma.In this thesis,the effectiveness of this algorithm was verified on two public data sets,and the average value of 99.25% was obtained for Dice and 98.32% was obtained for Io U.The final experimental results demonstrate the feasibility of the proposed 3D segmentation algorithm,and the proposed algorithm has better segmentation performance compared with other deep learning-based methods.(3)To address the problems of end vessel breakage or poor recognition in pulmonary vascular segmentation,this thesis introduces a skeleton extraction module on top of the proposed 3D segmentation network model,by calculating the "gap" between the pulmonary vascular segmentation results and the real sample label skeleton structure,and using this "gap By calculating the "gap" between the lung vessel segmentation results and the real sample label skeleton structure,and using this "gap" as part of the loss of the model to jointly train the overall network model,the topological continuity and integrity of the lung vessel segmentation results are maintained.Finally,the effectiveness and feasibility of the skeleton extraction module are verified through relevant experiments,and the proposed segmentation algorithm has better segmentation performance on the pulmonary vascular segmentation task compared with other deep learning-based methods.