2.5D Lung CT Images Segmentation Based On Gabor Convolutional Neural Network

Lung is an important respiratory organ and hematopoietic organ of human body.However,the long-term contact of lungs with the outside world through the respiratory tract and the influence of environmental pollution and unhealthy human lifestyle have led to a continuous increase in the incidence and mortality of lung cancer,which makes lung cancer become the main threat to human health.Using computer vision and deep learning technology to segment lung parenchyma and pulmonary nodules in lung CT images is the first and crucial step in the diagnosis of lung cancer.Due to the large volume of lung parenchyma,it is easy to be disturbed by the images of surrounding organs.In addition,pulmonary nodules are very similar to blood vessels in the lungs,and vary in location,shape and size.In this case,it is necessary to study a computer-aided diagnosis system to improve the accuracy of diagnosis and reduce the working pressure of clinicians.Therefore,this paper proposes a 2.5D lung CT images segmentation method based on Gabor convolutional neural network.Firstly,a computational model of 2D lung CT images segmentation based on Gabor convolutional neural network is proposed.Gabor convolution kernel has the advantage of matching the visual perception characteristics of human visual primary cortex cells,this model uses Gabor convolution kernel to replace the first and second layer traditional convolution kernel of U-Net,which can extract effective features in different directions and at different scales.In addition,parallel arous convolution and different-kernel pooling module are added to the bottom of U-Net.These two modules extract more abundant feature information by using multi-scale parallel arous convolution and multiple different-kernel pooling,and it ties the context information together tightly.Secondly,a 2.5D lung CT images segmentation method based on multi-sectional information fusion is proposed.In this method,Gabor convolution neural network is used to segment the 2D image sequences of the original 3D CT images in three sections(transverse section,sagittal section and coronal section),and the 2D prediction probability map of each section images segmentation is obtained.Then,the 2D prediction probability maps of the three sections are stacked to construct three 3D probability images with the same size.On this basis,the weight coefficients of the segmentation probability map of each section on the whole segmentation are obtained by classical network learning.Finally,the multi-information of the three sections are fused to obtain a more accurate 3D prediction map,so as to achieve multi-category segmentation of lung parenchyma and pulmonary nodules in lung CT images.Finally,in order to verify the effectiveness of 2.5D CT lung images segmentation method proposed in this paper,LUNA16(Lung Nodule Analysis 2016)dataset is selected for performance testing.The experimental results show that the 2.5D lung CT images segmentation method based on Gabor convolutional neural network proposed in this paper not only helps to improve the segmentation accuracy,but also greatly reduces the burden of medical personnel,which is expected to meet the needs of clinical application.