Pulmonary Nodule Segmentation Algorithm Based On Deep Learning

Lung cancer is the disease with the highest morbidity and mortality in China.In recent years,the morbidity rate has continued to increase,which has seriously affected the health of people.There is no obvious sign of lung cancer in the early stage.When the symptoms appear,most patients are already in the advanced stage.Even if a lot of medical resources are spent,the prognostic survival rate is still low.The early stage of lung cancer mainly exists in the form of pulmonary nodules on imaging.Early screening for lung cancer can detect the condition as early as possible,and lower medical costs in exchange for a higher survival rate of patients.Therefore,early screening is extremely important for the treatment of lung cancer.Aiming at the characteristics of pulmonary nodules and the shortcomings of existing deep learning networks,this paper proposes a new pulmonary nodules segmentation model with the research of high-precision pulmonary nodules segmentation algorithms and uses different types of pulmonary nodules in the LIDC-IDRI to verify the segmentation performance of the model.The main contents are as follows:(1)The preprocessing method of lung CT image is proposed.First,the original image is denoised by median filtering,and the image is divided into the training set and verification set,and the training set is augmented by translation,mirroring,rotation,zooming,and other methods,according to the XML file structure and reading principle extract nodule contours marked by experts to provide experimental data for subsequent experiments.Secondly,according to the characteristics of CT images,a series of algorithms such as image standardization,K-Means algorithm,and morphological operations are used to obtain the lung parenchyma Mask,and the lung parenchyma Mask is used to obtain the complete lung parenchymal area.(2)Res Wnet is proposed to solve the problem that U-Net and its deformed lung nodule segmentation algorithm have more continuous downsampling times,which makes the contour feature information of lung nodules lost.By adjusting the encoding-decoding structure of a single 4 consecutive downsamplings to two sets of encoding-decoding network structures of 2 consecutive downsamplings,more feature information is retained in the network and deeper semantic information is extracted;use the image pyramid module inputs feature images of different scales from the left side of the network to improve the sensitivity of the network to nodules of different scales;fusion and segment the feature images of different levels,scales,and receptive fields,and deepen the network’s recognition of nodules and non-nodules.The perception of the knot area.Use the LIDC-IDRI data set to perform ablation experiments on the network composed of each module,and determine the final structure of the network.Comparing the segmentation results of different types of nodules by U-Net and Res Wnet,the results show that the segmentation performance of Res Wnet for various nodules is better than U-Net.It proves the effectiveness of improving the segmentation accuracy of lung nodules by reducing the number of consecutive downsamplings and increasing the encoding and decoding paths.(3)Res Blocknet is proposed to solve the problems of U-Net and its deformed lung nodule segmentation algorithm,which have fewer feature image propagation paths and a low utilization rate of feature images.Compared with U-Net and its deformed structure,Res Blocknet has more feature propagation paths,and can make full use of context feature information.The image pyramid module inputs feature images of different scales into the network to improve the sensitivity of the network to nodules;the improved residual block can fuse local feature information at different levels;the fusion 1 module uses the idea of Dense jump connection to fuse the features,It can make full use of the relevance of the context feature information to learn the feature information of nodules and non-nodules;using the Fusion 2 module can improve the network’s segmentation effect on low-contrast or fuzzy-edge nodules.Use the LIDC-IDRI data set to perform ablation experiments on the network structure composed of each module to determine the final structure of Res Blocknet.By analyzing Res Blocknet’s segmentation results of different types of nodules,it is verified that the network has a good segmentation effect on various types of lung nodules.At the same time,by segmenting the same CT image containing multiple different types of nodules,it is verified that the network has strong robustness to the segmentation of CT images containing multiple different types of nodules.