Research On Detection,segmentation And Classification Of Benign And Malignant Pulmonary Nodules Based On Deep Learning

Lung cancer is a cancer with the highest morbidity and mortality in the world.Its early symptoms are lung nodules.Usually,medical workers are screened by observing Computed Tomography（CT）pictures,for early intervention.However,with the development and popularization of medical equipment,the accuracy and quantity of tomographic images are increasing day by day,and the demand for detecting pulmonary nodules based on lung images has skyrocketed,resulting in the need for a large number of medical workers with medical backgrounds to spend a lot of time and energy Observing the tomographic images of patients.Not only imposes a great burden on medical workers,but also easily leads to missed diagnosis of pulmonary nodules.Therefore,Computer Aided Diagnosis（CAD）,which can perform efficient and accurate automatic detection of pulmonary nodules,has become a promising method to solve the current predicament.The processing method of CAD is mainly a deep learning method based on Deep Neural Networks（DNN）.Due to the limitation of hardware computing power,two-dimensional tomographic CT slice images and three-dimensional CT small-volume block images are mostly used in recent years to deal with segmentation or detection problems.Especially for detection tasks,2D detection will face complex processing problems of slice screening and 3D reconstruction,and the independent processing of slices will lead to the loss of sequence information between slices and affect the accuracy of the model.3D detection is not a detection task in essence,but an alternative solution proposed due to the difficulty of 3D detection frame extraction,which requires complex post-processing to draw segmentation map detection frames.Therefore,this paper proposes an attempted solution for a new 3D CT image detection method.The system proposed in this paper includes four functional components: lung parenchyma extraction,lung nodule detection,local fine segmentation,and automatic classification of benign and malignant lung nodules.The backbone network is selected and improved for each functional component task,so that the overall network It is more efficient and improves the applicability of the task.The main research contents are as follows:（1）Study of lung parenchyma segmentation.The MF-VNet lung parenchyma segmentation network proposed in this chapter aims at the problem of insufficient multi-scale information fusion in the baseline V-Net network,and proposes the M-Block multi-scale block method and the mesh adaptive scale feature design method.The multi-depth route and mesh network design solve the applicability problem of target segmentation tasks of different sizes.The Dice index of lung parenchyma segmentation is 0.983,which is 0.02 higher than that of the baseline network,and the regression rate is0.987,which is 0.08 higher than that of the baseline network.The advanced network has been improved,but due to the maximum interpolation scaling operation during preprocessing,the edge features are more obvious and cannot be compared in parallel.（2）Pulmonary nodule detection research.The 3D-Center Net lung nodule detection network proposed in this chapter aims at the convergence failure problem in the baseline Center Net three-dimensional network,and proposes a positive and negative sample division strategy and agglomeration melting method of the M-Dice Loss function.By changing the ratio of positive and negative samples,The convergence failure problem of the Center Net three-dimensional model is solved,and the small-core average pooling is used to adaptively highlight the center point position,and the precision rate is 0.87 and the recall rate is 0.91,which is comparable to other advanced networks.But the detection speed has improved.（3）Partial study of lung nodule segmentation.The MF-VNet lung parenchyma segmentation network proposed in this chapter proposes the M-Block multi-scale block method and the mesh adaptive scale feature design method to solve the problem of unclear segmentation edges of small objects in the baseline V-Net network.The multi-level information compensation design solves the problem of over-segmentation/under-segmentation of small target segmentation tasks to a certain extent.The Dice coefficient of the lung nodule segmentation network is 0.831,which is0.04 higher than that of the baseline network,and the sensitivity is 0.805,which is higher than that of the baseline network.The baseline network improves by 0.05.（4）Classification of benign and malignant pulmonary nodules.This chapter uses the transformer feature extractor-based Vi T lung nodule classification network and the CNN feature extractor-based 3D-Res Net18 lung nodule classification network to conduct a comparative experiment,and obtains the Vi T classification index based on a small amount of data.The degree m PA is 0.398,which is lower than 0.049 of the 3D-Res Net18 network.The average accuracy m PA of benign and malignant grade 2 classification is0.755,which is lower than 0.122 of 3D-Res Net18 network.