Research On Detection And Recognition Of Pulmonary Nodules Based On 3D Convolutional Network

Lung cancer is a global high-risk disease,especially in men.Lung cancer is the most common cancer,and early lung cancer is manifested in the form of pulmonary nodules.Low-dose lung CT screening for pulmonary nodules provides an effective method for early diagnosis,greatly reducing lung cancer mortality.However,with the increase of information data,it brings a huge burden to doctors’ work.Therefore,it is necessary to use computer-aided lung nodule detection technique,which is beneficial to reduce the diagnosis time and help doctors improve work efficiency.Most of the methods used in traditional lung nodule detection are based on machine learning algorithms.Combining professional medical domain knowledge and CT image features,manually extracting features and selecting useful features for training in classifiers,which not only is the process cumbersome,but also the generalization ability of the model is limited.Therefore,it is not directly applicable to the role of adjuvant therapy.With the application of deep convolutional neural network in the field of image processing,deep learning algorithms have gradually become a research hotspot.Deep convolutional neural network no longer need the steps of feature extraction design in traditional methods,but are automatically extracted by the network in convolution mode.The characteristics extracted by deep network are more specific and discerning.Therefore,more and more people are beginning to integrate deep learning algorithms into lung nodule image research.The existing methods mainly combine traditional machine learning algorithms with deep learning,firstly,traditional algorithms are applied in segmentation of suspected pulmonary nodules,then,deep convolutional neural network is used to reduce false positives in lung nodules.However,the resulting model tends to be poorly generalized and cannot be accurately segmented during the extraction of suspected nodules.There are also some methods using 2D convolutional neural network that do not make full use of the 3D spatial features,which makes the false positives of the extracted suspected nodule regions higher,thus affecting the overall classification effect.Aiming at the shortcomings of the existing methods,this paper designs a set of lung nodule detection and recognition methods based on deep learning algorithm.The whole process is automated,including effectively eliminating non-nodular regions,automatically segmenting suspected nodule regions,and extracting spatial features of three-dimensional image blocks to achieve classification of true and false nodules.The main work of this paper is as follows:(1)The lung nodules were segmented based on the 3D ResUnet network to extract suspected nodules.The lung parenchyma is extracted from the CT image of the whole lung nodule.First,the nodule coordinates in the annotation file are read,and the stereo image block is intercepted near the region centered on the coordinate,then,the generated stereo block is resampled and data augument technology is used to increase the samples.This paper combines the original UNET network with the residual structure to form a new network ResUnet,which fully combines the advantages of the two networks.First,the residual structure simplifies the training of the network.Second,the jump connection in the residual unit,information dissemination between low-level and high-level networks both will promote information propagation without gradient disappearance.In the semantic segmentation of lung nodule images,the input is expanded into three dimensions.The generated data is divided into training set and validation set by 7:3,the training is performed in a three-dimensional ResUnet network.Finally,the lung nodule segmentation result is obtained through the training model and compared with other methods,the results show that the recall rate of this method is97.68%,which can effectively eliminate non-nodular areas and lay a good foundation for the classification of true and false positive nodules.(2)In order to better represent spatial information and texture information,3D Inception deep convolutional neural network is used to realize the classification of true and false nodules.The model is mainly composed of 3D convolution layer,3D pooling layer,Inception module and softmax layer.Finally,the classification of the model is evaluated by various aspects such as accuracy,sensitivity,specificity and recall rate,the accuracy rate of 87.3% and the false positive rate of 22.8% were obtained.The results showed that the model had a higher probability of nodule detection and a lower false positive rate,which had a good effect on reducing false positives.The set of lung nodule detection and recognition algorithms proposed in this paper simplifies the cumbersome steps of traditional methods,realizing automatic diagnosis,and is conducive to early screening of lung cancer,providing an idea for assisting doctors to detect pulmonary nodules.