Research And Implementation Of Pulmonary Nodule Classification Model Based On Attention Mechanism

According to the latest statistics from the National Cancer Center,lung cancer has the highest incidence and mortality in China.Many patients lost their precious lives because they could not get a timely diagnosis at an early stage and missed the best time for treatment.If it can be found early,it can save many lives.With the development of medical imaging,computerized tomography(CT)for early diagnosis can help doctors diagnose benign and malignant lung nodules,so that lung cancer can be detected in time.However,the diagnosis of doctors often depends on subjective experience.Different doctors may draw different conclusions on the same CT scan,and screening of CT scans requires a lot of time and energy,which is not conducive to the timely diagnosis of patients.With the rise of artificial intelligence,CT scans of the lungs are processed to achieve classification of benign and malignant lung nodules,thereby helping doctors to assist in diagnosis,which has become a research focus of computer vision.The traditional method is to segment the lungs or manually design features and classify them by machine learning algorithms,which requires a lot of time and human resources,and it is easy to introduce errors when segmenting or manually designing features.Because CT scans of the lungs are scanned along an axis,the classification of lung nodules requires consideration of axial slices.At the same time,the size of the different lung nodules is different,and the area occupied by the CT scan is small,and the number of slices is small.These issues have brought great challenges to the accurate classification of lung nodules.In response to the above issues,this paper proposes a new model of 3D dual-path bidirectional feature level fusion network for lung nodule classification based on attention mechanism.The main work of this article is as follows:1.A new 3D dual-path network structure is designed.Different from the traditional methods of manually design features,this paper uses deep neural networks to extract and learn high-dimensional features from CT scans of the lungs.Aiming at the problem of the axial direction of the CT scan of the lung,the number of slices along the axial direction is taken as the third dimension to perform 3D convolution extraction features.Due to the large amount of 3D convolution parameters,in order to reduce the model size and required memory,a dual path network connection(DPN)was added to the 3D network as the basic network structure,and a 3D DPN network structure was designed.2.A bidirectional feature level fusion network model based on attention mechanism is designed.The method based on feature level fusion can solve the problem that the information of different scales may be lost during the feature fusion and affect the classification effect.It has two network paths,with CT scans of different scales as input,and 3D DPN as the basic network module.Unidirectional feature level fusion uses one of the network paths as the main network and the other as the auxiliary network.Features extracted through different convolution modules of the auxiliary network are provided to the main network to achieve the purpose of unidirectional feature level fusion.The bidirectional feature level fusion is based on the unidirectional feature level fusion.The two network paths assist each other.When one network serves as the main network,the other network serves as the auxiliary network to provide features for it.Correlation is added between the two scales,and features of both scales are used comprehensively.3.Based on the public dataset LIDC-IDRI and LUNA16,the validity of the proposed classification model is verified.