Research On Detection And Classification Method Of CT Images Of Pulmonary Nodules Based On Deep Learning Network

Accurately identifying malignant lung nodules on computed tomography images is the key to early detection of lung cancer.But this is a tedious and difficult task because the radiologist needs to manually mark the nodule location.With the introduction of deep learning technology,a computer-assisted nodule detection system based on convolutional neural networks was developed for the early detection and classification and screening process of lung cancer,which significantly improved the quality and efficiency of radiologists’ diagnosis.Based on 3D convolutional neural networks,this subject focuses on the detection and classification of lung nodules.The research work is mainly divided into the following two parts:(1)CNN based on dual-path U-Net network lung nodule detection model.This paper proposes a two-stage convolutional neural network(URCNN)for nodule detection.The first stage uses a combination of U-Net network design and DPN module to extract features of lung nodules.There is a shortcut connection between the down-sampling and up-sampling paths,which can provide more spatial information,and at the same time use the DPN module to explore potential new features and feature sharing between the convolutional layers of the same module.Then,the best fine-grained feature map is input to the second stage nodule detection network.The first step is to use the feature map to generate candidate nodules of different shapes and sizes,and then the second step is to classify the candidate nodules and bounding boxes return.Compared with the original Faster R-CNN model,experiments can prove that the model detection effect of this chapter is significantly improved.A large number of experiments show that the model can achieve a sensitivity of 96.3% when the number of FPs per scan is 4.0,and the CPM score is 0.899.In addition,the segmentation performance of the U-Net sub-network part of the first stage was tested,and the Dice score was 82.36%.It is proved that this model can effectively improve the detection accuracy and reduce the false positive rate.(2)Nodule classification model based on channel attention dual-path network.Lung nodules usually have a complicated background environment and unrelated tissues with a high degree of similarity.In response to this problem,the model proposes a novel dualpath network structure that integrates channel attention mechanisms,which can avoid the disappearance of gradients and the effective transmission of context information between network layers through dual-path residual connections and dense connections,and then combine The squeeze-stimulus module realizes the perception of important channel information in three-dimensional space to mine the semantic information of lung nodules and their surrounding environment,thereby efficiently distinguishing lung nodules of different resolutions and sizes.The introduction of the attention mechanism can automatically retain the beneficial features of the dual-path network layer output,while ignoring irrelevant features to enhance the ability of the fusion part of the network to use these features.On the LIDC-IDRI dataset,this method obtains an average sensitivity of97.42% and an AUC score of 98.60%.The experimental results show that the model has outstanding performance in the classification task of pulmonary nodules.