Research On Detection Algorithm Of Pulmonary Nodules In Chest CT Images Based On Convolutional Neural Network

Due to the great success of deep learning technology and deep convolutional neural network in the field of image,by applying convolutional neural networks to CT image pulmonary nodule detection,a lot of research results have been obtained to improve detection rate and to reduce false positives.However,the pulmonary nodules in CT images have a wide variety of structures with different sizes and locations,and some false positive candidate pulmonary nodules have similar morphology with true nodules,and are easy to adhere to tissues such as trachea and blood vessels in the lungs,all of which greatly increases the difficulty of detecting pulmonary nodules.Although the convolutional neural network pulmonary nodule detection algorithm has achieved high performance,how to accurately and quickly detect and identify early pulmonary cancer nodules is still the focus and difficulty of the research based on the clinical application perspective.Therefore,in-depth study and development of convolutional neural network pulmonary nodule detection algorithm has theoretical significance and application value.On the basis of analyzing and studying the existing representative convolutional neural network algorithm for detection of pulmonary nodules,in order to reduce false positive rate and improve detection accuracy,the leading algorithm for universal target detection was applied,and RefineDet algorithm and SENet algorithm were integrated to design a convolutional neural network-based algorithm for detection of pulmonary nodules.(1)Reading of literature and research on related knowledge.Through the literature review,the CT image,pulmonary nodules and their detection related knowledge are deeply understood,and the theory and method of convolutional neural network applied to pulmonary nodule detection are deeply studied.Combined with the characteristics of pulmonary nodules,a pulmonary nodule detection algorithm based on convolutional neural network was designed based on the leading algorithm.(2)Data set preprocessing.Due to the format problem of the original image of the dataset,the format conversion is firstly performed,the relevant annotation information is extracted,and the rectangular frame is marked according to the annotation information.Experiment byselecting appropriate data through filtering,and noise processing of the data to increase the number of samples and improve the generalization ability of the model.(3)The design of the network structure.The network design mainly refers to the RefineDet network,removing the last two layers of convolution and its associated layers.For the detection of pulmonary nodules in this paper,because the target is small and difficult to be detected,the SENet module is added to the network to increase the image feature parameters,integrate more context information,and improve the model learning ability.(4)Experimental verification.The effectiveness of the proposed algorithm is verified by experiments from both qualitative and quantitative aspects.The FROC,AUC,and CPM indicators were used to quantify the effects,and the algorithms in this paper were compared with algorithms such as SSD,RefineDet.Innovations:(1)Improve the leading algorithm of general target detection and construct a suitable CNN network structure for CT image pulmonary nodule detection.The candidate box is filtered by a two-step cascading strategy.The anchor box refinement module is used to remove the negative candidate frame,and the position and size of the candidate frame are adjusted.The target detection module is used to regress the accurate object position and predict the nodule category.(2)The two-level feature extraction structure is adopted to make the features richer.Aiming at the small target problem of pulmonary nodules,the SENet module was introduced to enhance the feature,enhance the feature transfer and reuse,and adaptively learn feature weights by feature recalibration to suppress the features that are of little use in the current task,thus improving the performance of the model.Inadequacies:(1)The spatial information of the CT image is not considered,and the relationship between the CT image sequence before and after is not processed accordingly.In the future work,by designing a three-dimensional convolutional neural network,the spatial information can be effectively utilized to improve the network robustness.(2)In the training of the network model,the loss function for detecting the general target is used,and the problem of small target of the pulmonary nodule is not improved.In the future work,the lung nodules are detected in a targeted manner by improving the loss function.