Research On Lung Nodule Detection And Incremental Learning Methods Based On Deep Learning Algorithm

In recent years,the incidence and mortality rate of lung cancer are tending to be high due to various factors such as smoking,air pollution and genetics,and lung tumors are seriously endangering people’s health.Lung nodules as an early symptom of lung cancer,early detection of lung nodules and timely monitoring are important to reduce the incidence and mortality of lung cancer.With the development of deep learning,computer-aided lung nodule detection has made great progress.Using deep learning methods for lung nodule detection can achieve higher accuracy and sensitivity,which is important for early diagnosis and treatment.Deep learning detection models can automatically detect lung nodules in a large number of CT images,greatly reducing the workload of physicians.And it can process a large amount of data in a short time,improving the speed and efficiency of detection.Compared with the traditional method of manual production of features,the deep learning model can automatically learn nodule features,which can effectively improve the accuracy of the model.Based on the various advantages of deep learning in lung nodule detection,the main research of this paper is as follows:(1)Existing 3D detection methods require more computational resources and training time,and 2D detection methods are efficient but not accurate.How to balance computational resources and detection accuracy is the problem faced by clinical aid diagnosis of pulmonary nodules.To address this problem,an anchorless lung nodule detection method with multidimensional and multi-scale feature fusion is proposed.The feature extraction phase of this detection method is divided into two-dimensional feature extraction and multi-view feature extraction,and firstly,multi-scale semantic features are extracted using a two-dimensional convolutional network.Secondly,based on the similarity between self-attention and 3D convolution,the Transformer encoder is used to extract multi-view features instead of 3D convolution in order to reduce the computational effort.Finally,the extracted multi-scale semantic features and multi-view features are fused by a multi-dimensional feature fusion module to obtain a fused feature map with multi-resolution,rich semantic features and accurate location information.An anchorless detection head oriented to bio-vision is designed in the detection stage,and a circular convolution kernel is used in the detection head to obtain a circular perceptual field more adapted to the morphological features of nodules,and a centroid branching plus radius branching more in line with physician labeling is designed.Experimental validation on the LUNA16 dataset shows that the method achieves a competitive performance metric(CPM)of 0.850,which is higher than the current mainstream 2D detection methods.the value of FLOPs is as low as 9.7G,the detection time of single sample sequence is 1.59 s,and the computational resource consumption is lower than the mainstream 3D detection methods,achieving an effective balance between computational resources and detection accuracy.(2)Two convolutional and self-attention fusion modules are proposed to address the phenomenon that convolutional neural networks have only local sensory fields in lung nodule detection and high false positive detection rate.Firstly,3D convolution is introduced into Transformer as a global C-T module to capture the global contextual connections of lung nodules.Where convolution is used to model local spatial location information and selfattention is used to collate global information to improve the network’s ability to discriminate false positive nodules.The self-attentiveness is then introduced into the residual module as a local C-T module to improve the network generalization ability.A spatial attention is proposed for the difficulty of calculating the rich information of 3D features,and the proposed TM calculation is used to calculate the Q,K,and V matrices.In order to keep the nodal position information in the 3D features unchanged,the spreading operation is no longer performed,but is calculated directly on the 3D matrices.Cross-validation experiments were performed on the LUNA16 dataset,and the results showed that the FROC curve of the method was overall higher than the mainstream methods,and the competitive performance index(CPM)reached 0.877,which has high sensitivity and provides effective reference information for the clinicianassisted physician detection.(3)When a new sample arrives,the existing model incremental update methods suffer from catastrophic forgetting and high update time and space costs.To overcome the above problems,an incremental learning method for lung nodule detection based on elastic weight consolidation(EWC)and feature distillation is proposed.Firstly,the lung parenchyma segmentation is performed on CT images;secondly,the original model is optimized based on the elastic weight integration method by simplifying the Fisher information matrix into the parameter weight matrix to reduce the computational effort,and then the regularization penalty term is introduced into the loss function based on the parameter weight matrix in the updating process to limit the changes of important parameters to obtain the optimized model after incremental updating;finally,the original model is used as the teacher network and the optimized model is used as the Finally,the original model is used as the teacher’s network and the optimized model is used as the student’s network to calculate the feature distillation loss to improve the performance of the optimized model and achieve a full fit to the new sample without forgetting the learned knowledge.The method is cross-validated on the LUNA16 dataset,and the results show that the method improves the incremental learning ability of the model,achieves better results in terms of sensitivity and accuracy,and achieves low false positives with low spatio-temporal overhead.(4)Based on the innovative method in this paper,a lung nodule assisted detection system is designed,which uses Vue as the front-end framework,Flask server as the model server,and Orthanc server for storing sliced data.The system achieves the detection and recognition of pulmonary nodules.In summary,this paper proposes two lung nodule detection methods and an incremental learning method for lung nodule detection to address some problems in deep learning-based lung nodule detection,and verifies the effectiveness of the methods through extensive experiments.And a pulmonary nodule detection system is designed based on the anchor-free lung nodule detection method with multidimensional and multiscale feature fusion,which can improve the efficiency of doctors.