Research On Lung Nodule Recognition Method Based On Reverse Connection And Multi-view Feature Fusion

Lung cancer is a major disease that endangers the health and safety of all human beings.Early screening and diagnosis of lung nodules can effectively reduce lung cancer mortality.Therefore,the research on lung nodule recognition is of great significance.CT(Computed Tomography)scan images are more advantageous in the research of lung nodule recognition due to their advantages such as fast imaging and clear images.At present,the lung nodule recognition algorithm based on deep learning has the following problems: the CT image environment is complex,the small lung nodules are missing in the high-level depth feature map,and the false positive rate of detection results is high.Therefore,this paper studies a two-stage lung nodule recognition algorithm based on reverse connection and multi-view feature fusion: In the first stage,a lung nodule detection method based on reverse connection and multi-level features realizes CT lung nodule detection.The second stage,the lung nodule classification method based on multi-view depth feature fusion to reduce the false positive rate of test results.And based on the above research results,an automatic identification and auxiliary diagnosis system for lung nodules based on CT image sequences has been developed.The main work of this paper is as follows:(1)Aiming at the problem of the complex detection environment of CT image sequences,this paper performs lung CT image preprocessing steps based on image processing methods of denoising,enhancement,and lung parenchymal segmentation.First,the image noise is removed based on the Gaussian filtering method;then,the contrast of the lung nodules is enhanced based on the image enhancement algorithm;finally,the lung parenchymal region is segmented based on the combination of threshold segmentation and edge convex hull.Experiments show that the preprocessing method of CT image sequence based on denoising,enhancement,and lung parenchymal segmentation can effectively remove most of the unrelated areas of lung nodule detection and reduce the amount of calculation for lung nodule detection.(2)Aiming at the problem of the loss of small lung nodule targets in the high-level feature map and the wide distribution of lung nodule sizes,in the first stage,the lung nodule detection method based on reverse connection and multi-level features realizes the detection of lung nodules Detection.First,the low-level feature map is used to obtain high-level semantic information based on the reverse connection method;then,the lung nodule candidate regions are extracted based on the multi-level deep feature map and RPN;finally,the false positive lung nodules are initially screened and removed based on feature mapping and fully connected classification..Experiments show that the lung nodule detection method based on reverse connection and multi-level features has higher sensitivity and recall rate for lung nodules.(3)In view of the high false positive rate of lung nodule detection results in the first stage,the lung nodule classification method based on multi-view depth feature fusion in the second stage reduces the false positive rate of the first stage detection results.First,based on the first stage of the deep feature extraction network to obtain the three-view depth feature map of the lung nodule area;then,based on the channel fusion method to reduce the feature information redundancy of the lung nodule depth feature map,and the three views are merged by full connection Finally,based on the method of mining difficult samples of lung nodules and the Focal loss function to solve the problem of sample imbalance in the classification of lung nodules/non-pulmonary nodules.Experiments show that the algorithm in this chapter has a higher accuracy rate for the classification of lung nodules/non-pulmonary nodules,and can effectively reduce the false positive rate of lung nodule detection results.(4)Based on the above content,this paper designs and develops a computer-aided recognition system for lung nodules based on CT image sequences.The system contains four modules: CT image sequence reading and display,automatic detection of lung nodules,identification of lung nodule recognition results,and saving of lung nodule recognition results.Through this system,automatic auxiliary identification and diagnosis of lung nodules can be realized.Experiments show the use process of the system in this paper,and verify the actual performance of the algorithm in this paper.Experimental results show that the algorithm in this paper is highly sensitive to the detection of lung nodules,and can effectively reduce the false positive rate of lung nodule detection results.The algorithm in this paper can effectively detect lung nodules based on CT image sequences and assist doctors in identifying and diagnosing lung nodules,which has clinical significance.