Research On Lung Nodule Detection Technology Based On 3D Neural Network

Lung cancer is the cancer type with the highest morbidity and mortality in my country and even the world,and it is showing a growing trend.Because the initial symptoms of lung cancer are not obvious,many patients miss the best time for treatment.Low-dose lung computer tomography(CT)scan images can be effectively used in the early diagnosis of lung cancer.Therefore,how to efficiently extract key lesion features from lung CT images and use the features for target detection has become an effective aid in the early diagnosis of lung cancer,and it is also one of the key technologies in medical images and computer vision.The early lesion features of lung cancer are manifested in the form of lung nodules,so the detection of lung nodules is the first step in lung cancer screening.At present,there are still many challenges in the detection of lung nodules:(1)The lung CT image has spatial continuity,and the internal structure of the lung parenchyma is more complicated,and the morphology and location of the lung nodules are diverse.The traditional 2D CNN will lose Spatial information in the image.(2)The automatic detection system for lung nodules is used in clinical diagnosis and must have the characteristics of high efficiency,high precision and low false positives.However,the current automatic detection system still has the problems of low detection accuracy,high false positives,and high time cost caused by high network complexity.With the rapid development of deep learning in the field of image processing,convolutional neural networks(CNN)have been widely used in the field of medical image research and have achieved good results.In order to solve the above problems,this paper proposes a lung nodule detection method based on 3D CNN.Based on this method,a high-precision lung nodule detection method based on multi-scale input is proposed to improve detection accuracy and save time and cost.And the lightweight lung nodule detection network based on MobileNet,and the corresponding experimental analysis was carried out.The main work includes the following 4 aspects:(1)Construct a practical training set.This paper performs preprocessing operations such as lung parenchymal segmentation and resampling on the lung nodule analysis2016(LUNA16)data set,which reduces the impact of irrelevant noise on the detection effect,reduces the search space during network detection,and maximizes the model training effect.(2)Propose a single-stage lung nodule detection method based on 3D CNN.Based on the UNet++ network,this algorithm adopts the flexible nesting mode of residual blocks to strengthen the feature reuse.And innovatively combine the improved UNet++-like architecture with the region proposal network(RPN).This paper refers to this model as the R-UNet++ model.Since both the 3D UNet++ network and the residual block have multi-feature fusion characteristics,the design of this algorithm is to enhance the model’s ability to extract features of the lesion.Experiments show that the R-UNet++ model proposed in this paper has an average sensitivity of 87% in the false positive screening based on the LUNA16 data set,which is an increase of 7.5%compared to the UNet++ network;when the number of candidate nodules is 48,the sensitivity The degree is as high as 95.5%,an increase of 7.4% compared to the VGG16 network.It can be seen that the R-UNet++ model can significantly improve detection sensitivity and reduce false positives,which can provide a theoretical reference for clinical applications.(3)Propose a high-precision lung nodule detection method based on multi-scale input.On the basis of the R-UNet++ detection model,a detection mode with multiscale input is proposed.This method is called the MR-UNet++ model in this paper.Three sizes are designed as the input of 3D CNN,and finally the classification output obtained by these three input sizes are merged,and the final nodule determination result is obtained.The final experimental results prove that the average detection accuracy of the MR-UNet++ model reaches 87.3%,and the detection sensitivity is 96.2%.(4)Propose a lightweight lung nodule detection network based on MobileNet.On the basis of the R-UNet++ detection network,the lightweight network MobileNet is introduced,and the depth-level separable convolution operation of the MobileNet network is used to replace the standard convolution operation in the R-UNet++ model.This method is called the LR-UNet++ model in this paper.The LR-UNet++ model basically maintains the existing detection accuracy,while optimizing the network parameters and the running speed of the model.The experimental comparison before and after optimization shows that the optimized network parameters are 58.2% less than the R-UNet++ model parameters,and the detection sensitivity is 94.0%.This shows that the introduction of MobileNet can make the detection model more efficient.