Research And Deployment Of Lung Cancer Detection Model Based On Convolutional Neural Network

Lung cancer is one of the highest mortality cancers in the world.Although computed tomography(CT)can provide valuable information for the diagnosis of lung diseases,because the types of lung cancer are not single,and often a patient has a large number of CT scanning images,it is easy for imaging doctors to make false detection and increase the treatment cost due to fatigue in the process of observing images for a long time.In order to reduce the burden of doctors and reduce false detection to a certain extent,computer-aided diagnosis(CAD)system is applied to the detection of lung lesions.In the current research on the detection of lung lesions,most of them focus on the detection of early lung nodules and the determination of benign and malignant.There is less research on the detection and classification of lung cancer.Lung cancer has the characteristics of different types,and the treatment methods of different types of lung cancer are also different.Therefore,it is also of great significance to identify different types of lung cancer.Based on the Lung-PET-CT-Dx dataset provided by The Cancer Image Archive(TCIA)and the cooperative hospital dataset,this paper implements an end-to-end lung cancer detection and classification network.The following is the main research content of this paper:1.Collect samples and establish data sets.There is an extreme imbalance between lung cancer categories in the public data set Lung-PET-CT-Dx.Although the data set can be reconstructed by sampling technology to achieve relative balance,the data set constructed by sampling technology is different from the real data set in distribution.Therefore,in order to verify the performance of the model under the sample balanced data set,we constructed a lung cancer medical image data set with the cooperative hospital.2.A lung cancer identification model based on improved detection,classification and false positive reduction of yolox network is proposed.Based on the detection model of yolox,combining the shallow characteristics of the network with the coordinates of regression,an end-to-end false positive reduction and detection network is proposed,which reduces the false detection rate of the model to a certain extent.In order to further enhance the accuracy of the network for lung cancer detection,the spatial and channel attention modules are introduced into the backbone network,and the loss function of the network classification branch is modified.The experimental results show that compared with the basic model,the m AP of CT and PET in Lung-PET-CT-Dx dataset increases from 0.82 and 0.928 to 0.905 and 0.942 respectively;The cooperative hospital dataset m AP increased from 0.952 to 0.987.3.Although the deep learning model greatly improves the upper limit of lung cancer detection accuracy,different equipment has a great impact on the reasoning speed of the model,and the equipment with fast reasoning speed is often more expensive.Therefore,this paper deploys the above trained model on Jetson Xavier NX,and accelerates the model with Tensor RT,which not only reduces the cost,but also maintains a high reasoning speed.In addition,in order to improve the ease of use of CAD system,a set of graphical visual interface is also written in this paper.