Research On Detection Algorithm Of Lung Nodules Based On CT Images

Lung cancer is the second most common type of cancer in the world,and its five-year survival rate is only 19% due to the lack of early detection and precise methods.Therefore,it is of great significance to use some automatic detection technology in addition to the doctor’s diagnosis to assist the doctor in the diagnosis in order to detect lung cancer early and improve the survival rate of patients.In recent years,with the hot development of deep learning and the popularity of lung nodule detection research,more and more scholars use two-dimensional and three-dimensional deep learning methods to automatically detect lung nodules to assist in the early diagnosis of lung cancer.In this paper,considering the long training time of 3D deep learning methods and excessive network parameters,which are highly dependent on GPU computing resources,we will use 2D deep learning methods within limited computing resources and make full use of the 3D features of lung CT images to make the two-dimensional deep learning method can reach the lung nodule detection performance of the three-dimensional deep learning method under the condition of shorter training time and fewer parameters.The main research results of this article are as follows:First,analyze the characteristics of the size distribution of nodules in lung CT images.Because lung nodules are similar to circular targets that can be framed by squares on two-dimensional images and are small targets.The Faster-R-CNN network is used as the backbone network,and a more suitable anchor frame is obtained by modifying the ratio of generated anchor frames in the network.At the same time,a deconvolution layer is added after the convolution layer to restore some shallow features to expand the feature map,and finally enhance The network’s ability to detect lung nodules has increased from 67.71% before modification to 82.94% after modification.Finally,the detection performance bottleneck of the current network is analyzed,and a two-stage lung nodule target detection framework is proposed.Then,on the basis of the two-stage lung nodule target detection framework,a specific network was designed in order to achieve the two-stage detection task.First,in order to reduce false positives in the first stage,a U-net segmentation network is used for mask image generation.In the second stage,since the feasibility of the single-stream Faster R-CNN network has been verified,a two-stream Faster R-CNN network structure is proposed for the backbone network for lung nodule detection.While inputting the original lung image,the network provides the corresponding image containing the mask of the suspected lung nodule generated by the U-net network.The original data set,the dilated data set,and the simulated data set were used for verification,and CPMs of 83.78%,86.3%,and 91.18% were obtained,which verified the performance of the two-stream Faster R-CNN network.Finally,in order to make the entire detection process of lung nodules end-to-end,there is no need to manually provide mask image information.The U-net network and the two-stream Faster R-CNN network are combined to build an end-to-end two-stream Faster R-CNN network for lung nodule detection.In the end,the experiment proved that the network can complete end-to-end network detection,the detection performance is 10 pictures per second,and the CPM is 85.65%.