Detection Of Pulmonary Nodules In CT Images Based On Deep Convolution

Lung cancer is one of the cancers with the highest morbidity and mortality in China.The early lesions appear in the form of pulmonary nodules,so the early detection of lung cancer is the detection of pulmonary nodules.In actual medical work,the number of radiologists and the number of image data vary greatly,which aggravates the work burden of radiologists.Meanwhile,the pulmonary nodules in CT images are often easily missed by naked eye.Therefore,it is of great significance to apply computer technology to assist doctors to detect and identify pulmonary nodules and provide a second opinion for doctors’ diagnosis.With the rapid development of deep learning and the emergence of multiple publicly available medical data sets,the development of artificial intelligence in the medical field has been effectively promoted.Based on deep learning algorithms,this paper proposes two models of lung nodule detection in CT images based on deep convolution:A deep 2D convolution based on CT image lung nodule detection model is proposed,which adopts the idea of transfer learning to transfer the object detection model based on natural image to the lung nodule detection task.Using VGG16 as the backbone network of the model.Importantly,for the first time,a deformable convolution module has been introduced into the lung nodule detection task.By learning the input image,the sampling point can be moved appropriately,extending to a non-grid shape,so that the convolution window tends to be around the object.This reduces the impact of irregular pulmonary nodules and blurring of the margins between nodules and parenchyma on feature extraction.At the same time,multi-level feature fusion pyramid module was added to adapt to the small object pulmonary nodules,and the effect of small object on feature extraction was reduced by deconvolution fusion of feature map.Compared with other methods based on twodimensional convolutional neural network,this method has obvious improvement,with an average accuracy of 82.7%.2D CNN has good feature extraction ability for images.However,3d CT images were sliced to obtain cross-sectional images,which were treated as separate nodules and destroyed the correlation between the images.Meanwhile,if 3D CNN is adopted,more complex operation time and more storage space are needed due to the addition of one dimension.Based on these two characteristics,a deep mixed convolution based on CT image lung nodule detection model is proposed,and a mixed convolution rule is designed and implemented.Use3 D CNN on shallow network,the same as the input image data formats,can extract the spatial characteristics of pulmonary nodules.Use 2D CNN on deep network,the feature of pulmonary nodules was extracted from the perspective of 2D convolution,which strengthened the feature learning of the single slice image.At the same time,the deconvolution fusion module verified in the previous stage is expanded into a threedimensional form to further improve the detection accuracy of the model.Experiments on the LUNA16 data set showed that the proposed method had a sensitivity of 92.4% with an average of 8 false positives per scan.The two models presented in this paper have achieved good performance on the public data set,and the influence of different local network modules on feature extraction is designed and discussed,which is of great reference significance and practical value for the research of lung nodule detection.