Segmentation Of Lung Parenchyma And Detection Of Lung Nodules Based On Feature Pyramid

Lung cancer is one of the diseases with the highest morbidity and mortality in the world,and it is extremely harmful to human health.Pulmonary nodules are important early clinical manifestations of lung cancer.Accurate detection of lung nodules is crucial during the prevention and treatment of early lung cancer.At present,the detection and screening of lung nodules is mainly done by means of computed tomography(CT).Due to the large amount of medical imaging data,doctors scrutinize the lung nodules by observing Computed Tomography images with the naked eye,which is not only time-consuming and wasteful but also easy to missing and misdiagnose.The computer-aided diagnosis system not only helps relieve the doctor’s reading pressure,but also helps improve the doctor’s clinical diagnosis efficiency and accuracy.At present,many deep learning-based methods have been used in automated aided diagnostic systems,and great progress has been made in the detection and identification of lung nodules.Traditional lung nodule detection algorithms usually rely on feature extraction.The detection process is complicated and has many steps.In recent years,with the development of deep learning method,convolutional neural networks have achieved excellent performance in the field of computer vision,which has brought new directions for the automatic detection and recognition of lung nodules.Till now,most methods based on convolutional networks only extract single-scale feature from convolutional networks,resulting in certain limitations in the recognition performance of the network.The feature pyramid network is reconstructed on the basis of the convolutional network,which can effectively integrates and extracts the multi-scale features of the image.Therefore,this thesis is to reconstruct the feature pyramid network based on the convolutional network,propose an automatic lung parenchyma segmentation extraction and lung nodule detection recognition through the feature pyramid network,and improve the sensitivity of the automatic detection of lung nodules quickly.The main research contents and innovations of this article are as follows:(1)Conducted a systematic study on conventional convolutional networks,summarized the advantages of convolutional networks,and analyzed the problems in the recognition tasks of convolutional networks in small-scale and multi-scale targets(especially lung nodules).At the same time,the traditional multi-scale feature extraction method and the convolution network-based multi-scale feature extraction method are introduced to provide a basis for the improvement research in the following.(2)Rebuild Feature Pyramid Networks(FPN)based on Residual Network(ResNet50),and design a semantic segmentation network combined with feature pyramid network to quickly and accurately segment and extract lung parenchyma.A comparative analysis was conducted to verify the effectiveness of the lung parenchyma segmentation network in this paper.Good segmentation of the lung parenchyma is conducive to the subsequent lung nodule detection network for more efficient and accurate detection.(3)A new network based on Faster R-CNN algorithm is presented.First,add a deconvolution layer on the basis of the traditional feature pyramid network to obtain more scale and higher resolution features,thereby improving the model’s ability to obtain small-scale and multi-scale target features.Then,the improved feature pyramid network is introduced into Faster R-CNN.This paper proposes a multi-scale prediction method to effectively improve the performance of Faster R-CNN network.Finally,to slove the problem that the positive and negative sample imbalances in the lung nodule data set,which often cause the model to be insensitive to real nodules,this paper uses the Focal Loss function to improve the sample imbalance based on the proposed region(Region Proposals Networks,RPN).In order to qualitatively and quantitatively evaluate the performance of the improved method proposed in this article,experimental analysis was implemented on the LUNA16 public dataset.The final experiment achieved a detection accuracy of 86.9%,which was 7.5 percentage points higher than the Faster R-CNN baseline based on the ResNet50 backbone network.It was also compared with other methods of lung nodule detection based on the LUNA16 dataset.The proposed method achieves a detection sensitivity of 95.7% when the average number of candidate nodules per scan is 40.6,which verifies the effectiveness of the proposed method.