Research On Key Techniques Of Pulmonary Nodule Detection Based On Feature Fusion

In recent years,due to environmental problems and food safety reasons,the incidence and mortality of lung cancer in China are increasing rapidly.The early symptom of lung cancer is often manifested as lung nodules.If the lung nodules are found in the early stage,and the appropriate treatment is applied,the patients can generally recover.Therefore,lung nodule detection is currently a hot research topic in the field of medical image processing.The current lung nodule detection is facing some difficulties.Firstly,it may cause image distortion,artifacts and noise in the process of filming and storage for medical images,which will affect the detection accuracy of the nodules.Secondly,the blood vessels and other tissues in the lung CT images are similar to nodules.The presence of these tissues increases the difficulty for the detection of nodules.Thirdly,the traditional algorithm for the detection of pulmonary nodules has high false positive rate for pulmonary parenchyma segmentation and it is possible to lose the pulmonary nodules in the segmentation results.Finally,it is difficult for the traditional algorithm to adapt to different morphologies of parenchyma in the stage of detection of nodule,so it is urgent to study new algorithm to solve these problems.To solve these problems,the key technologies of lung nodule detection are studied in this paper.(1)In the traditional GrabCut algorithm,it is necessary to select the target area manually.In this thesis,we propose an improved GrabCut algorithm based on target object localization for lung parenchyma segmentation.The proposed algorithm can adapt to different morphologies of lung parenchyma.While ensuring accuracy,the recall rate is increased by 6.8% and 3.3% respectively compared with threshold-based segmentation and region growing based segmentation,which effectively improves the accuracy of the segmentation.(2)This thesis proposed a suspected lung nodule detection algorithm based on U-Net and fully connected conditional random fields.In the network,we use focal loss as the loss function to overcome the disadvantage of classical U-Net in small object segmentation.The experimental results show that the dice coefficient in the network with focal loss is increased by 37.5% and 30.9% compared to the cross entropy loss function and the dice loss function.In order to refine the extraction result,we use the fully connected conditional random field to optimize the segmentation result based on U-Net.(3)This thesis proposes an algorithm which combines traditional handcraft features and deep learning features.Compared with the algorithm of pure convolution neural network,random forest and support vector machine,the accuracy of the algorithm is increased by 1.1%,4.6% and 7.1% respectively.In this thesis,we combine handcraft features and deep learning features to improve the accuracy of classification.The above algorithms in this thesis have a certain theoretical and practical reference value for the further research of lung nodule detection.