Research On Algorithm Of Segmentation And Benign-Malignant Classification Of Pulmonary Nodules Based On Deep Convolutional Network

Lung cancer is the leading cause of the cancer-related deaths,and the number of new lung cancer patients every year remains high.The screening,diagnosis and treatment of early lung cancer can effectively improve the five-year survival rate of lung cancer patients.Computed Tomography(CT)is the most widely used technique for early lung cancer screening.Early lung cancer symptoms are not obvious and usually manifest as lung nodules.Therefore,in order to improve the survival rate of lung cancer patients,lung nodules segmentation and benign-malignant classification based on lung CT images is of great significance for the early screening of lung cancer.With the increasing number of patients with lung cancer,the number of CT images has also risen sharply,which undoubtedly poses a challenge to radiologists.Due to the increasing workload,radiologists are more probably to misdiagnosis and missed diagnosis when judging lung CT images.Therefore,the use of computer-aided diagnosis(CAD)systems to provide radiologists with reference opinions and assist doctors in the diagnosis of lung nodules can effectively reduce the rate of misdiagnosis and missed diagnosis,and help improve the survival rate of lung cancer patients.In recent years,with the continuous development of deep learning,convolutional neural network(CNN)has achieved good performance in medical image processing tasks due to its strong feature learning ability and representation ability.Based on this,a CNN-based method is proposed to segment and classify benign-malignant lung nodules.The main work of this thesis as follow:(1)The significance and importance of CAD system for early lung cancer screening were introduced.Methods about lung nodules segmentation and the benign-malignant classification were fully investigated and studied.At the same time,the methods of medical image segmentation and classification were summarized.(2)This thesis propose a lung nodule segmentation method,which based on dual encoding fusion.In order to solve the problems of high granularity in CT images and lung nodules usually with small size,irregular shape,and there are atypical nodules such as cavity nodules and ground glass nodules in lung CT images,Dual Encoding Fusion Network(DEF-NET)is proposed to segment lung nodules.Specifically,to avoid the influence of high granularity in CT image,high-level smoothing representation was generated by pyramid-upsampling.Secondly,different dilated residual blocks(DR-Block)are designed and embedded into different encoders for different levels of representations to extract the features with different characteristics.Finally,a U-Net like decoder is proposed to fuse the extracted features to obtain the final segmentation results.Comparison experiments and ablation studies were performed on LIDC-IDRI dataset,which demonstrated the effectiveness of the proposed methods,and the segmentation accuracy reached 86.12%(3)This thesis propose a benign-malignant lung nodule classification method,which based on Hyper-Attention Guided by Deep Feature Network.For the problem that lung tissues are similar to lung nodules,blood vessels and lung nodules may form pull,will import interference to the model.And results of benign-malignant lung nodule classification from current methods based on the CNN are not interpretable,those methods are difficult to use in clinical application,Hyper-Attention Guided by Deep Feature Network(HGDF-Net)is proposed to classify benign-malignant lung nodules.Firstly,Hyper Attention Mechanism(HAM)is proposed to avoid the interference of the model caused useless information such as tissues and blood vessels in CT images.Secondly,Semantic Learning(SL)is proposed to classify the attribute characteristics while classifying the benign-malignant lung nodules simultaneously to make the final results interpretable.Experimental results demonstrate that the proposed HGDF-Net can not only obtain interpretable classification results,but also achieve better performance of benign-malignant lung nodule classification,with a classification accuracy of93.52%.