Intelligent Detection And Auxiliary Diagnosis Of Pulmonary Nodules Based On Deep Learning

Lung cancer is one of the main causes of death in the world,which seriously affects the safety of human life.Early diagnosis and timely treatment of lung cancer can effectively improve the survival rate of lung cancer patients.Effective detection of pulmonary nodules and timely and accurate judgment of their benign or malignant in chest computed tomography(Computed Tomography,CT)is the key to early diagnosis of lung cancer.With the rapid development of artificial intelligence and the urgent needs of medical computer,the computer-aided diagnosis technology based deep learning has the advantages of end-to-end,self-adaptability and high detection rate,which can achieve rapid screening of pulmonary nodules.However,there are still thorn problems such as low detection sensitivity,high false positive rate and low classification accuracy.Therefore,this dissertation carried out in-depth research in three aspects: candidate nodule detection,false positive reduction and benign and malignant classification of pulmonary nodules.The specific research work is as follows:(1)Candidate nodule detection network based on improved residual U-Net is designed.Due to the heterogeneous,complex in shape and varied in size of pulmonary nodules in CT images,it often confused with blood vessels and other tissues.It is particularly important to extract rich and representative pulmonary nodules features.Therefore,3D context-guided module containing low-dimensional,high-dimensional local information and spatial context information is designed by using the dilated convolution of different dilation rates to extract rich 3D spatial features of pulmonary nodules based on 3D Residual U-Net.Considering the extraction of feature contains redundant information,channel attention mechanism is adopted to dynamically adjust the channel characteristics of the feature map.It can suppress redundant features and strengthen key features to further enhance the generalization performance of the network.Experimental results show that the proposed algorithm improves the detection sensitivity of network within the acceptable false positive rate.(2)Multi-branch false positive reduction network for multi-task learning is designed.After the candidate nodule detection,many false positive nodules will be appeared,so it is necessary to further reduce the false positive.Considering that the microscopic details of pulmonary nodules are very important and easy to lose in medical images,a multi-branch classification network for multi-task learning is designed.Multi-task includes image classification task and reconstruction task.The image reconstruction task is designed to assist the classification task,and the main purpose is to recover more microscopic nodule feature information from the CNN level.Multi-branch is designed to utilize the feature maps containing different receptive fields at different depths in the fine-tuned feature extraction network to perform multi-scale fusion and improve the ability to identify nodules of different scales.Experimental results show that the proposed algorithm can effectively reduce the false positive rate of the intelligent diagnosis system for pulmonary nodules,and greatly improve the classification accuracy of pulmonary nodules.(3)Multi-branch benign and malignant classification network based on multimode fusion is designed.Considering that the existing methods only improve the classification performance through model improvement in the diagnosis of benign and malignant of pulmonary nodules,but the clinical data,such as laboratory examination and radiology data is not considered enough to judge of patients’ condition.Therefore,a multi-mode fusion multi-branch classification network is proposed in this dissertation.The effective channel attention mechanism network based on multi-branch fusion is proposed to extract features from 3D CT Patch of unstructured data of nodules.The feature maps of different levels with different receptive fields are utilized to fuse and obtain features of multi-scale unstructured pulmonary nodules.At the same time,3D ECA-Res module is used to effectively adjust the extracted features adaptively and dynamically to make them more discriminant.Then,the clinical radiological data features(internal structure,calcification,etc.)of pulmonary nodules are used to construct radiological structural modal feature vectors.Finally,multi-modal fusion of structured and unstructured features is used to distinguish benign and malignant pulmonary nodules.Experimental results show that the proposed multi-branch classification network of multi-modal fusion has good generalization performance and greatly improves the classification accuracy.In conclusion,this dissertation proposes an intelligent pulmonary nodule diagnosis algorithm with excellent performance,which effectively improves the sensitivity and accuracy of pulmonary nodule diagnosis based on the LUAN16 and LIDC-IDRI database.The research work provides an effective auxiliary means for early screening of lung cancer and greatly relives the pressure faced by radiologists.