Pulmonary Nodules Detection And Diagnosis From CT Scans Based On Deep Learning

Lung cancer is the main cause of cancer death worldwide,which seriously endangers human health.Pulmonary nodules are the early manifestation of lung cancer.The early detection of pulmonary nodules plays a vital role in improving the survival rate of patients.Computed Tomography(CT)is an effective method for the diagnosis of lung cancer because of its high scanning speed,high image definition and the ability to capture small lesion areas.Computer Aided Diagnosis(CAD)technology is developed to improve the Diagnosis accuracy of lung cancer.However,pulmonary nodules vary in size and shape,and contain many similar tissues and organs around them,leading to the problem of missed detection and false detection in existing detection algorithms.Nowadays,deep network has achieved excellent results in the field of benign and malignant classification of pulmonary nodules.In view of the above problems,we use deep neural network to perform algorithm research and system prototype design of pulmonary nodules detection and diagnosis based on CT images.The specific works are as follows:(1)For the problem that the existing technology has a high rate of missed and wrong detection of pulmonary nodules,we propose a new algorithm ADR-CNN based on Faster R-CNN.By replacing the feature extraction network of Faster R-CNN with the InceptionRes Net-V2,features that are more conducive to the detection of pulmonary nodules can be extracted.The size and number of anchor points in the RPN network are modified to solve the problem of smaller size of pulmonary1 nodules.The experimental results show that the average sensitivity of this algorithm is 0.86,which is 0.013 higher than that of the state-of-art methods,reduces the detection of false positive nodules and improves the detection accuracy.(2)Referring to the habit of radiologists in the diagnosis of pulmonary nodules,a new algorithm Ms Mi-Dense Net is proposed to improve Dense Net.Because the tissues around pulmonary nodules affect the classification of benign and malignant,the data of multi-window-width and window-level nodules of two sizes are used,and the prior characteristics of pulmonary nodules are added to the Ms Mi-Dense Net to train the classification model.The experimental results showed that the sensitivity,specificity and accuracy of the network are 96.65%,92.93% and 94.17%,respectively,and the ROC curve area is 0.9820.Compared with the state-of-art methods,the proposed framework improves the ROC area by 0.01 and the interpretability of deep networks' diagnosis results.(3)Finally,the prototype of pulmonary nodule detection and diagnosis system is designed based on the pulmonary nodule detection and diagnosis model,which can provide effective help for doctors in clinical diagnosis of lung cancer.This thesis studies the improved algorithm of detection and diagnosis of pulmonary nodules,which improves the accuracy of detection and diagnosis of pulmonary nodules.The prototype of computer aided diagnosis system is designed,which is beneficial to the popularization of CAD system in clinical medicine.