Research On Key Technologies For Pulmonary Nodule Diagnosis Based On Deep Learning

Lung cancer is one of main cancers faced by human beings,and pulmonary nodules are the main manifestations of lung cancer in the early stage.Therefore,accurate diagnosis of pulmonary nodules is of great significance for the prevention and treatment of lung cancer.With the development of computer image processing technology,Computer-Aided Diagnosis(CAD)system has gradually become the main tool to assist doctors in effectively diagnosing pulmonary nodules.By simulating the diagnosis method of doctors,the CAD system often diagnoses pulmonary nodules according to the following steps: extract the lung parenchyma that may contain the lesions from the CT image firstly,detect the pulmonary nodules in the lung parenchyma,then use the local CT image containing the pulmonary nodule as input to segment the pulmonary nodule for obtaining its boundary positions.Finally,based on the segmentation result of the pulmonary nodule,extract the morphological features to predict benign and malignant.Based on deep learning technology,this paper explores the problems existing in the key links of the diagnosis process,and conducts in-depth research work.The main results of this dissertation are listed as follows:(1)In order to obtain accurate lung parenchyma regions from CT images,this paper proposes a boundary awareness-based lung parenchyma segmentation method.The method is divided into two stages: in the first stage,the segmentation model is trained with dataset after data augmentation and filling the outer region of the thoracic cavity to improve the awareness to the boundaries of the lung parenchyma,and then obtain segmentation results that fit the lung wall.In the second stage,the improved convex hull method was used to repair the pulmonary wall adhesion nodules missed in the segmentation results,while ensuring the accuracy of the segmentation results in other regions.The comparison experiments with other classical lung parenchyma segmentation methods show that the method in this paper has better lung parenchyma segmentation effect because it reduces the influence of surrounding substances such as trachea and bronchi on lung parenchyma segmentation.(2)In order to improve the detection performance of pulmonary nodules,a pulmonary nodule detection method based on candidate nodule detection and false positive reduction is proposed.Firstly,3D Faster R-CNN is used to detect candidate nodules above 3mm.Secondly,a multi-circle detection method based on heuristic optimization algorithm is proposed to detect tiny pulmonary nodules below 3mm.Finally,in order to improve the efficiency of false positive reduction and alleviate overfitting.Based on the principle of unsupervised anomaly detection,a generative adversarial network is constructed,in which the discriminator acts as a candidate nodule classifier,and the generator amplifies the differences between nodules and false positives by reconstructing the candidate nodules to support the discriminator for better classification.On this basis,the discriminator is supervised fine-tuned with reconstructed candidate nodules for final false positive reduction.Comparative experiments with other representative pulmonary nodule detection methods show that the proposed detection method has high detection sensitivity and low false positive rate.(3)Aiming at the morphological features of pulmonary nodules such as low contrast,blurred boundary,and adhesion to other tissues,a pulmonary nodule segmentation method based on object attention consistency is proposed.Firstly,based on background-change semantic data augmentation,the pulmonary nodules in the initial dataset are pasted into different regions to generate a set of images containing the same pulmonary nodules,and the learning effect is improved by simulating pulmonary nodule segmentation in different environments.Secondly,in order to improve the activation of the pulmonary nodule region,an image-level classifier is introduced into the segmentation model,and the nodule is used as classification attribute for classification training.Finally,an object attention consistency is proposed,which considers the same pulmonary nodule in different environments should have the same attention and be constrained by a consistency loss function.Compared with other classical pulmonary nodule segmentation methods,the proposed method achieves higher Dice similarity coefficient,sensitivity and precision due to enhancing the attention of pulmonary nodules with low contrast,blurred boundaries and adhesion to other tissues.(4)To improve the prediction effect for benign and malignant pulmonary nodules,a method for predicting benign and malignant pulmonary nodules based on hierarchical relationship is proposed.Firstly,the pulmonary nodule detection result is used as the prediction input,and the pulmonary nodule region is enlarged by synthesizing class response maps,extracting sparse attention,and resampling input image to achieve the prediction effect based on nodule features and combining location information.Secondly,in order to reduce the prediction error,the degree of benign and malignant is refined into five grades,and the ranking loss of five types of prediction probability is proposed to make the prediction result close to the real class.In order to reduce the influence of surrounding tissue on pulmonary nodule feature extraction,a prediction model is trained with pulmonary nodule segmentation results as input,and the prediction results are output by extracting accurate morphological features of pulmonary nodules.Finally,the mean of the predictions of the two prediction models is taken as the final prediction result.Comparison experiments with other classical benign and malignant prediction methods show that the proposed method has high sensitivity,specificity and accuracy.(5)Based on the proposed key technology of diagnosing pulmonary nodule by computer,a pulmonary nodule CAD system was built using Py Qt5.The whole system runs according to the diagnostic process including lung parenchyma segmentation,pulmonary nodule detection,pulmonary nodule segmentation,and benign and malignant prediction,which not only supports one-click diagnosis by users with CT data as input,but also allows viewing of the diagnostic process through step-by-step execution.At the same time,the system provides users with convenient diagnostic functions and rich visualization effects.Through the function and performance test of the system and the comparison with the expert diagnosis level,the results show that the system has a fast and robust diagnosis process and a good diagnosis effect,and can be well used in the auxiliary diagnosis of pulmonary nodules.