Research On Key Technologies Of Intelligent Analysis And Diagnosis Of Pulmonary CT Images

In the clinical imaging screening of lung cancer,the detection of pulmonary nodules in CT image is very crucial for the diagnosis of lung cancer,which can enhance the patient survival rate.However,the burden and fatigue of radiologists caused by the numerous CT images,heavily affect inspection results.The misinterpretations and detection errors of pulmonary nodules make the computer-aided detection(CAD)system very important and necessary to aid the radiologists.Due to the heterogeneity and similarity of the pulmonary nodules and the existence of a large number of similar tissues,the detection of the pulmonary nodule is still difficult and challenging.In this dissertation,based on the characteristics of CT image and pulmonary nodules,the investigation on key technologies of CAD is conducted after the full analysis of the CAD system.To improve the performance of CAD system,this dissertation devotes to three key technologies of the CAD system: 1)pulmonary parenchyma segmentation;2)nodule candidate detection;3)false positive reduction.Specifically,the main contents of this dissertation include the following aspects:(1)The introduction of the basic theory of CAD system for pulmonary nodules in CT images.This dissertation investigates the characteristics of CT imaging and pulmonary nodules,analyzes the key technologies,and evaluation methods and standards of CAD system.Based on the characteristics of CT images and pulmonary nodules,the limitations of key technologies of existing CAD systems are discussed.The investigation and analysis on characteristics of CT imaging and pulmonary nodules,provide a theoretical foundation for the following research of key technologies on CAD system,and for this dissertation as well.(2)Adaptive segmentation of pulmonary parenchyma in CT image based on morphology is proposed.Pulmonary parenchyma segmentation is a basic part of CAD system,which aims to concentrate the research on the lung area.The existing segmentation method is difficult to obtain a good trade-off between accuracy and generalization.According to the characteristics of CT images,distribution and morphological diversity of different pulmonary tissues are studied in this dissertation.A multi-stage method of lung parenchyma segmentation is proposed,combining the morphology,threshold,and connectivity analysis.The verification and comparison experiments of the clinical CT images indicate that the proposed method has high efficiency,high accuracy and preferable robustness.(3)A pulmonary nodule detection methodology using Teager-Kaiser main energy in generalized S-transform domain is proposed.Nodule candidate detection is a crucial part of CAD system,which determines the maximum performance of the whole CAD system.The existing methods utilize spatial or statistical features of nodules,nevertheless,the frequency-dependent features of nodules are not fully considered in traditional CAD systems.Therefore,in this dissertation,the time-frequency attributes of nodules are investigated by time-frequency analysis.The solid nodules are described by an energy attribute in time-frequency domain,which is applied to nodule candidate detection.Hence,a nodule detection method based on Teager-Kaiser main energy in generalized Stransform domain is proposed.By the analysis and comparison of experimental data,the attribute feature in time-frequency domain can effectively describe the solid nodules,and the proposed method has better detection performance.(4)A pulmonary nodule detection approach based on spectrum analysis using the optimal fractional S-transform(OFr ST)is proposed.Based on the excellent timefrequency resolution of the OFr ST,this dissertation uses it to explore the frequency components of the nodules.Then,according to the frequency components of nodules,a nodule detection method based on spectral analysis is proposed,which combines the energy attribute in time-frequency domain.Through the comparison and analysis of experimental results,it is verified that the difference of frequency components can effectively distinguish nodules and false-positive nodules and the proposed approach receives a good detection performance.(5)An attention-embedded complementary-stream CNN for false positive reduction is proposed.False positive reduction is the last key technology in CAD system,the nodules are detected after removing the false positives caused by nodule candidate detection.There are still two limitations in the existing CNN-based approaches: 1)the contributions of multi-scale inputs to the prediction are not fully considered;2)the significance of the different regions in the input is ignored.In this dissertation,an attention-embedded complementary-stream CNN is proposed,which utilizes the attention mechanism network to overcome the aforementioned two limitations.The attention mechanism is applied to guide the network to focus on the features of important areas and automatically weight multi-scale features.Therefore,false positive nodules could be suppressed by the proposed network.The results of experiments on open datasets and comparison of other methods,validate the feasibility and effectiveness of the proposed method and indicate the preferable performance of the proposed method in false positive reduction.In this dissertation,four methods based on the characteristics of CT images and pulmonary nodules are proposed for the corresponding three key technologies of CAD system.The experimental results and evaluation results on extensive datasets validate that the four methods not only can overcome the limitations of the existing CAD research,but also improve the performance of key technologies in the CAD system,and enhance the detection performance of CAD system for heterogeneous nodules in CT images.Moreover,the research on the key technologies of CAD system in this dissertation is also helpful for the clinical application.