Research On Detection Methods Of Lung Common CT Imaging Signs

Lung cancer is a leading type of cancers threatening human health,which has the largest number of patients and the highest mortality rate.The mortality of lung cancer can be reduced significantly if the cases are detected and treated earlier.Computer-aided detection of lung lesion regions has practical significance for improving the diagnostic accuracy of early lung disease.Radiologists focus primarily on “CT imaging signs”,a fine particle size,in clinical practice.Therefore,computer aided detection of lung lesion based on CT imaging signs is consistent with radiologist's experience.For these reasons,computer aided detection of GGO imaging signs in lung nodule has received more attention in recent years.Howerver,many other CT imaging signs inside or outside the range of lung nodule has not received particular attention.We review and classify the achieved progress of automatic detection methods of lesion regions in lung CT images,then we analyze the existed problems and challenges ahead.And on this basis we study the computer aided detection technology for lobulation,GGO and cavity imaging signs.The main research contents and innovations of this thesis include the following four parts.(1)This thesis analyzes pros and cons of the exsiting lung medical image databases,designs and constructs a publicly available database of lung common CT imaging signs.Data is the basics of researches and experiments,so the constructed database may help to accelerate detection technology researches of lung lesion regions based on CT imaging signs.At the same time,the database can assist radiologist to make decisions,and radiologists may retrieve CT images with corresbonding annotated information similar to what they are reading.Furthermore,the database may bring convenience to young doctors training.(2)We propose a lobulation detection method based on local features and bending energy.The method carries out lobulation detection directely unter the sliding-window framework,and avoids the negative effects of nodule(or mass)incomplete segmentation.Based on the essential curving characteristic of lobulation contour,we apply the bending energy method to lobulation detection and the experiment result shows that bending energy is high-efficiency for false positives elimination.We find that the global information and shape description ability are very important for lobulation detection,and also find the limitation of detection only rely on local features.Therefore,it is indispensable to utilize more information in lobulation detection work,such as tissue distribution and statistical distribution information of lobulation signs.(3)This thesis proposes a computer aided detection method of 3D GGO imaging signs by deep learning technology based on hybrid resampling and layer-wise fine-tuning scheme.According to the 3D shape characteristic of GGO signs,our hybrid resampling is performed on multi-views and multi-receptive fields,which reducing the risk of missing small or large GGOs by adopting representative sampling panels and processing GGOs with multiple scales simultaneously.We incorporate layer-wise fine-tuning scheme into multi-model fusion architecture and offer a practical way to reach the best finetune experiment.The former has ability to obtain the optimal fine-tuning model and the latter obtains better performance than any single trained model.(4)This thesis proposes an automatic detection method of cavity imaging signs in lung CT images through hybrid resampling and multi-feature fusion strategies.Our hybrid resampling is based on multi-receptive-field and multi-window settings,which processes cavities with multiple scales simultaneously and reserves context information of multiply CT windows more compactly.For multi-feature fusion,the deep CNN feature and classical feature(HOG and LBP)are combined to improve classification performance,and the experimental results show that fusion feature has better discriminative capability than any single feature.