Research On Segmentation And Classification Methods Of Lung Nodules From CT Images Based On Radiomics And Machine Learning

Lung cancer is one of the most threatening malignant tumors to human life and health.Moreover,it has the highest morbidity and mortality rate among all kinds of cancers.Early detection and diagnosis of lung cancer is pivotal to improving the survival rate of patients.Early manifestation of lung cancer is lung nodule.Therefore,early detection,accurate segmentation and classification of lung nodules are very critical to assist doctors in better diagnosis and treatment for lung cancer.As a noninvasive method,computed tomography(CT)has the advantages of rapid acquisition,highest sensitivity,low cost and wide availability,and is a commonimaging modality for the detection and diagnosis of lung nodules.Lung CT image is an important application of particle physics and nuclear physics in the field of medical image research.It can provide doctors with the physiological structure,functional state and pathological information of lung tissues for accurate diagnosis.However,with the acceleration of CT acquisition speed and the increase of the amount of data collected,CT images of a patient can reach hundreds.Therefore,reading images manually has become a heavy burden for radiologists,which is easy to cause the fatigue of reading,resulting in false negative and false positive.In order to reduce the workload of doctors and provide them with objective marks of suspicious nodule candidates,computer-aided detection(CAD)systems have been developed to help doctors accurately and quickly diagnose patients’ conditions.Since lung nodules from CT images can have different shapes,intensities,contours or locations and may be attached to other tissues,such as neighboring blood vessels or pleural surface,accurate detection and segmentation of different types of lung nodules are still complex and challenging in the research field of CAD system.In addition,some non-nodules(such as false positive nodule candidates)have similar intensities and morphological appearances to nodules.It is difficult for radiologists to classify all suspicious nodule candidates accurately,which may lead to misjudgment.Hence,accurate classification of nodules and non-nodules is very essential and crucial in CAD system.Centering on the core issues of segmentation and classification of lung nodules in CAD system,segmentation and classification of lung nodules from CT images based on radiomics and machine learning are deeply studied in this thesis.The research content is mainly divided into the following four parts:1.Parsed diagnostic annotations in XML files,read information from DICOM files and processed image data in the database.Diagnostic annotations and image data are the premise and key of the research on segmentation and classification of lung nodules from CT images.For XML and DICOM files in the public LIDC/IDRI database,this paper obtained slice locations and diagnostic annotations of lung nodules labeled by at least three radiologists by parsing XML files.According to slice locations,relevant information and CT image data in corresponding DICOM files could be read.Besides,image data could be processed.Moreover,positions and contours of lung nodules were determined according to diagnostic annotations.These work laid a foundation for the subsequent research on segmentation and classification methods of lung nodules from CT images.2.Established the segmentation method based on Otsu thresholding and the α-hull contour correction algorithm.In order to segment different types of lung nodules from CT images,avoid the problems of under-repair and over-repair in lung contour correction effectively,and improve the segmentation performance of different types of lung nodules,this paper improved the α-hull algorithm by comparing Hausdorff distance(HSD)and the area of lung before and after lung contour correction.For different types of lung nodules,the results manifested that the α-hull contour correction algorithm could obtain optimal a values adaptively,thus various lung contours were corrected effectively.In addition,a variety of metrics were used to evaluate the segmentation performance quantitatively,which could not only evaluate the segmentation performance from multiple dimensions,but also comprehensively analyze the causes of the segmentation results.Moreover,this segmentation method achieved good performance in Jaccard index(JI)and Dice similar coefficient(DSC).3.Studied the segmentation method based on improved U-Net network in depth to increase the segmentation performance.On the basis of the α-hull contour correction algorithm,batch normalization(BN)was introduced to improve U-Net network so that the segmentation performance could be improved effectively.DSC was higher than those of the currently widely used segmentation methods.Compared with the segmentation method based on Otsu thresholding and the α-hull contour correction algorithm,the segmentation method based on improved U-Net network had better segmentation effect and performance.The results of these segmentation methods provided an effective way to assist radiologists to segment lung nodules accurately.4.Developed two methods to study the classification of nodules and non-nodules,including the classification method based on radiomics and the classification method based on DenseNet convolutional network.In the former method,555 radiomics features were calculated.Then 3 feature selection methods with different numbers of features were applied for feature dimension reduction.Finally,based on different ratios between training dataset and testing dataset,12 machine learning based models weretrained.The results showed that when the feature selection method was recursive feature elimination(RFE),the number of features was 15 and the ratio was 9:1,the classification performance of random forest was the best.In the latter method,DenseNet convolutional network was used to classify nodules and non-nodules.The results manifested that DenseNet network had good classification performance,which also verified the feasibility of this network in the classification problem of nodules andnon-nodules.In the end,these two classification methods were compared.Both methods had good performance for the classification of nodules and non-nodules.These results had certain application value for assisting radiologists to accurately classify nodules and non-nodules.In summary,a series of method improvements and application studies were carried out in this paper in view of the existing problems in the segmentation and classification of lung nodules from CT images,so that they would be helpful to assist radiologists to segment and classify lung nodules accurately.