Research On The Key Technologies Of Computer Aided Diagnosis For Lung Cancer In Computed Tomography Images

According to the investigations of many cancer research centers and health organizations,lung cancer has been the leading cancer with the highest death rate in the world.Earlier Detecting,diagnosis,and therapy are the most efficient methods for raising the living rate of lung cancer.High resolution CT is used in clinical practice widely,and becomes the first choice for the diagnosis of lung cancer by imaging.But,reading vast amount CT data increases the doctor's burden.Computer-aided Detection(CADe)for pulmonary nodules using CT imaging can be used as “second opinion” to reduce the reading time and improve the correction of diagnosis.Once detecting pulmonary nodules,distinguishing the benign pulmonary nodules from malignant nodules can be used to reduce the pain of patients from lung biopsy and save medical cost.In this dissertation,the research work is focusing on feature extracting and classifying of the lung cancer CAD system.The research works are as the following:(1)The LIDC-IDRI is the largest public database of pulmonary nodule imaging in the world,which supports four radiologists' annotations about the location of nodules in CT imaging.Aiming to the variability of the annotations in LIDC-IDRI,the dissertation study the method of creating “gold standard” based on multiple expert annotations.An improved STAPLE algorithm is developed to deal with the imbalance between the nodule region and the background region in CT imaging,which lead to larger probability value of the creating gold standard.The improved STAPLE algorithm creates a balance data using data undersampling method,and then uses the STAPLE algorithm to create gold standard based on multiple expert annotations.(2)For extracting the nodule image feature to reduce the false positive nodules,a novel descriptor for the characterization of pulmonary nodules in CT imaging is proposed,named histogram of oriented surface normal(HoSN).The HoSN creates a window adaptively based on the center of a candidate nodule firstly,and then creates the histogram of surface normal orientation in the window.The HoSN feature makes nodule characterization independent of accurate nodule segmentation.We validate the HoSN using the data from LIDC-IDRI,and we compare it with state of the art approaches for 3-D shape description in medical imaging and computer vision,namely 3-D SIFTS.Experimental results show that the sensitivity is 97.2%,and the false positive rate is 6.45FPs/Scan using HoSN features of the confidant nodules extracted by multiple scale dot filters.(3)The dissertation studies the nodule semantic feature extraction method based on classifiers.Aiming to the variability of expert annotations of the LIDC-IDRI,a nodule semantic feature extracting method,called semi-supervised co-forest,is developed.The method marks the nodules with consistent annotations as marked sample and the ones with inconsistent annotations as unmarked samples,and then it trains random forest to build a mapping model from module image features to semantic features using marked samples.The results show that the semi-supervised co-forest method can improve the accuracy of model compared to decision tree,and random forest only using the marked samples.(4)To reduce the false positive nodules,a weighted border synthetic minority over-sampling technique(WBSMOTE)algorithm is presented.The WBSMOTE algorithm calculates the weights of minority samples according to their Euclidean distance from the nearest majority class samples,the density of them,and the density of their neighbors of majority samples.The higher weight samples have the more chance to be chosen as the seed samples to synthetic new samples.The results based on the reducing false positive nodules show that the WBSMOTE algorithm is better than or comparable with some other existing methods in terms of various assessment metrics,such as G-mean,area under curve,the sensitivity and the false positive rate.(5)The nodule features with different type of resources usually are used to distinguish the benign pulmonary nodules from malignant nodules.These nodule features include morphology,geometric,density,texture,which depict pulmonary nodule on various faces.To solve the difficulty of making the classification decision with the same classifier on the complicated heterogeneous data,a novel SVM ensemble algorithm based on grouped feature is proposed to distinct the benign nodules from malignancy nodules.Fist,various types of nodule features were extracted based on the segmentation of nodule including the morphology,geometric,density,the two-dimensional Haralick texture features,and the Gabor texture features based on the nodules;gray images.The features were grouped by their different type of resource,and feature subsets were randomly selected from each group to train base classifiers;then,based classifiers subsets were selected in according with the strategy of improving the accuracy and diversity of the base classifiers;finally,integration of the selected based classifiers under the framework of ensemble learning of weighted voting.The proposed method takes advantage of the difference and complementarities between the diversified features.The experimental results on LIDC-IDRI demonstrated that the method outperforms single SVM classifier,and ensemble learning classifiers without grouping features.