Design And Implementation Of Computer Aided Diagnosis System For Lung Nodules Based On CT Images

Numerous kinds of cancers have become one of the leading causes of humandeath, among which lung cancer is the most deadly type for its high incidence andhigh mortality. Early detection of lung cancer is the best way to treat lung cancer,while CT is one of the best methods for inspection of lung cancer. As the accuracyof CT technology becomes increasingly higher, resulting images with betterresolutions, the amount of data is greatly increased, making doctors overwhelmed.The fast and tireless computers thus become an effective tool to help doctors fordiagnosing. This dissertation focused on the research of lung nodules detectiontechnology. A computer-aided diagnosis system for lung nodules was designed andimplemented.As benign shadow and malignant shadow have different types with varied sizeand shape, both of them have low degree of similarity within each class if thetraining samples are divided only into two major categories, benign and malignant.In order to solve this problem and the potential defect in the Euclidean distanceused by many classifiers, a new classifier connected Fuzzy-c means algorithm andMahalanobis distance is presented. This classifier improves the classificationaccuracy by further subdividing the two categories, benign and malignant, toincrease the similarity of the subclasses and to calculate the similarity of thesamples with subclasses, by using Mahalanobis distance instead of Euclideandistance.The package of ITK and VTK tools was used to build a system platform withMFC. The whole system was consist with seven modules which were image reading,lung region segmentation, lung suspected shadow detection, discrimination ofsuspected shadow, image displaying, image tagging and interaction by researchingthe process of lung nodules diagnosis. All of the seven modules were designed andimplemented.The image reading module, image display module, image tagging andinteraction module were implemented with the package of ITK and VTK tools. Thelung region segmentation module was implemented with optimal threshold algorithm, regional connectivity labeling algorithm, hole filling algorithm andinpainting algorithm. The variable quoit filter was used in the lung suspectedshadow detection modules. At last, the benign shadow and malignant shadow wereclassified using the new classifier. The system has good sensitivity by testing withthe CT images of real patients.