Malignant-Benign Identification For Pulmonary Nodule With Semantic Features From Computed Tomography Images

Lung cancer is regarded as one of the most common malignant tumor both in incidence and mortality,which has threatened human health in the worldwide.Pulmonary nodule is potential manifestation of lung cancer and it has high probabilities to become malignant nodule.Early diagnosis of pulmonary nodule could improve survival rate of patients.At present,CAD(Computer Aided Diagnosis)system could diagnose pulmonary nodule according to low-level feature of lung images.But clinicans usually use semantic features.In order to eliminate this semantic gap between low-level features and semantic features,we map the low-level features of pulmonary nodule images to semantic features which we can better classify the benign or malignant of pulmonary nodule.In this thesis,we use computer technology,combined with pattern recognition and machine learning methods to construct CAD of pulmonary nodule as a “second reader”.The specific works and contributions are as follows:(1)We segment the region of interest(ROI)of pulmonary nodule using a semi-automated region growing algorithm.And then several low-level features are computed to represent the pulmonary nodule.In order to eliminate the influence of dimensions of different features on the classification,the extracted low-level features are normalized.(2)In order to improve the recognition rate when the data is unbalanced,we propose a new method named Improved Asymmetric Bagging(IM-ASYBagging).The algorithm can form a "complex" boundary between the majority and minority classes through the sampling mechanism so that the two groups of data can be well separated.(3)The doctor clinical diagnosis is based on advanced semantic attributes,and computer aided diagnosis is based on low-level feature.The low-level features can be understood by machine but it’s difficult for human,so it can’t be widely used in clinical settings.In order to overcome this shortcoming,we map the low-level features to semantic feature which are widely applied in clinic diagnoses.In addition,different low-level has different contributions for every semantic features,so feature selection is a significant pre-processing step before constructing predict model.(4)Lobulation is an important semantic feature which is widely used in clinical diagnosis.In order to better represent a pulmonary nodule,we have introduced three new features that are calculated by the minimum circumscribed circle and maximum inscribed circle.Experiment result has show that our proposed three new features based on nodule’s shape are effective for lobulation’s classification.