Lung cancer is the leading cause of cancer death in the United States. The five-year survival rate is only 15% because most patients present with advanced disease. If lung cancer is detected and treated at its earliest stage, the five-year survival rate has been reported to be as high as 92%. Computed tomography (CT) has been shown to be more sensitive than chest radiography in detecting abnormal lung lesions (nodules), especially when they are small. However, each thin-slice thoracic CT scan can contain hundreds of images. Large amounts of image data, radiologist fatigue, and diagnostic uncertainty may lead to missed cancers or unnecessary biopsies. We address these issues by developing a computer-aided diagnosis (CAD) system that would serve as a second reader for radiologists by analyzing nodules and providing a malignancy estimate using computer vision and machine learning techniques. To segment the nodules, we extended the active contour (AC) model to 3D by adding new energy terms. The classification accuracy, quantified by the area ( A_z) under the receiver operating characteristic curve, was used as the figure-of-merit to guide segmentation parameter optimization. The effect of CT acquisition parameters on 3DAC segmentation was systematically studied by imaging simulated nodules in chest phantoms. We conducted simulation studies to compare the relative performance of feature selection and classification methods and to examine the bias and variance introduced due to limited training sample sizes. We also designed new feature descriptors to describe the nodule surface, which were combined with texture and morphological features extracted from the nodule volume and the surrounding tissue to characterize the nodule. Stepwise feature selection was used to search for the subset of most effective features to be used in the linear discriminant analysis classifier. The CAD system achieved a test A_z of 0.86±0.02 in a leave-one-caseout resampling scheme for 256 nodules from 152 patients. We conducted an observer study with six thoracic radiologists and found that their average Az in assessing nodule malignancy increased significantly (p<0.05) from 0.83±0.03 without CAD to 0.85±0.04 with CAD. These results indicate the potential usefulness of the CAD system as a second reader for radiologists in characterizing lung nodules.