Three-dimensional computer-aided diagnosis schemes for classification of benign and malignant pulmonary nodules

We present an example-based assisting approach for classifying pulmonary nodules in 3-D thoracic CT images. The technique represents the internal and surrounding structures of the nodule by means of the distribution pattern of CT density and 3-D curvature indexes. When given an unknown nodule image, the images of lesions with known diagnoses (e.g. malignant vs. benign) are retrieved from a 3-D nodule image database. The malignant likelihood of the unknown case is estimated by the difference between the representation patterns of the unknown case and the retrieved lesions. In the present study,we adopt the Mahalanobis distance as the difference measure and then, explore the feasibility of the classification based on patterns of similar lesion images.