Sensitivity of Template Matching for Pulmonary Nodule Detection: A Case Study

Template matching is a common approach for pulmonary nodule detection from CT scans which may take different shapes, sizes and density distributions. The process of nodule detection is essentially two steps: isolation of candidate nodules, then elimination of false positives. The process of outlining the detected nodules and their classification complete the CAD system for early detection of lung nodules. This paper is concerned with the template design and evaluation of the effectiveness of the first step in the nodule detection process. Only parametric templates are considered and a model-based approach is used for selecting the nodule size and gray level distribution. The effectiveness of the template matching is investigated by cross validation with respect to the ground truth provided by radiology experts, and is described by receiver-operating-characteristic (ROC) curves indicating the probability of detection as a function of shape, size and orientation of the templates. We use samples from the ELCAP lung images database in our experiments. Experiments show that variations in the template shape and/or orientation has minimal effect on well-circumscribed nodule detection. Also the semi-circular template best simulates juxta-pleural, pleural-tail and vascular nodules. It is found that smaller template radii yield higher sensitivity, while the performance tends to decrease for larger template sizes. The rate of performance decay depends on template shape and nodule type.