Computerized detection of lung nodules in computed tomography scans

Computed tomography (CT) is the most sensitive imaging modality for lung nodule diagnosis. A single CT examination, however, may result in the acquisition of over 30 images. A computerized scheme has been developed to detect lung nodules within the vast amount of image data comprising a CT scan. Gray-level thresholds are determined to first segment the thorax from the background and then to segment the lungs within the thorax. Multiple gray-level thresholding is used to extract candidates from the lung regions. These candidates represent both nodules and normal pulmonary vasculature. For each candidate, geometric features, such as candidate area, compactness, circularity, and distance from the lung boundary, are calculated. To distinguish between nodules and non-nodules, the values of these features are used as input to an artificial neural network (ANN). This method was applied to 17 helical thoracic CT cases. The ability of the ANN to distinguish nodules from non-nodules was evaluated using receiver operating characteristic (ROC) analysis, which yielded an area under the ROC curve of 0.90. This automated method coupled with an ANN demonstrates promising performance in its ability to detect lung nodules in thoracic CT images.