Computer-aided techniques to characterize solitary pulmonary nodules imaged on CT

The purpose of this research is to measure quantitative features of solitary pulmonary nodules (SPNs) imaged on high resolution CT to characterize these nodules as benign or malignant. High resolution CT images and definitive diagnoses of 31 patients with SPNs (14 benign and 17 malignant) were obtained. Software tools were developed to perform the classification task. Semi-automated contouring was used to isolate each nodule. Features related to the size, shape, attenuation and texture of each nodule were computed. A Student's t-test was used to identify the features whose mean values differed significantly between benign and malignant cases. A stepwise feature selection process chose a small set of features that separated benign from malignant cases. Linear discriminant analysis using only selected features was used to evaluate the ability to predict the classification for each nodule. A jackknifed procedure was also performed to provide a less biased estimate of performance. Linear discriminant analysis using only two texture features correctly classified 28/31 cases (90.3%); the jackknifed procedure yielded the same results. In conclusion, features extracted from high resolution CT images can be used to distinguish between benign and malignant nodules with a high degree of accuracy. Future work involves including contrast-enhancement and three dimensional measures from volumetric CT scans.