| Lung cancer is the second most common cancer in the United States and it is the primary cause of cancer related death. Prognosis for the patient is greatly affected by the stage at diagnosis. Often, the first sign of primary lung cancer is the solitary pulmonary nodule (SPN), but in order to deliver appropriate treatment, the nodule must be efficiently detected and characterized.;Machine learning provides high-volume, high dimension pattern recognition with powerful statistical machinery to assess the usefulness of features. By exploiting these qualities, image features can be evaluated in a novel fashion.;In this report, we investigate machine learning approaches for the development and assessment of image features with applicability to malignancy status of pulmonary nodules on computed tomography with the following aims: (1) Develop a set of candidate image features related to morphology, attenuation, and enhancement; (2) Assess the effect of image resolution on classification performance and feature emphasis; (3) Assess the performance of classifiers developed from (1) and (2) on images commonly acquired in clinical practice. My hypotheses are: there are novel quantitative features important to nodule classification as benign or malignant; and features that correlate to descriptors used by radiologists are also important to nodule classification. |
