Computer-aided lung nodule diagnosis using a simple classifier

Fifty-four solitary pulmonary nodules with diagnosis established with pathological or 2 years of radiographic followup were imaged using CT with a high-resolution protocol without contrast. A thoracic radiologist manually outlined each nodule on a single slice that they chose as,most representative of it. Nineteen two-dimensional attenuation, size, and shape features were calculated based on this slice and were input into a OneR and logistic regression classifier. Classifier performance was measured using resubstitution, leave-one-out, and a split training set (42 cases) and test set (12 cases). When testing the classifier using resubstitution OneR achieved 90.9% sensitivity, 71.4% specificity, and 83.3% accuracy while the logistic regression classifier achieved 81.8% sensitivity, 85.7% specificity, and 83.3% accuracy. The OneR classifier in a leave-one-out testing methodology was able to achieve 90.9% sensitivity, 66.7% specificity, and 81.4% accuracy compared to 63.6% sensitivity, 52.4% specificity, and 59.3% accuracy of the logistic regression classifier. When separate training/testing sets were made OneR resulted in 85.7% sensitivity, 60% specificity, and 75% accuracy while logistic regression resulted in 71.4% sensitivity, 40% specificity, and 58.3% accuracy. Thus a simple classifier like OneR can perform better than a more complex classifier. (C) 2004 CARS and Elsevier B.V. All rights reserved.