Can Computed Morphological Features of Solitary Pulmonary Nodules Match the Observations of Radiologists for Diagnosis?
Radiologists have been paying attention to the contour features of pulmonary nodules such as smooth, ragged, and spiculated for their differential diagnosis. In this paper, we aimed to confirm the feasibility of using the morphological features of the circumference of solitary pulmonary nodules for a differential diagnosis. We studied pulmonary nodules that were clinically confirmed as primary lung cancers, metastatic pulmonary nodules, or other benign nodules, obtained from 91 computed tomography examinations that included less than five pulmonary nodules (5-30 mm). Two board-certificated radiologists graded and recorded the conspicuiry of the features on a 5-point scale. Using the same nodules, the contour element spectrum (CES) of the circumference was computed using 28 structure elements and summarized in four spectra. A percentage of each CES was used for computed inference. Diagnosis was inferred using a Bayesian network that was programmed in the conspicuity of features and the confirmed diagnoses of all other nodules. The spectra showed good correlation (R2 > 0.8) with the subjective ratings by radiologists. There were clear differences between smooth and rough contours on the correlation coefficients. The diagnostic accuracy of the radiologists was 67.6％, that of Bayesian inference with subjective ratings was 61.3％, and that of the Bayesian network using the CES was 63.6％. Computer-based inference using computed nodule attributes might be feasible as a second opinion.
pulmonary nodule computed inference morphological feature computer-assisted diagnosis
K. Sakai N. Sugimoto M. Yakami K. Togashi M. Kawagishi
Graduate School of Medicine Kyoto University Kyoto, JAPAN Medical Informatics and Computer Vision Canon Inc. Tokyo, JAPAN
国际会议
三亚
英文
26-29
2011-03-25（万方平台首次上网日期，不代表论文的发表时间）