Jason M. Zhao MD, PhD, Abstract Co-Author: Nothing to Disclose

Shoshana Ginsburg MS, Presenter: Nothing to Disclose

Stephen Humphries, Abstract Co-Author: Nothing to Disclose

Kunihiro Yagihashi MD, Abstract Co-Author: Nothing to Disclose

David Augustine Lynch MBBCh, Abstract Co-Author: Research support, Siemens AG Scientific Advisor, PAREXEL International Corporation Consultant, Boehringer Ingelheim GmbH Consultant, InterMune, Inc Consultant, Gilead Sciences, Inc Consultant, F. Hoffmann-La Roche Ltd Consultant, Veracyte, Inc Research support, Johnson & Johnson Research support, AstraZeneca PLC

Joyce Denise Schroeder MD, Abstract Co-Author: Research Grant, Siemens AG

Centrilobular nodularity (CN) and centrilobular emphysema (CLE) are important early markers of smoking related lung injury on CT. However, visual detection of these lesions is subject to substantial observer variation. This study aims to develop a relatively simple method for automated detection and quantification of CN and CLE by designing a set of visual features that are intuitive to understand and highly predictive of the presence of CLE and CN.

1247 circular ROIs (35-pixel diameter) from the inspiratory CT scans of 40 smoking and 19 nonsmoking subjects enrolled in the COPDGene study were manually selected and labeled by one chest radiologist and independently confirmed by another. Of these ROIs, 463 depicted normal lung, 374 contained CN, and 410 depicted CLE patterns. Within each ROI, our algorithm identified clusters of low and high attenuation areas (LAAs and HAAs) and extracted 18 visual features including cluster amount, area, intensity, background, contrast, and gradient. Feature selection was performed to identify the best performing features. Sensitivities were calculated using a Logistic Linear Regression classifier cross-validated on randomly selected patient sample sets (60% for training, 40% for testing).

Our preliminary analysis shows sensitivities of 78%, 83%, 85% for detecting Normal, CN, CLE ROIs, respectively. Misclassification of CN for Normal or vice versa occurred at a rate of 12-14%; misclassification between CLE and CN or between CLE and Normal occurred less frequently at 4-8%. The three top-performing features were the intensity range of HAA clusters, the mean intensity of LAA clusters, and the mean background of LAA clusters. The sensitivities of using only these three features reached 70%, 80%, 83% for Normal, CN, CLE ROIs, respectively.

This novel set of visual features provides good sensitivity and specificity for automatically classifying Normal, CN and CLE ROIs on CT scans.

The visual feature based approach simplifies automated detection of CN and CLE in CT scans and may lead to automated quantification of CN and CLE burden.

Zhao, J, Ginsburg, S, Humphries, S, Yagihashi, K, Lynch, D, Schroeder, J, Visual Features for Automated Detection of Centrilobular Nodularity and Emphysema.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14011731.html