Abstract Archives of the RSNA, 2012
LL-CHS-TH4A
Quantitative CT Assessment of Centrilobular and Paraseptal Emphysema and “Smoker’s Lung” Using Texture Features: Automated Classification of Smoking-related Lung Injury
Scientific Informal (Poster) Presentations
Presented on November 29, 2012
Presented as part of LL-CHS-TH: Chest Lunch Hour CME Posters
Shoshana Ginsburg MS, Presenter: Nothing to Disclose
Joyce Denise Schroeder MD, Abstract Co-Author: Research Grant, Siemens AG
David Augustine Lynch MD, Abstract Co-Author: Research support, Siemens AG
Scientific Advisor, Perceptive Informatics, Inc
Consultant, Actelion Ltd
Consultant, InterMune, Inc
Quantification and subtyping of smoking-related lung injury may be helpful in phenotyping disease and measuring treatment response. This study shows an automated texture-based method to identify the extent and pattern of emphysema and/or “smoker’s lung” (centrilobular nodularity, respiratory bronchiolitis) on CT images.
Chest CT scans were obtained from 6 non-smokers with no emphysema (NNE), 6 smokers with no emphysema (SNE), 6 smokers with centrilobular emphysema (CLE, GOLD stages 1-3), and 6 smokers with paraseptal emphysema (PSE, GOLD stages 1-3) in the COPDGene™ Study. Regions-of-interest (ROIs) of 35-pixel diameter representing normal tissue in the NNE scans, centrilobular nodularity in the SNE scans, CLE in the CLE scans, and PSE in the PSE scans were manually selected. In total, 678 NNE ROIs, 438 SNE ROIs, 292 CLE ROIs, and 166 PSE ROIs were selected.
30 texture features were computed for each ROI and employed in a two-stage decision tree classifier that first classified ROIs as emphysematous or non-emphysematous and subsequently discriminated between CLE and PSE among the ROIs deemed emphysematous and between normal and “smoker’s lung” textures among the remaining ROIs. A 6-fold, randomized cross-validation procedure was used to train and evaluate the classifier.
84.2% of ROIs from NNE scans were classified as normal, while 13.6% were classified as “smoker’s lung” and 2.2% as CLE (p < .001). Of SNE ROIs, 79.6% were classified as “smoker’s lung”, while 19.2% were classified as normal and 1.1% as CLE (p < .001). While 81.8% of CLE ROIs were classified correctly, 5.2% were classified as “smoker’s lung” and 10.3% as PSE (p < .001). Of PSE ROIs, 91.0% were correctly classified, while 6.6% were classified as CLE and 2.4% as non-emphysematous (p < .001).
Near-complete discrimination between emphysema and non-emphysematous tissue is achieved by our automated scheme. Although there was overlap between normal and “smoker’s lung” textures, as well as between CLE and PSE textures, the four classes of texture are clearly distinct. Thus, disease type and extent can be quantified in patients with smoking-related lung injury.
Objective identification and quantification of smoking-related lung injury may be helpful in directing treatment and monitoring patient response to therapy.
Ginsburg, S,
Schroeder, J,
Lynch, D,
Quantitative CT Assessment of Centrilobular and Paraseptal Emphysema and “Smoker’s Lung” Using Texture Features: Automated Classification of Smoking-related Lung Injury. Radiological Society of North America 2012 Scientific Assembly and Annual Meeting, November 25 - November 30, 2012 ,Chicago IL.
http://archive.rsna.org/2012/12021949.html