Mahdi Orooji PhD, Presenter: Nothing to Disclose

Mirabela Rusu DPhil, MENG, Abstract Co-Author: Nothing to Disclose

Prabhakar Rajiah MD, FRCR, Abstract Co-Author: Institutional Research Grant, Koninklijke Philips NV

Michael Yang, Abstract Co-Author: Nothing to Disclose

Frank Jacono, Abstract Co-Author: Nothing to Disclose

Robert C. Gilkeson MD, Abstract Co-Author: Research Consultant, Riverain Technologies, LLC Research support, Koninklijke Philips NV Research support, Siemens AG

Philip Aaron Linden, Abstract Co-Author: Nothing to Disclose

Anant Madabhushi MS, Abstract Co-Author: Research partner, Siemens AG Research partner, General Electric Company Research partner, F. Hoffmann-La Roche Ltd Founder and President, IbRiS, Inc

Radiographic characteristics to reliably define the degree of invasion of early Non-Small Cell Lung nodules with ground glass opacity (GGO) components on CT have yet to be reliably defined. Our goal is to identify quantitative computer extracted image texture features to distinguish GGO nodules with no/minimal invasion from those with frank invasion on pre-operative CT. Computer-extracted texture features quantitatively describe the spatial arrangement of intensities in an image and have been shown to distinguish benign from malignant nodules. In this study we evaluate the utility of computer extracted texture features in distinguishing GGO with no/minimal and frank invasion.  

We used a retrospective cohort of 33 slices (15 no/minimal and 18 frank) of in vivo lung CT from patients who had surgical resection. All nodules measured less than 16 mm in diameter. The size of the invasive component was utilized to stratify the nodules in the no/minimal (<5mm) or invasive category (>5mm invasion). A total of 63 of computer extracted texture features including gray-level statistical, steerable Gabor, Haralick, and Laws were obtained on CT from the manually delineated nodule. Following feature extraction, the total number of features was reduced from 63 to 3 via principal component analysis.

Three texture features, Inertia, Correlation and Difference Entropy, were selected by the classifier, providing an area under the receiver operating characteristic curve (AUC) of 0.92 for distinguishing on CT the no/minimal invasion nodules from the frank invasion tumors. By comparison, Laws features provided an AUC of 0.61 and Gabor features yielded an AUC of 0.68.

Texture analysis of CT scan showed reasonable discrimination of level of invasion in the context of GGO cancerous lung nodules.

Computerized image analysis of in vivo CT may allow for identification of computer extracted CT features associated with no/minimal and frank invasion in GGO lung nodules. It has the potential to impact clinical, economic, and societal burden of lung cancer by increasing average 5-year survival rate from early detection of invasive nodules, significant economic benefits to the health care system by reduction in unnecessary interventions, better image analytics can potentially reduce dependence on repeat or higher resolution CT exams, and noninvasive means of assessing response to targeted therapies.

Orooji, M, Rusu, M, Rajiah, P, Yang, M, Jacono, F, Gilkeson, R, Linden, P, Madabhushi, A, Computer Extracted Texture Features on CT Predict Level of Invasion in Ground Glass Non-Small Cell Lung Nodules.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14014130.html