Ye Xu MS, Presenter: Nothing to Disclose

Millan Sonka PhD, Abstract Co-Author: Nothing to Disclose

Edwin Vanbeek, Abstract Co-Author: Nothing to Disclose

Geoffrey McLennan MD, PhD, Abstract Co-Author: Nothing to Disclose

Junfeng Guo, Abstract Co-Author: Nothing to Disclose

Eric A. Hoffman PhD, Abstract Co-Author: Nothing to Disclose

Computer-aided detection algorithms for analysis of MDCT data have the potential to significantly alter clinical practice. Focus has been on lung nodule detection,pulmonary emboli and quantitation of emphysema. However, quantification of interstitial lung diseases (ILD) can be equally important in early detection and assessment of treatment response. In this project, we utilize our newly extended 3D-based computer-aided tissue characterization to evalute parenchymal properties important to ILD.

We performed MDCT (SIMENS Sensation 16/64 120kV, B50f Convolution kernel, and ≤0.75mm slice thickness) on 15 human volunteers. A total of 905 volumes of interest (VOIs; 21x21 pixels in plane with z dimension selected for cubic VOIs) were marked by a senior radiologist and a senior pulmonaligist using the Iowa PASS software: emphysema(n=230); ground glass(137); honeycomb(137), normal non-smokers(238), and “normal” smokers(163). We calculated 24 volumetric features including statistical features (first order features and run-length and co-occurrence features), histogram and model-based features on VOIs. Bayesian methods were used for classification and was compared with a Support Vector Machine(SVM) using 5-fold and 10-fold cross validation and leave-one-VOI-out validation methods.

Using the Bayesian method on 10-fold cross validation yielded a sensitivity of 93, 89, 96, 90 and 84% for emphysema, ground glass, honeycomb, normal non-smokers and “normal” smokers, respectively. The specificity was 97, 97, 98, 96 and 98%, respectively. The overall accuracy for all five characteristics combined for Bayesian method on 5-fold cross validation, 10-fold cross validation and leave-one-out method was 90, 91 and 91%, respectively. The overall accuracy for these validations was 91, 91, and 91% using SVM respectively.

We conclude that volumetric features including statistical, histogram and model-based features can be successfully used in differentiation of both emphysema and interstitial lung diseases. Our system is highly sensitive and specific in detecting early abnormalities both via a Bayesian and SVM classifier.

M.S.,G.M.,J.G.,E.A.H.: VIDA Diagnostics, LLC

Computer-aided detection algorithms for analysis of MDCT data have the potential to significantly alter clinical practice. Focus has been on lung nodule detection,pulmonary emboli and quantitation of emphysema. However, quantification of interstitial lung diseases (ILD) can be equally important in early detection and assessment of treatment response. In this project, we utilize our newly extended 3D-based computer-aided tissue characterization to evalute parenchymal properties important to ILD.

We performed MDCT (SIMENS Sensation 16/64 120kV, B50f Convolution kernel, and ≤0.75mm slice thickness) on 15 human volunteers. A total of 905 volumes of interest (VOIs; 21x21 pixels in plane with z dimension selected for cubic VOIs) were marked by a senior radiologist and a senior pulmonaligist using the Iowa PASS software: emphysema(n=230); ground glass(137); honeycomb(137), normal non-smokers(238), and “normal” smokers(163). We calculated 24 volumetric features including statistical features (first order features and run-length and co-occurrence features), histogram and model-based features on VOIs. Bayesian methods were used for classification and was compared with a Support Vector Machine(SVM) using 5-fold and 10-fold cross validation and leave-one-VOI-out validation methods.

Using the Bayesian method on 10-fold cross validation yielded a sensitivity of 93, 89, 96, 90 and 84% for emphysema, ground glass, honeycomb, normal non-smokers and “normal” smokers, respectively. The specificity was 97, 97, 98, 96 and 98%, respectively. The overall accuracy for all five characteristics combined for Bayesian method on 5-fold cross validation, 10-fold cross validation and leave-one-out method was 90, 91 and 91%, respectively. The overall accuracy for these validations was 91, 91, and 91% using SVM respectively.

We conclude that volumetric features including statistical, histogram and model-based features can be successfully used in differentiation of both emphysema and interstitial lung diseases. Our system is highly sensitive and specific in detecting early abnormalities both via a Bayesian and SVM classifier.

M.S.,G.M.,J.G.,E.A.H.: VIDA Diagnostics, LLC

Xu, Y, Sonka, M, Vanbeek, E, McLennan, G, Guo, J, Hoffman, E, MDCT-based Computer-aided Detection of Interstitial Lung Disease Using 3-Dimensional Texture Features.  Radiological Society of North America 2005 Scientific Assembly and Annual Meeting, November 27 - December 2, 2005 ,Chicago IL. http://archive.rsna.org/2005/4411256.html