Namkug Kim PhD, Presenter: Nothing to Disclose

Youngjoo Lee MS, Abstract Co-Author: Nothing to Disclose

Joon Beom Seo MD, PhD, Abstract Co-Author: Speaker, Siemens AG

David Augustine Lynch MB, Abstract Co-Author: Consultant, Actelion Ltd Research support, Siemens AG Consultant, Gilead Sciences, Inc Consultant, Novartis AG Scientific Advisor, Perceptive Informatics, Inc

To investigate the effect of HRCT images from different vendors on classification accuracy for regional disease pattern analysis of diffuse lung disease.

Using HRCT images from vendor A and B, two sets of 600 rectangular regions of interest (ROIs) with 20x20 pixels, comprising of each 100 ROIs representing six regional disease patterns (normal, GGO, reticular opacity, honeycombing, emphysema, and consolidation) were marked by two experienced radiologists with consensus at HRCT images of various diffuse lung diseases. Each ROI image, then, was represented by 28 features (histogram, gradient, run-length, GLCM, LAA cluster, and top-hat transform). For automatic classification, support vector machine was employed to generate a classification model. Firstly, classification accuracies using each vendors' data were estimated by a 10-folding cross validation with 20 repetitions (internal-vendor study). Secondly, A and B data were used for training and testing, respectively, and vice versa (cross-vendor study). Finally, all ROI data were integrated, trained and tested by 2-folding cross validation with 20 repetitions (integrated data study). All experiments were repeated with feature selection to investigate the contribution of each feature.

On internal-vendor study, the classification accuracies of vendor A and B were 89.4±0.98% and 89.8±0.79% , respectively. The classification accuracies of cross-vendor study were 45.8±0.85% when training with vendor A data and testing with vendor B data, and 42.7±0.81% vice versa (t-test, p=0.000). On integrated data study, classification accuracy was 88.3±1.48%. Selected features were different between internal-vendor and cross-vendor studies.

When the training and testing data were obtained from HRCT images of different vendors, the classification accuracy significantly decreases in comparison with using HRCT of a same vendor. On the other hand, the accuracy of the automatic classifier trained with the integrated data set was similar to the case of the same vendor. Therefore, it is recommended to train automatic classifiers for each vendor separately or to build an automatic classifier with integrated data set if needed.

This study can provide a guideline to apply an automatic classifier for the quantification of regional disease pattern of diffuse lung disease at HRCT from different CT vendors.

Kim, N, Lee, Y, Seo, J, Lynch, D, Automatic Classification of Regional Disease Pattern of Diffuse Lung Disease at HRCT: Cross-Vendor Study.  Radiological Society of North America 2010 Scientific Assembly and Annual Meeting, November 28 - December 3, 2010 ,Chicago IL. http://archive.rsna.org/2010/9009609.html