Yongjun Chang, Abstract Co-Author: Nothing to Disclose

Jangpyo Bae MS, Presenter: Nothing to Disclose

Namkug Kim PhD, Abstract Co-Author: Stockholder, Coreline Soft, Inc

Jung Won Moon, Abstract Co-Author: Nothing to Disclose

Ho Yun Lee MD, Abstract Co-Author: Nothing to Disclose

Joon Beom Seo MD, PhD, Abstract Co-Author: Nothing to Disclose

To develop a computer-aided diagnosis (CAD) system to investigate the possible usefulness of 2D and 3D features by measuring the accuracy of regional pattern classification of diffuse interstitial lung disease (DILD) from volumetric chest Dual Energy CT (DECT) images.

Twenty eight patients with suspected DILD were enrolled from February 2010 to August 2011. All patients underwent surgical biopsy within 3 days from DECT scanning (Somatom Definition Flash) with the dual-energy technique. Preprocessing including noise filtering and threshold-based segmentation of lung and airway using a rib detection and inverse level set algorithm were performed to extract lung with threshold of -130 HU at virtual non-contrast (VNC) images. For training with features, circular regions of interest (ROI) with 10-pixel diameter including normal, ground-glass opacity (GGO), reticular opacity, and consolidation patterns were picked by the two chest radiologists with consensus. For testing, a hundred 2D images randomly selected were manually divided into the four classes in the similar way. For training with 2D and 3D features, 10x10 reticular and 10x10x10 cubic regions were selected respectively. The characteristics of each ROI were represented by thirteen textural features and eleven shape features in 2D and 3D features were extracted. Support vector machine (SVM) classifier were used. Twenty repetitions with five-fold cross-validation were performed to evaluate overall accuracy of these classifiers.

The overall accuracy is 90.47 ± 4.62% for the whole lung classification with the combination of 2D and 3D features, which is significantly enhanced by 8.69% compared with that of 2D features only (paired t-test, p<0.05). 

We proposed an automatic classification method based on SVM with 2D and 3D image features for DILD in DECT imaging. The experiments using twenty five-fold cross-validations result in 90.47% in mean accuracy, which demonstrates the effectiveness of the proposed CAD system for classifying DILD in DECT images.

This method is useful in computer aided differentiation and quantification of regional disease patterns of diffuse infiltrative lung disease in DECT images, which could be helpful for lessening radiologists’ workload by the initial diagnosis of the possible DILD.

Chang, Y, Bae, J, Kim, N, Moon, J, Lee, H, Seo, J, Automatic Classification of Regional Patterns of Diffuse Interstitial Lung Disease with 3D Features from the Volumetric Chest CT Images.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14045762.html