Ted Win Way MS, Presenter: Nothing to Disclose

Lubomir M. Hadjiiski PhD, Abstract Co-Author: Nothing to Disclose

Berkman Sahiner PhD, Abstract Co-Author: Nothing to Disclose

Heang-Ping Chan PhD, Abstract Co-Author: Nothing to Disclose

Philip Neil Cascade MD, Abstract Co-Author: Nothing to Disclose

Naama R. Bogot MD, Abstract Co-Author: Nothing to Disclose

Ella Annabelle Kazerooni MD, Abstract Co-Author: Nothing to Disclose

Jeffrey A Fessler, Abstract Co-Author: Nothing to Disclose

Zhanyu Ge PhD, Abstract Co-Author: Nothing to Disclose

et al, Abstract Co-Author: Nothing to Disclose

To develop an automated nodule classification method based on 3D morphological and texture features extracted from an active contour segmentation of CT scans.

Regions of interest containing lung nodules were identified by experienced radiologists. The CT voxels were interpolated to be isotropic in all directions. The nodule boundary was automatically extracted using a 3D active contour model. Morphological features were extracted from the segmented nodule. The rubber band straightening transform (RBST) was applied to a band of voxels surrounding the nodule on each of the eight oblique planes that cut along the longitudes of the nodule surface with two sets of cuts (one through the north-south pole and the other through an east-west pole), in addition to slices parallel to the CT planes. Texture features based on run-length statistics were extracted from the Sobel-filtered RBST images and averaged over the slices of each group, providing 3D texture information around the nodule. In this preliminary study, we used a data set of 86 nodules (42 malignant and 44 benign) from 56 patients. The nodules were determined to be malignant or benign either by biopsy or by 2-year follow up. Stepwise linear discriminant analysis with simplex optimization was used to select the most effective features and formulate a feature classifier in a leave-one-case-out resampling scheme. Classifier performance was evaluated by ROC analysis.

Four 3D texture features, along with the maximum CT value, were consistently selected: latitude long run emphasis (LRE), north-south longitude short run (SR) low gray-level emphasis, east-west longitude LRE, and latitude SR emphasis. The classifier achieved a training Az of 0.88±0.04 and a test Az of 0.84±0.04. If texture features were extracted only from the planes of CT slices, the training Az was 0.82±0.04 and test Az was 0.79±0.05.

Our study indicates that the texture features around a nodule are important for characterizing a nodule as malignant or benign. The addition of the 3D texture extracted from oblique planes around the nodule can improve the classification accuracy in comparison to using the 2D texture on the plane of CT slices alone.

Way, T, Hadjiiski, L, Sahiner, B, Chan, H, Cascade, P, Bogot, N, Kazerooni, E, Fessler, J, Ge, Z, et al, , Classification of Pulmonary Nodules Using Automated 3D Segmentation and Feature Classification for Computer-aided Diagnosis on CT Scans.  Radiological Society of North America 2004 Scientific Assembly and Annual Meeting, November 28 - December 3, 2004 ,Chicago IL. http://archive.rsna.org/2004/4411901.html