Hsien-Chi Kno, Presenter: Nothing to Disclose

Maryellen L. Giger PhD, Abstract Co-Author: Stockholder, Hologic, Inc Shareholder, Quantitative Insights, Inc Royalties, Hologic, Inc Royalties, General Electric Company Royalties, MEDIAN Technologies Royalties, Riverain Technologies, LLC Royalties, Mitsubishi Corporation Royalties, Toshiba Corporation Researcher, Koninklijke Philips NV Researcher, U-Systems, Inc

Ingrid Reiser PhD, Abstract Co-Author: Nothing to Disclose

Karen Drukker PhD, Abstract Co-Author: Royalties, Hologic, Inc

John M. Boone PhD, Abstract Co-Author: Research Grant, Siemens AG Research Grant, Hologic, Inc Consultant, Varian Medical Systems, Inc

Karen K. Lindfors MD, Abstract Co-Author: Research Grant, Hologic, Inc

Kai Yang PhD, Abstract Co-Author: Nothing to Disclose

Alexandra V. Edwards, Abstract Co-Author: Nothing to Disclose

To develop a new quantitative feature of lesion spiculation by 3D lesion surface analysis for use in computer-aided diagnosis system (CADx) on dedicated breast CT (bCT) images.

Patient images in this study were generated from a cone-beam CT scanner dedicated to breast imaging. Voxel dimensions range between 190 and 390 μm in coronal planes and 200 to 700 μm in the anterior-posterior direction. The data set included 116 non-contrast dedicated bCT images with 129 masses (80 malignant, 49 benign). For each lesion, the center was labeled by the radiologist and used as a seed point in a previously-developed two-stage lesion segmentation algorithm. A tissue map indicating fibroglandular versus adipose tissue was calculated in the lesion center neighborhood; both lesions and spiculations are labeled as “fibroglandular” tissue. A new 3D spiculation feature was developed, based on the number of connecting points between the segmented lesion surface and the surrounding “fibroglandular-labeled” regions. The performance in the task of distinguishing benign from malignant lesions was investigated in conjunction with other morphological and texture features utilizing stepwise feature selection, leave-one-out linear discriminant analysis, and ROC analysis. Feature selection was performed with and without inclusion of the proposed spiculation feature, thus 2 different set of the most dominant features were obtained. The results of ROC analysis with the two feature sets were compared.   

The new spiculation feature was consistently selected during feature selection when included in the feature pool. The area under the ROC curve (AUC) improved from 0.81, without the spiculation feature, to 0.85 with spiculation feature (p-value ≪ 0.001) in the task of breast mass classification.

The results demonstrated that our proposed 3D spiculation feature was able to significantly improve the performance of breast mass classification on dedicated bCT. 

Without superimposition effect existing in dedicated bCT, further investigation of parenchymal pattern in lesion neighborhood is suggested for the future CAD use on 3D imaging modalities.

Kno, H, Giger, M, Reiser, I, Drukker, K, Boone, J, Lindfors, K, Yang, K, Edwards, A, Development of a New 3D Spiculation Feature for Enhancing Computerized Classification on Dedicated Breast CT.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14008189.html