Share Email Print
cover

Proceedings Paper

Accuracy and variability of texture-based radiomics features of lung lesions across CT imaging conditions
Author(s): Yuese Zheng; Justin Solomon; Kingshuk Choudhury; Daniele Marin; Ehsan Samei
Format Member Price Non-Member Price
PDF $14.40 $18.00

Paper Abstract

Texture analysis for lung lesions is sensitive to changing imaging conditions but these effects are not well understood, in part, due to a lack of ground-truth phantoms with realistic textures. The purpose of this study was to explore the accuracy and variability of texture features across imaging conditions by comparing imaged texture features to voxel-based 3D printed textured lesions for which the true values are known. The seven features of interest were based on the Grey Level Co-Occurrence Matrix (GLCM). The lesion phantoms were designed with three shapes (spherical, lobulated, and spiculated), two textures (homogenous and heterogeneous), and two sizes (diameter < 1.5 cm and 1.5 cm < diameter < 3 cm), resulting in 24 lesions (with a second replica of each). The lesions were inserted into an anthropomorphic thorax phantom (Multipurpose Chest Phantom N1, Kyoto Kagaku) and imaged using a commercial CT system (GE Revolution) at three CTDI levels (0.67, 1.42, and 5.80 mGy), three reconstruction algorithms (FBP, IR-2, IR-4), four reconstruction kernel types (standard, soft, edge), and two slice thicknesses (0.6 mm and 5 mm). Another repeat scan was performed. Texture features from these images were extracted and compared to the ground truth feature values by percent relative error. The variability across imaging conditions was calculated by standard deviation across a certain imaging condition for all heterogeneous lesions. The results indicated that the acquisition method has a significant influence on the accuracy and variability of extracted features and as such, feature quantities are highly susceptible to imaging parameter choices. The most influential parameters were slice thickness and reconstruction kernels. Thin slice thickness and edge reconstruction kernel overall produced more accurate and more repeatable results. Some features (e.g., Contrast) were more accurately quantified under conditions that render higher spatial frequencies (e.g., thinner slice thickness and sharp kernels), while others (e.g., Homogeneity) showed more accurate quantification under conditions that render smoother images (e.g., higher dose and smoother kernels). Care should be exercised is relating texture features between cases of varied acquisition protocols, with need to cross calibration dependent on the feature of interest.

Paper Details

Date Published: 9 March 2017
PDF: 7 pages
Proc. SPIE 10132, Medical Imaging 2017: Physics of Medical Imaging, 101325F (9 March 2017); doi: 10.1117/12.2255806
Show Author Affiliations
Yuese Zheng, Duke Univ. (United States)
Justin Solomon, Carl E. Ravin Advanced Imaging Labs., Duke Univ. (United States)
Duke Univ. (United States)
Kingshuk Choudhury, Duke Univ. (United States)
Daniele Marin, Duke Univ. (United States)
Ehsan Samei, Carl E. Ravin Advanced Imaging Labs., Duke Univ. (United States)
Duke Univ. Health System (United States)


Published in SPIE Proceedings Vol. 10132:
Medical Imaging 2017: Physics of Medical Imaging
Thomas G. Flohr; Joseph Y. Lo; Taly Gilat Schmidt, Editor(s)

© SPIE. Terms of Use
Back to Top