(Tuesday, 7/31/2018) 11:00 AM - 12:15 PM
Room: Davidson Ballroom B
Purpose: Tumor heterogeneity is an important characteristic in diagnosis/prognosis. Texture feature is currently being used for tumor heterogeneity quantification and classification. We studied the robustness of heterogeneity classification by texture feature.
Methods: Raw projection data of scans of 17 patients undergoing low dose lung cancer screening CT exams were acquired from an MDCT scanner under IRB approval. Calibrated noise was added to the raw projection data to simulate low dose scans at 50%, 25%, and 10% of the original dose. Images were reconstructed at 1mm slice thickness using wFBP(medium kernel), and iterative reconstruction(SAFIRE I44/3). A bilateral filter was used to create denoised images for all reconstructions, resulting in total of 16 images for each nodule. wFBP images at 100% dose without denoising were identified as the reference condition. An in-house CAD tool segmented nodules at the reference condition; the segmentations were then mapped to images under all other conditions. GLCM-based entropy texture feature was calculated for nodules under all conditions. To classify nodules into groups of heterogeneous/homogenous texture, a cut-off value of 4.57, extracted from the literature, was applied to the entropy feature. For each case, the classification under the reference condition served as the reference category; comparisons were made to texture classifications based on entropy feature value under all other conditions and assessed using the Kappa agreement index.
Results: Texture classifications using the GLCM entropy feature and a specific threshold demonstrated only slight variations in classification decisions for wFBP images across a range of dose values (0.8≤Kappa≤1). Classifications using SAFIRE images had less agreement to the reference (0.4≤Kappa≤1).
Conclusion: Texture classifications by GLCM across a wide range of dose levels were quite robust for wFBP images and denoised wFBP images. The smoothing effect observed for the images using denoising or iterative reconstruction resulted in poor agreements to reference classifications.
Funding Support, Disclosures, and Conflict of Interest: Funding support was provided by the National Cancer Institute Quantitative Imaging Network (QIN grant U01-CA181156). Disclosures: The UCLA Department of Radiological Sciences has an Institutional Master Research Agreement with Siemens Healthineers (formerly Siemens Healthcare, Erlangen, Germany).
Keywords
CAD, Quantitative Imaging, Texture Analysis
Taxonomy
IM- CT: Quantitative imaging/analysis
Contact Email
