TY - JOUR T1 - <strong>Robustness of radiomic features in <sup>18</sup>F-FDG PET/CT imaging of Nasopharyngeal Carcinoma: impact of parameter settings on different feature matrices</strong><strong/> JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 1319 LP - 1319 VL - 58 IS - supplement 1 AU - Wenbing Lv AU - Lijun Lu AU - Jun Jiang AU - Wei Yang AU - Jianhua Ma AU - Qianjin Feng AU - Arman Rahmim AU - Wufan Chen Y1 - 2017/05/01 UR - http://jnm.snmjournals.org/content/58/supplement_1/1319.abstract N2 - 1319Objectives: Radiomic features, increasingly explored as biomarkers of disease, can be affected by multiple factors. The impact of parameter setting on different matrices and correspondingly radiomic features still needs to be further investigated. The purpose of this study was to investigate effect of parameter settings including symmetry, direction, averaging strategy, etc. on different feature matrices.Methods: 32 patients with nasopharyngeal carcinoma were enrolled in the study and underwent 18F-FDG PET/CT scans. Tumors were segmented using consensus of 2 expert physicians as a volume of interest (VOI), and VOI SUVs were discretized with a constant resolution bin size of 0.1. For each VOI, radiomic GLCM, GLRLM, GLSZM and NGTDM matrices were generated. Both symmetric (S) or asymmetric (A) GLCM matrices were generated, considering 13 directions with (13) or without (1) an averaging step, with distance fixed as 1 voxel. This resulted in four strategies noted as 1S, 1A, 13S and 13A. Secondly, 1S with 10 distances (D=1, 2, 3,&amp;#8943;, 10 voxels) were considered to extract GLCM. GLRLM was also acquired by considering above-mentioned two strategies (1 and 13) with or without an averaging step. GLSZM was constructed with 6, 18 and 26 neighborhoods (N). NGTDM was created by considering five different window sizes (i.e. W=3, 5, 7, 9 and 11 voxels). After the matrices were generated, 26, 13, 13 and 5 features were extracted from GLCM, GLRLM, GLSZM and NGTDM respectively. The robustness of the 57 features was evaluated by intra-class correlation coefficient (ICC) and Spearman correlation (R).Results: (1a) 5/26 GLCM based features (Entropy, DifEntropy, IMC1, IMC2 and IDN) displayed wide ICCs ranging from 0.13 to 0.99, and 3/26 features (IMC1, IMC2 and IDN) exhibited wide R ranging from 0.24 to 0.99 between the four strategies (1S, 1A, 13S and 13A). (1b) Nearly half of all the GLCM based features demonstrated wide range of ICCs (0.01-0.99) between 1 voxel distance and other nine distances, but these features exhibited narrow range R (higher than 0.80) except for ClusterShade (ICCs: 0.06-0.66, R: 0.43-0.97). (2) For all of the GLRLM based features, ICC and R between strategy 1 and 13 were higher than 0.8 (lowest value was 0.86 and 0.85 respectively) (3) Only two GLSZM based features (HGZE and SZLGE) showed ICC higher than 0.8 between reference neighborhood 26 and neighborhood 6 and 18, but all the features showed R higher than 0.8. (4) 2/5 NGTDM based features (Complexity and Coarseness) exhibited ICC lower than 0.8 between window size of 3 voxels (W3) and other four individual window size (5, 7, 9, 11 voxels).Conclusion: The symmetry, direction and averaging steps of constructing GLCM have slight influence on 5 features (Entropy, DifEntropy, IMC1, IMC2 and IDN), while distance consistency exhibits the largest impact on most features. GLRLM based features were independent of the averaging step. GLSZM based features varied between different neighborhoods. The window size of NGTDM only notably impacted Complexity and Coarseness. Overall, there needs to be careful standardization of parameter settings on the construction of matrices prior to generation of radiomic features, while a number of features are quite robust to some parameter variations, as demonstrated in this work. Research Support: This work was supported by the National Natural Science Foundation of China under grants 61628105, 81501541, 31371009, 81371544, the Natural Science Foundation of Guangdong Province under grants 2014A030310243, 2016A030313577, the Science and Technology Planning Project of Guangdong Province under grant 2015B010131011, and the Program of Pearl River Young Talents of Science and Technology in Guangzhou under grant 201610010011. ER -