Abstract
1780
Objectives Delineation of volume of interest (VOI) for FDG PET of brain tumor is difficult due to non-zero brain uptake in the surrounding tissues, making it difficult to use volume-based analysis, such as metabolic tumor volume (MTV) and total lesion glycolysis (TLG). Since intratumoral metabolic heterogeneity reflects pathological characteristics, texture analysis is an alternative approach, but only if it is not affected by VOI definition or tumor size. We aimed to evaluate inter-operator reproducibility and tumor volume dependency of the parameters from texture analysis.
Methods FDG PET images from 25 brain tumor patients were used in this study. Images were acquired with a Siemens EXACT HR+ scanner, and reconstructed with filtered backprojection (voxel size, 2.6×2.6×2.4 mm). Two nuclear physicians independently defined volume of interest (VOI) to enclose the entire tumor. Histogram analysis and higher order texture analysis using 4 kinds of matrices (co-occurrence matrix, gray-level run length matrix, gray-level size zone matrix, and neighborhood gray-level different matrix) produced a total of 35 texture feature parameters for each patient.
Results Thirteen parameters were strongly (either positively or negatively) correlated with tumor volume (|R| ≥ 0.8), indicating significant influence of tumor size. Using intra-class correlation (ICC) analysis, 10/ 35 parameters achieved very good agreement between two operators (ICC ≥ 0.9), whereas 9 parameters had poor reproducibility (ICC < 0.7). Among all parameters, “standard deviation” and “coarseness” were both less correlated with tumor volume (R = -0.38 (P<0.01), -0.28 (P=NS), respectively) and highly reproducible (ICC = 0.99, 0.95, respectively).
Conclusions These data suggested that, while some texture parameters were affected by tumor volume, “standard deviation” and “coarseness” were found to have both high inter-operator reproducibility and tumor volume independency, indicating their possible usefulness for featuring brain tumor.