Abstract
504
Objectives: Tumor heterogeneity is of interest as research indicates that it negatively impacts therapeutic outcome [1]. Numerous radiomics features have been proposed for assessing the heterogeneity of tumor images [2], but to be of value they need to be sensitive to true underlying heterogeneity and insensitive to other factors (controllable and uncontrollable) of the image data acquisition. This may be a particular issue with PET where factors such as high noise and low resolution inhibit measurement. In this work, we quantify the relative sensitivity of selected heterogeneity features to controlled variations in the true underlying heterogeneity using a phantom. Selected features have been previously identified as robust to PET methodological factors [3,4,5].
Methods: A bespoke PET and MR visible head sized phantom was designed using materials with soft tissue equivalent attenuation coefficients and consisting of: six cylindrical compartments of varying volumes of interest (VOIs of 6-100 ml) which model tumors, with heterogeneity introduced using plastic spheres of 3.97, 6.35 and 7.94 mm in diameter surrounded by radioactive agar gel; and a warm background with a contrast ratio across the cylinders of ~8:1. Four PET scans (3 sphere sizes + a homogeneous gel solution) were conducted with the phantom containing 20.7 +/- 2.3 MBq of 18F and data acquired for 4 hours on a Siemens Biograph TruePoint TrueV PET-CT. The resulting data were reconstructed multiple times (+/- resolution modelling (RM); 2 and 4 mm voxel sizes; 0, 3, 4, 5, 6 mm FWHM Gaussian post filter; 2, 4, 10 iterations) using 3D OP-OSEM as implemented in e7-tools. For each scan and reconstruction, the location of each cylinder was segmented and selected radiomics features calculated using IBEX [6]. The relative sensitivity for each feature was quantified using univariate analysis of variance to partition the sum of squares of the 1440 derived values (SPSS v23).
Results: In 7 of 9 metrics the sphere size was the largest single source of metric variability but this factor only contributed at most 39% of the total sum of squares for low gray level run emphasis. Voxel size and VOI had the next largest independent impact on heterogeneity features. Interactions between the parameters examined accounted for a larger proportion of the metric variability than sphere size in 7 of 9 metrics, contributing a maximum of 69% of the total sum of squares for entropy.
Conclusion: PET image reconstruction variations and tumor size contribute significantly to the variability of radiomics features when compared to their sensitivity to true underlying heterogeneity. It is therefore important to compare methodological variability to heterogeneity sensitivity and not to mean values. Standardisation of reconstruction may therefore be necessary to reduce this contaminating variability in applications. Future work will focus of examining the impact of uncontrollable factors such as contrast, count level, and replication. Other types of heterogeneity will also be examined as some features may be insensitive to the sphere size alterations. Research Support: Cancer Research UK