Abstract
Purpose:
Despite increased use of heterogeneity metrics for PET imaging, standards for metrics such as textural features have yet to be developed. We evaluated the quantitative variability caused by image acquisition and reconstruction parameters on PET textural features.
Methods:
PET images of the NEMA IQ phantom were simulated with realistic image acquisition noise. 35 features based on intensity histograms (IH), co‐occurrence matrices (COM), neighborhood‐difference matrices (NDM), and zone‐size matrices (ZSM) were evaluated within lesions (13, 17, 22, 28, 33 mm diameter). Variability in metrics across 50 independent images was evaluated as percent difference from mean for three phantom girths (850, 1030, 1200 mm) and two OSEM reconstructions (2 iterations, 28 subsets, 5 mm FWHM filtration vs 6 iterations, 28 subsets, 8.6 mm FWHM filtration). Also, patient sample size to detect a clinical effect of 30% with Bonferroni‐corrected α=0.001 and 95% power was estimated.
Results:
As a class, NDM features demonstrated greatest sensitivity in means (5–50% difference for medium girth and reconstruction comparisons and 10–100% for large girth comparisons). Some IH features (standard deviation, energy, entropy) had variability below 10% for all sensitivity studies, while others (kurtosis, skewness) had variability above 30%. COM and ZSM features had complex sensitivities; correlation, energy, entropy (COM) and zone percentage, short‐zone emphasis, zone‐size non‐uniformity (ZSM) had variability less than 5% while other metrics had differences up to 30%. Trends were similar for sample size estimation; for example, coarseness, contrast, and strength required 12, 38, and 52 patients to detect a 30% effect for the small girth case but 38, 88, and 128 patients in the large girth case.
Conclusion:
The sensitivity of PET textural features to image acquisition and reconstruction parameters is large and feature‐dependent. Standards are needed to ensure that prospective trials which incorporate textural features are properly designed to detect clinical endpoints.
Supported by NIH grants R01 CA169072, U01 CA148131, NCI Contract (SAIC‐Frederick) 24XS036‐004, and a research contract from GE Healthcare.
