PET radiomics can reveal a lot of clinical information but with many technical and physical variables to be tackled. The aim of this study was to find out the most reproducible and robust PET radiomics that can be used as benchmark for future clinical studies. Materials and Methods: Two phantom datasets were retrieved from the large Cancer Imaging Archive repository (NIH, USA) and employed for further radiomics feature extraction (i.e. 108 features) and analysis including calculation of coefficient of variation (CV) and percent deviation. Further testing was carried out to investigate effect of sphere size and contrast level as well as imaging scanner on feature reproducibility. Feature correlation with tumor TNM stage was also performed using Spearman correlation and Kendall tau statistical tests. Principle component analysis was used to reduce the large feature number to principle components using the eigen vector and eigen value. Results: The (CV) and percent deviation revealed 56 features out of 108 that have less than or equal to 10% variability. When features were compared in terms of high and low image contrast, there was 46 features that showed less sensitivity to different concentrations. Inter-scanner variability testing has reduced the number of features to 36 out of the 46 features. Analysis of PCA results showed that 4 components can account for 90.3%, namely, SUV max, inverse difference moment normalized; size zone non-uniformity normalized, and short run low gray level emphasis. When correlating the 108 features with patient status, two features only showed significant correlation with tumor stage namely maximal correlation coefficient and flatness. Conclusions: The 108 features were reduced to a lower percentage of features while maintaining large percentage of data variance and feature reproducibility. The combination of texture feature of robust technical qualifications to those of clinical value would finally improve the clinical decision model.