A novel methodology for assessing reproducibility of heterogeneity metrics in PET radiomics using noise-equivalent count rate, Monte Carlo simulation and 3D-printed patient-specific tumour phantoms

Abstract

Aim/Introduction: The advent of accessible machinelearning software has inspired a great interest in research around radiomics in PET. There is little of this research dedicated to establishing the reproducibility of the hundreds of metrics extracted using radiomics software. This work suggests a framework for validating these metrics, specifcally those quantifying heterogeneity, by comparison with how they and the noise-equivalent count rate (NECR) are afected by increasing activity within the feld of view (FOV). Materials and Methods: The NECR is a measurement of collected PET data after removing the efects of scatter and random coincidences. Comparing the characteristic curve of the NECR against activity in the scanner FOV against similar curves for heterogeneity metrics can provide a robust method for determining whether any given heterogeneity metric is exaggeratedly afected by noise in the data. The S285 Eur J Nucl Med Mol Imaging (2021) 48 (Suppl 1): S1–S648 NECR plotted against activity exhibits a characteristic peak, beyond which any increased activity disproportionately increases random coincidences; if a heterogeneity metric is unduly infuenced by data noise, its curve against activity will manifest a stationary point at this NECR peak. Physical PET data is collected from phantoms; included in this study are a 20 cm diameter cylindrical phantom, the NEMA Image Quality phantom and a series of 3D-printed tumour phantoms created from patient data. These phantoms are flled with an 18F solution of activity around 500 MBq and data is acquired for 12 hours. Once validated, a Monte Carlo simulation is used to predict the behaviour of metrics at a range of activity values to determine repeatability and uncertainty, with suggestions for clinical applications. Results: Pilot studies using the cylindrical and NEMA IQ phantoms have been conducted and compared to simulation results. Data is to be acquired for the 3D-printed tumour phantoms and presented. Conclusion: The NECR provides inspiration for a novel methodology for verifying the susceptibility of heterogeneity-based image metrics to noise. The hypotheses and initial work presented here have an exciting potential for guiding a more reliable use of radiomics metrics in the clinical environment.