Repeatability of diffusion and perfusion MRI derived radiomic features in glioblastoma: a test-retest study

An understanding of the repeatability of imaging biomarkers is key to their implementation as clinical tools. In this study we calculate the repeatability and inter-correlation of radiomic features derived from quantitative MRI (qMRI) of Glioblastoma (GBM) patients and assess the effect of image standardisation methods on these factors. We analysed scan-rescan Diffusion Weighted MR Images (DWI) and Dynamic Contrast Enhanced MR Images (DCE) from 36 GBM patients obtained from The Cancer Imaging Archive (TCIA). These included 17 patients, from the QIN-GBM-Treatment-Response patient cohort, scanned post surgery and prior to chemo-radiation therapy and 19 patients, from the RIDER Neuro MRI patient cohort, scanned at diagnosis of tumour recurrence. For both patient cohorts, two sets of scans were taken 2-6 days apart. Each of these patient cohorts was analysed independently to determine if findings were consistent across different acquisition parameters. Parametric maps of Apparent Diffusion Coefficient (ADC) and Cerebral Blood Volume (CBV) were obtained from DWI and DCE data, respectively. Intensity normalisation and noise filtering were applied to the parametric maps in multiple permutations to give 7 distinct standardisation methods. Shape, first order and second order radiomic features for the parametric maps were calculated within the Gross Tumour Volume (GTV). The Intraclass Correlation Coefficient (ICC) was calculated between the feature value at each imaging timepoint. The ICC of first and second order features derived from images with each standardisation method was compared to the ICC of corresponding features derived from images without standardisation. Based on the average ICC of features derived from ADC images without image standardisation, first order features were the most repeatable in both patient cohorts. For ADC derived features in the QIN cohort, shape features were the second most repeatable followed by second order features. For ADC derived features in the RIDER cohort, second order features were the second most repeatable followed by shape features. In CBV images, shape features were the most repeatable followed by second order and then first order in both patient cohorts. No image standardisation method implemented in this study was found to significantly increase the repeatability of ADC-derived first or second order features. For first order CBV features z-score normalisation without noise filtering produced a significant improvement in feature repeatability in both patient cohorts. Radiomic feature repeatability is impacted by feature class. Image standardisation methods implemented in this study were not found to be effective at improving the repeatability of ADC-derived features and had limited utility for improving CBV derived features. Future radiomic studies should consider feature repeatability as an important factor in feature selection.