Association Between MRI Radiomic Features and 18F-FDG-PET Image-Derived Metrics: Radio-Metabolomics Analysis in Head and Neck Cancer

Introduction: This study investigated the association between radiomic features derived from magnetic resonance images (MRI) and metabolic indices extracted from ¹⁸F-FDG positron emission tomography (PET) to interpret the biological underpinnings of the observed correlations in head and neck squamous cell carcinoma (HNSCC).

Methods: Thirty-three histologically-proven HNSCC patients who underwent multiparametric MRI including T1- and T2-weighted and diffusion-weighted sequences with B1000 and apparent diffusion coefficient (ADC) maps along with ¹⁸F-FDG PET/CT were enrolled in this study. Two experienced radiologists delineated manually by consensus the contours on the lesions separately on each modality. MR images were preprocessed through bias field correction followed by histogram normalization whereas PET images were converted to SUV indices. Prior to radiomics features extraction, all images were interpolated into isotropic voxel spacing (SUV, ADC, B1000 1×1×1 mm³ and T1 and T2 0.5×0.5×0.5 mm³). The intensity levels in regions of interest in MRI images were discretized into 64 bins (fixed bin numbers) while SUV images were discretized by a fixed bin size of 0.25 unit interval. A comprehensive set of features (morphological, intensity-based and texture-based) were extracted from MRI datasets (218 features) while first-order metabolic-related indices were derived from SUV PET images (53 features). The correlation between MRI radiomic features and SUV-derived features were analyzed through Spearman’s correlation.

Results: Among a total of 218×53 extracted features and metabolic indices, 280 pairs of features demonstrated a meaningful correlation (<m:omath>ρ <m:scr m:val="roman"><m:sty m:val="p"></m:sty></m:scr></m:rpr>≥ 0.6, p < 0.05</m:omath> ). For T1 and T2 image sets, 131 and 115 features showed significant correlation with SUV images, respectively, while for ADC and B1000 sequences, 18 and 16 pairs of features passed the criteria. In diffusion-based images, shape and size features mainly showed correlation with energy-based and momentum invariant SUV metabolic features. In T1- and T2-weighted images, shape, size and texture features (Co-occurrence information correlation, Neighboring grey level dependence non-uniformity) showed correlation with histogram-based and momentum invariant SUV features.

Conclusions: Significant correlation of MRI-based radiomics features with PET-derived features was observed. This study warrants further investigation to elucidate the biological underpinning of the observed correlations.

• Download figure
• Open in new tab
• Download powerpoint