Expanding the Range of Extracted Dynamic FET Features As Candidate Glioblastoma Response Biomarkers
Abstract
Purpose
To assess the feasibility of using radiomics features from dynamic FET imaging parameterizations as potential biomarkers of glioblastoma (GBM) treatment response by evaluating the repeatability of parametric-map radiomic features from dynamic 18F-FET PET data under test–retest conditions.
Methods
Dynamic 18F FET-PET data from 8 test–retest glioblastoma patients (16 datasets) were processed using a standardized pipeline to generate voxel-wise Logan parametric maps. The pipeline uses automated selection of the linear fitting start time with model-fit quality control. Tumour ROIs were applied per-scan, with optional voxel-level R² gating to exclude poorly modelled voxels. Conventional radiomics features were extracted from parametric maps using a fixed, IBSI-aligned preprocessing configuration (3D resampling, fixed bin widths, and defined image types). Feature repeatability was assessed using ICC(1,1) and complementary agreement metrics (within-subject coefficient of variation and median absolute percent difference). Sensitivity analyses evaluated stability under multiple gray-level discretization settings (bin widths 0.02–0.08 in map units) and across common radiomics filter families (Original, Gaussian/LoG, intensity transforms, wavelets). Volume coupling and proportional bias were screened to identify features driven primarily by ROI size or systematic effects.
Results
A consistent subset of parametric-map features demonstrated high test–retest repeatability across discretization settings (≈ 50–52 features meeting ICC ≥ 0.85 per bin width, with 44 stable across all tested bin widths). Filter sensitivity analysis showed many stable features persisted across Original/Gaussian/transform families, while wavelet-derived features were substantially less repeatable overall.
Conclusion
Dynamic PET parametric-map radiomics can yield a robust core of repeatable features, but stability depends on discretization, filter choice, and ROI variability. These preliminary results support the feasibility of standardized dynamic PET feature extraction and motivate future work on explicitly temporal feature trajectories and modelling for downstream clinical prediction.