Quantitative Reality: A Physical Ground Truth to Assess Heterogeneity Beyond Visual Limits in Molecular Imaging
Authors
Affiliations (1)
Affiliations (1)
- Universite De Toulouse
Abstract
BackgroundAccurate quantification of intratumoral heterogeneity is essential for validating radiomics features and AI models in oncology. However, standard physical phantoms typically utilize homogeneous compartments, lacking the spatially complex uptake patterns required to rigorously benchmark these metrics against a known reality. To address this, we introduce porous 3D-printed inserts for simulating controlled spatial heterogeneity. MethodsModular grid inserts (17, 22, 28 mm targets) were developed with contrast of 4:1. Nine heterogeneity models, simulating progressions from uniform uptake to multifocal to necrosis, were imaged on a 32cm Omni Legend PET/CT system. Two different tracers: [18F]FDG ([18F]-2-fluoro-2-deoxy-d-glucose) and [68Ga] (68-Gallium) were used on different days. Quantitative performance was assessed using conventional Standardized Uptake Values (SUV) and Target-to-Background Ratios (TBR), alongside the Gini Index (GI) to quantify spatial heterogeneity and radiotracer distribution inequality. ResultsThe phantom demonstrated high consistency, with reference solution variability remaining under 2.2% across layers. In 28 mm inserts, experimental TBRmax values fell within 10% of the theoretical 4:1 design ratio. GI successfully quantified structural complexity, ranging from 0.05 in homogeneous regions to 0.30 in necrotic targets. Cross-tracer analysis demonstrated robust reproducibility with a mean bias of <10%. While Partial Volume Effects reduced visual distinctness in 17 mm geometries, quantitative metrics successfully preserved the intended pathological trends. ConclusionThis study establishes a foundational ground truth for heterogeneous uptake in PET. By providing defined physical standards validated by quantitative metrics, it enables the robust validation of segmentation algorithms and radiomics features even when spatial resolution limits visual assessment.