PET-Guided CT Radiomics for Characterizing Multiscale Intratumoral Heterogeneity In Lung Cancer
Abstract
Purpose
To investigate whether PET-defined metabolic subregions within lung tumors exhibit distinct CT radiomic characteristics, enabling a spatially imaging-based assessment of intratumoral heterogeneity.
Methods
PET/CT scans from 169 lung cancer patients were analyzed retrospectively. Tumors were segmented into whole tumor, high uptake (SUV ≥ 50% of SUVmax), and low uptake (SUV < 50% of SUVmax). Using 3D slicer and PyRadiomics, 1,130 CT radiomic features per scan were extracted from original (107), wavelet (744), and Laplacian-of-Gaussian (LoG) filtered images (279). Habitat contrasts were defined as Δ = HighSUV − LowSUV. Paired Wilcoxon signed-rank tests were performed for each feature, and Benjamini-Hochberg false discovery rate (FDR) correction was applied.
Results
Our preliminary analysis showed that of 1,130 features, 479 (42.4%) differed between high- and low-uptake habitats at FDR ≤ 0.05, increasing to 620 (54.9%) at FDR ≤ 0.10. At FDR ≤ 0.05, significant contrasts were present across image types: 57/107 (53.3%) original, 341/744 (45.8%) wavelet, and 81/279 (29.0%) LoG features. By feature family, significant differences were concentrated in first-order and texture metrics: first-order 107/216 (49.5%), GLCM 141/288 (49.0%), GLRLM 87/192 (45.3%), GLDM 63/168 (37.5%), GLSZM 62/192 (32.3%), and NGTDM 9/60 (15.0%). Most robust examples (FDR q = 0.029 with consistent contrast direction across patients) included wavelet-LLL GLCM ClusterShade, wavelet-LLL first-order Minimum, original first-order 10Percentile, original GLDM LargeDependenceEmphasis, LoG-sigma-5 GLSZM LargeAreaLowGrayLevelEmphasis, and LoG-sigma-3 GLSZM SizeZoneNonUniformity.
Conclusion
PET-defined metabolic subregions exhibit distinct CT radiomic signatures that persist across unfiltered and multiscale filtered feature sets, with strongest effects in first-order and texture families. PET-guided CT radiomics provides a spatially resolved, physics-informed approach to quantifying intratumoral heterogeneity and motivates validation in larger cohorts and predictive response modeling.