Enhancing radiomic feature selection in PET-based heterogeneity analysis through tumor-volume-weighted normalization

Accurate assessment of intratumoral heterogeneity is critical for tumor characterization and monitoring, but variations in tumor size make it difficult to determine whether radiomic differences reflect true biological heterogeneity or mere volume effects. To address this, we developed Tumor-Volume-Weighted Normalization (TVWN), a framework that minimizes size-related biases while preserving intrinsic image heterogeneity.

TVWN was evaluated using: (i) computer-simulated tumors (185-800  mm³), (ii) heterogeneous, tumor-mimicking phantoms (10-170  mm³), and (iii) longitudinal PET scans of prostate tumor-bearing mice (35.2-1184  mm³) acquired at four time points. We analyzed 97 radiomic features, including texture metrics from GLRLM and GLSZM, before and after normalization to assess and correct volume dependence.

TVWN consistently reduced the influence of tumor volume across various acquisition and reconstruction conditions, including different noise levels, Gaussian filtering (2  mm), number of iterations and spatial resolutions (0-6.0  mm FWHM). For example, the volume dependence ofGlcm_IDNwas effectively reduced (Spearman ρ: 0.290 ± 0.097 to 0.007 ± 0.103 at α = 0.01). In phantom experiments, TVWN reduced the CoV ofGlrlm_REfrom 17.75 % to 3.18 % in filtered data and from 20.14 % to 3.87 % in unfiltered data. It also enhanced heterogeneity discriminability (p = 0.0286). In vivo, TVWN decoupled radiomic features from tumor size progression, ensuring that observed changes reflected biological rather than volumetric effects.

TVWN offers a simple, scalable approach to normalizing radiomic features against tumor volume effects, improving their reproducibility and biological interpretability in longitudinalstudies.