Multi-scale Radiomic Fingerprint: Quantifying Spatial Changes in Biology

Traditional radiomic studies build texture matrices using single-voxel increments. However, useful information may emerge when radiomic features are instead evaluated across multiple spatial scales. Moreover, basing these scales on physical units may produce results that are more interpretable to clinicians. We propose a multi-scale radiomic approach that defines texture distances in millimeter-based units to capture a more inclusive range of texture information, promote reproducibility, and improve clinician interpretability. We examine the variance in quantified radiomics across multiple spatial scales and diseases, including venous malformations, gliomas, Alzheimer's disease, brain metastases, and multiple sclerosis. We subsequently generated anisotropic counterparts to originally isotropic datasets to compare their performance in clinical predictive modeling. Finally, we evaluate differences between radiomic features captured at millimeter and voxel units. We discovered that the radiomic features captured at different millimeter scales were almost always statistically different ( $p<$ 0.05) across five diseases. Predictive modeling revealed that models trained on radiomics extracted from multiple millimeter scales consistently had a higher mean F1 across folds compared to those built from voxel scales. Roughly 93%, 90%, and 88% of texture metrics were statistically different between millimeter and voxel scales for venous malformations, Alzheimer's, and gliomas, respectively, suggesting that variations in spatial scale may capture differences in biology. We demonstrate that a multi-scale, millimeter-based alternative to fixed-distance voxel-based radiomics captures previously unacquired textural information while remaining clinically interpretable. This approach may have broad implications in all applications of clinical radiomic analysis, including disease diagnosis, monitoring, and treatment evaluation.