Multiscale Radiopathomic Fusion Identifies Topological Drivers of Immune Exclusion and Clinical Progression In Oropharyngeal Carcinoma
Abstract
Purpose
Personalized oncology requires bridging the gap between macro-scale imaging and micro-scale cellular architecture. We developed a radiopathomic fusion framework to integrate PET/CT metabolic heterogeneity with graph-based topological modeling of the tumor microenvironment (TME) to identify the structural drivers of treatment resistance in Oropharyngeal Carcinoma (OPC).
Methods
A multiscale pipeline integrated pre-treatment PET/CT scans with digitized H&E slides. Pathomic features were extracted via deep-learning nuclei segmentation and graph spectral analysis, quantifying TME integration through Fiedler values (lambda) and H1 persistent homology. To preserve scale-specific biological signals, we performed intensity discretization benchmarking: an optimization framework evaluated a range of metabolic intensity windows (FBN 32, 64, 128, and 256). Feature selection was performed using Partial Least Squares regression and Leave-One-Out Cross-Validation to identify the scales that maximized the association between modalities and clinical progression
Results
Discretization-specific radiomic features serve distinct roles. Coarse-scale discretization (FBN32) was the primary prognostic signature; High Gray Level Size Emphasis stratified progression-free survival (plog-rank = 0.0412), representing the global metabolic behavior of the bulk tumor. Conversely, fine-scale discretization (FBN256) served as the structural mechanistic link to the microenvironment: Small Size Low Gray Level Emphasis demonstrated the highest global importance (0.499) and was spatially co-located with pathomic exclusion gaps (Moran I = 0.65, p < 0.001). These optimized features mapped directly to a fragmented-resistant pathomic phenotype (low Fiedler values, lambda < 10-5) where high H1 persistence identified stable immune-cold vacuoles across scales.
Conclusion
The selection of discretization is an empirical determinant of biological scale. Emergence FBN32 and FBN256 through cross-validated optimization suggests the global prognostic signature and the micro-scale topological fragmentation are coupled to discretization-specific radiomic phenotypes.This multiscale approach provides a potential framework for stratifying patients at risk of a treatment failure from standard-of-care chemoradiotherapy.