Physics-Informed Multiscale Decoding of Tissue Microstructure: The Gray Level Affinity Metrics (GLAM) Framework

Drawing inspiration from statistical mechanics, which provides a rigorous framework for understanding how microscopic interactions give rise to macroscopic structures, we propose a novel family of physics-inspired texture descriptors termed Gray Level Affinity Metrics (GLAM). Unlike conventional radiomics, which relies on localized co-occurrence statistics, GLAM treats image voxels as interacting particles and utilizes Radial Distribution Functions to characterize spatial organization across a continuous range of length scales, yielding physics-inspired structural analogues of the macroscopic tumor architecture. Quantitative evaluation reveals that GLAM possesses significantly higher intrinsic dimensionality than standard texture metrics, capturing substantial non-redundant information that drives superior variance in multi-parametric imaging space. In a multi-center high-grade glioma cohort, this informational density translated to targeted, subgroup-specific prognostic performance. Under rigorous Leave-One-Center-Out cross-validation and 2000 independent bootstrap iterations, the models achieved statistically significant risk stratification. In the treatment-responsive MGMT promoter-methylated phenotype, standalone GLAM emerged as the optimal framework (Mean Test C-Index 0.646, yielding a 668-day median survival separation between risk tiers). Conversely, in the highly aggressive MGMT promoter-unmethylated phenotype, a Combined model synergizing GLAM and conventional radiomics established a mathematically sound risk gradient (Mean Test C-Index 0.643, 267-day separation) while achieving the smallest cross-institutional stability (Δ 0.067 performance spread). Ultimately, this differential feature selection underscores a critical biological synergy: while conventional radiomics capture short-range, localized intensity fluctuations, the GLAM framework characterizes overarching multiscale architecture using statistical physics analogs, working in concert to maximize precision prognostication and domain resistance across diverse molecular cohorts.