Paired Structural Characterization of Trabecular Bone Patterns in Panoramic Dental Radiographs Using Monofractal, Surface, Multifractal, and Topological Descriptors

Quantitative analysis of trabecular bone patterns in routine 2D panoramic dental radiographs (orthopantomograms / OPGs) offers potential non-invasive insights into regional jaw micro-architecture. However, existing texture analysis literature exhibits limited reproducibility due to inconsistent image preprocessing, variable region-of-interest (ROI) selection, uncontrolled dataset heterogeneity, and a lack of cryptographically sealed, patient-level out-of-sample holdout evaluations. This study evaluates whether mathematically defined image descriptors under a fixed geometric spatial reference framework yield reproducible paired structural measurements across independent patient cohorts in panoramic dental radiographs, while systematically assessing multi-cohort robustness, descriptor collinearity, and anatomical interpretability limitations. A total benchmark sample of N = 2,461 unique 2D panoramic dental radiographs across three cohorts was evaluated: Roboflow Development (N = 657 images / 643 unique patients), Zenodo (N = 1,611 unique images), and an independent cryptographically sealed Holdout cohort (N = 193 images / 162 unique patients; SHA-256 hash 85e3a856...). Four mathematically validated descriptors—box-counting fractal dimension (D0), Minkowski surface blanket dimension (DM), multifractal spectrum width (Delta alpha), and Betti-0 connected-component count (beta0)—were locked under frozen preprocessing pipelines. Two deterministic fixed spatial references (64 x 64 pixels) were positioned at Crest (0.50w, 0.45h) and Basal (0.50w, 0.80h) locations. All four locked descriptors demonstrated statistically significant paired structural contrasts across cohorts. In the independent Holdout cohort, effect sizes were: D0 = -0.2321 (95% CI [-0.4071, -0.0857]), DM = -0.8656 (95% CI [-1.0559, -0.6984]), Delta alpha = +0.5893 (95% CI [+0.4800, +0.6979]), and beta0 = +0.3303 (95% CI [+0.2614, +0.4674]). Variance Inflation Factors in paired-difference representations were all below 2.50 (D0: 1.62, DM: 2.15, Delta alpha: 1.84, beta0: 1.41), indicating limited linear multicollinearity. Delta alpha preserved the positive direction across cohorts, with a smaller effect magnitude in Zenodo (dz = +0.2783) than in Roboflow Development (dz = +0.8753). Blinded expert evaluation confirmed usable trabecular tissue in 88.0% of Crest readings and 95.0% of Basal readings, while identifying root overlap interference in 19.0% of Crest readings. This study establishes out-of-sample reproducibility of paired regional structural contrasts in 2D panoramic dental radiographs under a locked computational framework. This is a 2D regional characterization study; no clinical diagnostic utility, disease classification, or anatomical independence is claimed.