Biphasic remodeling and trabecular-cortical uncoupling in the jawbone during antiresorptive therapy and medication-related osteonecrosis: A quantitative imaging analysis

Abstract

Background/purpose

Although antiresorptive therapy (ART) is crucial for managing oncology-related skeletal complications, it carries a risk of medication-related osteonecrosis of the jaw (MRONJ). Because conventional radiography relies on subjective interpretation, detecting early necrotic transition remains challenging. This study aimed to establish objective, quantitative imaging criteria for jawbone remodeling and MRONJ onset by evaluating 2D fractal dimension (FD) and 3D CBCT cortical thickness under zoledronic acid or denosumab therapy.

Materials and methods

Panoramic radiographs and CBCT scans of cancer patients receiving zoledronic acid (ZA, n = 155) or denosumab (Dmab, n = 168), and healthy controls (n = 95) were analyzed. Longitudinal FD and contralateral and lesion-site cortical thicknesses were quantified to differentiate systemic remodeling from localized necrosis.

Results

Systemic ART induced widespread jawbone osteosclerosis, characterized by significant longitudinal FD increases (ZA: P < 0.05; Dmab: P < 0.001) and contralateral cortical thickening (P < 0.05). Conversely, MRONJ onset triggered a distinct trabecular-cortical uncoupling: a paradoxical reduction in FD (osteolysis) at the necrotic site (P < 0.05) concurrent with reactive periosteal cortical thickening, most pronounced on the lingual aspect (P < 0.001). ZA induced diffuse sclerosis, whereas Dmab-induced changes were localized to the anterior mandible.

Conclusion

Jawbone trabecular and cortical compartments exhibit distinct remodeling patterns during ART and osteonecrosis. Combining quantitative assessments of these compartments provides an objective, non-invasive method for monitoring skeletal burden and detecting early necrotic transitions before clinical bone exposure occurs.

Introduction

Antiresorptive agents, primarily bisphosphonates and receptor activator of nuclear factor kappa-B ligand (RANKL) monoclonal antibodies such as denosumab, are cornerstones in the clinical management of metabolic and oncological bone disorders.1,2 These therapeutics are extensively employed to treat osteoporosis3,4 and to prevent skeletal-related events (SREs)—including pathological fractures, spinal cord compression, and bone pain—in patients presenting with multiple myeloma or bone metastases from solid tumors.5–7 Despite their shared clinical objective of suppressing osteoclast-mediated bone resorption, these classes of drugs exert their pharmacological effects through distinct molecular pathways. Bisphosphonates exhibit a high affinity for hydroxyapatite crystals in the bone matrix, where they are internalized by active osteoclasts, leading to cellular apoptosis and a biological half-life that can exceed ten years.8 Conversely, denosumab targets circulating RANKL directly, preventing osteoclast maturation and survival without matrix incorporation, resulting in a reversible therapeutic effect and a short clearance half-life of only a few weeks to months.9

Although highly effective in preserving systemic skeletal integrity, the profound inhibition of bone remodeling by high-dose antiresorptive therapy compromises the jawbone’s capacity for physiological repair, creating a risk for medication-related osteonecrosis of the jaw (MRONJ).10,11 Typically triggered by dentoalveolar trauma, extraction, or localized infection, MRONJ clinically presents as exposed necrotic bone in the oral cavity, often accompanied by persistent pain, paresthesia, and secondary soft-tissue infection.12 In oncology cohorts receiving high-dose antiresorptive regimens, the incidence of MRONJ can reach up to 5% and is strongly correlated with both cumulative dosage and duration of therapy.12–14

Detecting the early transition toward osteonecrosis remains a major diagnostic hurdle. While established MRONJ is diagnosed via clinical criteria, early or atypical presentations (such as Stage 0) are difficult to identify. Traditional radiographic evaluations, including panoramic radiography and computed tomography, often reveal signs of drug-induced sclerosis, such as lamina dura thickening15 or generalized osteosclerosis,16 which overlap significantly with the normal physiological bone remodeling response to these medications. Consequently, qualitative radiographic assessments frequently fail to differentiate drug-induced adaptive sclerosis from the early, pre-clinical stages of active bone necrosis, limiting their utility in clinical decision-making.17

To introduce objectivity into radiographic bone assessment, investigators have explored quantitative imaging biomarkers. Fractal dimension (FD) analysis, utilizing the box-counting method on digital radiographs, offers a mathematically robust and non-invasive approach to quantify the architectural complexity of trabecular bone networks.18 The validity of FD analysis has been established in predicting dental implant stability,19 monitoring periapical lesion healing, 20 and screening for systemic osteoporosis with high sensitivity and specificity.21 While preliminary investigations have applied FD or textural parameters to evaluate jawbone alterations in patients on antiresorptives,22–24 these studies have generally been constrained by small sample sizes, cross-sectional designs, or a focus on single imaging modalities, leaving the relationship between distinct anatomical compartments uncharacterized.

We propose that the biomechanical and metabolic adaptation of the jawbone under antiresorptive pressure cannot be fully captured by looking at either the trabecular compartment or the cortical shell in isolation. The structural integrity of the mandible depends on a coordinated relationship between the trabecular network and the outer cortical plate.25 Therefore, the objective of this study was to systematically evaluate 2D panoramic FD (representing the trabecular micro-architecture) and 3D CBCT dimensions (representing the cortical shell) in a large cohort of cancer patients receiving zoledronic acid or denosumab. By defining this biphasic remodeling model and its spatial characteristics, we aimed to establish objective, quantitative radiographic criteria to monitor skeletal burden and detect the onset of osteonecrosis. Here, the term “trabecular-cortical uncoupling” is used to describe the divergent macro-structural or compartmental radiographic response of the trabecular (osteolysis) and cortical (reactive thickening) compartments at the necrotic site, distinct from the cellular-level coupling between osteoclastic resorption and osteoblastic bone formation.

Materials and methods

Study design

This retrospective cohort study evaluated radiographic records of oncology patients receiving high-dose ART who presented to the Department of Oral and Maxillofacial Surgery at a tertiary referral center between 2010 and 2022. The study protocol was approved by the Institutional Review Board (Approval No. 202312084RINA), and the requirement for informed consent was waived due to the retrospective nature of the database review.

Patient selection

To systematically separate systemic drug-induced remodeling from localized pathological necrosis, we structured the evaluation into two imaging cohorts:

1. Fractal dimension cohort (Panoramic radiography): A total of 323 oncology patients were evaluated. To track systemic drug effects longitudinally, we analyzed patients who did not develop MRONJ (zoledronic acid [ZA] subgroup, n = 91; denosumab [Dmab] subgroup, n = 75) using pre-medication and post-medication radiographs. To characterize the transition at disease onset, we analyzed patients who developed MRONJ (ZA subgroup, n = 64; Dmab subgroup, n = 93) using radiographs obtained at the time of clinical diagnosis.

2. Cortical thickness cohort (CBCT imaging): Due to clinical protocol, CBCT scans were exclusively obtained for patients presenting with suspected or established MRONJ. Thus, this cohort comprised patients with established MRONJ (ZA group, n = 64; Dmab group, n = 93). To evaluate the systemic drug-induced cortical changes while avoiding the localized inflammatory confounding of the osteonecrotic lesions, measurements were taken at the contralateral healthy sites of these patients. These were compared with a baseline control group of completely healthy, non-medicated, age-matched patients (n = 95) without any history of cancer, chemotherapy, or boneactive medications, who underwent CBCT for dental implant planning. Localized pathological cortical changes were evaluated by comparing the active lesion site directly with the contralateral healthy site within the same patient.

Image acquisition and standardization

All panoramic radiographs were captured using standardized exposure parameters adjusted for patient size on a single unit. Images were exported as uncompressed, high-resolution TIFF files (8-bit grayscale) for digital image analysis. CBCT scans were obtained using a standardized protocol on a 3D Accuitomo 170 unit (J. Morita Mfg. Corp., Kyoto, Japan) operated at 89.0 kVp and 5.5 mA, with an exposure time of 17.5 s, voxel size of 0.25 mm, and a slice thickness of 1.0 mm. Cross-sectional reconstructions were standardized at the level of the maximum mental foramen opening, perpendicular to the curvature of the mandibular arch.

Fractal dimension (FD) analysis

FD analysis was performed using ImageJ software (U.S. National Institutes of Health, Bethesda, MD, USA). Six standardized square regions of interest (ROIs), measuring 50 × 50 pixels, were selected to evaluate trabecular micro-architecture (Fig. 1A, left panel): (1) Mandibular anterior (ROI-1, ROI-2): Located bilateral and mesial to the mental foramina, positioned in the trabecular bone area between the roots of the premolars and the lower border of the mandible. (2) Mandibular posterior (ROI-3, ROI-4): Located bilateral in the periapical regions of the mandibular first molars, positioned to exclude root apices, lamina dura, and the inferior alveolar canal. (3) Maxillary posterior (ROI-5, ROI-6): Located bilateral in the maxillary tuberosity regions, positioned to avoid the maxillary sinus floor and adjacent root apices. The image processing protocol followed the method described by Yasar and Akgünlü,26 which includes Gaussian blur application, image subtraction, binary conversion, erosion and dilation, inversion, skeletonization, and calculation (Fig. 1A, right panel). The box-counting algorithm in ImageJ software was used to calculate FD.

Figure 1.

Methodologies for quantitative imaging analysis of trabecular and cortical bone compartments. (A) Standardized regions of interest (ROIs) and image processing steps for fractal dimension (FD) analysis. The left schematic displays a panoramic view with six standardized ROIs (mandible: 1–4; maxilla: 5–6). The right panel shows representative images pre- and post-zoledronic acid therapy, along with the sequential ImageJ processing workflow (Gaussian blur, subtraction, binarization, erosion/dilation, inversion, skeletonization). (B) Standardized landmarks and parameters for 3D CBCT cortical bone thickness measurements. A: Vertical bone thickness (VBT); B: Vertical cortical bone thickness (VCBT); C: Transverse bone thickness (TBT); D: Buccal cortical bone thickness (BCBT); E: Lingual cortical bone thickness (LCBT) (C–D) Representative slices show the contralateral healthy site and the active MRONJ lesion site with localized lingual cortical thickening (LCBT).

Cortical bone thickness measurement

CBCT paraxial slices at the mental foramen level were used to measure five primary anatomical parameters and two derived ratios, including vertical bone thickness (VBT), vertical cortical bone thickness (VCBT), transverse bone thickness (TBT), buccal (BCBT) and lingual (LCBT) cortical bone thicknesses, vertical cortical bone ratio (VCBR) and transverse cortical bone ratio (TCBR); the schematic definitions of these standardized parameters are illustrated in Fig. 1B. For patients with MRONJ, measurements were taken at both the active lesion site and the contralateral healthy site. The contralateral healthy site represented the systemic drug effect, while the lesion site captured the localized pathological changes, as demonstrated by representative CBCT cross-sections in Fig. 1C and D, respectively.

Reliability and general statistical analysis

All imaging measurements were performed by a single, trained examiner (Lee JC). To evaluate intra-observer reliability, a random subset of 10% of the images was re-measured after a one-month interval. Reliability was assessed using the intraclass correlation coefficient (ICC; two-way mixed-effects, absolute agreement, single-rater model), which yielded 0.812 for panoramic FD measurements and 0.972 for CBCT cortical thickness measurements, indicating good to excellent reproducibility. Statistical analyses were performed using SPSS software (MAC version 26.0, IBM Corp., Armonk, NY, USA) and Python (version 3.13, using pandas, numpy, and scikit-learn libraries). Continuous variables are presented as mean ± standard deviation (SD). Longitudinal changes in FD (pre-vs. post-medication) were evaluated using paired ttests. Cross-sectional differences in cortical thickness among the control, ZA, and Dmab groups were analyzed using one-way ANOVA, followed by Tukey’s HSD (for homogeneous variances) or Games-Howell post-hoc tests (where variance homogeneity was violated). Localized differences (lesion site vs. contralateral healthy site) were assessed using paired t-tests. To control for baseline demographic imbalances in sex and primary diagnoses within the Dmab cohort, a multivariable linear regression analysis was conducted. The model utilized the post-medication FD value of the anterior mandible (ROI-1) as the dependent variable, with MRONJ status, sex, and age entered as independent covariates.

Receiver operating characteristic (ROC) and diagnostic model analysis

To determine optimal clinical diagnostic thresholds, receiver operating characteristic (ROC) curve analysis was performed. A self-controlled paired diagnostic model was constructed where each patient’s active MRONJ lesion site represented the positive class (ONJ = 1) and their own contralateral healthy site represented the negative control class (ONJ = 0). ROC curves and area under the curve (AUC) were computed for 2D FD (hypothesizing lower FD at the lesion site) and 3D CBCT lingual cortical bone thickness (LCBT, hypothesizing higher thickness representing reactive thickening). Optimal thresholds were determined by maximizing the Youden index (J = sensitivity + specificity - 1). Additionally, joint serial (AND) and parallel (OR) diagnostic rules were evaluated, and a combined diagnostic score was generated using multivariable logistic regression. Statistical significance was set at P < 0.05.

Results

Demographic and clinical characteristics

The demographic characteristics of the FD analysis cohort are summarized in Supplementary Table 1 (https://jds.ads.org.tw/cgi/editor.cgi?article=3698&window=additional_files&context=journal). Within the ZA cohort, patients who developed MRONJ received a significantly higher cumulative drug dose compared to those who did not develop the disease (31.16 ± 22.00 vs. 13.60 ± 14.84 doses; P < 0.001) and were slightly older (69.11 vs. 65.43 years; P = 0.044). In the Dmab cohort, significant demographic heterogeneity was observed between subgroups: the non-MRONJ subgroup was predominantly male (92.0%) with a primary diagnosis of prostate cancer (86.7%), whereas the MRONJ subgroup exhibited a more balanced distribution of sex (P < 0.001) and primary diagnoses (P < 0.001), reflecting the higher clinical inclusion of female breast cancer and lung cancer patients who developed MRONJ. The demographics of the cortical thickness cohort (Supplementary Table 2 (https://jds.ads.org.tw/cgi/editor.cgi?article=3698&window=additional_files&context=journal)) showed no significant differences in mean age among the control, ZA, and Dmab groups (P = 0.920). However, the antiresorptive-treated groups exhibited a significantly higher frequency of concurrent corticosteroid use compared to the control group (P = 0.001). Additionally, the mean duration of therapy was significantly longer in the ZA group than in the Dmab group (39.06 ± 27.81 vs. 26.57 ± 23.75 months; P = 0.004).

Longitudinal changes in trabecular FD: the systemic phase

Longitudinal analysis of the non-MRONJ subgroups revealed a distinct anatomical pattern of bone remodeling following antiresorptive exposure (Table 1). Mandibular trabecular complexity significantly increased in both the ZA group (mandible pooled: 1.431 ± 0.105 to 1.456 ± 0.107; P < 0.001) and the Dmab group (Mandible Pooled: 1.430 ± 0.081 to 1.451 ± 0.072; P < 0.001). Conversely, no significant longitudinal alterations were detected in the maxilla for either drug (P > 0.05). We observed drug-specific spatial differences in mandibular remodeling. In the ZA group, significant increases in FD occurred in both the anterior mandible (ROI-1, P = 0.013; ROI-2, P = 0.007) and the posterior mandible (ROI-4, P = 0.003). In contrast, denosumab-induced sclerosis was localized to the anterior mandible (ROI-1, P < 0.001), while the posterior mandibular regions remained structurally stable (P > 0.05).

Table 1.

Longitudinal changes in fractal dimension (FD) pre- and post-medication in non-MRONJ subgroups.

ROI location	Zoledronic acid (ZA) group (n = 91)				Denosumab (Dmab) group (n = 75)
	na	Pre-medication	Post-medication	P-valueb	na	Pre-medication	Post-medication	P-valueb
ROI-1 (ant. mandible, R)	91	1.436 ± 0.099	1.460 ± 0.106	0.013	75	1.419 ± 0.070	1.459 ± 0.056	<0.001
ROI-2 (ant. mandible, L)	91	1.432 ± 0.095	1.456 ± 0.103	0.007	75	1.437 ± 0.071	1.455 ± 0.059	0.053
ROI-3 (post. mandible, R)	89	1.440 ± 0.105	1.455 ± 0.102	0.164	74	1.439 ± 0.081	1.448 ± 0.083	0.423
ROI-4 (post. mandible, L)	90	1.416 ± 0.119	1.453 ± 0.118	0.003	75	1.426 ± 0.098	1.443 ± 0.087	0.179
ROI-5 (maxilla, R)	84	1.395 ± 0.145	1.411 ± 0.152	0.315	67	1.380 ± 0.106	1.371 ± 0.118	0.575
ROI-6 (maxilla, L)	86	1.372 ± 0.171	1.365 ± 0.150	0.703	68	1.373 ± 0.125	1.367 ± 0.118	0.730
Pooled analysis c
Maxilla (total sites)	170	1.383 ± 0.159	1.388 ± 0.152	0.450	135	1.376 ± 0.115	1.369 ± 0.117	0.650
Mandible (total sites)	361	1.431 ± 0.105	1.456 ± 0.107	<0.001	299	1.430 ± 0.081	1.451 ± 0.072	<0.001
Overall pooling	531	1.416 ± 0.126	1.434 ± 0.127	<0.001	434	1.414 ± 0.096	1.426 ± 0.096	0.019

Abbreviations: ROI, region of interest; Ant., anterior; Post., posterior; R, right; L, left.

^an represents the number of measurable sites (excludes files with insufficient image quality).

^bCalculated using paired t-tests.

^cRepresents the aggregation of all measurable ROI within the respective anatomical zones (Maxilla: ROI 5–6; Mandible: ROI 1–4).

Localized trabecular collapse at MRONJ onset

In patients who developed MRONJ, X-rays obtained at disease onset revealed a significant reduction in trabecular FD values compared to the post-treatment values of patients who did not develop MRONJ (Supplementary Table 3 (https://jds.ads.org.tw/cgi/editor.cgi?article=3698&window=additional_files&context=journal)). For the ZA cohort, this reduction was widespread, showing statistical significance across all mandibular regions (ROI-1 to ROI-4; P < 0.05). For the Dmab cohort, the reduction was localized to the anterior mandible (ROI-1; P = 0.001). To verify that the association between MRONJ onset and reduced FD was not confounded by the demographic imbalances in the Dmab group (such as sex and age), we performed a multivariable linear regression (Supplementary Table 4 (https://jds.ads.org.tw/cgi/editor.cgi?article=3698&window=additional_files&context=journal)). The model confirmed that the presence of MRONJ remained a significant independent predictor of reduced FD values in the anterior mandible (standardized β = −0.214, P = 0.011), while age (P = 0.630) and sex (P = 0.105) exerted no significant influence.

Systemic drug effects on cortical bone thickness

To quantify systemic, drug-induced cortical changes, CBCT measurements at the contralateral healthy sites of the medicated cohorts were compared with the non-medicated control group (Table 2). Both antiresorptive groups demonstrated significant cortical thickening compared to controls. Post-hoc pairwise comparisons (Supplementary Table 5 (https://jds.ads.org.tw/cgi/editor.cgi?article=3698&window=additional_files&context=journal)) revealed that both the ZA and Dmab groups exhibited significantly greater vertical cortical bone thickness (VCBT), vertical cortical bone ratio (VCBR), and transverse cortical bone ratio (TCBR) compared to controls (P < 0.05), with no significant differences in these vertical parameters between the two drug groups. However, differences were observed in the transverse dimension: the ZA group demonstrated significantly greater buccal cortical thickness (BCBT, mean difference = 0.021 mm, P = 0.002) and lingual cortical thickness (LCBT, mean difference = 0.028 mm, P < 0.001) compared to the Dmab group.

Table 2.

Comparison of mandibular cortical bone thickness among study groups (Systemic medication effect evaluated at contralateral healthy sites in MRONJ patients).

Measurement parameter	Medications
	Control (n = 95)	Zoledronic acid (n = 64)	Denosumab (n = 93)	P-valuea
Measurement sitesb, n	188	115	167
Vertical measurements, mean ± SD
VBT (mm)	1.01 ± 0.13	1.06 ± 0.15	1.00 ± 0.14	0.002
VCBT (mm)	0.37 ± 0.09	0.43 ± 0.10	0.42 ± 0.09	<0.001
VCBR (ratio)	0.37 ± 0.09	0.41 ± 0.10	0.43 ± 0.10	<0.001
Transverse measurements, mean ± SD
TBT (mm)	1.11 ± 0.15	1.16 ± 0.14	1.10 ± 0.18	0.015
BCBT (mm)	0.22 ± 0.05	0.27 ± 0.06	0.25 ± 0.05	<0.001
LCBT (mm)	0.21 ± 0.04	0.25 ± 0.07	0.23 ± 0.05	<0.001
TCBR (ratio)	0.39 ± 0.07	0.45 ± 0.09	0.43 ± 0.07	<0.001

Abbreviations: VBT, vertical bone thickness; VCBT, vertical cortical bone thickness; VCBR, vertical cortical bone ratio; TBT, transverse bone thickness; BCBT, buccal cortical bone thickness; LCBT, lingual cortical bone thickness; TCBR, transverse cortical bone ratio.

^aCalculated using one-way ANOVA.

^bFor the control group, measurement sites represent bilateral measurements from 95 non-medicated subjects (with 2 sites excluded due to artifacts). For the zoledronic acid and denosumab groups, measurement sites represent all available healthy (non-lesion) mental foramen sites from these patients; since many patients had lesions located outside the mental foramen region, both left and right sites could be included as healthy representative sites of systemic drug exposure.

Localized cortical changes: pathological periosteal reaction

Comparing the active MRONJ lesion site directly to the contralateral healthy site within the same patient revealed localized cortical bone changes (Table 3). In the ZA group, the MRONJ lesion site exhibited significant thickening across nearly all cortical parameters, including VCBT, VCBR, BCBT, LCBT, and TCBR, compared to the healthy contralateral side (P < 0.001). In the Dmab group, the localized response was anatomically specific: significant localized thickening at the lesion site was observed in lingual cortical bone thickness (LCBT, 0.24 ± 0.06 vs. 0.20 ± 0.04 mm; P < 0.001) and transverse bone thickness (TBT, 1.23 ± 0.17 vs. 1.16 ± 0.13 mm; P = 0.016). In contrast, buccal cortical parameters (BCBT and TCBR) showed only borderline significance (P = 0.059 and P = 0.060, respectively), indicating that the localized reactive cortical process in denosumab-treated patients is primarily focused on the lingual aspect.

Table 3.

Comparison of cortical bone thickness at medication-related osteonecrosis of the jaw (MRONJ) lesion site vs. contralateral healthy site within patients.

Measurement parametersa	Zoledronic acid (ZA) group (n = 47 pairs)			Denosumab (Dmab) group (n = 60 pairs)
	ONJ site	Contralateral	P-valueb	ONJ site	Contralateral	P-valueb
Vertical measurements, mean = SD
VBT (mm)	0.97 ± 0.15	0.96 ± 0.16	0.769	0.90 ± 0.16	0.90 ± 0.15	0.809
VCBT (mm)	0.43 ± 0.09	0.35 ± 0.10	<0.001	0.40 ± 0.09	0.38 ± 0.10	0.310
VCBR (ratio)	0.45 ± 0.09	0.37 ± 0.11	<0.001	0.44 ± 0.10	0.43 ± 0.11	0.374
Transverse measurements, mean ± SD
TBT (mm)	1.26 ± 0.16	1.18 ± 0.17	0.016	1.23 ± 0.17	1.16 ± 0.13	0.016
BCBT (mm)	0.28 ± 0.06	0.20 ± 0.05	<0.001	0.26 ± 0.06	0.24 ± 0.05	0.059
LCBT (mm)	0.28 ± 0.07	0.19 ± 0.05	<0.001	0.24 ± 0.06	0.20 ± 0.04	<0.001
TCBR (ratio)	0.45 ± 0.08	0.33 ± 0.08	<0.001	0.41 ± 0.08	0.38 ± 0.07	0.060

Abbreviations: VBT, vertical bone thickness; VCBT, vertical cortical bone thickness; VCBR, vertical cortical bone ratio; TBT, transverse bone thickness; BCBT, buccal cortical bone thickness; LCBT, lingual cortical bone thickness; TCBR, transverse cortical bone ratio; ONJ, osteonecrosis of the jaw.

^aValues are expressed as mean ± SD.

^bCalculated using paired t-tests. Note that n values represent the complete paired patient-level data where both contralateral and ipsilateral measurements were achievable.

Receiver operating characteristic (ROC) and combined diagnostic models

To establish clinically applicable diagnostic thresholds, ROC curve analyses were conducted on paired patient-level data (n = 107 pairs with complete imaging and clinical records; Table 4 and Fig. 2). For the ZA cohort (n = 47 pairs), the optimal cutoff for 2D FD was FD < 1.3610 (AUC = 0.6388; sensitivity = 36.17%, specificity = 91.49%), indicating high specificity for bone micro-architectural collapse (Fig. 2A). For CBCT lingual thickness, the optimal cutoff was LCBT >0.2400 mm (AUC = 0.7311; sensitivity = 78.72%, specificity = 68.09%). Integrating both variables in a multivariable logistic regression model yielded a combined AUC of 0.7266 with the formula: logit (P) = 1.0265 – 0.9039 × FD + 0.9680 × LCBT. Evaluation of joint decision rules revealed that a serial “AND” logic (FD < 1.460 and LCBT >0.235 mm) provided a balanced sensitivity of 68.09% and specificity of 72.34%. A parallel “OR” logic (FD < 1.300 or LCBT >0.235 mm) maximized screening sensitivity to 82.98% (specificity = 65.96%). A strict clinical rule requiring FD < 1.465 and local lingual cortical thickening >5.0% relative to the contralateral healthy side achieved a diagnostic specificity of 100.00% in this study cohort.

Table 4.

Receiver operating characteristic (ROC) and joint diagnostic decision rules for zoledronic acid and denosumab cohorts.

Cohort	Diagnostic modality/rule	Optimal cutoff	AUC	Sensitivity	Specificity
Zoledronic acid (n = 47 pairs)	2D panoramic FD (lower)	FD < 1.3610	0.6388	36.17%	91.49%
	CBCT LCBT (thickening)	LCBT >0.2400 mm	0.7311	78.72%	68.09%
	Combined score (logistic regression)	P > 0.5008a	0.7266	65.96%	74.47%
	Serial joint rule (AND)	FD < 1.460 AND LCBT >0.235 mm	–	68.09%	72.34%
	Parallel joint rule (OR)	FD < 1.300 OR LCBT >0.235 mm	–	82.98%	65.96%
	Strict confirmatory rule (AND)	FD < 1.465 AND thickening >5.0%b	–	59.57%	100.00%
Denosumab (n = 60 pairs)	2D panoramic FD (lower)	FD < 1.3660	0.5968	31.67%	86.67%
	CBCT LCBT (thickening)	LCBT >0.2700 mm	0.5796	28.33%	90.00%
	Combined score (logistic regression)	P > 0.5152c	0.6261	38.33%	90.00%
	Serial joint rule (AND)	FD < 1.475 AND LCBT >0.260 mm	–	23.33%	93.33%
	Parallel joint rule (OR)	FD < 1.345 OR LCBT >0.285 mm	–	43.33%	88.33%
	Modified parallel rule (OR)	FD < 1.345 OR thickening >5.0%b	–	66.67%	91.67%

Abbreviations: FD, fractal dimension; LCBT, lingual cortical bone thickness; AUC, area under the curve.

^aProbability equation: logit (P) = 1.0265 – 0.9039 × FD + 0.9680 × LCBT.

^bLocal lingual cortical thickening percentage calculated as: [(LCBT at lesion site - LCBT at contralateral healthy site)/LCBT at contralateral healthy site] x 100%. The contralateral healthy site has a baseline thickening of 0%.

^cProbability equation: logit (P) = 0.9404 – 0.7406 × FD + 0.5146 × LCBT.

Figure 2.

Receiver operating characteristic (ROC) curves comparing diagnostic performance of independent imaging biomarkers and combined logistic regression models. (A) ROC analysis for the zoledronic acid cohort (n = 47 patient-level pairs). (B) ROC analysis for the denosumab cohort (n = 60 patient-level pairs). The blue dashed line represents 2D panoramic FD, the green dotdashed line represents CBCT LCBT (lingual cortical bone thickness), and the solid red line represents the multivariable combined model. Circles, triangles, and squares indicate the optimal threshold cutoffs determined by maximizing the Youden index.

For the Dmab cohort (n = 60 pairs), the optimal FD cutoff was FD < 1.3660 (AUC = 0.5968; sensitivity = 31.67%, specificity = 86.67%; Fig. 2B). The optimal LCBT cutoff was LCBT >0.2700 mm (AUC = 0.5796; sensitivity = 28.33%, specificity = 90.00%). The combined logistic regression model yielded an AUC of 0.6261 with the formula: logit (P) = 0.9404 – 0.7406 × FD + 0.5146 × LCBT. A joint serial “AND” decision rule (FD < 1.475 and LCBT >0.260 mm) achieved a specificity of 93.33%. A parallel “OR” decision rule (FD < 1.345 or LCBT >0.285 mm) yielded a sensitivity of 43.33% and a specificity of 88.33%. Using a modified parallel rule combining FD < 1.345 or local lingual cortical thickening >5.0% compared to the contralateral healthy side yielded a screening sensitivity of 66.67% and a specificity of 91.67%.

For the pooled cohort (n = 107 pairs), the combined logistic regression model yielded an AUC of 0.6698 with the formula: logit (P) = 1.6675 – 1.3989 × FD + 1.3283 × LCBT, with an optimal probability cutoff of P > 0.5138 (sensitivity = 51.40%, specificity = 79.44%).

Discussion

By evaluating 2D fractal dimension (FD) and 3D CBCT cortical thickness in a large cohort, this study characterizes the temporal and spatial radiographic alterations associated with antiresorptive therapy and MRONJ. The findings define a biphasic radiographic progression model: an initial therapeutic phase of systemic drug-induced osteosclerosis, followed by a localized pathological phase of trabecular-cortical uncoupling at MRONJ onset (Fig. 3). Initially, systemic antiresorptive exposure drives widespread jawbone osteosclerosis, characterized by longitudinal increases in trabecular complexity and generalized cortical thickening. This systemic phase aligns with previous cross-sectional studies reporting increased bone density and texture alterations under antiresorptives.22,27 Biologically, the suppression of osteoclast-mediated bone resorption by zoledronic acid or denosumab leads to the accumulation of trabecular micro-structures,18 and the slow thickening of the outer cortical shell.14,28 Our clinical data support this, showing that MRONJ patients received significantly higher cumulative doses, which corroborates the dose-dependent risk profile described in the literature. Consequently, our longitudinal findings demonstrate that jawbone adaptation represents a slow, cumulative “skeletal burden” of long-term therapy.12

Figure 3.

Schematic illustration of the proposed radiographic progression model depicting changes in quantitative FD and cortical bone thickness during antiresorptive therapy and MRONJ development. The progression is categorized into three phases. Left panel: Baseline/healthy bone. This phase represents the normal radiographic state before medication exposure, with baseline trabecular complexity and cortical thickness. Middle panel: Therapeutic phase -sclerosis & burden. Following the administration of antiresorptive therapy, the jaws exhibit systemic sclerosis, characterized by an increase in trabecular complexity (increased FD) and generalized thickening of the cortical plates, reflecting accumulated skeletal burden. Right panel: MRONJ phase - paradoxical lysis & reactive thickening. Upon the onset of MRONJ predisposing factors, a paradoxical pattern emerges. Localized necrotic osteolysis results in a sharp decrease in FD. Concurrently, cortical thickness continues to increase due to periosteal reaction. Notably, this phase reveals drug specificity in cortical thickening patterns: zoledronic acid tends to cause widespread (buccal and lingual) thickening, whereas denosumab induces preferentially profound lingual thickening. The hallmark diagnostic feature of MRONJ onset: reduced FD combined with continued cortical thickening.

However, the transition to active osteonecrosis is marked by a localized trabecular-cortical uncoupling. While previous investigations evaluated trabecular networks22,27 or cortical shells29,30 in isolation, our model characterizes their concurrent, paradoxical response at the necrotic site. Specifically, the local trabecular network collapses due to osteolysis (evidenced by a drop in FD), while the outer cortical shell undergoes extreme, reactive periosteal thickening. This uncoupling represents a major departure from the coordinated remodeling seen in the therapeutic phase. Recognizing this paradoxical divergence addresses a key gap in early MRONJ detection, shifting focus from individual parameters to the structural relationship between bone compartments. It is important to emphasize that this “uncoupling” represents a divergence in macro-structural compartment behaviors at the pathological site, rather than a disruption of cellular-level osteoclast-osteoblast coupling.

The localized cortical thickening at the MRONJ lesion site—predominantly on the lingual aspect—reflects a reactive periosteal response, functionally analogous to involucrum formation in osteomyelitis.31 Biologically, as the mandible’s centrifugal blood supply from the inferior alveolar artery is compromised during necrosis,32 the jawbone becomes dependent on centripetal perfusion from the periosteum. The highly vascularized lingual periosteum undergoes reactive hyperplasia to wall off the necrotic sequestrum, manifesting radiographically as localized cortical thickening.33 Our findings provide quantitative support for this biological defense mechanism across both zoledronic acid and denosumab cohorts.

We also demonstrated clear, drug-specific spatial differences in mandibular remodeling. Zoledronic acid induced diffuse, widespread sclerosis and greater cortical thickening throughout the mandible, whereas denosumab-induced alterations were localized to the high-stress anterior mandible. This spatial divergence is explained by their different pharmacological profiles: zoledronic acid binds mineral avidly with a skeletal half-life exceeding ten years, driving generalized remodeling suppression. Conversely, denosumab has a short clearance half-life and does not incorporate into the matrix.8,9 Its actions are thus concentrated at sites of high mechanical load and bone turnover, such as the anterior mandibular symphysis. 34 Radiographically, the anterior mandible also represents the region with the least anatomical superimposition and projection distortion in panoramic imaging,26,35 providing a high signal-to-noise ratio that facilitates the detection of these localized, subtle alterations. This pharmacological and radiographic insight can guide clinicians in interpreting radiographic changes based on the specific agent administered.

Our results show that the maxilla remains radiographically stable compared to the mandible. Biologically, the mandible has a higher baseline bone turnover rate, making it more sensitive to antiresorptives.34 Anatomically, the maxilla possesses a thinner cortical shell and a more porous, less organized trabecular network compared to the mandible.36 Radiographically, maxillary analysis on 2D panoramic films is also limited by anatomical superimpositions (e.g., zygomatic process, nasal cavity).37 Therefore, the mandible serves as a more reliable anatomical site for quantitative imaging monitoring of antiresorptive effects.

These findings address a major clinical gap in current MRONJ guidelines. The American Association of Oral and Maxillofacial Surgeons (AAOMS) diagnostic criteria rely heavily on the visual detection of exposed necrotic bone,12 which represents a late stage of the disease. Early-stage or “Stage 0” cases, presenting with pain but no exposed bone, remain diagnostically challenging.24 This clinical challenge has driven interest in radiomics and quantitative texture analysis to identify early trabecular micro-structural alterations.38,39 By applying our ROC-derived joint decision rules, clinicians can utilize routine panoramic and CBCT imaging to identify early-stage lesions. This drug-specific selection of diagnostic rules reflects the different clinical challenges and biological profiles of the two cohorts. In patients treated with zoledronic acid—who exhibit pronounced, widespread sclerosis—the priority is to confirm necrotic transition while avoiding false positives. Consequently, a serial confirmatory rule (AND) is clinically optimal, yielding a specificity of 100.00% in this cohort. Conversely, in the denosumab cohort—where radiographic changes are highly localized and subtle—the primary clinical challenge is avoiding missed diagnoses of early lesions. Thus, a parallel screening rule (OR) is preferred, boosting sensitivity to 66.67% while maintaining a high specificity of 91.67%. Furthermore, the diagnostic performance of the individual and combined models for the denosumab cohort was relatively modest compared to the zoledronic acid cohort (combined AUC of 0.6261 vs. 0.7266). This difference is likely due to the localized and subtle radiographic signature of denosumab-induced alterations compared to the diffuse, long-lasting sclerosis of zoledronic acid. Therefore, these quantitative radiographic criteria should be utilized as supplementary diagnostic aids to support clinical decision-making, rather than as standalone diagnostic tools.

Several limitations should be noted. First, the retrospective design prevented standardized timing for radiographic follow-up. Second, the FD cohort lacked an untreated cancer control group; however, the significant longitudinal increase in FD observed here contrasts with the age-related decline or stability reported in untreated populations,21,40 supporting a drug-induced effect. Third, because the control group consisted of healthy individuals, we cannot completely rule out the potential confounding effects of underlying primary malignancies or concurrent cancer treatments (e.g., chemotherapy, corticosteroids) on baseline bone dimensions in the study groups. Fourth, the denosumab cohort exhibited baseline demographic imbalances in sex and primary diagnoses. We addressed this by performing a multivariable regression analysis (Supplementary Table 4 (https://jds.ads.org.tw/cgi/editor.cgi?article=3698&window=additional_files&context=journal)), which confirmed that the association between MRONJ and reduced FD in the anterior mandible remained significant independent of age and sex. Finally, this study evaluated active antiresorptive users; future research should investigate the radiographic changes during drug holidays, particularly given the rapid rebound in bone turnover associated with denosumab discontinuation.41,42

In summary, this study proposes a quantitative progression model of jawbone remodeling under antiresorptive therapy (Fig. 3). While systemic drug exposure drives widespread trabecular complexity and cortical thickening (therapeutic phase), the transition to active MRONJ is characterized by localized trabecular-cortical uncoupling (osteolytic collapse and reactive periosteal thickening). These quantitative parameters may serve as objective tools to monitor skeletal burden and identify early pathological transitions in patients receiving antiresorptives.

Funding Statement

This work was partially supported by the National Science and Technology Council, Taiwan, under grant [111-2314-B-002-063] to Ling-Ying Wei.