Radiomics-based quantitative assessment of the mandibular cortical index on panoramic radiographs

Abstract

This study aimed to quantitatively evaluate mandibular cortical changes related to the Mandibular Cortical Index (MCI) using radiomic features extracted from panoramic radiographs in order to improve the objectivity and reproducibility of MCI assessment. A total of 159 panoramic radiographs (C1 = 53, C2 = 53, C3 = 53) were retrospectively analyzed. The right mandibular cortex was segmented following standardized anatomical boundaries, and radiomic features were extracted using PyRadiomics in 3D Slicer. Five feature classes—shape, first-order, GLCM, GLRLM, and GLSZM—were evaluated. Inter-observer reproducibility was assessed using intraclass correlation coefficients (ICC), and only features with ICC > 0.75 were included. Statistical significance was set at p < 0.05. Several first-order features, including Energy (p = 0.006), Entropy (p = 0.001), IQR (p = 0.002), MAD (p = 0.002), rMAD (p = 0.001), and Variance (p = 0.001), differ significantly across groups, consistently distinguishing C1 from resorbed cortices. Texture analysis revealed significant differences in multiple GLCM features, such as Correlation (p < 0.001), Cluster Shade (p = 0.010), and Joint Entropy (p = 0.006). Gray Level Variance (GLRLM) (p = 0.001) and Zone Entropy (GLSZM) (p = 0.002) also demonstrated significant group effects. However, after Benjamini–Hochberg false discovery rate (FDR) correction for multiple comparisons, no radiomic feature remained significantly different between the moderately (C2) and severely (C3) resorbed cortices. Radiomics enables objective quantification of morphometric and microtextural differences associated with MCI categories and may enhance the objectivity of mandibular cortical assessment on panoramic radiographs.

Introduction

Osteoporosis (OP) is a systemic skeletal disorder with increasing global prevalence, largely driven by population aging, and it carries substantial clinical and public health implications¹. OP leads to reduced bone mineral density, resulting in cortical thinning and diminished structural strength, which substantially increases the risk of fractures². OP is a prevalent public health condition affecting 30–50% of women and 15–30% of men, and the increasing incidence of osteoporotic fractures in aging populations underscores the critical importance of early diagnosis at both individual and societal levels^3,4.

Given the substantial clinical burden of OP, early diagnostic modalities that are both accurate and widely accessible are essential for timely intervention. Dual-energy X-ray absorptiometry (DXA) remains the gold standard for evaluating bone mineral density; however, its high cost, limited availability, and limited efficiency for population-based screening considerably restrict its utility in routine clinical practice⁵. This has underscored the need for more accessible, cost-effective, and opportunistic screening approaches.

Previous research has shown a strong association between dental imaging findings and systemic bone quality^6–11. Panoramic radiograph (PR), which are routinely obtained, low-cost, and low-radiation imaging tools, therefore present a practical opportunity for opportunistic OP screening. The mandibular cortex visible on PR reflects systemic skeletal changes, and radiomorphometric indices such as the mandibular cortical index (MCI), cortical width, and the panoramic mandibular index have been associated with reduced bone mineral density^12–14. MCI has been widely used as a radiographic marker of low bone mineral density, its assessment relies on subjective visual interpretation, which introduces considerable inter- and intra-observer variability. These limitations have generated growing interest in quantitative, observer-independent approaches capable of capturing cortical changes more objectively and consistently. Recent studies further indicate that quantitative image-based analysis of the mandibular cortex may improve the ability of PR to identify radiographic patterns associated with reduced bone mineral density¹⁵.

Radiomics has become an increasingly valuable tool in bone research due to its ability to quantify textural patterns, structural heterogeneity, and cortical integrity—features that reflect bone quality beyond what can be visually assessed^16,17. Although radiomics was initially developed for oncologic applications, its use in maxillofacial imaging has expanded, with studies demonstrating its applicability to the characterization of bone structures and mandibular morphology^18–20. Radiomics-based AI models derived from PRs have shown promising predictive performance for OP, with some deep-learning methods outperforming classical radiomics feature sets^21,22. However, systematic reviews emphasize ongoing challenges, including image-quality variation, data heterogeneity, and insufficient methodological standardization, which highlight the need for further validation²³. Radiomics has also demonstrated strong predictive capability for OP and osteopenia in other imaging modalities such as quantitative computed tomography (QCT)²⁴. These quantitative descriptors provide insight into microarchitectural characteristics that may correlate with systemic skeletal health, thereby supporting the use of radiomics as a complementary biomarker in OP assessment.

Although PRs show potential for identifying reduced bone mineral density, studies that objectively analyze the MCI using radiomics features remain scarce. Integrating radiomics into MCI assessment may reduce observer dependency and provide a more objective characterization of mandibular cortical morphology across MCI categories. Additionally, existing AI studies indicate that methodological inconsistencies and limited standardization continue to pose challenges²³. Because DXA measurements were not available, the present study was not designed to diagnose osteoporosis but rather to quantitatively evaluate radiomic differences among established MCI categories. This study aimed to evaluate differences in radiomic features across Mandibular Cortical Index (MCI) categories (C1–C3) using panoramic radiographs, in order to improve the objectivity and reproducibility of traditional MCI assessment.

Materials and methods

Study design and ethical considerations

This cross-sectional retrospective study was conducted at the Department of Oral and Maxillofacial Radiology, Necmettin Erbakan University. The study protocol included digital PRs obtained for various diagnostic purposes between January 2023 and October 2025. Ethical approval was granted by the Necmettin Erbakan University Faculty of Dentistry Non-Drug and Non-Medical Device Research Ethics Committee (Approval No: 2025/696), and all procedures were conducted in strict accordance with the principles of the Declaration of Helsinki. Due to the retrospective nature of the study, the Necmettin Erbakan University Faculty of Dentistry Non-Drug and Non-Medical Device Research Ethics Committee waived the requirement for informed consent.

To ensure the study possessed adequate statistical power to detect radiomic differences among the MCI groups, a priori power analysis was performed using G*Power (version 3.1.9.7). The sample size calculation was performed using Cohen’s medium effect size of 0.25²⁵. Based on this effect size (f = 0.25), with a significance level of 5% (α = 0.05) and a statistical power of 80% (1–β = 0.80), the required number of participants per group was determined to be 53.

Image acquisition and initial screening

All PRs were acquired using a 2D Veraviewpocs digital panoramic system (J MORITA MFG Corp., Kyoto, Japan) operating with standardized exposure parameters of 70 kVp, 5 mA, and 15 s. Image evaluation was performed using i-Dixel software (J MORITA) on a 27-inch UltraSharp LED TFT monitor (Dell Inc., Round Rock, TX, USA) with a resolution of 2560 × 1440 pixels to ensure optimal diagnostic clarity.

The inclusion criteria were: (1) female individuals aged 50 years or older; (2) absence of systemic diseases or medications known to affect bone metabolism; and (3) no history of maxillofacial trauma, surgery, or reconstruction. Radiographs were excluded if they exhibited diagnostic limitations, including superimposition of anatomical structures (e.g., hyoid bone, cervical vertebrae), positioning errors, motion artefacts, or ghost images obscuring the inferior mandibular cortex.

Initial screening was performed by two experienced oral and maxillofacial radiologists (8 years experienced SU and 14 years experienced MT). The morphology of the mandibular inferior cortex was evaluated using the MCI as described by Klemetti et al.¹⁴ The cortical margin extending from the distal aspect of the mental foramen to the antegonial region was assessed and classified into three categories (Fig. 1):

Fig. 1.

Representative examples of MCI categories and corresponding ROI segmentations. (a) C1 (normal cortex) showing a smooth and well-defined endosteal margin; (b) C2 (moderately resorbed cortex) characterized by semilunar resorption cavities and thinning of the cortical border; (c) C3 (severely resorbed cortex) demonstrating a porous, irregular endosteal surface; (d–f) Segmented regions of interest (ROIs) corresponding to C1, C2, and C3 images, respectively, delineated according to the standardized anatomical boundaries used for radiomic analysis.

• C1 (Normal Cortex): The endosteal margin is smooth, even, and sharply demarcated on both sides (Fig. 1a).

• C2 (Moderately Resorbed Cortex): The endosteal margin shows semilunar-shaped resorption cavities with cortical remnants one to three layers thick (Fig. 1b).

• C3 (Severely Resorbed Cortex): The cortex presents a clearly porous, irregular appearance with thick endosteal residues, indicating advanced resorption (Fig. 1c).

Each side of the mandible was evaluated independently. Following standard recommendations, the final MCI classification for each subject was determined by the side exhibiting the greater degree of cortical deterioration (C3 > C2 > C1) to avoid underestimating cortical erosion¹⁴. To ensure methodological standardization for radiomic analysis, only the right mandibular cortex was designated as the region of interest (ROI) (Fig. 1a-c). To maintain complete correspondence between the radiomic features and the assigned MCI category, an additional filtering step was applied during cohort construction. In cases with bilateral asymmetry, radiographs were included only when the right mandibular cortex exhibited the highest (dominant) MCI category identified for that individual. Cases in which the overall MCI classification was driven exclusively by a more severely resorbed left cortex were excluded from radiomic analysis. This approach ensured that all extracted radiomic features originated from the same cortical side used for final group assignment (Fig. 2).

Fig. 2.

Summary flowchart of the analytical pipeline.

A total of 795 panoramic radiographs were screened, and those exhibiting superimposition of anatomical structures (e.g., hyoid bone, cervical vertebrae), positioning errors, motion artefacts, or ghost images obscuring the inferior mandibular cortex were excluded before performing the MCI assessment. Following the exclusion process, 200 radiographs were classified as C1, 150 as C2, and 100 as C3. To assess the reliability of MCI scoring, 20% of the radiographs were re-evaluated after one week by the primary observer (SU) and independently by a second observer (MT).

Following this, strict secondary exclusion criteria were applied to eliminate images in which anatomical structures, such as mandibular molar roots extending below the mandibular canal, periodontal ligament spaces, lamina dura, or impacted teeth, encroached upon the defined segmentation area. After all secondary exclusion criteria were applied, the remaining eligible radiographs were stratified into the three MCI categories. For each group, a subset of 53 radiographs was randomly selected to create a balanced study cohort for radiomic analysis, resulting in a final dataset of 159 PRs. This final sample size met the prespecified requirements for ensuring adequate statistical robustness in group comparisons.

Image processing and segmentation

The selected radiographs were exported in TIFF format with a matrix size of 1595 × 830 pixels and an 8-bit depth and subsequently imported into 3D Slicer (version 5.10.0) for analysis²⁶. Mandibular segmentation was performed using a semi-automatic protocol, representing a modified version of the method described by Konishi et al.²⁰, designed to capture both cortical and trabecular bone components essential for comprehensive radiomic characterization.

The ROI was delineated according to strict anatomical boundaries (Fig. 1d–f):

• Superior boundary: The superior margin of the mandibular canal.

• Inferior boundary: The inferior border of the basal mandibular cortex.

• Anterior boundary: The lateral margin of the mental foramen.

• Posterior boundary: The anterior border of the mandibular ramus extending into the gonial region.

All segmentations were performed by the primary radiologist (SU). To assess segmentation reproducibility, 20% of the images were re-segmented after a one-week interval by the primary observer and independently segmented by a second oral and maxillofacial radiologist (KD, nine years of experience). Intra- and inter-observer agreement were subsequently evaluated using intraclass correlation coefficients (ICC).

Radiomic feature extraction

Radiomic features were extracted using PyRadiomics (version 3.1.0a2) within 3D Slicer (version 5.10.0). Prior to feature extraction, gray-level discretization was performed using a fixed bin width of 25 (Fig. 3) within the PyRadiomics framework. Because all panoramic radiographs were acquired using the same imaging system and identical acquisition parameters, spatial resampling was not applied. A detailed overview of the radiomics extraction settings, including feature-class selection, discretization parameters, preprocessing configuration, and representative extracted features, is provided in Fig. 3. PyRadiomics is compliant with the definitions established by the Image Biomarker Standardization Initiative (IBSI)²⁶. The extracted features were exported as a tab-separated values (TSV) file for subsequent statistical analysis. Five feature classes were computed, capturing complementary aspects of mandibular bone morphology^27–29:

Fig. 3.

PyRadiomics settings used for radiomic feature extraction and representative examples of extracted features. PyRadiomics (version 3.1.0a2) configuration within 3D Slicer (version 5.10.0), highlighting the selected radiomic feature classes (yellow box), the absence of spatial resampling and filtering with retention of the original image spacing (pink box), the gray-level discretization setting using a fixed bin width of 25 (blue box), and the preprocessing configuration indicating that intensity normalization was not applied (normalize = False; orange box). Representative examples of radiomic features exported from the output table are also shown, including a shape feature (Maximum 2D Diameter Row; red box) and a first-order feature (Energy; green box).

1. Shape: Shape features quantitatively describe the two-dimensional and/or the three-dimensional geometry of the ROI, capturing global morphological characteristics of the mandibular cortex. These metrics are essential for assessing cortical thinning, deformation, and overall structural integrity.

2. Parameters used: Surface Area, Maximum 2D Diameter (Slice), Maximum 2D Diameter (Column), Maximum 2D Diameter (Row), Major Axis Length, Minor Axis Length.

3. First-Order (FO) Features: These features describe the distribution of pixel intensities within the ROI and quantify global grayscale patterns. They provide insights into cortical bone density variability, asymmetry, and heterogeneity.

   Parameters used: Energy, Entropy, Mean, Median, Variance, Skewness, Kurtosis, Mean Absolute Deviation (MAD), Interquartile Range (IQR), Robust Mean Absolute Deviation (rMAD).

4. Gray-Level Co-occurrence Matrix (GLCM) Features: GLCM features characterize second-order statistical relationships between pixel pairs. They are sensitive to microtexture changes and cortical coarseness, including trabecular complexity and cortical porosity—key indicators of bone fragility.

   Parameters used: Contrast, Correlation, Cluster Shade, Cluster Prominence, Joint Energy, Joint Entropy, Inverse Difference Moment (IDM), Sum Average, Sum Entropy.

5. Gray-Level Run Length Matrix (GLRLM) Features: These features quantify the length and uniformity of consecutive pixel runs with the same gray level. They aid in evaluating cortical continuity and detecting subtle structural degradation associated with MCI progression.

   Parameters used: Short Run Emphasis (SRE), Long Run Emphasis (LRE), Gray Level Non-Uniformity (GLN), Run Length Non-Uniformity (RLN), Gray Level Variance (GLV), Run Length Variance (RLV), Run Percentage (RP), Run Variance (RV), Low Gray Level Run Emphasis (LGRE), High Gray Level Run Emphasis (HGRE).

6. Gray-Level Size Zone Matrix (GLSZM) Features: GLSZM features measure homogenous zones within the ROI. They reflect architectural breakdown, distinguishing between intact (C1) and severely eroded cortices (C3).

   Parameters used: Small Area Emphasis (SAE), Large Area Emphasis (LAE), Gray Level Non-Uniformity (GLN), Size Zone Non-Uniformity (SZN), Zone Entropy (ZE), Zone% (ZP), Large Area High Gray Level Emphasis (LAHGLE), Small Area Low Gray Level Emphasis (SALGLE).

These feature classes collectively quantify bone density, microtexture, structural continuity, and cortical porosity, providing a multidimensional characterization of mandibular cortical morphology beyond subjective visual assessment.

These five feature classes were selected because they represent the most commonly used radiomic descriptors in jawbone radiomics studies. A recent systematic review by Santos et al. reported that shape, first-order, GLCM, GLRLM, and GLSZM features were among the most frequently utilized and diagnostically relevant feature families in maxillofacial imaging applications¹⁷. GLDM and NGTDM features were not included because they have been less consistently reported in the jawbone radiomics literature¹⁷.

Statistical analysis

The intraclass correlation coefficient (ICC) was calculated for all parameters to assess inter-observer reliability. IBM SPSS Statistics (version 25.0) was used for all statistical analyses. The normality of continuous variables was evaluated using the Shapiro–Wilk test. Descriptive statistics, including minimum, maximum, mean, standard deviation, and frequency distributions, were computed. Group differences among the three MCI categories (C1, C2, C3) were assessed using one-way ANOVA for normally distributed features and the Kruskal–Wallis test for non-normally distributed features. Post-hoc pairwise comparisons were performed using Tukey or Dunn tests, respectively. Statistical significance for the primary analyses was initially set at p < 0.05 using unadjusted p-values. To account for multiple testing across post-hoc comparisons, p-values were additionally adjusted using the Benjamini–Hochberg false discovery rate (FDR) procedure. Both unadjusted and FDR-adjusted p-values are reported. FDR-adjusted p-values < 0.05 were considered statistically significant after correction.

Results

A total of 159 panoramic radiographs were included in the study, with an equal number of cases allocated to each MCI category (C1 = 53, C2 = 53, C3 = 53). The intra- and inter-observer reliability values for the MCI scoring were high (ICC = 0.92 and 0.88, respectively). According to accepted thresholds, ICC values between 0.75 and 0.90 indicate good agreement, whereas values above 0.90 reflect excellent reliability³⁰. In addition, to evaluate the reproducibility of ROI segmentation and radiomic feature extraction, both intra- and inter-observer ICCs were assessed. Intra-observer ICC values ranged from 0.88 to 0.97, indicating excellent repeatability of the segmentation procedure. Inter-observer ICCs were calculated based on independently segmented ROIs, and only radiomic features with inter-observer ICC values greater than 0.75 were retained for further analysis to ensure adequate reproducibility of the extracted radiomic features (Table 1).

Table 1.

Inter-Observer ICC results of radiomic feature extraction.

Feature	ICC
Shape-based features
Major axis length	0.566
Max 2D diameter row	0.860
Max 2D diameter slice	0.497
Minor axis length	0.526
Surface area	0.497
First order features
Energy	0.783
Entropy	0.904
Interquartile range	0.758
Kurtosis	0.026
Mean absolute deviation	0.856
Robust mean absolute deviation	0.795
Skewness	0.579
Variance	0.884
GLCM features
Cluster prominence	0.800
Cluster shade	0.871
Correlation	0.967
Joint energy	0.937
Joint entropy	0.935
Sum entropy	0.917
GLRLM features
Gray level variance	0.823
GLSZM features
Zone entropy	0.794

Analysis of the shape-based radiomic parameters showed that only the Maximum 2D Diameter Row differed significantly across the groups (Table 2; p < 0.001). The remaining shape indicators—such as Surface Area, Maximum 2D Diameter in the slice or column direction, and the Major and Minor Axis Lengths—displayed a similar numerical pattern, with C1 tending to have the greatest values, followed by C2 and then C3. However, these differences were not statistically meaningful.

Table 2.

Radiomic features of the mandibular cortex and statistical differences among MCI groups.

	C1 Mean ± SD	C2 Mean ± SD	C3 Mean ± SD	P value
Shape-based features
Max 2D diameter row	141.57 ± 19.11	128.21 ± 23.13	120.34 ± 20.86	< 0.001a
First-order features
Energy	278727988.3±172768326.2	210349826.8±143575635.55	203778456.4±123517961.2	0.006a
Entropy	1.31 ± 0.34	1.02 ± 0.38	1.11 ± 0.43	0.001a
Interquartile range	20.30 ± 9.93	15.07 ± 7.41	16.28 ± 8.02	0.002a
Mean absolute deviation	11.11 ± 4.64	8.49 ± 3.68	9.59 ± 4.47	0.002a
Robust mean absolute deviation	8.69 ± 4.33	6.37 ± 2.89	6.90 ± 3.46	0.001a
Variance	215.49 ± 175.57	128.76 ± 119.10	171.01 ± 161.61	0.001a
GLCM
Cluster prominence	8.68 ± 12.67	4.72 ± 8.71	8.27 ± 13.21	0.001a
Cluster shade	0.25 ± 0.72	-0.16 ± 1.08	-0.45 ± 2.01	0.010a
Correlation	0.89 ± 0.05	0.84 ± 0.09	0.84 ± 0.10	< 0.001a
Joint energy	0.42 ± 0.11	0.52 ± 0.17	0.49 ± 0.17	0.009a
Joint entropy	1.68 ± 0.41	1.37 ± 0.50	1.50 ± 0.51	0.006a
Sum entropy	1.61 ± 0.40	1.30 ± 0.50	1.42 ± 0.49	0.003a
GLRLM
Gray level variance	0.62 ± 0.30	0.46 ± 0.23	0.53 ± 0.28	0.001a
GLSZM
Zone entropy	5.18 ± 0.33	4.89 ± 0.44	5.03 ± 0.37	0.002a

C1 = normal cortex; C2 = moderately resorbed cortex; C3 = severely resorbed cortex; SD= standard deviation; GLCM= gray-level co-occurrence matrix; GLRLM= gray-level run length matrix; GLSZM= gray-level size zone matrix.

^aStatistically significant (P < 0.05).

P values presented in this table represent the unadjusted overall group comparison (omnibus test). Pairwise unadjusted and Benjamini–Hochberg FDR-adjusted post-hoc p-values are presented in Table 3.

In the post-hoc comparisons, C1 was again the only group that separated clearly from C2 and C3 for the Maximum 2D Diameter Row (Table 3; p ≤ 0.001), whereas C2 and C3 did not differ from each other (p > 0.05). Taken together, these findings point to a gradual decrease in cortical dimensions from C1 toward C3, although Maximum 2D Diameter Row was the only shape measure capable of distinguishing the groups in a statistically reliable way.

Table 3.

Post-hoc pairwise comparisons of radiomic features across MCI groups showing unadjusted and FDR-adjusted p-values.

	C1–C2		C1-C3		C2–C3
	Unadjusted p value	FDR-adjusted p value	Unadjusted p value	FDR-adjusted p value	Unadjusted p value	FDR-adjusted p value
Shape-based features
Max 2D diameter row	0.001a	0.004	< 0.001a	< 0.001b	0.070	–
First-order features
Energy	0.004a	0.010b	0.008a	0.018b	0.987	–
Entropy	< 0.001a	< 0.001b	0.022a	0.039b	0.258	–
Interquartile range	0.002a	0.006b	0.003a	0.008b	0.500	–
Mean absolute deviation	0.001a	0.004b	0.020a	0.038b	0.308	–
Robus mean absolute deviation	0.001a	0.004b	0.002a	0.006b	0.667	–
Variance	< 0.001a	< 0.001b	0.016a	0.031b	0.332	–
GLCM
Cluster prominence	< 0.001a	< 0.001b	0.069	–	0.125	–
Cluster shade	0.008a	0.018b	0.009a	0.019b	0.793	–
Correlation	< 0.001a	< 0.001b	0.001a	0.004b	0.825	–
Joint energy	0.002a	0.006b	0.064	–	0.293	–
Joint entropy	0.001a	0.004b	0.122	–	0.141	–
Sum entropy	0.001a	0.004b	0.081	–	0.140	–
GLRLM
Gray level variance	< 0.001a	< 0.001b	0.015a	0.03b	0.232	–
GLSZM
Zone entropy	< 0.001a	< 0.001b	0.176	–	0.038a	0.062

C1 = normal cortex; C2 = moderately resorbed cortex; C3 = severely resorbed cortex; GLCM= gray-level co-occurrence matrix; GLRLM= gray-level run length matrix; GLSZM= gray-level size zone matrix; FDR = false discovery rate.

Values are reported as unadjusted p-value/Benjamini–Hochberg FDR-adjusted p-value. Superscript “a” indicates statistical significance based on the unadjusted p-value (p < 0.05). Superscript “b” indicates statistical significance after Benjamini–Hochberg FDR correction (adjusted p < 0.05).

The examination of the FO radiomic features showed that several parameters—Energy, Entropy, IQR, MAD, rMAD, and Variance—varied significantly among the three groups (Table 2; p ≤ 0.006). When the numerical profiles were considered, the C1 group generally appeared at the upper end of the scale for all of these measurements. Energy displayed a clear stepwise decline from C1 to C3 (C1 > C2 > C3), whereas the other metrics tended to follow a slightly different pattern, with C3 frequently recording values above those of C2 (C1 > C3 > C2).

Follow-up pairwise analyses indicated that C1 differed from both C2 and C3 across all first-order parameters that reached significance (Table 3; p ≤ 0.022). By contrast, even though C2 and C3 showed some numerical separation, none of these differences proved to be statistically significant (Table 3; p > 0.05). Taken together, these findings suggest that the first-order features examined were helpful in distinguishing C1 from the other subgroups but did not provide enough discriminatory power to separate C2 from C3.

Texture-based metrics captured microarchitectural changes in greater detail.

• GLCM Features: Several GLCM-derived measures—Cluster Prominence, Cluster Shade, Correlation, Joint Energy, Joint Entropy, and Sum Entropy—showed significant variation across the groups (Table 2; p ≤ 0.010). The numerical patterns of these features did not follow a single uniform trend. Cluster Prominence, Joint Entropy, and Sum Entropy tended to be highest in C1, intermediate in C3, and lowest in C2 (C1 > C3 > C2). Cluster Shade, however, showed a steady decline from C1 to C3 (C1 > C2 > C3). Joint Energy demonstrated the opposite direction, with C2 presenting the highest values, followed by C3 and then C1 (C2 > C3 > C1).

  Pairwise analyses indicated that C1 differed from C2 in all significant GLCM measures (Table 3; p ≤ 0.008). Comparisons between C1 and C3 identified meaningful differences only for Cluster Shade (p = 0.009) and Correlation (p = 0.001). As for the two resorbed cortex groups, none of the GLCM parameters distinguished C2 from C3 (p > 0.05).

• GLRLM Features: GLV emerged as the sole GLRLM feature exhibiting a significant overall group effect (Table 2; p = 0.001). GLV followed a descending pattern of C1 > C3 > C2, suggesting that C1 displayed the greatest variability in gray-level distribution. Post-hoc results demonstrated significant differences between C1 and C2 (p < 0.001) and between C1 and C3 (p = 0.015), whereas the C2–C3 comparison did not reach significance (p > 0.05).

• GLSZM Features: ZE was the only GLSZM feature showing statistically significant group variation (Table 2; p = 0.002). Like GLV, ZE also followed the trend C1 > C3 > C2, indicating a progressively reduced degree of textural irregularity from C1 toward C2. In the initial post-hoc analysis, ZE differed significantly between C1 and C2 (p < 0.001) and showed a marginal difference between C2 and C3 (p = 0.038). However, after applying the FDR correction, the C2–C3 comparison was no longer statistically significant (adjusted p = 0.062). Accordingly, none of the evaluated radiomic features remained capable of significantly differentiating the two resorbed cortical groups after correction for multiple comparisons.

Discussion

Opportunistic screening for OP using PRs has become a focal point in dentomaxillofacial radiology research, with the MCI serving as a fundamental tool due to its well-documented association with systemic bone turnover. Roebers and Schulze³¹, highlight the diagnostic value of this index, reporting specificities approaching 93% in certain studies; however, they also note that its traditional application relies on qualitative visual grading, which introduces observer dependency and variability. Importantly, the cortical segment in which MCI is assessed is also the anatomical region used for several mandibular radiomorphometric indicators related to low bone mineral density, including the mental index (MI), antegonial index (AI), antegonial angle (AA), antegonial depth (AD), and the panoramic mandibular index (PMI)³². Thus, the ROI selected in the present study was not intended to reflect MCI alone but to encompass a radiomorphometrically meaningful cortical corridor that has historically been central to multiple bone mineral density (BMD) related assessments. In this framework, our radiomics analysis does not rely on a single index category; rather, it provides a quantitative characterization of the cortical–endosteal interface underlying these established indicators.

Efforts to reduce the subjectivity of visual cortical assessment have prompted several image-processing strategies. Tassoker et al.¹⁵ reported that pseudocolored images derived from ImageJ were not successful in distinguishing osteoporotic changes among MCI classes, whereas histogram-based numerical density values demonstrated clear separability. This discrepancy suggests that diagnostic improvement is more likely to come from objective pixel-level quantification than from enhanced visualization—an approach radiomics implements through high-dimensional extraction of intensity-, texture-, and structure-related features.

Previous attempts to quantify mandibular bone quality, such as Fractal Dimension (FD) and Strut Analysis (SA), have produced variable results. A recent systematic review and meta-analysis by Cavalcante et al.³³ found that although FD demonstrated a pooled sensitivity of 86.17%, its specificity was lower (72.68%) and highly heterogeneous across studies (I² = 97.10%), limiting its consistency as a biomarker. Variation in anatomical targeting further contributes to inconsistency. For example, recent artificial intelligence (AI) driven studies, including Fanelli et al.²¹ adopted relatively broad ROIs extending along the inferior mandibular border and into posterior cortical regions to maximize contextual information for convolutional neural network (CNN) based models. While such an approach can be advantageous for end-to-end deep learning, it may also introduce anatomical noise and non-specific variation. In contrast, Hwang et al.³⁴ demonstrated through SA that isolated medullary ROIs often fail to differentiate osteoporotic from non-osteoporotic bone, whereas the endosteal margin—defined as the transitional zone incorporating both the inferior cortex and its adjacent trabecular layer—was most effective in detecting osteoporotic changes. Building on these observations, our study focuses specifically on this metabolically active cortical–trabecular interface. An additional consideration is that early osteoporotic changes tend to arise locally rather than diffusely; thus, concentrating on the cortical–endosteal boundary helps reduce anatomical noise and increases the likelihood of detecting subtle, biologically meaningful textural alterations. By applying radiomic feature extraction to this focused and biologically grounded region, we aim to capture the microarchitectural changes that occur at the earliest stages of bone loss without the confounding variability introduced by overly broad ROIs.

AI and deep learning (DL) approaches represent important developments in automated OP screening; however, clinical integration remains limited. Ghasemi et al.²³ meta-analysis demonstrated that although DL models show strong diagnostic potential, they also display extreme heterogeneity (I²> 94%), largely due to differences in model architectures and preprocessing strategies. Fanelli et al.²¹ similarly noted that end-to-end CNN models, including ResNet-based architectures, require large datasets to generalize effectively. In contrast, radiomics-based approaches—particularly those incorporating dimensionality reduction techniques—tend to be more data-efficient and methodologically transparent.

Furthermore, a significant methodological concern highlighted in recent AI-based studies, including the work of Fanelli et al.²¹ is the limited attention given to reproducibility when manually segmenting ROIs, which may introduce operator-dependent variability. To address this issue, our study conducted a detailed assessment of inter-observer agreement for all extracted radiomic features. As shown in Table 1, texture-based features demonstrated high reproducibility, with First Order Entropy (ICC = 0.904) and GLCM Correlation (ICC = 0.967) showing particularly strong consistency between observers. In contrast, shape-based measures exhibited considerably lower reproducibility, with ICC values between 0.497 and 0.566 for parameters such as surface area and axis lengths. This pattern aligns with findings from Traverso et al.³⁵ who reported that shape metrics are inherently more sensitive to the exact placement of the segmentation boundary, as small variations can meaningfully influence geometric measurements while having minimal influence on internal texture descriptors. A similar conclusion was drawn by Xue et al.³⁶ who noted that shape features are frequently the most affected feature class across different segmentation approaches and often show weaker reliability than first-order or GLCM features. For this reason, only features with ICC values above 0.75 were retained for subsequent analyses, ensuring that the radiomic evaluation was based on features demonstrating adequate reproducibility and methodological stability.

In the present evaluation of shape-based radiomic features, Maximum 2D Diameter Row was the only geometric parameter that demonstrated a statistically significant difference among the study groups. This measurement reflects the horizontal extent of the region of interest on PRs and may be particularly responsive to cortical boundary changes associated with bone demineralization. As bone density declines, the radiographic clarity of anatomical landmarks—especially the superior cortical margin of the mandibular canal and the borders of the mental foramen—may diminish. In the C2 and C3 categories, this reduced definition can result in less clearly delineated cortical margins, leading to a segmented region with decreased horizontal continuity when compared with the well-preserved cortex observed in C1 cases.

In contrast, other shape-related features, including Surface Area, Maximum 2D Diameter in the slice direction, and the major and minor axis lengths, did not reach statistical significance. Nevertheless, these parameters displayed a uniform numerical trend, with values progressively decreasing from C1 to C3. Although these differences were not statistically robust, the pattern itself is consistent with the biological progression described in traditional radiomorphometric research. Previous studies^32,37,38, have reported comparable reductions in mandibular linear indices and alveolar bone height in osteoporotic individuals, particularly among older women. The alignment between these established linear measurements and the volumetric shape trends observed in the current study supports the notion that progressive cortical resorption is accompanied by gradual reductions in mandibular cortical dimensions across MCI categories.

The analysis of FO radiomic features provided a quantitative overview of gray-level intensity distribution within the mandibular cortex. Among these parameters, Energy—defined as the sum of squared voxel intensities—showed a progressive decline from C1 to C3. As previously reported by Fanelli et al.²¹ reduced energy values in panoramic radiographs are associated with decreased mineral density. Accordingly, the stepwise reduction observed in the present study may serve as an indirect quantitative indicator of global bone density loss accompanying cortical erosion in OP. However, because Energy is a size-dependent radiomic feature, part of the observed group difference may also be influenced by variations in ROI extent among MCI categories.

Entropy and Variance, which describe the randomness and dispersion of gray-level values, offered additional insight into cortical complexity. In radiomics studies of odontogenic cysts, İçöz et al.³⁹ associated elevated entropy with structural irregularity in pathological tissues, while temporomandibular joint (TMJ) magnetic resonance imaging (MRI) based analyses by Duyan Yüksel et al.⁴⁰ reported that entropy reflects complex intensity distributions. OP, however, differs fundamentally from focal pathologies, as it involves a diffuse degradation of bone architecture rather than localized tissue infiltration. In this context, the higher entropy and variance values observed in the intact cortex (C1) may reflect the inherently rich microstructural organization of healthy mandibular bone, characterized by a dense cortical layer transitioning into trabecular bone. Conversely, the reduction of these metrics in eroded cortices may indicate a gradual homogenization process, in which architectural complexity diminishes, and the radiographic appearance becomes more uniform.

Robust dispersion measures, including IQR, MAD, and rMAD, further supported this pattern by consistently distinguishing the intact cortex from eroded groups. As highlighted by Duyan Yüksel et al.¹⁹ such intensity-based parameters are relatively insensitive to anatomical variability and may provide stable descriptors of global tissue alterations. Taken together, these findings suggest that FO features are effective in capturing the transition from intact to eroded cortical bone through changes in overall intensity distribution.

An additional observation concerns the non-linear numerical trend of heterogeneity-related metrics, with C2 showing lower values than C3 (C1 > C3 > C2). This pattern may be explained by differences in the underlying morphology of resorption stages. The C2 category is characterized by a generalized attenuation of trabecular structure, which may result in a relatively uniform and low-contrast intensity distribution. In contrast, advanced cortical erosion in C3 is often accompanied by pronounced endosteal porosity and residual calcified fragments. The coexistence of severely demineralized regions and remaining mineralized areas may introduce localized intensity contrasts, leading to a relative increase in entropy and variance compared with the more uniformly attenuated C2 cortex. This non-linear behavior suggests that FO metrics may reach a diagnostic plateau once erosion is established, supporting the notion that texture-based features—such as GLCM-derived parameters—are required for finer differentiation between advanced stages of cortical deterioration.

Despite these distinctions between intact and resorbed cortices, the lack of statistical differentiation between C2 and C3 across FO parameters suggests that intensity-based homogenization captured by first-order statistics may occur relatively early in the resorption process. As the C2 category within the MCI classification represents a transitional stage without a strict early–late subdivision, it is plausible that the present cohort included a substantial proportion of radiomically advanced C2 cases that closely resembled C3. This overlap may have limited the discriminatory power of FO features between these two resorbed groups.

Texture-based radiomic features were employed to capture microarchitectural alterations of the mandibular cortex that cannot be adequately described by intensity-based metrics alone. Previous radiomics literature has emphasized the value of texture analysis methods such as GLCM, GLRLM, and GLSZM for the objective characterization of bone-related changes, particularly in complex skeletal structures beyond subjective visual assessment¹⁷.

Within the GLCM framework, features reflecting spatial dependency and structural organization, most notably Cluster Shade and Correlation, consistently distinguished the intact cortex (C1) from both eroded groups (C2 and C3). GLCM-based measures have been described as reflecting spatial relationships between neighboring pixels rather than isolated gray-level values⁴¹. In line with this concept, Cluster Shade demonstrated a gradual decline from C1 to C3, which may indicate a progressive loss of textural asymmetry and structural organization as cortical erosion advances³⁹. Similarly, reduced Correlation values in the eroded groups suggest disruption of the interconnected trabecular pattern that characterizes healthy mandibular bone²¹. Because both moderate and severe resorption involve a breakdown of this spatial continuity, these parameters appear sensitive to the presence of disease rather than its exact stage.

In contrast, heterogeneity-related GLCM metrics, including Cluster Prominence, Joint Entropy, and Sum Entropy, exhibited a non-linear pattern (C1 > C3 > C2). Prior studies have noted that entropy- and variance-related features increase in the presence of marked intensity differences between adjacent voxels⁴¹. The observation that C2 showed the lowest values for these metrics may suggest that this stage represents a period of diffuse homogenization, during which trabecular detail becomes attenuated and gray-level contrast is reduced. Conversely, advanced erosion in C3 is often accompanied by pronounced porosity and residual calcified fragments, potentially reintroducing localized intensity variation and modestly elevating entropy-related values. This density-dependent behavior is consistent with findings by He et al.⁴² who demonstrated that radiomic feature values and stability are influenced by underlying tissue density.

This interpretation is further supported by the behavior of Joint Energy, a metric inversely related to textural randomness. Joint Energy followed the opposite trend (C2 > C3 > C1), reaching its highest values in the C2 group⁴¹. As higher energy reflects greater uniformity, this pattern suggests that the moderately eroded cortex may represent the most homogeneous structural phase, lacking both the organized complexity of intact bone and the fragmented heterogeneity seen in severe resorption.

Among higher-order texture descriptors, GLSZM-derived ZE showed an interesting pattern across MCI categories. Previous studies have suggested that GLSZM features are sensitive to complex spatial patterns and clustered tissue organization within osseous structures¹⁷. In the present study, ZE demonstrated a nominal difference between C2 and C3 in the unadjusted analysis; however, this relationship did not remain significant after FDR correction. Consequently, the apparent discrimination between moderately and severely resorbed cortices should be interpreted with caution and regarded as exploratory. The loss of statistical significance after correction highlights the substantial overlap between advanced stages of cortical resorption and suggests that radiomic features may reach a diagnostic plateau once marked cortical deterioration has occurred. Although GLSZM-derived descriptors may be sensitive to alterations in spatial heterogeneity and microarchitectural organization, the present findings do not provide sufficient evidence to support their reliable use for distinguishing C2 from C3 categories.

Several limitations should be acknowledged when interpreting the results of this study. The retrospective and cross-sectional design does not allow conclusions to be drawn regarding the temporal progression of cortical deterioration across MCI categories. In addition, the study population was restricted to women aged 50 years and older in order to reduce biological variability and to focus on the group at highest risk for OP. While this approach increased internal consistency, it may limit the generalizability of the findings to male individuals or younger populations. Future studies may benefit from evaluating a more homogeneous sample by controlling for dentition status and the potential effects of masticatory force. However, in the present study, excluding these variables would have resulted in an insufficient sample size, and the underlying causes of tooth loss (e.g., osteoporosis, dental caries) could not be clearly determined. Therefore, such stratification could not be performed.

Another consideration is related to imaging modality. Although strict inclusion and exclusion criteria were applied to minimize anatomical variability, the analysis relied on two-dimensional (2D) PRs, which are inherently affected by geometric distortion and superimposition. Radiomics can partially compensate for these limitations by extracting quantitative information from standardized regions; nevertheless, three-dimensional imaging techniques such as cone-beam computed tomography (CBCT) or QCT could provide complementary insight into cortical and trabecular microarchitecture.

The segmentation strategy also warrants consideration. The ROI was manually delineated using a rigorous protocol and validated through inter-observer reliability analysis, yet this process remains time-consuming and operator dependent. As such, immediate translation into routine clinical workflows may be limited until fully automated segmentation approaches become available. Furthermore, although MCI is a widely accepted surrogate marker of skeletal status, the present study focused on the quantitative characterization of MCI categories rather than on direct correlation with femoral or lumbar dual-energy X-ray absorptiometry (DXA) T-scores. Therefore, the present findings should be interpreted as reflecting radiomic differences among MCI categories rather than direct evidence of osteoporosis or low bone mineral density.

Radiomic analysis was restricted to the right mandibular cortex in order to ensure methodological standardization of ROI definition, segmentation, and feature extraction across all subjects. Although this approach reduced potential variability associated with bilateral ROI selection, possible side-related differences in cortical morphology were not evaluated. Future studies incorporating bilateral radiomic assessment may help clarify the influence of anatomical asymmetry and functional variation on radiomic feature behavior. Additionally, certain size-sensitive radiomic features may be influenced by ROI extent rather than exclusively reflecting underlying tissue characteristics. Specifically, Energy is known to depend partly on the number of pixels or voxels contained within the segmented region. Because segmentation was performed using anatomical boundaries rather than fixed-size ROIs, differences in cortical dimensions among MCI categories may have contributed to the observed Energy values. Therefore, Energy-related findings should be interpreted with appropriate caution.

An additional limitation concerns the scope of model development. The primary objective of this study was to establish a standardized and reproducible ROI framework and to systematically assess the behavior and stability of radiomic features across MCI categories. Given the relatively modest sample size and the emphasis on interpretability and robustness, advanced deep learning approaches such as CNN modeling were not pursued. Nevertheless, the demonstrated ROI standardization and feature reproducibility may provide a methodological foundation for future CNN-based applications. With larger, multi-center datasets and harmonized acquisition protocols, such models could further improve automation and potentially enhance the clinical utility of radiomics-based mandibular cortical assessment.

In essence, the findings suggest a stage-dependent role of radiomic features in assessing cortical deterioration. Intensity-based metrics, such as Energy and Correlation, appear effective for identifying the presence of cortical erosion, whereas higher-order texture features provide additional information related to disease severity. This pattern indicates that osteoporotic cortical changes involve more than a simple, linear reduction in bone density and instead reflect progressive alterations in underlying microarchitecture.

Conclusion

Radiomic analysis of panoramic radiographs offers a quantitative and reproducible approach for evaluating mandibular cortical characteristics, extending beyond the inherent subjectivity of visual grading methods. The findings of this study suggest that first-order radiomic metrics vary systematically across MCI categories, reflecting differences in mandibular cortical morphology. Distinguishing between moderately (C2) and severely (C3) resorbed cortices remains challenging, as no radiomic feature remained statistically significant after correction for multiple comparisons. By implementing a standardized segmentation protocol and systematically examining the behavior of selected radiomic features, this work represents an important methodological step toward more objective and reproducible cortical assessment. These findings may provide a methodological foundation for future machine learning–based investigations of mandibular cortical changes and for studies exploring the relationship between radiomic features, MCI classification, and skeletal bone status.

Author contributions

Conceptualization and methodology were developed by S.U., K.D., and M.T.Data collection, image segmentation, and radiomic measurements were performed by S.U.Radiomic feature extraction and software-based processing were carried out by S.U.Inter-observer segmentation and reproducibility analyses were conducted by S.U. and K.D.Statistical analysis was performed by M.T.Resources and study materials were provided by S.U., K.D., and M.T.The original draft of the manuscript was written by S.U. and K.D.Figures and tables were prepared by K.D. and S.U.Critical revision of the manuscript for important intellectual content was performed by M.T., S.U., and K.D.All authors read and approved the final manuscript.

Data availability

The datasets generated and/or analyzed during the current study are available from the corresponding author upon reasonable request.

Code availability

No custom code, bespoke computational tool, or new algorithm was generated or used in this study. All radiomic feature extraction was performed using the publicly available, open-source software 3D Slicer (version 5.10.0; https://www.slicer.org) with the PyRadiomics extension (version 3.1.0a2), and all statistical analyses were conducted using IBM SPSS Statistics (version 25.0). The complete PyRadiomics feature-extraction settings (fixed bin width of 25, IBSI-compliant configuration, and no spatial resampling) are reported in the Methods section and Fig. 3, enabling full reproduction of the analysis pipeline using these publicly available tools.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval

This retrospective archive study was conducted in accordance with the principles of the Declaration of Helsinki. Ethical approval was obtained from the Necmettin Erbakan University Faculty of Dentistry Non-Drug and Non-Medical Device Research Ethics Committee (Approval No: 2025/696). Due to the retrospective nature of the study, the requirement for informed consent was waived by the same ethics committee.