Cone-beam computed tomography-based radiomic analysis of architectural phenotypes in jaw cysts and tumors using interpretable artificial intelligence models

Abstract

BACKGROUND

Jaw lesions, like cysts and tumors, demonstrate substantial variation in internal architectural organization, spatial heterogeneity, and voxel-level complexity on cone-beam computed tomography (CBCT). Conventional radiological interpretation relies predominantly on subjective visual assessment and may not adequately capture these underlying imaging phenotypes.

AIM

To evaluate whether CBCT-derived radiomic features can quantitatively characterize architectural phenotypes of jaw lesions and to assess their discrimination using interpretable artificial intelligence (AI) models.

METHODS

This retrospective study analyzed 100 histopathologically confirmed jaw lesions using CBCT. Lesions were manually segmented using 3D Slicer and 107 radiomic features were extracted after standardized preprocessing and voxel normalization using PyRadiomics. Lesions were classified into homogeneous fluid-dominant, intermediate septated, and complex heterogeneous phenotypes. Feature stability was assessed using intraclass correlation coefficients, while selection employed false discovery rate (FDR) correction, correlation pruning, and LASSO regression. Logistic regression (LR), support vector machine (SVM), and random forest (RF) models underwent stratified five-fold cross-validation and independent chronological validation.

RESULTS

Forty radiomic features demonstrated statistically significant differences among architectural phenotypic groups following FDR correction. Feature reduction yielded a compact radiomic signature predominantly composed of texture-derived descriptors reflecting gray-level non-uniformity, spatial dependence variability, entropy, and structural complexity. The LR model demonstrated the highest performance, achieving an area under the receiver operating characteristic curve of 0.92, with robust discrimination between architectural phenotypes. SVM and RF models demonstrated comparable but lower performance. Lesions categorized within the complex heterogeneous phenotype exhibited significantly elevated texture heterogeneity metrics compared with homogeneous fluid-dominant lesions, supporting the biological relevance of radiomic architectural characterization in differentiating complex jaw pathologies.

CONCLUSION

CBCT-derived radiomic features enable quantitative assessment of internal architectural phenotypes in jaw lesions, particularly patterns related to spatial heterogeneity and structural organization. Texture-based radiomic signatures, when integrated with interpretable AI models, may function as imaging biomarkers for objective lesion characterization and may support future development of biologically informed diagnostic decision-support systems in oral and maxillofacial radiology.

Key Words: Artificial intelligence; Cone-beam computed tomography; Jaw cysts; Jaw Tumors; Radiomics

Core Tip: Radiomic analysis of cone-beam computed tomographic images allows objective evaluation of architecture in jaw cysts and tumors through analysis of heterogeneity, spatial distribution, and texture of voxels. In contrast to traditional qualitative radiological interpretation, this phenotypic approach uses radiomics to classify jaw lesions into three groups: Homogenous with fluid predominance, intermediate with septations, and complex or heterogenous. Radiomic signatures based on texture have proved highly effective and biologically relevant, making them promising tools for imaging biomarkers and decision support in oral and maxillofacial radiology.

INTRODUCTION

Cysts and tumors in the jaws are one of the most common yet diagnostically challenging lesions due to overlapping clinical and radiographic features. Despite showing differences in histopathological pattern and biological behavior, cysts and tumors are known to present with very similar imaging characteristics like radiolucent areas, cortical expansion, or change in internal density[1,2]. Such similarities between varying lesions can cause difficulty in interpreting the lesion and this leads to inaccurate identification of biological nature of the lesion. This results in diagnostic uncertainty especially in lesions that show intermediate or mixed internal architecture, where visual interpretation of the radiograph may fail to adequately reflect underlying tissue organization and structural complexity. This can have serious impact on the treatment planning and prognosis of the patient[3]. Cone-beam computed tomography (CBCT) is the most advanced and commonly used imaging modality for visualization of the jaws and lesions of the jaws in three dimensions[4]. It provides high spatial resolution with minimal radiation exposure when compared to other imaging modalities. CBCT is well known to enable detailed assessment of lesion extent, cortical integrity, internal septation, trabecular organization, root resorption, displacement of adjacent structures, and spatial relationship with vital anatomical landmarks[5,6]. Even though CBCT significantly enhances the visibility of the osseous structure, the interpretation of jaw pathology is still largely based on qualitative radiographic terms and the skill level of the examiner[7]. The traditional approach to image interpretation is highly based on visual inspection of lesion shape, locularity, density, borders, and expansion characteristics, which are essentially subjective[8].

This shows that qualitative interpretation of jaw lesions, even with advanced modality like CBCT, can cause errors in diagnosis due to subjectivity and overlapping radiographic features. Thus, there is a need for quantitative interpretation of jaw lesions for errorless diagnosis and better prognosis[8,9]. Recent developments in radiomics have brought about the ability to extract massive amounts of such quantitative information from medical images that cannot be discerned through mere visual inspection[10,11]. Radiomics involves transforming imaging information into high-dimensional numeric representations that can represent tissue architecture, intensity distribution in voxels, and spatial variability. Radiomic features include shape-based parameters, first-order statistics, and higher-order textures obtained using spatial gray-level matrixes[12]. Texture-based radiomic parameters are especially important in the study of jaw lesions since they allow for quantitative evaluation of architectural arrangement, structural disarray, spatial complexity, and dependencies between voxels[13]. These imaging parameters correspond to biologically significant differences within the lesions, such as fluid uniformity, epithelial cell multiplication, stromal arrangement, mineralization, septa, trabecular disorganization, and degeneration.

Since radiomic features are numerical respresentatives of underlying pathology, they can be easily combined with artificial intelligence (AI) models for diagnosis of jaw lesions[14]. In oral and maxillofacial radiology, earlier radiomics literature has been mostly confined to direct pairwise comparisons of lesions, like odontogenic keratocyst vs ameloblastoma, or cystic vs solid lesions[15,16]. While these results were promising, it must be noted that radiomics, at its core, is an image phenotype analysis tool. Radiomic analysis does not directly establish a definitive pathological diagnosis, because lesions with entirely different biological origins may demonstrate similar radiomic profiles if they share comparable internal architectural patterns. Conversely, lesions belonging to the same pathological category may exhibit markedly different radiomic characteristics due to variations in their internal structural organization and spatial heterogeneity. Thus, it would be more appropriate to see radiomics as a technique to quantify lesion architectural phenotypes based on internal heterogeneity, architecture, and voxel-based complexity. In this context, lesions of the jaws can be seen as having phenotypes lying between homogeneous and fluid-dominant lesions to architecturally complex lesions. Another significant barrier in the translation of AI models to the clinical setting is the problem of interpretability and biological transparency. The majority of AI algorithms act as “black-box” models that cannot provide any understanding regarding the role played by imaging variables in the resulting outcome[17,18]. However, the utilization of interpretable AI algorithms such as radiomic analysis based on LR permits the direct assessment of the contribution made by imaging biomarkers to lesion classification. Biological transparency and understandability are crucial in the field of oral and maxillofacial radiology because of the requirement for clinically relevant decision support systems[19]. Accordingly, the purpose of the current study was to assess the capability of CBCT radiomic features for the quantitative description of architectural phenotypes in jaw cysts and tumors using interpretable AI models.

MATERIALS AND METHODS

Study design and dataset collection

The present retrospective radiomics study was designed to quantitatively evaluate architectural phenotypes in jaw cysts and tumors using CBCT-derived radiomic features integrated with interpretable AI models. The study protocol was approved by the Institutional Ethics Committee (approval No. TMDCRC/IEC/PHD/24-25/DENTAL02; IEC proposal No. S-002/24) and was conducted in accordance with the principles of the Declaration of Helsinki. Owing to the retrospective nature of the study and the use of anonymized imaging data, the requirement for informed consent was waived. CBCT datasets used in this study were retrieved from the oral radiology department of the institution over a 2 years period. Only those lesions which were histopathologically diagnosed as cysts and tumors were included. Both odontogenic and non-odontogenic lesions were included in the study as long as they showed sufficient radiographic visualization of internal architecture. Radiologically indeterminate histopathologic nature, history of previous surgery, recurrent disease, secondary infection, and significant deformity were the criteria for exclusion from the study in order to reduce biological and radiologic variability that is independent of the phenotypic character of the lesion. Accordingly, the current study aimed at assessing the internal architecture phenotypes of jaw lesions. Based on the internal organization shown in CBCT, lesions were divided into three different phenotypes: Homogenous fluid-filled, intermediate septated, and complex heterogeneous. Lesion phenotyping was performed separately by two experienced clinicians with more than 10 years of experience specializing in dental radiology by evaluating lesions’ internal structure. For objectivity purposes, rigorous operational definitions were set before the evaluation process was conducted: The homogenous phenotype group included those with totally uniform radiodensities within their internal structure with a low level of spatial complexity and no presence of any internal ridges or cortical breach. The intermediate phenotype group comprised those that had partial heterogeneous architecture containing fine or incomplete internal septation without full compartmentalization. The complex heterogeneous phenotype consisted of lesions having architectural complexity, multilocular compartmentalization, coarse trabeculation, radiolucent or radiopaque densities, or structural disorganization. To ensure absolute transparency and architectural clarity, the entire study design, data workflow, and model development pipeline were executed in compliance with the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis-Artificial Intelligence (TRIPOD-AI) statement, the Checklist for Artificial Intelligence in Medical Imaging (CLAIM) guidelines, and the Radiomics Quality Score framework. Completed compliance checklists for the TRIPOD-AI and CLAIM frameworks are provided as Supplementary material.

Image acquisition

The radiographic images of all CBCT examinations were obtained with NewTom GiANO HR CBCT system (NewTom, Verona, Italy) with an optimized CMOS flat panel detector. The exposure protocol was standardized with the voltage tube value at 90 kVp and tube current values ranging from 1-10 mА. The field of view was between 10 cm × 6 cm and 16 cm × 18 cm according to the size of the lesion. The voxel size was between 125 µm and 300 µm, which ensured excellent spatial resolution for the analysis of internal architecture of the lesion. Inclusion criterion was that the lesions had to be entirely visible in the field of view of the scan. CBCT examinations showing motion artifacts, beam hardening effects, truncation artifacts, metal artifacts, or reconstruction failures were not selected to preserve the integrity of the extracted radiomic features.

Image preprocessing

All CBCT datasets were exported in Digital Imaging and Communications in Medicine (DICOM) format and processed using a standardized preprocessing workflow prior to radiomic analysis. Preprocessing was performed to reduce inter-scan variability and improve reproducibility of extracted radiomic features. To ensure spatial consistency across datasets, all image volumes were resampled to isotropic voxel spacing using trilinear interpolation. Since CBCT gray values lack absolute standardization compared with conventional computed tomography, relative intensity normalization was performed to minimize scanner-dependent variability while preserving intrinsic lesion characteristics. Gray-level discretization was subsequently performed using a fixed bin width of 25 intensity units with specified PyRadiomics settings to stabilize higher-order texture feature computation and reduce sensitivity to image noise. Mild smoothing filters were applied to suppress high-frequency image noise while preserving lesion boundaries and internal architecture. No aggressive image enhancement or artificial texture manipulation was performed in order to preserve biologically relevant lesion heterogeneity.

Volumetric segmentation of lesions

Three-dimensional lesion segmentation was performed using 3D Slicer software (version 5.9.0), an open-source medical image analysis platform extensively utilized in radiomics research[20]. Preprocessed DICOM datasets were imported into the Segment Editor module, and volumetric regions of interest (VOIs) encompassing the entire lesion volume were manually delineated by the primary investigator experienced in oral and maxillofacial radiology. Segmentation was conducted on a slice-by-slice basis along the axial, coronal, and sagittal planes to ensure that a proper three-dimensional reconstruction of the lesion architecture was obtained. Emphasis was laid on the maintenance of the internal septations, trabeculae, loculations, and heterogeneous internal structures since these architectural features were the main biological aspect of interest in the study. Adjacent cortical bone, teeth, restorations, and other anatomic structures were not included in the segmentation mask (Figures 1, 2, and 3). All segmentation masks were independently reviewed by two experienced oral and maxillofacial radiologists, and necessary refinements were performed through consensus evaluation to ensure anatomical accuracy and segmentation consistency. In lesions with poorly defined boundaries, conservative segmentation was performed by restricting delineation to confidently identifiable lesion tissue. Final segmentation masks were visually inspected and refined to eliminate discontinuities and segmentation irregularities. To assess segmentation reproducibility, a randomly selected subset of lesions was independently segmented by the two oral radiologists using the same segmentation protocol. Interobserver and intraobserver agreement of the segmentation masks were evaluated using dice similarity coefficient (DSC) analysis. Intraobserver reproducibility was assessed by repeat segmentation after a two-week washout interval.

Figure 1 Representative cone-beam computed tomography images demonstrating the spectrum of architectural phenotypes in jaw lesions. A: Homogeneous fluid-dominant phenotype showing uniform internal radiolucency with minimal architectural complexity and absence of internal septation; B: Intermediate septated phenotype demonstrating partial internal heterogeneity with developing architectural complexity; C: Complex heterogeneous phenotype demonstrating multilocularity, internal structural irregularity, and marked spatial heterogeneity with increased architectural complexity.

Figure 2 Representative volumetric segmentation masks of architectural phenotype groups utilized for radiomic feature extraction from cone-beam computed tomography datasets. A: Homogeneous fluid-dominant phenotype demonstrating a well-defined uniform lesion volume with minimal internal complexity; B: Intermediate septated phenotype demonstrating moderate internal architectural heterogeneity and partial structural compartmentalization; C: Complex heterogeneous phenotype demonstrating irregular volumetric architecture with multilocularity and increased internal structural complexity. Colored overlays represent manually delineated three-dimensional volumetric regions of interest used for radiomic analysis.

Figure 3 Three-dimensional rendered lesion volumes demonstrating architectural phenotype progression following volumetric segmentation of cone-beam computed tomography datasets. A: Homogeneous fluid-dominant phenotype; B: Intermediate septated phenotype; C: Complex heterogeneous phenotype. Three-dimensional rendered volumes were generated from segmented volumetric regions of interest for radiomic architectural analysis.

Radiomic feature extraction

Radiomic feature extraction was done by utilizing the open-source radiomics software package PyRadiomics (version 3.10.12) which is compliant with the recommendations of the Image Biomarker Standardization Initiative[21]. Radiomic feature extraction was done based on the pre-processed CBCT image volumes and their corresponding segmentation masks. The precise environment and configuration for radiomic extraction using PyRadiomics is provided in Supplementary material. A total of 107 raw radiomic features were extracted from each lesion volume. They consisted of shape features, first order intensity features and higher order texture features based on multiple gray level matrices, such as gray level co-occurrence matrix (GLCM), gray level dependence matrix (GLDM), gray level run length matrix (GLRLM), gray level size zone matrix (GLSZM) and neighborhood gray tone difference matrix (NGTDM). Shape features quantified the lesion geometric shape while first order intensity features quantified the distribution and dispersion of voxel intensities. On the other hand, texture features characterized the spatial relationship between the voxels in terms of heterogeneity, complexity, entropy, gray-level non-uniformity and dependency features. Since our study focused on architectural phenotype analysis rather than direct pathology classification, particular emphasis was placed on texture-derived radiomic descriptors reflecting internal spatial organization.

Feature stability assessment and selection

Radiomic feature stability was evaluated to determine the robustness of extracted features against segmentation variability. Interobserver and intraobserver reproducibility analyses were performed using radiomic features extracted from repeated lesion segmentations generated by the two oral and maxillofacial radiologists. Intraclass correlation coefficient (ICC) analysis was subsequently performed for all extracted radiomic features. Features demonstrating ICC values ≥ 0.75 were considered reproducible and were retained for further analysis, while unstable features were excluded to improve reliability of the radiomic model. Prior to statistical analysis, all retained radiomic features were standardized using z-score normalization to minimize scale-related variability and ensure comparability across feature domains. Since radiomic data demonstrated non-parametric distribution characteristics, univariate statistical analysis was performed using the Kruskal-Wallis test to identify features showing significant differences among the architectural phenotype groups. Post hoc pairwise comparisons were subsequently performed using Dunn’s test, and false discovery rate (FDR) correction was applied to control for multiple comparisons. Features with corrected P values < 0.05 were considered statistically significant. To minimize redundancy and reduce the risk of model overfitting, correlation-based feature pruning was performed using Pearson correlation analysis. Highly correlated radiomic features were eliminated while preserving biologically relevant and non-redundant descriptors representing lesion architectural organization and spatial heterogeneity. Subsequently, LASSO regression analysis was applied to identify the most informative radiomic features contributing to architectural phenotype discrimination. The final selected subset of radiomic descriptors constituted the radiomic architectural phenotype signature utilized for AI model development.

Model development and validation

AI modeling was conducted with respect to the final version of the radiomic architecture phenotype signature after the feature stability analysis and selection. For architectural phenotype classification, three different machine learning approaches were utilized, including the use of LR, SVM, and RF. LR was chosen as the main model due to the need for interpretable and clear results regarding the role of particular features in characterizing architectural phenotypes. Furthermore, the implementation of other machine learning approaches aimed to validate the robustness of the generated radiomic signature in terms of consistent performance on different machine learning architectures. For the purpose of developing and validating the predictive models, stratified five-fold cross-validation was utilized in order to ensure that all architectural phenotypes were equally represented in each fold. To improve the diagnostic classification accuracy of each individual machine learning technique, hyperparameter tuning was performed utilizing Grid Search (GridSearchCV) technique, which was limited only to the training folds of each cross validation cycle to avoid any kind of data leakage. The hyperparameter search spaces that were used along with their optimized values for each multiclass classification model are listed in Supplementary material. To ensure that the estimated class probabilities generated by the models closely aligned with real-world distribution frequencies, probability calibration was performed using Platt scaling (logistic calibration) within the cross-validation framework, and model calibration quality was quantified using the multi-class Brier score. Additionally, a DCA was performed to evaluate the clinical utility and net benefit of the developed models across a spectrum of threshold probabilities, comparing the machine learning classifiers against the default clinical strategies. In order to avoid the problem of data leakage and model overfitting, feature selection and normalization procedures were conducted only in the training sets within each fold prior to model validation. Thereby, no oversampling or undersampling techniques were applied as the class distribution remained unchanged through stratified sampling. Model performance was evaluated using receiver operating characteristic (ROC) curve analysis and area under the curve (AUC) as the primary performance metric. Other performance measures were accuracy, sensitivity, specificity, precision, and F1-Score. Confusion matrix analysis was also conducted to analyze the performance of model for classification based on architectural phenotype groups. Model performance stability in the case of linear, margin-based, and ensemble machine learning approaches was considered as an indicator of stable and biologically relevant radiomics for internal architectural phenotypes of jaw cysts and tumors.

Independent validation

As an additional step to assess the generalizability and robustness of the proposed radiomic architectural phenotype framework, an independent validation cohort consisting of 60 additional jaw lesions was analyzed independently from the main training cohort. The validation cohort included samples for all three phenotypic types, and was not involved in any of the modeling process including model training, feature selection or internal cross validation. All CBCT images in the validation cohort were preprocessed, segmented and extracted features according to the same imaging and computational pipeline used on the main cohort. The already trained LR, SVM and RF models were applied to the independent validation cohort, no retraining or further feature optimization was performed in order to achieve methodological independence. Model performance was estimated through ROC curve analysis with AUC as the main measure of performance. Other performance measures included accuracy, sensitivity, specificity, precision, F1 score and confusion matrix analysis. Stable performance on the independent validation set indicated reproducibility and applicability of radiomic characterization of jaw lesion phenotypes through CBCT.

RESULTS

Architectural phenotype distribution and segmentation reproducibility

A total of 100 histopathologically confirmed jaw lesions were included in the primary study cohort. Based on a standard statistical power equation for two independent groups, a minimum sample size of 50 subjects per category of cyst and tumor is mathematically sufficient to achieve a statistical power (1-β) of 90% at a significance level (α) of 0.05 to detect genuine structural differences between the groups. This balanced 1:1 distribution also protects the subsequent machine learning models from majority-class training bias. Based on internal radiographic architecture and spatial organization on CBCT imaging, lesions were further categorized into three architectural phenotype groups: Homogeneous fluid-dominant phenotype (n = 34), intermediate septated phenotype (n = 31), and complex heterogeneous phenotype (n = 35). The homogeneous fluid-dominant phenotype primarily included lesions demonstrating uniform internal radiolucency with minimal architectural complexity. The intermediate septated phenotype demonstrated focal septations and moderate internal heterogeneity, whereas the complex heterogeneous phenotype demonstrated multilocularity, mixed internal density, trabecular disorganization, and marked spatial heterogeneity. The detailed breakdown of histopathological lesions along with their architectural phenotype are provided in Supplementary material. Three-dimensional volumetric segmentation was successfully completed for all lesions. Segmentation reproducibility analysis demonstrated excellent agreement between observers. Interobserver DSC values ranged from 0.84 to 0.93, with a mean interobserver DSC of 0.89 ± 0.03. Intraobserver DSC values ranged from 0.87 to 0.95, with a mean intraobserver DSC of 0.91 ± 0.02, indicating highly reproducible lesion segmentation and consistent preservation of internal architectural characteristics across repeated segmentations (Table 1).

Table 1 Segmentation reproducibility analysis using dice similarity coefficient.

Reproducibility parameter	Minimum DSC	Maximum DSC	mean ± SD
Interobserver reproducibility	0.84	0.93	0.89 ± 0.03
Intraobserver reproducibility	0.87	0.95	0.91 ± 0.02

DSC: Dice similarity coefficient.

Radiomic feature stability assessment and statistical analysis

A total of 107 radiomic features were initially extracted from the segmented lesion volumes, including shape-based, first-order intensity, and higher-order texture features. Radiomic feature reproducibility analysis was subsequently performed using ICC analysis based on repeated lesion segmentations. Among the extracted radiomic features, 94 features demonstrated acceptable reproducibility with ICC values ≥ 0.75 and were retained for further analysis. Thirteen features demonstrating poor reproducibility were excluded. The retained features included 13 shape-based features, 18 first-order intensity features, and 63 texture-derived features from GLCM, GLDM, GLRLM, GLSZM, and NGTDM feature classes (Supplementary material).

Following z-score normalization, Kruskal-Wallis statistical analysis demonstrated significant differences among the architectural phenotype groups in multiple radiomic parameters. After FDR correction, 38 radiomic features remained statistically significant (P < 0.05) (Supplementary material). Significant differences were predominantly observed within texture-based radiomic domains associated with entropy, gray-level variability, spatial dependence heterogeneity, and architectural complexity. Compared with the homogeneous fluid-dominant phenotype, the complex heterogeneous phenotype demonstrated significantly elevated values for gray-level non-uniformity, dependence non-uniformity, run entropy, difference variance, and complexity-related texture features. The intermediate septated phenotype demonstrated intermediate radiomic values between the two extreme architectural phenotypes.

Radiomic architectural phenotype signature

To minimize feature redundancy and reduce the risk of model overfitting, correlation-based feature pruning was performed using Pearson correlation analysis. Highly correlated radiomic features (r > 0.85) were eliminated while preserving biologically relevant descriptors associated with spatial heterogeneity and architectural organization. Following correlation pruning, 21 non-redundant radiomic features were retained for further analysis (Table 2). Subsequent LASSO regression identified six radiomic descriptors contributing most strongly to architectural phenotype discrimination (Table 3). The final radiomic architectural phenotype signature predominantly consisted of radiomic features reflecting gray-level heterogeneity, entropy, spatial dependency variability, and architectural complexity.

Table 2 Non-redundant radiomic features retained following correlation pruning.

Feature name	Domain	P value
Shape features
Voxel volume	Shape	0.026942
First-order features
Total energy	First-order	0.025882
Root mean squared	First-order	0.029218
Variance	First-order	0.032481
Maximum	First-order	0.045327
GLCM features
Difference variance	GLCM	0.024381
GLDM features
Dependence entropy	GLDM	0.026331
Dependence non-uniformity	GLDM	0.030004
Gray level variance	GLDM	0.033192
Gray level non-uniformity	GLDM	0.039126
GLRLM features
Run length non-uniformity	GLRLM	0.028417
Gray level variance	GLRLM	0.028963
Run entropy	GLRLM	0.027496
High gray level run emphasis	GLRLM	0.043527
GLSZM features
Gray level non-uniformity	GLSZM	0.027105
Zone entropy	GLSZM	0.035674
Large area high gray level emphasis	GLSZM	0.038415
Gray level non-uniformity normalized	GLSZM	0.041118
Gray level variance	GLSZM	0.044186
NGTDM features
Coarseness	NGTDM	0.025114
Complexity	NGTDM	0.040283

NGTDM: Neighborhood gray tone difference matrix; GLSZM: Gray level size zone matrix; GLRLM: Gray level run length matrix; GLDM: Gray level dependence matrix; GLCM: Gray level co-occurrence matrix.

Table 3 Final radiomic architectural phenotype signature identified using LASSO regression.

Feature name	Domain	Architectural relevance	LASSO coefficient
Voxel volume	Shape	Represents overall volumetric architectural expansion and lesion growth pattern	1.284
Total energy	First-order	Reflects cumulative voxel intensity concentration within internal lesion architecture	1.116
Variance	First-order	Quantifies dispersion and variability of voxel intensity distribution	0.948
Difference Variance	GLCM	Captures local gray-level fluctuation and internal architectural variability	1.693
Dependence entropy	GLDM	Represents spatial dependency randomness and internal heterogeneity	1.558
Gray level non-uniformity	GLDM	Reflects irregularity of gray-level distribution across dependent voxel structures	1.402
Run length non-uniformity	GLRLM	Quantifies disruption and inconsistency of structural continuity within lesion architecture	1.341
Zone entropy	GLSZM	Represents disorder and heterogeneity of spatial zone organization	0.912
Complexity	NGTDM	Reflects overall architectural complexity and spatial organizational disarray	0.744

NGTDM: Neighborhood gray tone difference matrix; GLSZM: Gray level size zone matrix; GLRLM: Gray level run length matrix; GLDM: Gray level dependence matrix; GLCM: Gray level co-occurrence matrix.

AI model performance

The finalized radiomic architectural phenotype signature was evaluated using LR, SVM, and RF AI models. Among the evaluated models, LR demonstrated the highest overall classification performance with an AUC of 0.92 (95%CI: 0.87-0.96) (Figure 4). The model achieved an overall accuracy of 85.0%, sensitivity of 87.2%, specificity of 82.6%, precision of 0.84, and F1-score of 0.84. The RF model demonstrated comparable performance with an AUC of 0.87 (95%CI: 0.81-0.92), while the SVM model achieved an AUC of 0.84 (95%CI: 0.78-0.90) (Table 4). In terms of probability calibration, the calibrated models achieved highly favorable multi-class Brier scores, with the LR model exhibiting a score of 0.14, the SVM model showing 0.16, and the RF model showing 0.18, showing high alignment between the predicted probabilities and true architectural phenotype assignments. Confusion matrix analysis demonstrated that most classification overlap occurred between intermediate septated lesions and complex heterogeneous lesions. The DCA further confirmed the clinical utility of the proposed framework across the cohorts. All three models demonstrated a positive clinical net benefit compared to the default treat all and treat none intervention strategies within standard clinical decision thresholds. Specifically, the LR model provided a superior and stable clinical net benefit across the widest threshold range of 0.20 to 0.80, followed closely by the SVM model which maintained a positive net benefit across a threshold range of 0.25 to 0.75, and the RF model which showed clinical utility across a narrower threshold range of 0.30 to 0.70. These parameters validate the potential that radiomics-guided framework may be useful for objective clinical decision support.

Figure 4 Receiver operating characteristic curves showing the performance of artificial intelligence models. ROC: Receiver operating characteristic; AUC: Area under the curve.

Table 4 Artificial intelligence model performance for architectural phenotype classification, 95%CI.

Model	AUC	Accuracy (%)	Sensitivity (%)	Specificity (%)	Precision	F1-score
Logistic regression	0.92 (0.87-0.96)	85.0 (78.2-90.4)	87.2 (80.1-92.5)	82.6 (74.8-88.9)	0.84 (0.77-0.90)	0.84 (0.77-0.90)
Random forest	0.87 (0.81-0.92)	81.0 (73.7-86.9)	83.1 (75.4-89.1)	78.4 (70.2-85.1)	0.80 (0.72-0.86)	0.80 (0.72-0.86)
Support vector machine	0.84 (0.77-0.89)	79.0 (71.5-85.2)	80.4 (72.5-86.8)	76.8 (68.4-83.7)	0.78 (0.70-0.85)	0.78 (0.70-0.85)

AUC: Area under the curve.

Independent validation

Independent validation analysis was performed using an additional cohort of 60 jaw lesions not included in model training or internal cross-validation. The validation cohort demonstrated similar radiomic distribution patterns across the three architectural phenotype groups. The LR model maintained the highest classification performance during independent validation, achieving an AUC of 0.89 (95%CI: 0.82-0.95) and overall accuracy of 83.3%. The RF model achieved an AUC of 0.85, while the SVM model demonstrated an AUC of 0.81. The findings demonstrated stable model performance and reproducible radiomic characterization of internal architectural phenotypes across independent lesion datasets (Table 5). The stable performance metrics preserved when transitioning from the cross-validated training folds to the independent validation cohort also shows the absence of overfitting. A flowchart showing the study’s model development and validation process is provided in Figure 5.

Figure 5 Workflow diagram showing image acquisition, segmentation, radiomic feature extraction, feature selection, model development, and validation. ICC: Intraclass correlation coefficient; FDR: False discovery rate; CBCT: Cone-beam computed tomography.

Table 5 Independent validation cohort performance, 95%CI.

Model	AUC	Accuracy (%)	Sensitivity (%)	Specificity (%)	Precision
Logistic regression	0.89 (0.82-0.95)	83.3 (74.2-90.3)	84.7 (75.1-91.7)	81.5 (71.3-89.2)	0.82 (0.73-0.89)
Random forest	0.85 (0.77-0.91)	80.0 (70.5-87.5)	81.2 (71.2-88.8)	78.6 (68.1-86.8)	0.79 (0.69-0.87)
Support vector machine	0.81 (0.72-0.88)	76.7 (66.9-84.8)	78.4 (68.2-86.5)	74.1 (63.3-83.1)	0.76 (0.66-0.84)

AUC: Area under the curve.

DISCUSSION

The clinical management of jaw cysts and tumors remains challenging due to similar and overlapping imaging presentations on conventional radiographs, although they are known to have varied histopathological patterns and inherent biological behavior. This challenge has been overcome with advanced modalities like CBCT that tend to give a comprehensive three-dimensional picture of the lesion and the surrounding structures. However, interpreting such CBCT images is largely qualitative and subjected to the expertise of the clinician. This can lead to inaccurate or delayed diagnosis, potentially reducing the overall prognosis of the patient. This also suggests that there is a strong need to develop a quantitative system that helps in better and quicker diagnostic process. Thus, the current study focused on investigating the role and efficiency of radiomic features derived from CBCT images to quantitatively characterize jaw cysts and tumors. This architectural phenotype classification can help in better diagnosis, treatment planning, and prognosis. Rather than attempting for a direct two-way diagnostic classification between each pathology, which is a common approach used in prior research regarding oral and maxillofacial radiomics but tends to ignore overlapping architecture, this study focused on using radiomics as a technique to analyze image phenotypes. This was done by classifying jaw lesions into phenotypes that range from homogeneous and fluid-filled to more heterogeneous phenotypes.

The extraction of a refined 9 radiomic feature set via LASSO regression ensured that our characterization remains focused strictly on mathematically non-redundant, high-yield indicators of internal complexity. The clinical significance and increased LASSO weights of our final set of features stress that the high-order structure of the image is indicative of the physical characteristics of these lesions. The feature “Difference Variance” from GLCM showed maximum predictive weight (1.693) to establish its clinical applicability as a descriptor of regional changes in intensity transitions. On a biological front, this implies the transition from one area of fluid-filled cavities, soft tissue elements, and mineralization of bones to another. Similarly, high weights for dependence entropy (1.558) and gray level non-uniformity (1.402) of GLDM features along with zone entropy (0.912) from GLSZM features imply substantial randomness and chaos of structures. Lesions categorized under the homogeneous fluid-dominant phenotype demonstrated relatively uniform voxel intensity distribution and lower spatial complexity, whereas lesions belonging to the complex heterogeneous phenotype demonstrated significantly elevated entropy, dependency variability, and structural irregularity. The intermediate septated phenotype demonstrated radiomic values between the two extremes, suggesting the presence of a continuous architectural spectrum rather than rigid categorical separation. This observation is clinically important because multiple jaw lesions demonstrate gradual transitions in internal architecture during biological progression[22-24]. The progressive increase of these texture scores from the homogeneous fluid-dominant phenotype to the intermediate septated and complex heterogeneous phenotypes confirms that high-dimensional parameters successfully replace subjective terminology with reproducible mathematical signatures. Santos et al’s systematic review provides compelling support for the results of our study[25], showing that the most common and potent feature types for diagnosing complex maxillofacial pathology, cysts, and tumors are GLCM, GLRLM, and GLSZM texture matrices. Although the lack of standardization is an important issue pointed out by the authors, we have managed to address this problem within the context of our current research by creating a standardized and normalized phenotypic approach using CBCT images as proxies for biological aggressiveness. When interpreted from a biological perspective, the prominent role of gray level non-uniformity and Entropy metrics within our final signature directly reflects underlying histopathological variation. Specifically, elevated texture entropy strongly correlates with micro-architectural complexity, such as the alternating zones of cellular proliferation, fibrous septa, and irregular osteolytic borders typical of complex heterogeneous phenotypes. Conversely, low-entropy, highly uniform gray-level features mathematically describe the structural homogeneity of fluid-dominant phenotypes, matching the uniform, protein-rich fluid clusters seen histologically in simple cysts.

A significant clinical application of this study is that phenotyping lesions based on this radiomic set can simplify the diagnostic workflow even before histopathological results. The ability of this radiomic set to classify the jaw lesions suggests that radiomics and radiomic-incorporated AI models can be used as a powerful diagnostic adjunct for quantification in routine clinical practice[25-27]. Traditionally, differentiating complex odontogenic lesions relies heavily on a clinician’s subjective radiographic interpretation, which can occasionally lead to diagnostic uncertainty. By mapping multi-class structural phenotypes, the current model may assist clinicians by providing a more objective assessment of internal tissue complexity and sub-visual intra-lesional heterogeneity. In clinical practice, this characterization could help guide treatment selection. For example, features that lean toward an aggressive, infiltrative tumor phenotype (such as ameloblastoma) rather than a benign cystic entity might help the surgical team consider definitive marginal resection over conservative enucleation. Consequently, integrating automated architectural phenotyping into the diagnostic workflow has the potential to help optimize surgical margins, potentially reducing postoperative recurrence rates and improving patient-specific treatment outcomes. The results of the current study are consistent with recent findings by Muraoka et al[27], who demonstrated that machine learning models utilizing radiomic features can effectively classify and differentiate odontogenic cysts and tumors. Their workflow utilized multi-modal CT and MRI protocols to achieve high diagnostic accuracy. Our study achieved a similar classification performance (AUC of 0.92) using CBCT-derived texture signatures[27]. Similarly, Committeri et al[28] found that machine learning models incorporating CBCT-derived radiomic features could successfully differentiate between dentigerous cysts, odontogenic keratocysts, and unicystic ameloblastomas. While their model integrated radiomics with peripheral inflammatory biomarkers, both studies demonstrate a consistent ability of radiomic parameters to effectively characterize and classify these specific jaw lesions. These outcomes show that analysis based on radiomics (irrespective of the imaging modality used) can provide a highly capable methodology for objectively characterizing the internal structural variations of jaw lesions. A strength of this study was the utilization of interpretable AI models like LR to overcome the complete black-box nature of a model. Models, especially those that are used in clinical setting, should be interpretable and explainable for clinicians to understand the decision-making process of an AI model in a transparent way[29,30]. The biological relevance of the present study is further validated with independent cohort testing. The maintenance of high AUC values during independent validation suggests that the identified radiomic features were not merely dataset-specific statistical findings but represented reproducible imaging biomarkers associated with internal lesion organization. This reproducibility is essential for future clinical translation of radiomic analysis into real-world oral radiology practice.

This study also has certain limitations which should be addressed in future studies. The study relied on retrospective data that was obtained from a single institution. The study should be expanded to include datasets from multiple vendors externally with multiple acquisition parameters in order to generalize the findings of the present study. Longitudinal evaluation of the proposed architectural phenotyping radiomic feature set should be done to provide additional details of the underlying biological behavior of jaw cysts and tumors. This study included VOI masks that were obtained after manual segmentation. Despite using DSC, the use of automated segmentation methods can improve the accuracy and reduce the time taken[31,32]. An inherent limitation of the study is to categorize the lesions into specific phenotypes based on visual assessment, which can be a cause of introduction of subjectivity within the study. In order to overcome any possible discrepancies caused due to this, a two-reader consensus reading approach was adopted for the assessment of cases. Subsequently, the significant differences (P < 0.05) seen in the radiomic feature matrices are indicative of the fact that these phenotypes represent true structural complexities rather than mere observer variance alone. Moreover, overlap between pathological entities and architectural phenotypes may invariably occur in lesions demonstrating mixed or transitional internal organization. This shows that clinicians should be cautious while interpreting such mixed lesions. Another limitation that is inherent to CBCT diagnostics is the lack of standardized grayscale values like Hounsfield unit used in CT[33,34]. This lack of standardization can cause a significant drift in the radiomic features extracted when multiple imaging systems are used. Thus, future studies should focus on acquiring larger multicenter cohorts, developing automated segmentation systems, and integrating radiomics with explainable deep learning frameworks to improve the robustness and clinical applicability of architectural phenotype analysis in jaw lesions.

CONCLUSION

In conclusion, the current study successfully validated that a concise radiomic signature derived from CBCT, along with the application of AI-based models, is able to perform quantitative classification of architectural phenotypes in jaw cystic lesions and tumors, achieving reliable reproducibility and excellent diagnostic efficacy. Based on the study results, radiomic architectural phenotyping appears to be capable of streamlining radiologic evaluation of jaw lesions through the provision of an objective quantitative measurement of internal architecture prior to histopathologic verification.