CBCT-Based Clinico-Radiomic Nomogram Predicting Preoperative Mandibular Third Molar Difficulty: Development/Validation

Abstract

Objective

This study aimed to develop and validate a clinico-radiomic nomogram, integrating Cone-Beam Computed Tomography (CBCT) radiomic features with clinical characteristics, to predict the difficulty of mandibular third molar (MTM) extraction.

Methods

A retrospective cohort of 600 patients undergoing MTM extraction was divided into training and validation sets. Radiomic features were extracted from preoperative CBCT images. The Least Absolute Shrinkage and Selection Operator (LASSO) method was used to select key features and build a radiomic score (Rad-score). Multivariable logistic regression identified independent clinical predictors. A final nomogram was established by combining the Rad-score and clinical factors. The model's performance was assessed for discrimination (AUC), calibration (Hosmer–Lemeshow test), and clinical utility (Decision Curve Analysis - DCA).

Results

Eleven robust features were selected to construct a radiomic signature. The Rad-score was significantly higher in the high-difficulty group (P < .001). Five independent predictors were identified: Age, BMI, Pell & Gregory classification, Root Curvature, and the Rad-score. The combined clinico-radiomic nomogram demonstrated superior predictive performance in both training (AUC = 0.892) and validation (AUC = 0.865) cohorts, significantly outperforming models based solely on clinical or radiomic factors alone. The model showed excellent calibration between predicted and observed probabilities and demonstrated substantial clinical net benefit via DCA.

Conclusion

The novel CBCT-based clinico-radiomic nomogram provides a noninvasive, accurate, and visual tool for the preoperative stratification of MTM extraction difficulty. This facilitates more informed surgical planning and personalised patient counselling.

Introduction

The extraction of impacted mandibular third molars (IMTMs) is one of the most common procedures in oral and maxillofacial surgery. These teeth have a high prevalence of impaction and may fail to erupt into a normal functional position, which often necessitates surgical removal to prevent or treat complications such as pericoronitis, caries, root resorption of adjacent teeth, periodontal defects, and cystic lesions.^1,2 Given the frequency of this procedure, accurately predicting surgical difficulty is clinically important. A precise preoperative assessment can support clinical decision-making, improve patient counselling regarding potential risks such as inferior alveolar nerve injury, optimise surgical scheduling, and help allocate appropriate surgical expertise and resources.^2,3

Traditionally, the preoperative assessment of IMTM extraction difficulty has relied on classification systems based on 2-dimensional panoramic radiographs, particularly the Pell and Gregory and Winter classifications, with later quantification attempts such as the Pederson Difficulty Index.⁴ However, the predictive performance of these indices has been inconsistent, and validation studies have reported limited reliability and only moderate discriminatory value for surgical difficulty prediction.^4,5 This limitation is partly attributable to the inherent deficiencies of 2-dimensional panoramic imaging, including magnification, distortion, and anatomical superimposition. More importantly, panoramic radiographs cannot adequately capture essential 3-dimensional determinants of extraction difficulty, including fine root morphology, bucco-lingual bone thickness, localised trabecular and cortical bone characteristics, and the true spatial relationship between the tooth roots and the mandibular canal.^5,6 The absence of these dimensions restricts the accuracy of conventional preoperative assessments.

Cone-beam computed tomography (CBCT) provides high-resolution 3-dimensional visualisation of dentomaxillofacial structures and can overcome several geometric and anatomical limitations of panoramic radiography.^5,6 In selected IMTM cases, CBCT can more accurately depict root configuration, impaction depth, surrounding alveolar bone, and the relationship between the tooth and the mandibular canal, thereby offering a more complete anatomical basis for surgical planning.^5,6 Beyond visual interpretation, radiomics has emerged as a quantitative imaging approach that converts standard medical images into high-dimensional mineable data.⁷ By extracting shape, first-order, and texture features from defined regions of interest, radiomics can capture subtle patterns of tissue heterogeneity and structural organisation that are not readily appreciable by the naked eye.^7,8 Recent reviews have emphasised that radiomics is increasingly being extended to CBCT imaging and dentomaxillofacial applications, including jaw lesions, temporomandibular joint disorders, dental implant planning, periapical bone assessment, and other maxillofacial diagnostic tasks.8, 9, 10 These studies suggest that CBCT-based radiomics may provide objective imaging biomarkers that complement conventional anatomical assessment by quantifying local bone texture, density-related heterogeneity, and microarchitectural features relevant to oral and maxillofacial decision-making.8, 9, 10

In the context of IMTM extraction, surgical difficulty is determined not only by macroscopic anatomical factors such as impaction depth, angulation, root curvature, and canal proximity, but also by the biomechanical resistance of the surrounding bone. Conventional classifications mainly describe tooth position and angulation, whereas CBCT-derived radiomic features may quantify the local tooth–bone microenvironment more comprehensively. The integration of clinical variables, conventional radiographic features, and radiomic signatures therefore provides a rational framework for developing individualised prediction models. Recent reviews have shown that radiomics and deep learning are increasingly being applied in dentomaxillofacial imaging, including CBCT-based assessment, jaw lesions, temporomandibular joint disorders, implant-related decision-making, and maxillofacial disease diagnosis.8, 9, 10 In addition, recent oral and maxillofacial imaging studies have further demonstrated the feasibility of using quantitative imaging biomarkers for preoperative prediction and individualised risk stratification. For example, a CBCT-based multitask deep learning radiomics nomogram has been developed to predict implant failure risk after sinus floor elevation,¹¹ ultrasound-based deep learning radiomics has been used to predict cervical lymph node metastasis in major salivary gland carcinomas,¹² and contrast-enhanced CT-based deep learning and habitat radiomics have been applied to predict cervical lymph node metastasis and pathological subtypes in oral squamous cell carcinoma.¹³ These findings collectively support the rationale for applying CBCT-derived radiomic features to enhance individualised prediction of IMTM extraction difficulty.

Therefore, this study aimed to develop and validate a CBCT-based clinico-radiomic nomogram for the preoperative prediction of IMTM extraction difficulty. By integrating key clinical and radiographic variables with objective radiomic features extracted from preoperative CBCT images, we hypothesised that the proposed model could provide a more accurate and individualised assessment of surgical difficulty than conventional clinical assessment alone, thereby supporting surgical planning, risk communication, and personalised patient management.

Methods

Study design and ethical statement

This single-centre, retrospective cohort study was conducted to develop and validate a comprehensive clinico-radiomic nomogram for predicting the surgical difficulty of MM3 extraction. This study was approved by the human subjects ethics board of Wuxi Stomatological Hospital and was conducted in accordance with the Helsinki Declaration of 1975, as revised in 2013. Given the retrospective nature of the study, which utilised deidentified historical data from electronic medical records and radiographic archives, the requirement for written informed patient consent was formally waived by the institutional review board. Clinical and imaging data were systematically retrieved from the Electronic Medical Records (EMR) and Picture Archiving and Communication Systems (PACS) of the Department of Oral and Maxillofacial Surgery at *, a tertiary academic dental centre, encompassing the period from January 1, 2020, to December 31, 2024.

Study population

A systematic query was performed to screen all adult patients who underwent MM3 extraction at the Department of Oral and Maxillofacial Surgery of our tertiary academic dental centre. To ensure data integrity and cohort suitability, the initial list of potential participants underwent a rigorous, multistage validation process. First, an automated algorithm filtered the cohort based on structured electronic health record (EHR) and Picture Archiving and Communication System (PACS) data, using procedure codes for MM3 surgery and confirming the availability of a preoperative cone-beam computed tomography (CBCT) scan. Second, the medical and radiographic records of the shortlisted patients were independently reviewed by 2 calibrated investigators to verify all eligibility criteria. Finally, any discrepancies regarding diagnosis, image quality, or outcome data were resolved by a senior maxillofacial surgeon through consensus.

Patients were required to meet all of the following inclusion criteria to be enrolled in the study: (1) aged between 18 and 60 years at the time of the procedure, a range selected to focus on a population with completed dental and skeletal development while minimising the confounding effects of systemic diseases more prevalent in older age¹⁴; (2) a definitive diagnosis of an impacted MM3 requiring surgical extraction; (3) complete root development, confirmed radiographically by closed root apices on the preoperative CBCT scan¹⁵; and (4) availability of complete baseline clinical data and follow-up records, including precise documentation of surgical duration. The exclusion criteria were applied to minimise confounding variables and included any of the following: (1) the presence of odontogenic cysts, tumours, osteomyelitis, or other jawbone lesions that could substantially alter local bone architecture and introduce lesion-related radiomic heterogeneity unrelated to extraction difficulty¹⁶; (2) a history of head and neck radiotherapy, which induces microvascular damage and fibroatrophic changes that compromise bone healing¹⁷; (3) a history of treatment with antiresorptive or antiangiogenic medications (eg, bisphosphonates), which are known risk factors for medication-related osteonecrosis of the jaw¹⁸; or (4) surgery performed during an active phase of acute pericoronitis, as clinical practice guidelines recommend resolving the acute infection prior to elective extraction to avoid increased postoperative morbidity.¹⁹

The sample size was determined based on the events per variable (EPV) criterion. Considering that difficult extractions represent a substantial proportion of cases, with studies using the Pederson difficulty index categorising approximately 25% to 30% of MM3 removals as moderately difficult,²⁰ based on preliminary data and similar radiomics studies, we anticipated that the number of candidate predictors (including clinical variables and selected radiomic features) for the final multivariable model might be around 10.²¹ Therefore, targeting an EPV of 10 for the less frequent outcome (difficult extraction) was deemed appropriate to ensure model stability. Assuming a conservative event rate of 25%, a minimum of 400 patients would be required (10 variables × 10 EPV/0.25 event rate). To ensure sufficient statistical power and to account for potential data attrition, a target sample size of 600 patients was established.²² The final eligible cohort was then randomly partitioned into a training set (70%, n = 420) and a validation set (30%, n = 180) using a computer-generated random seed. Stratified sampling was employed to ensure a balanced distribution of the outcome event between the 2 datasets.

Outcome definition

The primary outcome of this study was surgical difficulty, which was operationalised using surgical duration as a robust and objective surrogate marker. While subjective assessments like the visual analogue scale (VAS) are susceptible to interoperator variability and bias, procedure time is widely recognised as a highly reproducible and direct indicator of the technical challenges encountered during an extraction.²³ Surgical duration was rigorously defined as the interval from the initial mucosal incision with a scalpel to the placement and completion of the final suture knot. This interval explicitly excludes the time for local anaesthesia administration and onset of effect. This time, precise to the minute, was systematically extracted from the standardised electronic surgical records. To categorise surgical difficulty for binary logistic regression analysis, a predefined cut-off value was established. In line with findings from studies that have correlated extraction time with postoperative complications and perceived difficulty,²⁴ cases with a surgical duration exceeding 20 minutes were classified into the “Difficult Group,” while those completed in 20 minutes or less were assigned to the “Non-Difficult Group.” This threshold effectively distinguishes routine extractions from more complex procedures typically requiring extensive bone removal or tooth sectioning. To minimise the confounding effect of operator skill, we included only procedures performed by attending oral and maxillofacial surgeons or by senior residents under direct supervision, with the primary operator having a minimum of 5 years of postgraduate surgical experience. Cases performed solely by junior residents were excluded. This criterion ensures that the recorded surgical time predominantly reflects the intrinsic difficulty of the case rather than the operator’s level of training or skill, a methodological refinement supported by previous studies on surgical difficulty.²⁵

Data collection and feature extraction

Clinical and conventional radiographic variables

Clinical and radiographic data were retrospectively extracted from the EHR and PACS. Two board-certified oral and maxillofacial radiologists (blinded to surgical duration) independently evaluated the images. Interobserver agreement for categorical variables was assessed using Cohen’s Kappa (κ), and discrepancies were resolved by consensus. The collected variables included: (1) Demographics: Age, gender, and Body Mass Index (BMI) were recorded. Higher BMI and age are established risk factors associated with limited surgical access and bone ankylosis, respectively.²⁶ (2) Anatomical Classification: Impacted teeth were categorised using the standard Pell & Gregory (depth/ramus relationship) and Winter (angulation) classifications.²⁷ (3) Root Morphology: Root curvature was quantitatively classified based on the angle between the root axis and the crown long axis: straight (< 10 degrees), curved (10–40 degrees), or severely curved/hooked (> 40 degrees).²⁸ (4) Periodontal Ligament (PDL) Space: The PDL interface was assessed as normal or narrowed/obliterated. The latter was strictly defined as the loss of the continuous radiolucent line around the root on at least 2 orthogonal CBCT slices, serving as a proxy for ankylosis. (5) Relationship to Mandibular Canal: Signs of nerve proximity (eg, darkening of roots, interruption of the cortical white line) were recorded.²⁹ (6) Alveolar Bone Thickness: Buccal and lingual cortical bone plate thickness was measured perpendicular to the bone surface at 2 standardised locations: the cemento-enamel junction (CEJ) and the root apex, using axial CBCT slices.³⁰

Radiomics workflow

The radiomics workflow strictly adhered to the Image Biomarker Standardization Initiative (IBSI) guidelines.³¹ Image Preprocessing: All CBCT scans were acquired using a single device (KaVo 3D eXam, KaVo Dental GmbH, Biberach, Germany) with a strictly standardised clinical acquisition protocol to prevent scanner-induced variability: tube voltage of 120 kV, tube current of 8.9 mA, field of view (FOV) of 16 × 13 cm, and an original voxel size of 0.3 mm, reconstructed using the standard filtered back-projection algorithm. To further guarantee data uniformity for radiomics analysis in accordance with IBSI guidelines,³² a 2-step pipeline was applied. First, all DICOM images were resampled to an isotropic voxel size of 1.0 × 1.0 × 1.0 mm³ using B-spline interpolation. Second, image intensity was normalised using Z-score standardisation on a per-image basis (x' = [x - mean]/SD), rescaling grey levels to a zero mean and unit variance. Segmentation and Reliability: Segmentation was performed using ITK-SNAP (v3.8.0). A dual-ROI strategy was employed: a Tooth ROI covering the entire tooth structure, and a Bone Microenvironment ROI generated by a 3-mm morphological dilation from the root surface to capture periradicular bone texture.³³ To ensure reproducibility, 50 randomly selected cases were resegmented. Segmentation reproducibility accuracy was evaluated using the Dice Similarity Coefficient (DSC). Feature extraction stability was assessed using the Intraclass Correlation Coefficient (ICC); only features with an ICC > 0.75 were retained. Feature Extraction: Quantitative features were extracted using the PyRadiomics platform (v3.0.1).³⁴ Features included: (1) Shape-based (eg, Sphericity); (2) First-order statistics (eg, Entropy); (3) Texture features (GLCM, GLRLM, GLSZM, NGTDM, GLDM) to quantify trabecular heterogeneity³⁵; and (4) Transformed features. To capture anatomical details at multiple scales, we applied Wavelet decompositions and Laplacian of Gaussian (LoG) filters with sigma values of 1.0, 3.0, and 5.0 mm, thereby capturing texture patterns at multiple spatial scales. A total of 1316 features were extracted per patient.

Statistical analysis

All statistical computations were performed using R software (version 4.3.2) with specific packages including glmnet, rms, pROC, and dca.R. A 2-sided P value of less than .05 was considered statistically significant. The prevalence of missing data was assessed. Given the low overall missing rate (< 5% for any variable), we employed multiple imputation by chained equations (MICE) to handle all missing values under the missing-at-random assumption, creating and pooling results from 5 imputed datasets. To rigorously evaluate the randomisation balance between the training and validation cohorts, Standardized Mean Differences (SMDs) were quantified for all baseline variables, with an absolute SMD value of less than 0.1 considered to indicate adequate balance, thereby overcoming the potential sample-size dependency of traditional hypothesis testing. Complementary statistical tests were also performed for descriptive purposes: continuous variables were assessed for normality using the Shapiro-Wilk test; normally distributed data were expressed as mean ± standard deviation (SD) and compared using the Student's t-test, whereas non-normally distributed data were presented as median (interquartile range, IQR) and compared using the Mann–Whitney U test. Categorical variables were reported as frequencies (percentages) and analysed using the Chi-square test or Fisher's exact test.

For the high-dimensional radiomic data, a multistep dimensionality reduction strategy was employed to prevent overfitting and eliminate redundancy. Initially, features demonstrating poor reproducibility (Intraclass Correlation Coefficient < 0.75) were discarded. To further compress the high-dimensional space and eliminate highly redundant variables prior to modelling, a Pearson correlation matrix was constructed; for any pair of features exhibiting a correlation coefficient greater than 0.90, the feature with the lower mean absolute correlation with the remaining features was removed. The remaining robust and independent features were then subjected to Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression in the training set. To select the most predictive features, 10-fold cross-validation was performed to determine the optimal regularisation parameter (lambda); specifically, the value that minimised the binomial deviance (lambda.min) was selected to maximise predictive accuracy. A radiomics signature, termed “Rad-score,” was then calculated for each patient via a linear combination of the selected features weighted by their respective nonzero coefficients.

To identify independent predictors of surgical difficulty and construct the final prediction models, a systematic screening process was conducted in the training cohort. Univariate analysis was first applied to all clinical variables (eg, age, BMI, Pell & Gregory classification) and the Rad-score; variables achieving a P value less than .10 were advanced to multivariable analysis. To verify the incremental value of the radiomic signature, 3 distinct multivariable logistic regression models were constructed and compared: (1) a Clinical Model, built using only the significant clinical predictors identified via backward stepwise regression based on the Akaike Information Criterion (AIC); (2) a Radiomics Model, consisting solely of the Rad-score; and (3) a Combined Nomogram Model, integrating the Rad-score with the independent clinical predictors. To ensure the robustness of the final model, the Variance Inflation Factor (VIF) was calculated for all included predictors, with a strict threshold of VIF < 5 applied to exclude multicollinearity. Furthermore, a sensitivity analysis was performed by treating “surgeon identity” as a covariate in the multivariable model to confirm that the predictive performance was driven by patient-specific factors rather than operator variability.

The predictive performance of the developed models was comprehensively validated in both the training and independent validation sets. Discrimination was quantified using the Area Under the Receiver Operating Characteristic (ROC) Curve (AUC), with the DeLong test used to statistically compare the Combined Nomogram against the Clinical and Radiomics models. Calibration was assessed by plotting calibration curves and performing the Hosmer–Lemeshow goodness-of-fit test, where a P value greater than .05 indicated satisfactory agreement between predicted probabilities and observed outcomes. Finally, to evaluate clinical utility, Decision Curve Analysis (DCA) was performed to calculate the net benefit of the model across a range of threshold probabilities, determining whether the nomogram-guided decision-making offers superiority over default strategies.

Results outline

Baseline characteristics and univariate analysis of difficulty factors

A total of 842 patients who underwent MM3 extraction were initially screened. After applying the strict inclusion and exclusion criteria, 242 patients were excluded due to age incompatibility (< 18 or > 60 gt; 60 years), incomplete radiographic records, or the presence of confounding pathological lesions (eg, cysts). Consequently, a final cohort of 600 eligible patients was included in the analysis. The study cohort was randomly partitioned into a training set (n = 420) and an independent validation set (n = 180). Table 1 summarises the baseline demographic, clinical, and radiographic characteristics of patients in both datasets. The mean age of the total population was 26.4 ± 5.8 years, with a slight female predominance (53.5%). Statistical comparison revealed no significant differences between the training and validation sets across all analysed variables (all P > .05). Furthermore, the SMD for all covariates remained below 0.10, quantitatively confirming that the randomisation process achieved an excellent balance of baseline characteristics and minimised selection bias between the 2 cohorts.

Table 1.

Baseline characteristics of patients in the training and validation sets.

Characteristic	Training set (n = 420)	Validation set (n = 180)	SMD	Statistic	P-value
Demographics
Age (years), mean ± SD	26.5 ± 5.9	26.2 ± 5.6	0.052	t = 0.584	.559
Gender, n (%)			0.024	χ² = 0.045	.832
Male	195 (46.4)	84 (46.7)
Female	225 (53.6)	96 (53.3)
BMI (kg/m²), mean ± SD	22.8 ± 3.1	23.0 ± 2.9	0.067	t = –0.742	.458
Side, n (%)			0.018	χ² = 0.028	.867
Left	218 (51.9)	92 (51.1)
Right	202 (48.1)	88 (48.9)
Anatomical classification
Pell & Gregory class, n (%)			0.041	χ² = 0.412	.814
I	118 (28.1)	52 (28.9)
II	220 (52.4)	97 (53.9)
III	82 (19.5)	31 (17.2)
Pell & Gregory position, n (%)			0.035	χ² = 0.285	.867
A	115 (27.4)	51 (28.3)
B	242 (57.6)	103 (57.2)
C	63 (15.0)	26 (14.4)
Winter's classification, n (%)			0.048	χ² = 1.152	.886
Vertical	105 (25.0)	44 (24.4)
Mesioangular	188 (44.8)	79 (43.9)
Horizontal	78 (18.6)	35 (19.4)
Distoangular	38 (9.0)	18 (10.0)
Inverted/Other	11 (2.6)	4 (2.2)
Radiographic features
Root number, n (%)			0.029	χ² = 0.125	.939
Single/Conical	142 (33.8)	62 (34.4)
Two roots	256 (61.0)	109 (60.6)
Multiple (> 2)	22 (5.2)	9 (5.0)
Root curvature, n (%)			0.038	χ² = 0.354	.838
Straight (< 10°)	168 (40.0)	74 (41.1)
Curved (10–40°)	195 (46.4)	82 (45.6)
Severe/Hooked (> 40°)	57 (13.6)	24 (13.3)
PDL space, n (%)			0.015	χ² = 0.018	.893
Normal	365 (86.9)	157 (87.2)
Narrowed/Obliterated	55 (13.1)	23 (12.8)
Relation to mandibular canal, n (%)			0.031	χ² = 0.092	.761
Separated	285 (67.9)	124 (68.9)
Contact	135 (32.1)	56 (31.1)
Alveolar bone thickness (mm)
Buccal (CEJ level)	1.42 ± 0.45	1.45 ± 0.42	0.069	t = –0.782	.435
Lingual (CEJ level)	1.85 ± 0.52	1.82 ± 0.55	0.056	t = 0.615	.539
Buccal (Apex level)	2.10 ± 0.65	2.15 ± 0.62	0.078	t = –0.892	.373
Lingual (Apex level)	1.35 ± 0.48	1.32 ± 0.45	0.064	t = 0.715	.475
Outcomes
Surgical duration (min), median [IQR]	15 [11, 28]	16 [10, 29]	0.045	Z = –0.412	.68
Difficulty group, n (%)			0.009	χ² = 0.005	.942
Low-to-moderate (≤ 20 min)	312 (74.3)	134 (74.4)
High (> 20 gt; 20 min)	108 (25.7)	46 (25.6)

In the training set, patients were further stratified based on surgical duration into the Low-to-Moderate Difficulty group (≤ 20 minutes, n = 312) and the High Difficulty group (> 20 gt; 20 minutes, n = 108). As presented in Table 2, univariate analysis demonstrated distinct differences between the 2 groups. Patients in the High Difficulty group were significantly older (Mean: 28.6 vs 25.8 years, P < .001), suggesting reduced bone elasticity, and had a higher BMI (23.3 vs 22.6 kg/m², P = .038), which may correlate with restricted surgical access. Anatomically, the High Difficulty group exhibited a significantly higher prevalence of deep impactions (Pell & Gregory Class III: 44.4% vs 10.9%, P < .001; Position C: 34.3% vs 8.3%, P < .001). Additionally, complex root morphologies, such as severe curvature or hook-like apices, were more frequent in difficult cases (33.3% vs 6.7%, P < .001). Radiographic signs of ankylosis, indicated by a narrowed or obliterated PDL space, were also significantly enriched in the High Difficulty group (29.6% vs 7.4%, P < .001). Regarding alveolar bone, buccal bone thickness at the apical level was significantly greater in the High Difficulty group (P = .021).

Table 2.

Univariate analysis of characteristics associated with high extraction difficulty in the training set (n = 420).

Characteristic	Low-to-moderate difficulty (n = 312)	High difficulty (n = 108)	Statistic	P-value
Demographics
Age (years), mean ± SD	25.8 ± 5.2	28.6 ± 6.5	t = –4.125	< .001
Gender, n (%)			χ² = 0.582	.445
Male	148 (47.4)	47 (43.5)
Female	164 (52.6)	61 (56.5)
BMI (kg/m²), mean ± SD	22.6 ± 2.9	23.3 ± 3.4	t = –2.084	.038
Anatomical classification
Pell & Gregory class, n (%)			χ² = 72.415	< .001
I	108 (34.6)	10 (9.3)
II	170 (54.5)	50 (46.3)
III	34 (10.9)	48 (44.4)
Pell & Gregory position, n (%)			χ² = 58.632	< .001
A	102 (32.7)	13 (12.0)
B	184 (59.0)	58 (53.7)
C	26 (8.3)	37 (34.3)
Winter's classification, n (%)			χ² = 38.921	< .001
Vertical	88 (28.2)	17 (15.7)
Mesioangular	155 (49.7)	33 (30.6)
Horizontal	42 (13.5)	36 (33.3)
Distoangular	22 (7.0)	16 (14.8)
Inverted/Other	5 (1.6)	6 (5.6)
Radiographic features
Root curvature, n (%)			χ² = 48.552	< .001
Straight (<10°)	145 (46.5)	23 (21.3)
Curved (10–40°)	146 (46.8)	49 (45.4)
Severe/Hooked (> 40°)	21 (6.7)	36 (33.3)
PDL Space, n (%)			χ² = 34.128	< 0.001
Normal	289 (92.6)	76 (70.4)
Narrowed/Obliterated	23 (7.4)	32 (29.6)
Relation to mandibular canal, n (%)			χ² = 9.845	0.002
Separated	225 (72.1)	60 (55.6)
Contact	87 (27.9)	48 (44.4)
Alveolar bone thickness (mm)
Buccal (CEJ level)	1.41 ± 0.43	1.45 ± 0.48	t = –0.785	0.433
Lingual (CEJ level)	1.84 ± 0.51	1.89 ± 0.56	t = –0.842	0.401
Buccal (Apex level)	2.05 ± 0.62	2.22 ± 0.68	t = –2.315	0.021
Lingual (Apex level)	1.34 ± 0.47	1.36 ± 0.51	t = –0.368	0.713

Radiomic feature selection and Rad-score construction

To ensure the reproducibility and robustness of the radiomic signature, an interobserver stability analysis was first conducted. Of the 1316 initially extracted features, 892 features (67.8%) demonstrated satisfactory stability with an ICC greater than 0.75 and were retained for subsequent dimensionality reduction.

The LASSO logistic regression algorithm was then applied to the training set to identify the most potent predictors of surgical difficulty from the high-dimensional data. As illustrated in Figure 1A, the coefficient profiles of the radiomic features were plotted against the log(lambda) sequence. The optimal penalisation parameter (lambda) was determined using 10-fold cross-validation. As shown in Figure 1B, the value of lambda that minimised the binomial deviance (lambda.min = 0.042) was selected. Consequently, 11 non-zero features with non-zero coefficients were screened out to construct the radiomics signature. These key features primarily consisted of texture features (n = 7) and first-order statistics (n = 3), along with one shape-based feature. Specifically, features such as original_glcm_ClusterProminence and wavelet-LHL_glrlm_LongRunEmphasis were identified, suggesting that the spatial heterogeneity and coarseness of the trabecular bone in the retromolar region are critical determinants of extraction resistance. Additionally, original_firstorder_90Percentile was selected, reflecting the high-density cortical bone component that impedes surgical access.

Fig. 1.

Radiomic feature selection using the Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression model. (A) LASSO coefficient profiles of the 892 stable radiomic features. Each coloured line represents the coefficient of a specific feature. As the penalty parameter lambda increases (x-axis, log scale), coefficients shrink towards zero. (B) Tuning parameter selection via 10-fold cross-validation.

The Rad-score for each patient was calculated using a linear combination of these 11 features weighted by their respective LASSO coefficients plus the intercept, as detailed in Supplementary Table S1. The quantitative distribution of the Rad-score is visualised in the boxplots in Figure 2. In the training set, the Rad-score was significantly higher in the High Difficulty group compared to the Low-to-Moderate Difficulty group (Median [IQR]: 1.25 [0.68, 1.82] vs –0.42 [–1.05, 0.21], P < .001). This discriminative capability was robustly validated in the validation set, where the High Difficulty group also exhibited a significantly elevated Rad-score (Median [IQR]: 1.18 [0.62, 1.75] vs –0.35 [–0.98, 0.25], P < .001). These results confirm that the constructed Rad-score serves as a potent and independent biological marker for stratifying surgical complexity.

Fig. 2.

Distribution of the Radiomics Score (Rad-score) by surgical difficulty group. Boxplots comparing the Rad-score between the low-to-moderate difficulty group (blue) and the high difficulty group (red) in the (A) Training set (n = 420) and (B) Validation set (n = 180). The central line represents the median, the box limits indicate the interquartile range (IQR), and the whiskers extend to 1.5 times the IQR. Statistical significance was determined using the Mann–Whitney U test (***P < .001 in both sets).

Predictive model construction and comparison

To determine the independent predictors of surgical difficulty, variables that demonstrated statistical significance (P < .10) in the univariate analysis were entered into a multivariable logistic regression model using a backward stepwise selection method. As detailed in Table 3, 5 variables were identified as independent risk factors. Specifically, deep impaction was a strong predictor: compared to Position A, Position C showed a substantially increased risk (OR = 4.52, 95% CI: 1.98–10.55), while Position B exhibited a trend towards higher difficulty (OR = 1.92, P = .065) and was retained in the final model to preserve the overall fit. Other independent predictors included older Age (OR = 1.12), higher BMI (OR = 1.15), complex Root Curvature (Severe vs Straight: OR = 3.85), and a higher Rad-score (OR = 2.78, 95% CI: 1.85–4.25). Collinearity diagnostics revealed that the VIF for all included predictors ranged from 1.12 to 1.85 (all < 5), indicating no severe multicollinearity.

Table 3.

Multivariable logistic regression analysis of risk factors for high-difficulty extraction in the training set.

Variable	β (coefficient)	SE	Wald χ²	P-value	OR (95% CI)	VIF
Intercept	–4.852	1.125	18.605	< .001	-	-
Age (years)	0.112	0.032	12.25	< .001	1.12 (1.05–1.19)	1.15
BMI (kg/m²)	0.14	0.062	5.098	.024	1.15 (1.02–1.30)	1.22
Pell & Gregory position
Level A (Ref)	-	-	-	-	1.00 (Reference)	-
Level B	0.652	0.354	3.392	.065	1.92 (0.96–3.84)	1.45
Level C	1.508	0.425	12.589	< .001	4.52 (1.98–10.55)	1.68
Root curvature
Straight (Ref)	-	-	-	-	1.00 (Reference)	-
Curved	0.585	0.285	4.214	.04	1.80 (1.03–3.15)	1.35
Severe/Hooked	1.348	0.405	11.076	.001	3.85 (1.75–8.64)	1.52
Rad-score	1.022	0.212	23.235	< .001	2.78 (1.85–4.25)	1.85

Variable selection was based on backward stepwise regression using the Akaike Information Criterion (AIC). Although Pell & Gregory Position Level B showed borderline significance (P = .065), it was retained in the final model as part of the categorical variable set to preserve overall model fit.

A sensitivity analysis was performed to assess the potential confounding effect of operator variability. When “Surgeon Identity” was introduced as a covariate, it did not achieve statistical significance (P = .682), and the regression coefficients of the primary predictors remained stable (changes < 5%). This confirms that within our cohort of experienced surgeons, the difficulty prediction model is robust and primarily driven by patient-specific factors.

To evaluate the incremental value of the radiomic signature, 3 predictive models were constructed: (1) a Clinical Model (Age, BMI, Pell & Gregory, Root Curvature); (2) a Radiomics Model (Rad-score only); and (3) a Combined Nomogram Model. The performance of these models is summarised in Table 4. In the training set, the Combined Nomogram achieved the highest discrimination with an AUC of 0.892 (95% CI: 0.854–0.930). Crucially, pairwise comparisons using the DeLong test demonstrated that the Combined Nomogram significantly outperformed both the Clinical Model (AUC = 0.745, P < .001) and the Radiomics Model alone (AUC = 0.812, P = .004). This superior performance was replicated in the validation set (AUC = 0.865), where the Combined Model again showed statistically significant improvement over the single-modality models (vs Clinical: P < .001; vs Radiomics: P = .032). These results underscore the complementary nature of macroscopic anatomical features and microscopic bone texture information.

Table 4.

Comparison of predictive performance.

Model	Dataset	AUC (95% CI)	Sensitivity (%)	Specificity (%)	Accuracy (%)	P-value (vs Clinical)	P-value (vs Radiomics)
Clinical model	Training	0.745 (0.698–0.792)	68.5	72.4	71.4	Reference	< .001
	Validation	0.728 (0.655–0.801)	65.2	70.9	69.4	Reference	.042
Radiomics model	Training	0.812 (0.770–0.854)	75.9	78.2	77.6	.015	Reference
	Validation	0.795 (0.730–0.860)	73.9	76.1	75.6	.042	Reference
Combined nomogram	Training	0.892 (0.854–0.930)	82.4	85.3	84.5	< .001	.004
	Validation	0.865 (0.805–0.925)	82.6	79.1	80	< .001	.032

P-values were calculated using the DeLong test for pairwise comparison of AUCs. The Combined Nomogram demonstrated statistically significant improvement over both single-modality models in both datasets.

Abbreviations: AUC, area under the receiver operating characteristic curve; CI, confidence interval.

Visualisation and calibration of the nomogram

Based on the independent predictors identified in the final multivariable model (Age, BMI, Pell & Gregory Position, Root Curvature, and Rad-score), a quantitative clinico-radiomic nomogram was constructed to facilitate individualised preoperative risk assessment (Figure 3). In this graphical tool, each predictor is assigned a specific point value on a scaled axis (top rows) corresponding to its regression coefficient magnitude. By summing the points from all 5 variables, a “Total Points” score is obtained, which can be vertically projected onto the bottom scale to estimate the probability of a high-difficulty extraction (surgical duration > 20 gt; 20 minutes). To illustrate its clinical application, consider a representative case of a 30-year-old patient (approximately 20 points) presenting with a Pell & Gregory Position C impaction (65 points), a severe root curvature (45 points), a BMI of 24 kg/m² (10 points), and a high Rad-score of 1.5 (40 points). The sum of these values yields a Total Score of 180 points, which corresponds to a predicted risk of difficulty exceeding 85%. This quantifiable risk estimation allows for more informed surgical planning and patient counselling.

Fig. 3.

The Radiomics-based nomogram for predicting the probability of high-difficulty mandibular third molar extraction. The nomogram integrates 5 independent risk factors: Age, BMI, Pell & Gregory Position, Root Curvature, and Rad-score. To use the nomogram, locate the value of each variable on its respective axis and draw a vertical line upwards to the “Points” axis to determine the score. Sum the points for all variables to get the “Total Points.” Finally, locate the Total Points on the bottom axis and draw a vertical line downwards to estimate the “Risk of Difficulty (> 20 gt; 20 minutes).”

The discriminatory performance and calibration accuracy of the nomogram were further evaluated (Figure 4). The ROC curves for the training (Figure 4A) and validation (Figure 4B) sets graphically confirm the superior performance of the Combined Nomogram, with its curve located closest to the upper-left corner compared to the single-modality models, reflecting the high AUC values reported in Table 4. Furthermore, the calibration curves (Figure 4C, D) demonstrated excellent agreement between the nomogram-predicted probabilities (x-axis) and the actual observed frequencies of high-difficulty extractions (y-axis). The bias-corrected curves in both datasets closely tracked the ideal 45-degree diagonal line, indicating high reliability. This goodness-of-fit was statistically confirmed by the Hosmer-Lemeshow test, which yielded non-significant P-values for both the training set (χ² = 8.45, P = .391) and the validation set (χ² = 6.12, P = .634), suggesting that the model possesses no significant overestimation or underestimation bias.

Fig. 4.

Discrimination and calibration assessment of the combined nomogram. (A, B) Receiver operating characteristic (ROC) curves comparing the Clinical Model (green), Radiomics Model (blue), and Combined Nomogram (red) in the training set (A) and validation set (B). The Combined Nomogram exhibits the largest Area Under the Curve (AUC) in both datasets. (C, D) Calibration curves of the Combined Nomogram in the training set (C) and validation set (D).

Clinical decision analysis

To evaluate the clinical translatability of the prediction model, DCA was performed to quantify the net benefit of using the Combined Nomogram for surgical planning. Figure 5 illustrates the decision curves for the training and validation sets. The x-axis represents the threshold probability (Pt), which reflects the surgeon's or patient's preference regarding the trade-off between the benefit of correctly identifying a difficult case and the cost of unnecessary preparation. The y-axis represents the net benefit, calculated by subtracting the proportion of false-positive patients from the proportion of true-positive patients, weighted by the relative harm of foregoing intervention compared to the negative consequences of an unnecessary intervention.

Fig. 5.

Decision Curve Analysis (DCA) for the Combined Nomogram. (A) Decision curve in the training set. (B) Decision curve in the validation set.

In the analysis, the black horizontal line represents the “Treat-none” strategy (assumption that no extraction will be difficult, resulting in zero net benefit), while the gray curve represents the “Treat-all” strategy (assumption that all extractions will be difficult and require advanced preparation). As shown in the figure, the decision curve of the Combined Nomogram (red line) consistently lies above both the “Treat-none” and “Treat-all” reference lines across a broad range of threshold probabilities, specifically from 10% to approximately 80% in both the training (Figure 5A) and validation (Figure 5B) sets. This indicates that if the threshold probability set by the surgeon falls within this wide range, using the Nomogram to decide whether to implement intensified surgical protocols (eg, scheduling a senior surgeon, preparing piezosurgery units, or allocating extended operating room time) provides a greater net benefit than either default strategy. Within this range, the model effectively maximises the detection of true high-difficulty cases while minimising the waste of medical resources on routine cases, thereby optimising clinical efficiency without compromising patient safety.

Discussion

In this study, we successfully developed and validated a novel clinico-radiomic nomogram for the preoperative prediction of surgical difficulty in MM3 extractions. The final model, integrating 5 independent predictors which include age, BMI, Pell & Gregory position, root curvature, and a robust radiomics signature (Rad-score) derived from the tooth and its surrounding bone microenvironment, demonstrated substantial predictive accuracy. Our principal finding is that this combined model, achieving an AUC of 0.892 in the training set and 0.865 in the validation set, significantly outperforms models based on either clinical variables or radiomic features alone. Furthermore, decision curve analysis confirmed the model’s clinical utility, showing a net benefit across a wide range of clinically relevant risk thresholds. This suggests that a holistic assessment combining macroscopic anatomical risk factors with microscopic bone texture information can provide a more accurate and reliable prediction of surgical complexity, moving beyond traditional assessment methods.³⁶

The multivariable analysis identified several well-established clinical factors as independent predictors of increased surgical difficulty. Older age was associated with a higher likelihood of a difficult extraction, a finding consistent with literature.³⁷ Ideally, elective prophylactic extractions are performed following orthodontic treatment before the age of 24. However, our cohort's mean age (∼26.4 years) reflects the reality of a tertiary referral centre where many patients present later with symptomatic impactions. Importantly, from the early twenties up to the age of approximately 30, the human mandible progressively reaches its peak bone mass (PBM). This physiological accrual results in increasingly denser, less elastic bone. This relationship is thus attributed to these physiological changes—decreased bone elasticity due to approaching PBM, a narrowed periodontal ligament space, and an increased incidence of ankylosis—all of which profoundly heighten resistance during surgical removal.³⁸ Similarly, our model confirmed that a higher BMI is a significant predictor. This can be explained by the increased volume of soft tissue in the buccal and cervical regions, which can severely limit both physical access and direct visualisation of the surgical field, complicating the procedure.³⁹ Unsurprisingly, anatomical impaction depth, specifically Pell & Gregory Position C, was one of the strongest predictors. This corroborates decades of clinical experience and numerous studies, as deeper impactions necessitate more extensive bone removal and create challenging instrument angulation.⁴⁰ Likewise, severe root curvature was identified as a critical risk factor, as hooked or dilacerated roots are prone to fracture and require meticulous, often multi-step, extraction techniques.⁴¹

Furthermore, while our model primarily predicts procedural duration, the critical issue of violating the mandibular canal and its contents—the IAN—must be highlighted. High-difficulty extractions often involve deep impactions and complex roots in close proximity to the canal.²⁹ Prolonged, forceful luxation in these dense bone scenarios drastically increases the risk of iatrogenic nerve injury. By utilising our clinico-radiomic nomogram to accurately forecast surgical difficulty preoperatively, surgeons can more rationally weigh the risks and consider alternative, nerve-sparing techniques, such as coronectomy, for high-risk patients, thereby minimising severe neurologic complications.³⁹

A central innovation of our study lies in the incorporation of a quantitative radiomics signature derived from a dual-region segmentation strategy. By creating a 3-mm dilation around the tooth root to define a “Bone Microenvironment” region of interest, we captured not only the tooth’s features but also the texture of the surrounding trabecular bone. The 3-mm margin was empirically chosen to encompass the immediate periradicular bone compartment that is critically involved in resisting tooth elevation and luxation forces, based on the typical dimensions of extraction sockets and the zone of bone remodelling.⁴² Specifically, 3 mm was deemed the optimal distance to capture sufficient stress-bearing trabecular heterogeneity while strictly avoiding unintended geometric overlap with adjacent confounding structures, such as the second molar roots and the mandibular canal.^10,42 Although this fixed margin was based on these anatomical rationales rather than an experimental comparison of multiple dilation radii in the current study, it provides a pragmatic balance. Future studies should investigate whether dynamic or multi-radius ROI dilations could further optimise predictive performance. The LASSO regression selected 11 key features, predominantly texture-based metrics like GLCM_ClusterProminence and GLRLM_LongRunEmphasis. These features quantify the heterogeneity and organisation of the bone microarchitecture, aspects invisible to the naked eye. Elevated values of these texture features likely correspond to denser, more complex, and less organised trabecular bone patterns, which would increase the physical resistance encountered during luxation and elevation.^43,44 The inclusion of first-order features such as the 90th percentile further reflects the influence of high-density cortical bone. This biological rationale is supported by studies in other fields, where radiomic texture analysis of bone has been successfully used to predict outcomes like implant stability or fracture risk, linking quantitative imaging features to underlying bone quality.^45,46 Thus, the Rad-score serves as a surrogate marker for the biomechanical properties of the periradicular bone, explaining its strong predictive power.

The incremental value of integrating radiomics with clinical data was a key finding. Our results showed that while the clinical-only model had fair performance (AUC ≈ 0.74), it was significantly improved by adding the Rad-score (Combined AUC ≈ 0.89). Traditional scoring systems like the Pederson Difficulty Index, while useful, often rely on subjective radiographic interpretation and typically demonstrate only moderate predictive accuracy, low sensitivity and specificity in predicting the difficulty of surgery for impacted mandibular 3rd molar.⁴⁷ The radiomics-only model performed better than the clinical model, highlighting that bone quality can sometimes be a more critical determinant than conventional anatomical factors. However, the superior performance of the combined nomogram underscores the complementary nature of these 2 data types. Macroscopic anatomy (eg, Pell & Gregory class) dictates the surgical strategy and access, while microscopic bone texture (the Rad-score) reflects the physical resistance that will be met. This synergy between macro and microlevel features allows for a more comprehensive and accurate risk stratification than either can provide alone.⁴⁸ The robustness of our model was further supported by a sensitivity analysis showing that surgeon experience did not significantly confound the results, indicating that the predictions are driven by patient-specific factors.

The clinical implications of our findings are significant. The developed nomogram is a user-friendly, graphical tool that translates a complex statistical model into an individualised probability score for each patient. This tool has multiple applications in clinical practice. It can facilitate evidence-based surgical planning; for instance, cases predicted to have a high probability of difficulty could be scheduled for longer time slots, assigned to more experienced surgeons, or have specialised equipment such as piezoelectric units prepared in advance.⁴⁹ It can also aid in resident training and tiered clinical privileging. Most importantly, it serves as an objective communication aid for patient counselling. By visually demonstrating the factors contributing to the surgical risk, clinicians can better manage patient expectations regarding procedure duration and potential postoperative sequelae, fostering shared decision-making and potentially reducing medicolegal risk.⁵⁰ The results of the decision curve analysis strongly support this clinical utility, demonstrating that using the nomogram to guide decision-making provides a clear net benefit over default strategies of treating all patients as either high or low risk across a wide spectrum of threshold probabilities.

This study has several strengths, including the use of an objective, reproducible primary outcome (surgical duration), adherence to rigorous radiomics standards (IBSI guidelines), and the robust statistical methodology employed, which included multivariable analysis, 10-fold cross-validation, and decision curve analysis. The innovative dual-ROI approach to capture the bone microenvironment is also a key methodological advantage. However, we must also acknowledge several limitations. First, as a retrospective, single-centre study, our findings are subject to potential selection bias and require external validation in a larger, multi-centre cohort to confirm generalisability. Second, the segmentation of the regions of interest was performed manually, which is a time-consuming process. Future work should focus on developing and validating automated segmentation algorithms to improve workflow efficiency for clinical implementation. Third, although we controlled for surgeon experience, other unmeasured variables such as patient anxiety, mouth opening limitation, or specific surgical techniques used could have influenced the outcome. Fourth, while isotropic resampling via B-spline interpolation is an IBSI-recommended prerequisite to ensure rotation invariance and consistent distance metrics for 3D texture matrix computation, this process inherently introduces a smoothing effect. Such interpolation may inadvertently alter the native texture features and noise characteristics uniquely inherent to CBCT images. Finally, this model predicts only surgical duration as a surrogate for difficulty and does not directly predict postoperative complications, which should be a focus of future investigations incorporating sensitivity analyses of various interpolation algorithms.

Conclusion

This study demonstrates that a clinico-radiomic nomogram integrating standard clinical risk factors with a quantitative radiomics signature of the bone microenvironment is a powerful tool for predicting the difficulty of MM3 extraction. This approach provides a more personalised and accurate risk assessment than is possible with clinical or radiomic data alone. By translating complex radiographic data into an actionable prediction, this nomogram has the potential to optimise surgical planning, enhance patient communication, and ultimately improve the quality of care in oral and maxillofacial surgery. Further prospective validation and automation are warranted to facilitate its widespread clinical adoption.

Ethics approval and consent to participate

This study was approved by the Ethics Committee of Wuxi Stomatological Hospital. Given the retrospective nature of the study, which utilised deidentified historical data from electronic medical records and radiographic archives, the requirement for written informed patient consent was formally waived by the institutional review board. All methods were carried out in accordance with Declaration of Helsinki.

Data availability

The experimental data used to support the findings of this study are available from the corresponding author upon request.

Author contributions

Mingchen Xu and Yijun Wu made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Conflict of interest

None disclosed.