Construction of a CNN Model Based on Lateral Cephalometric Imaging Features and Its Application in Surgery-First Treatment Decision-Making for Skeletal Class III Malocclusion

Construction of a CNN Model Based on Lateral Cephalometric Imaging Features and Its Application in Surgery-First Treatment Decision-Making for Skeletal Class III Malocclusion | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Construction of a CNN Model Based on Lateral Cephalometric Imaging Features and Its Application in Surgery-First Treatment Decision-Making for Skeletal Class III Malocclusion Jin Li, Yingzhu Bao, Haochen Zhang, Keying Qi, Le Yang, Yufan Chen, and 6 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9389538/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 7 You are reading this latest preprint version Abstract Background A deep learning model using four convolutional neural networks (CNNs) was trained based on lateral cephalograms from patients with skeletal Class III malocclusion to assist in selecting between orthodontic-first and surgery-first approaches. Methods A total of 511 skeletal Class III patients undergoing orthodontic–orthognathic treatment were analyzed. Twelve cephalometric variables were assessed, and multivariate analysis identified independent predictors of treatment selection. Subsequently, the lateral cephalograms were split (7:1.5:1.5) into training, validation, and test sets. Four deep learning models were evaluated using five-fold cross-validation. Results Multivariate analysis identified L1–MP was an independent predictor for the selection of SFA. All four deep learning models achieved accuracies above 0.850. MobileNetV2 performed best, with an AUC of 0.949, accuracy of 0.953, sensitivity of 0.941, specificity of 0.957, precision of 0.889, and an F1 score of 0.914. Conclusion MobileNetV2 demonstrated high accuracy in classifying lateral cephalograms, providing significant decision support for the selection between orthodontic-first and surgery-first approaches in skeletal Class III malocclusion patients. Skeletal Class III malocclusion Surgery-first approach Lateral cephalogram Deep learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Introduction In the combined orthodontic–orthognathic treatment of skeletal Class III malocclusion, the selection between surgery-first approach (SFA) and orthodontics-first approach(OFA) represents a complex clinical decision‑making process [1][2][3] .While SFA has gained significant attention for its potential to shorten total treatment duration and provide immediate facial esthetic improvement, its application requires careful patient selection [4][5] . Inappropriate treatment pathways may lead to increased postoperative orthodontic complexity and compromised long-term stability [6] . Currently, this decision relies primarily on comprehensive evaluation of lateral cephalometric parameters, dental model, clinician experience, etc. However, this process largely depends on manual analysis and lacks objective, standardized criteria [11] , which is necessary for consistent preoperative assessment across different institutions. Recent advancements in digital orthodontics have introduced new possibilities for enhancing this decision-making process through deep learning, which excels at efficient feature extraction and automated prediction [7][8] . Nevertheless, the clinical adoption of these models is often hindered by their "black-box" nature and a perceived lack of interpretability [9][10] . In contrast, standard analysis offers established, theory-based criteria but is constrained by manual processing and a narrow range of predefined variables [12] . While several studies have attempted to integrate these two modalities—utilizing CBCT-based nomograms or multi-modal data such as lateral cephalograms combined with 3D dental scans —to assist SFA selection, certain considerations remain [11][12] . For instance, while CBCT provides a distinct three-dimensional advantage, its routine application may be tempered by concerns regarding higher radiation exposure [13] . In contrast, lateral cephalograms remain the most routine and available imaging modality, yet the overall decision-relevant morphological information they contain has not been fully explored. Furthermore, because clinical indications inherently vary across institutions, models trained exclusively on single-center data inevitably face limitations in generalizability. [14] . Nevertheless, while individual studies are often initially constrained to a single center, utilizing universally accessible imaging modalities allows for data pooling across multiple centers in the future [15] . This enables the gradual construction of expansive, diverse databases, progressively mitigating selection bias and approaching more generalizable outcomes [16] . Although previous studies have shown that candidates for SFA and OFA exhibit distinct maxillofacial, vertical skeletal, and dental compensation patterns, synthesizing these morphologic characteristics into systematic clinical guidelines remains challenging [1][7][8] . In traditional practice, surgical indications rely on implicit morphological criteria and individual clinical experience, making the decision-making process difficult to standardize [17] . In this context, the morphological decision-related information embedded within accessible 2D lateral cephalograms remains underexplored by modern deep learning approaches. Therefore, AI-driven imaging modeling is meaningful even within single-center settings, because it successfully translates unquantifiable clinical experience into an objective, mathematical framework [18] . By automatically extracting the core anatomical determinants from routine radiographs, the deep learning model explicitly clarifies and quantifies the shared morphological characteristics driving these complex surgical decisions [19] . This critical transformation of implicit clinical heuristics into a baseline provides evidence to support the development of more consistent and generalizable decision-making frameworks. To address these issues, we developed and validated a CNN-based decision-support framework for analyzing routinely acquired lateral cephalograms. By focusing on standard 2D imaging, we investigated whether the global morphological information it contains can provide reliable guidance for treatment selection without the need for 3D imaging or manual annotation. Among the evaluated models, MobileNetV2 was utilized for its balance between computational efficiency and classification performance. To enhance clinical interpretability, multivariate statistical analysis was incorporated to validate the independent predictors identified by the model, thereby linking automated efficiency with clinical insight. This study represents an independent evaluation of AI-driven decision-making in a new clinical setting and supports the development of more consistent and accessible treatment planning strategies. METHODS Patients Patients with skeletal Class III malocclusion who underwent orthognathic surgery at the × between 2019 and 2024 were consecutively enrolled. A total of 530 patients were screened, and 511 were finally included, comprising 77 males and 114 females in the SFA group and 145 males and 175 females in the OFA group. Preoperative treatment plans were determined jointly by orthodontists and oral and maxillofacial surgeons, and patients were grouped according to the surgical approach. The study was approved by the Ethics Committee of the × (GHKQ-202401-L2), and all participants provided informed consent. Details of data selection and grouping are shown in Figure 1 and Table 1. The exclusion criteria included: (1) a history of orthodontic treatment or orthognathic surgery; (2) dentition defects, defined as the absence of two or more teeth per quadrant, excluding third molars; (3) craniofacial developmental deficiencies or congenital deformities (e.g., cleft lip and palate), as well as a history of infection, trauma, or tumors; and (4) image quality issues that could compromise diagnostic validity, including motion artifacts, metal artifacts, image blurring, or insufficient resolution. Table 1 Distribution of imaging data Lateral cephalograms（N） Man Woman SFA 77 114 OFA 145 175 OFA orthodontic-first approach, SFA surgery-first approach Variables of structural parameters on lateral cephalograms The definitions or descriptions of the variable parameters used are provided in Table 2. Table 2. The definitions or descriptions of the variable measurements Variable measurement Description SNA(°) Angle formed between sella, nasion and point A SNB(°) Angle formed between sella, nasion and point B ANB(°) Angle formed between point A, nasion and point B SN-MP(°) Sella–Nasion to Mandibular Plane Angle FH-MP(°) Frankfort Horizontal to Mandibular Plane Angle U1-SN(°) Angle formed by SN-plane and the axis of the maxillary incisor (U1) in midsagittal plane U1-NA(°) Angle between long axis of the maxillary central incisor and nasion–point A line L1-MP(°) Mandibular Incisor to Mandibular Plane Angle Wits（mm） Drawn perpendiculars from points A and B onto the occlusal plane and measured the distance between these two points Po-NB（mm） Pogonion to NB Line Distance Co-A（mm） Distance between condylion and point A Co-Gn（mm） Distance between condylion and gonion Image Preprocessing The dataset was randomly divided into training, validation, and test sets at a ratio of 7:1.5:1.5, ensuring independence to avoid data leakage. All images were resized to 224 × 224 pixels and converted into tensor formats (Fig.2). To improve generalization, data augmentation was applied only to the training set, including random translation within ±5 pixels, rotation of 0–5°, and ±10% adjustments in contrast and brightness. No augmentation was applied to the validation or test sets to ensure unbiased evaluation. CNN classification workflow based on lateral cephalograms Four pretrained models—DenseNet121, InceptionV3, MobileNetV2, and VGG16—were trained on lateral cephalograms. Four backbone architectures were trained using cross-entropy loss, ReLU activation, and the Adam optimizer. All models were evaluated using five‑fold cross‑validation with a batch size of 64. DenseNet121, InceptionV3, and MobileNetV2 were trained for 100 epochs, while VGG16 was trained for 300 epochs, with an initial learning rate of 0.001 and weight decay of 2 × 10⁻⁴. Learning rate reduction and early stopping were applied according to validation performance. The workflow is illustrated in Fig. 3. Data cleaning and dataset construction: The dataset was first cleared of invalid, low-quality, or redundant images. Only qualified lateral cephalograms were retained for subsequent analysis. Image preprocessing and data partitioning: The dataset was resized to 224 × 224 pixels and converted into vector formats compatible with the deep learning framework, with grayscale normalization and other preprocessing steps applied. Data augmentation: Common augmentation techniques, such as rotation, scaling, and flipping, were applied to expand sample diversity and improve model generalization for the subsequent classification tasks. Model construction and optimization: The four models were selected as model backbones. Network parameters were tuned through learning rate adjustment, weight decay regularization, and data augmentation to enhance classification performance. Model performance evaluation: The trained deep learning models processed the test set data, and metrics such as accuracy, recall, F1 score, and AUC evaluated their generalization and clinical applicability in classifying SFA versus OFA. Model output evaluation and statistical methods All statistical analyses were performed using SPSS version 25.0. Descriptive statistics were used to summarize the baseline parameters of the SFA and OFA groups, followed by independent-samples t-tests for between-group comparisons of continuous variables. Univariate analysis was conducted to identify potential predictive factors, and multivariate logistic regression analysis was performed to determine independent predictors. All statistical tests were two-sided, with a P-value < 0.05 considered statistically significant. Intra-observer reliability of cephalometric measurements was assessed using the intraclass correlation coefficient (ICC) with a two-way random effects model and absolute agreement, based on repeated measurements of 511 patients by the same examiner at three time points with two-week intervals. The classification performance of the model for distinguishing SFA from OFA was evaluated by standard diagnostic metrics, including accuracy (ACC), sensitivity (Sen), specificity (Spec), precision, F1 score, positive predictive value (PPV), negative predictive value (NPV), and AUC; ROC curves and confusion matrices were generated, and 95% confidence intervals were estimated using 1,000 bootstrap resampling iterations. RESULTS The means and standard deviations of the twelve cephalometric parameters for both groups from the 511 patients are presented in Table 3. The SFA group demonstrated smaller SN–MP and FH–MP angles, increased labial inclination and protrusion of the maxillary incisors, and more pronounced labial inclination of the mandibular incisors. In addition, the intraclass correlation coefficients (ICCs) for the 12 variables ranged from 0.973 to 0.989, indicating excellent intra-observer reliability. Table 3. Lateral cephalometric parameters in skeletal Class III patients under two treatment approaches OFA Group SFA Group p-value Mean SD Mean SD SNA 83.78 3.55 84.05 3.74 0.430 SNB 86.77 4.35 87.02 4.38 0.521 ANB -2.98 3.01 -3.01 2.81 0.901 SN-MP 33.96 7.54 32.27 7.24 0.01 3 ※ FH-MP 27.59 6.79 26.09 6.53 0.01 5 ※ U1-SN 114.51 7.11 117.55 7.46 ＜ 0.01 ※※ U1-NA 30.59 5.99 33.41 6.79 ＜ 0.01 ※※ L1-MP 80.13 8.78 83.00 7.99 ＜ 0.01 ※※ Wits -10.77 4.94 -9.97 5.25 0.085 Po-NB 1.34 1.07 1.51 1.39 0.115 Co-A 81.95 6.37 82.38 8.66 0.518 Co-Gn 122.78 10.84 122.96 14.05 0.878 ※p<0.05, ※※p<0.001 The results of the univariate analysis are presented in Table 4, which identified five variables with statistical significance: SN–MP, FH–MP, U1–SN, U1–NA, and L1–MP. Following these findings, multivariable analysis was conducted on the variables that showed significance in the univariate analysis. The results demonstrated L1–MP as an independent predictive factor, as detailed in Table 5. Table 4. Results of univariate analysis in patients with skeletal Class III malocclusion Variable β OR 95%CI P SNA -0.02 0.98 0.93 ~ 1.03 0.429 SNB -0.01 0.99 0.95 ~ 1.03 0.520 ANB 0.004 1.004 0.94 ~ 1.07 0.901 SN-MP 0.03 1.03 1.01 ~ 1.05 0.0 14 ※ FH-MP 0.03 1.03 1.01 ~ 1.06 0.0 15 ※ U1-SN -0.06 0.94 0.92 ~ 0.97 ＜ 0.0 0 1 ※※ U1-NA -0.07 0.93 0.91 ~ 0.96 ＜ 0.0 0 1 ※※ L1-MP 0.04 1.04 1.02 ~ 1.06 ＜ 0.0 0 1 ※※ Wits -0.03 0.97 0.93 ~ 1.01 0.086 Po-NB -0.12 0.89 0.77 ~ 1.03 0.116 Co-A -0.08 0.99 0.97 ~ 1.02 0.520 Co-Gn -0.001 1.00 0.98~ 1.01 0.878 ※p<0.05, ※※p<0.001 Table 5. Results of multivariate analysis in patients with skeletal Class III malocclusion Variable β OR 95%CI P SN-MP -0.012 0.988 0.91 ~ 1.07 0.768 FH-MP 0.023 1.023 0.80 ~ 1.43 0.661 U1-SN -0.023 0.978 0.92 ~ 1.04 0.452 U1-NA -0.043 0.958 0.90 ~ 1.02 0.176 L1-MP -0.033 0.968 0.95 ~ 0.99 0.00 5 ※ ※p<0.05, ※※p<0.001 Following the standard statistical analyses, we evaluated classification performance using deep learning to automatically distinguish between SFA and OFA. Five-fold cross-validation was used to evaluated the classification performance of DenseNet121, InceptionV3, MobileNetV2, and VGG16. In the training set, the mean accuracies were 97.73% for DenseNet121, 99.56% for InceptionV3, 100% for MobileNetV2, and 90.25% for VGG16. In the test set, the corresponding accuracies were 87.01%, 87.27%, 89.35%, and 80.78%, respectively. In the validation set, the accuracies were 87.50%, 89.69%, 89.69%, and 82.50%, respectively. Taken together, the overall mean accuracies across all datasets were 90.74% for DenseNet121, 92.17% for InceptionV3, 93.01% for MobileNetV2, and 84.51% for VGG16. In the training set, the standard deviations of the models were 1.67% for DenseNet121, 0.28% for InceptionV3, 0.00% for MobileNetV2, and 2.08% for VGG16. In the test set, the standard deviations were 4.64% for DenseNet121, 3.22% for InceptionV3, 3.53% for MobileNetV2, and 2.08% for VGG16. In the validation set, the standard deviations were 2.96% for DenseNet121, 2.53% for InceptionV3, 3.22% for MobileNetV2, and 3.03% for VGG16. The overall standard deviations of the four models were 2.04%, 1.22%, 0.36%, and 1.22%, respectively. All models demonstrated strong discriminative ability and high stability during the training process. The above results indicate that MobileNetV2 achieved consistently high accuracy across all datasets with relatively small standard deviations, suggesting stronger learning capability and greater stability. InceptionV3 and DenseNet121 also attained high accuracies with moderate standard deviations. In contrast, VGG16 showed lower accuracies and larger standard deviations across the datasets, indicating the need for further optimization and refinement (Fig. 4). Building upon the evaluation of model stability, overall performance was analyzed across multiple classification metrics. As presented in Table 6 , Fig 5 and Fig 6, all major metrics exceeded 0.870 for the four deep learning models in distinguishing SFA from OFA. MobileNetV2 achieved the best performance (AUC = 0.949, accuracy = 0.953, sensitivity = 0.941, F1 = 0.914), outperforming the other models; its negative predictive value was 0.978, indicating a strong ability to exclude non-SFA cases. Model Evaluation metrics ACC Sen Spec Precision F1 PPV NPV DenseNet121 0.906 0.885 0.921 0.885 0.885 0.885 0.921 InceptionV3 0.938 0.833 0.978 0.938 0.882 0.938 0.938 MobileNetV2 0.953 0.941 0.957 0.889 0.914 0.889 0.978 VGG16 0.875 0.810 0.907 0.810 0.810 0.810 0.907 Table 6 Prediction performance of each model Computational parameters of each model were evaluated to further explain the differences in model performance, including network depth, parameter count, and inference time (Table 7). The results showed that MobileNetV2 achieved the best balance of lightweight design and efficiency, reducing hardware dependence and making it particularly suitable for small-dataset settings while helping to mitigate overfitting. VGG16, with its simpler architecture, achieved faster inference, which may partly explain its lower accuracy. In contrast, DenseNet121 and InceptionV3 had a larger number of parameters and longer inference times, reflecting the greater computational cost of dense connectivity and multi-branch structures. This more complex design allows them to strike a balance between performance and efficiency. Therefore, model selection should be guided by both computational resources and dataset size. Table 7 Performance of each model Model Network depth Number of parameters Inference time (milliseconds) DenseNet-121 431 7563330 52 InceptionV3 315 22852898 39 MobileNetV2 158 2914882 31 VGG-16 23 14978370 11 Considering the influence of the aforementioned architectural characteristics on model learning, the convergence behavior of each model was analyzed. As presented in Fig 7, MobileNetV2 exhibited the fastest increase in training accuracy, with validation accuracy improving steadily and stabilizing around epoch 57, demonstrating rapid convergence and strong overall performance. DenseNet121 and InceptionV3 also performed well, reaching stability at epochs 76 and 84, respectively, though their validation accuracy gains gradually slowed, indicating mild overfitting. In contrast, VGG16 maintained high training accuracy but showed more variable and slower validation performance, consistent with the results in Table 4 and likely due to its deeper architecture. Overall, MobileNetV2 achieved high and stable accuracy in fewer epochs, demonstrating the best performance among the models. Finally, we evaluated the practical performance of the models in the classification task by analyzing the confusion matrices derived from a single fold of data. As presented in Fig 8, the numbers of correctly classified SFA cases were 16 for DenseNet121, 15 for InceptionV3, 16 for MobileNetV2, and 17 for VGG16, whereas the numbers of correctly classified OFA cases were 42, 45, 45, and 39, respectively. MobileNetV2 demonstrated balanced identification of true positives (SFA) and true negatives (OFA), reflecting good discriminative performance. DISCUSSION In the treatment of skeletal Class III malocclusion, the choice between SFA and OFA can substantially affect treatment duration and facial aesthetic outcomes. Although SFA offers advantages such as immediate correction and a shorter treatment course [20] , clear criteria for patient selection remain lacking in routine clinical practice. To address this issue, we developed and evaluated four deep learning models and, in parallel, performed multivariate statistical analysis, which identified L1–MP as an independent predictor of SFA selection. Among the evaluated models, MobileNetV2 showed superior discriminative performance in distinguishing candidates for SFA and OFA, highlighting its potential as a valuable clinical decision-support tool. In recent years, the integration of artificial intelligence in orthodontic-orthognathic treatment planning has been extensively explored, with various models proposed to assist in selecting between OFA and SFA. These decision-support systems generally rely on high-dimensional clinical data, often combining traditional 2D cephalometric measurements with 3D intraoral scans, or utilizing advanced deep learning architectures on pure 3D imaging. For instance, in a previous work, several advanced deep learning architectures (including Simple-CNN, DenseNet121, MobileNetV2, EfficientNet-B0, ResNet10, and SEResNet50) were applied to classify OFA and SFA using 3D imaging data, achieving a high degree of performance with an accuracy of 0.909 and an AUC of 0.896 [11] . Similarly, another representative study developed a multimodal ResNet34 model combining standard lateral cephalograms with 3D maxillary and mandibular scans, achieving an accuracy of 0.906 [12] . While 3D maxillary models can achieve an accuracy of up to 0.970, the lack of mandibular information limits comprehensive evaluation of skeletal Class III malocclusion. Against the backdrop of increasingly complex, multi-dimensional methodologies, the present study sought to determine whether comparable or superior performance could be achieved using only routine 2D lateral cephalograms. Through architectural optimization with MobileNetV2, our 2D-based model achieved an accuracy of 0.953. This performance approaches or even exceeds that of the aforementioned 3D and multimodal models. Achieving high predictive robustness using standard 2D imaging—which constitutes a foundational, low-radiation, and highly accessible diagnostic tool universally available across clinical settings—highlights the immense clinical translatability of this approach. From a machine learning perspective, this may relate to the tendency of high-dimensional features to overfit under limited data, thereby constraining generalization [21] . This further indicates that lightweight architectural designs allow MobileNetV2 to efficiently extract essential discriminative features from 2D radiographs, preserving dependable generalization without the absolute necessity for resource-intensive 3D or multimodal data [22] . Furthermore, architectural variations, particularly network depth and parameter complexity, were associated with marked differences in model performance. MobileNetV2 demonstrated superior overall efficacy compared to heavier architectures like DenseNet121, InceptionV3, and VGG16. Theoretically, excessively deep networks may suffer from diminished generalization due to vanishing gradients, increased susceptibility to overfitting, and reduced backpropagation efficiency [23][24] . For example, although the dense connectivity of DenseNet121 and the multi-branch architecture of InceptionV3 facilitate feature fusion, their relatively high parameter counts (7.56M and 22.85M, respectively) under limited data conditions may lead to overfitting and reduced generalization performance [25][26] . Although the VGG16 model has fewer layers (16 layers) and a relatively simple architecture, its large number of parameters and complex fully connected layer design (14.98 million parameters) result in high computational cost and relatively weak resistance to overfitting [27][28] . In contrast, MobileNetV2 employs depthwise separable convolutions and linear bottleneck techniques to maintain network depth while compressing the parameter count to 2.91 million, thereby reducing computational complexity and memory consumption. This lightweight design enables stronger generalization capability in small-sample scenarios and is particularly well suited for applications with limited data availability [29][30] . At the feature extraction level, the linear bottleneck design of MobileNetV2 better preserves global information and captures skeletal features relevant to surgical decision-making. As a result, minimizing information distortion or loss may be particularly important for improving the accuracy of orthognathic surgery approach prediction [31][32] . A major barrier to the clinical adoption of deep learning in medicine is its “black-box” nature. To mitigate this, the present study positions deep learning as a decision-support tool rather than a replacement for clinical judgment. The consistency between statistical findings and model performance provides a degree of indirect interpretability. Multivariate analysis identified lower incisor compensation (L1–MP) and vertical skeletal characteristics as key determinants. Meanwhile, the CNN achieved high classification accuracy using only raw lateral cephalograms, without explicit anatomical landmark annotation, suggesting that it implicitly learns relevant biomechanical features [7] . Moreover, MobileNetV2’s streamlined architecture promotes the progressive accumulation of broader spatial relationships rather than convoluted feature entanglements. Consequently, such lightweight networks minimize redundant feature assimilation, inherently reducing the likelihood of the model relying on erroneous or uncontrollable latent artifacts for its predictions [33] . Despite the promising findings, several limitations should be acknowledged. This study was a single-center retrospective analysis with a relatively limited sample size, and the generalizability of the results requires validation using large-scale, multicenter datasets. In addition, the analysis was based solely on lateral cephalograms, lacking multimodal data, which may limit comprehensive evaluation of three-dimensional structures. Future studies should incorporate multimodal imaging and explainable artificial intelligence techniques (e.g., Grad-CAM) to further enhance the model’s interpretability and clinical applicability. CONCLUSIONS A CNN-based framework utilizing routine lateral cephalograms demonstrated robust clinical potential for predicting the optimal treatment approach (SFA vs. OFA) in patients with skeletal Class III malocclusion. Among the evaluated architectures, MobileNetV2 exhibited consistently superior performance, achieving an accuracy of 0.953 and an AUC of 0.949, coupled with sF generalization across training and validation cohorts. By bridging computational efficiency with high predictive accuracy on standard 2D radiographs, this deep learning framework provides an accessible, evidence-based decision-support tool for orthodontic–orthognathic treatment planning, offering a valuable methodological foundation for future translational investigations. Declarations Ethics approval and consent to participate This study was approved by the Ethics Committee of Hospital of Stomatology, Sun Yat-sen University, with a reference number of (GHKQ-202401-L2), and the conduct of the study adheres to the Declaration of Helsinki and relevant ethical standards. Informed consent was obtained from all participants prior to their involvement in the study. Clinical trial number is not applicable. Consent for publication Not applicable. Competing interests The authors declare that they have no competing interests. Funding This study was supported by IOF Young Research Grant (IOF2023Y09) and National Natural Science Foundation of China (82201060). Author Contribution XW and XL was responsible for the study design, administration, drafting of the manuscript, critical revision, and final approval of the article. YB and JL contributedto the literature search, performed data acquisition and data analysis. HZ ,KQ and YCperformed data acquisition and data analysis. LY,CZ and WW were responsible for the drafting, data interpretation and manuscript editing. BB and GZ were responsible for critical revision. All authors read and approved the final manuscript. Acknowledgements Not applicable. Availability of data and materials De-identified data are available upon request to qualified researchers. Access requires (1) approval of a research proposal by the institutional ethics committee and (2) a signed data use agreement. Requests should be directed to corresponding author email. Unsectioned Paragraphs Eslami S, Faber J, Fateh A, Sheikholaemmeh F, Grassia V, Jamilian A. Treatment decision in adult patients with class III malocclusion: surgery versus orthodontics. Prog Orthod. 2018 Aug 2;19(1):28. doi: 10.1186/s40510-018-0218-0. PMID: 30069814; PMCID: PMC6070451. Alhammadi MS, et al. Orthodontic camouflage versus orthodontic-orthognathic surgical treatment in borderline class III malocclusion: a systematic review. Clin Oral Investig. 2022;26:6443–6455. Beccuti ML, et al. “Surgery First” vs “Traditional Sequence” Surgery: A qualitative study of health experiences. J Maxillofac Oral Surg. 2022;21:1267–1278. Sharma V, Yadav K, Tandon P. An overview of surgery-first approach: Recent advances in orthognathic surgery. J Orthod Sci. 2015;4(1). Khalil AS, Alrehaili RS, Bajunaid M, Alhazmi M, Alshami A, Alharthy B, Fakhry O, Olfat Y, Taher A, Alotaibi R, Alrefai M, Barashid AA. Does Surgery-First Orthognathic Approach Improve Quality of Life of Orthodontic Patients With Skeletal Class III Malocclusion? A Systematic Review Following Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) Guidelines. Cureus. 2025 Mar 29;17(3):e81433. doi: 10.7759/cureus.81433. PMID: 40296934; PMCID: PMC12037207. Chen YC, et al. Dental occlusion characteristics for treatment decision-making regarding surgery-first approach. J Clin Med. 2023;12(18). Lo Giudice A, Ronsivalle V, Spampinato C, Leonardi R. Fully automatic segmentation of the mandible based on convolutional neural networks (CNNs). Orthod Craniofac Res. 2021 Dec;24 Suppl 2:100-107. doi: 10.1111/ocr.12536. Epub 2021 Dec 1. PMID: 34553817. Choi HI, Jung SK, Baek SH, Lim WH, Ahn SJ, Yang IH, Kim TW. Artificial Intelligent Model With Neural Network Machine Learning for the Diagnosis of Orthognathic Surgery. J Craniofac Surg. 2019 Oct;30(7):1986-1989. doi: 10.1097/SCS.0000000000005650. Erratum in: J Craniofac Surg. 2020 Jun;31(4):1156. doi: 10.1097/SCS.0000000000006531. PMID: 31205280. Khanagar SB, Al-Ehaideb A, Maganur PC, Vishwanathaiah S, Patil S, Baeshen HA, Sarode SC, Bhandi S. Developments, application, and performance of artificial intelligence in dentistry - A systematic review. J Dent Sci. 2021 Jan;16(1):508-522. doi: 10.1016/j.jds.2020.06.019. Epub 2020 Jun 30. PMID: 33384840; PMCID: PMC7770297. Bağ İ, Bilgir E, Bayrakdar İŞ, Baydar O, Atak FM, Çelik Ö, Orhan K. An artificial intelligence study: automatic description of anatomic landmarks on panoramic radiographs in the pediatric population. BMC Oral Health. 2023 Oct 17;23(1):764. doi: 10.1186/s12903-023-03532-8. PMID: 37848870; PMCID: PMC10583406. Bao Y, Qi K, Yang L, Chen Y, Zhou C, Wang W, Bao B, Long X, Zheng G, Wang X. CBCT radiomic features-based machine learning prediction models and nomogram for treatment decision-making regarding surgery-first approach in skeletal Class III malocclusion. BMC Oral Health. 2025 Sep 26;25(1):1464. doi: 10.1186/s12903-025-06702-y. PMID: 41013431; PMCID: PMC12465188. Chang JS, Ma CY, Ko EW. Prediction of surgery-first approach orthognathic surgery using deep learning models. Int J Oral Maxillofac Surg. 2024;53:942–949. Litjens G, et al. A survey on deep learning in medical image analysis. Med Image Anal. 2017;42:60–88. Krois J, Garcia Cantu A, Chaurasia A, Patil R, Chaudhari PK, Gaudin R, Gehrung S, Schwendicke F. Generalizability of deep learning models for dental image analysis. Sci Rep. 2021 Mar 17;11(1):6102. doi: 10.1038/s41598-021-85454-5. PMID: 33731732; PMCID: PMC7969919. Zaborowicz, K.; Zaborowicz, M.; Cieślińska, K.; Biedziak, B. Artificial Intelligence Methods in Cephalometric Image Analysis—A Systematic Narrative Review. J. Clin. Med. 2026, 15, 1920. https://doi.org/10.3390/jcm15051920 Chang, Q., Bai, Y., Wang, S. et al. Automated orthodontic diagnosis via self-supervised learning and multi-attribute classification using lateral cephalograms. BioMed Eng OnLine 24, 9 (2025). https://doi.org/10.1186/s12938-025-01345-0 Durão AP, Morosolli A, Pittayapat P, Bolstad N, Ferreira AP, Jacobs R. Cephalometric landmark variability among orthodontists and dentomaxillofacial radiologists: a comparative study. Imaging Sci Dent. 2015 Dec;45(4):213-20. doi: 10.5624/isd.2015.45.4.213. Epub 2015 Dec 17. PMID: 26730368; PMCID: PMC4697005. Schwendicke F, Samek W, Krois J. Artificial Intelligence in Dentistry: Chances and Challenges. J Dent Res. 2020 Jul;99(7):769-774. doi: 10.1177/0022034520915714. Epub 2020 Apr 21. PMID: 32315260; PMCID: PMC7309354. Schwendicke, F., Chaurasia, A., Arsiwala, L. et al. Deep learning for cephalometric landmark detection: systematic review and meta-analysis. Clin Oral Invest 25, 4299–4309 (2021). https://doi.org/10.1007/s00784-021-03990-w Uribe FA, Farrell B. Surgery-First Approach in the Orthognathic Patient. Oral Maxillofac Surg Clin North Am. 2020 Feb;32(1):89-103. doi: 10.1016/j.coms.2019.08.009. Epub 2019 Nov 1. PMID: 31685343. ]Parekh V, Jacobs MA. Radiomics: a new application from established techniques. Expert Review of Precision Medicine and Drug Development. 2016;1(2):207–226. Vinayahalingam S, Kempers S, Limon L, et al. The Automatic Detection of Caries in Third Molars on Panoramic Radiographs Using Deep Learning: A Pilot Study[J].2021 Hawkins DM. The problem of overfitting. J Chem Inf Comput Sci. 2004 Jan-Feb;44(1):1-12 Yu XH. Can backpropagation error surface not have local minima. IEEE Trans Neural Netw. 1992;3(6):1019-21 LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015 May 28;521(7553):436-44 Zhang C, Bengio S, Hardt M,et al. Understanding deep learning requires rethinking generalization[J]. 2016 Omiotek Z, Kotyra A. Flame Image Processing and Classification Using a Pre-Trained VGG16 Model in Combustion Diagnosis. Sensors (Basel) 2021, 21(2) Kloster M, Langenkämper D, Zurowietz M, Beszte et al. Deep learning-based diatom taxonomy on virtual slides. Sci Rep. 2020 Sep 2;10(1):14416 Xie X, Liu Z, Wang Y, et al. Puppet Dynasty Recognition System Based on MobileNetV2. Entropy (Basel) 2024, 26(8) Yong L, Ma L, Sun D, et al. Application of MobileNetV2 to waste classification. PLoS One 2023, 18(3): e0282336 Kim YH, Park JB, Chang MS, et al. Influence of the Depth of the Convolutional Neural Networks on an Artificial Intelligence Model for Diagnosis of Orthognathic Surgery. J Pers Med. 2021 Apr 29;11(5):356 Krizhevsky A, Sutskever I, Hinton G.E. ImageNet classification with deep convolutional neural networks. Commun. ACM. 2017; 60:84–90 Sandler M, Howard A, Zhu M, Zhmoginov A, Chen LC. MobileNetV2: Inverted residuals and linear bottlenecks. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2018:4510–4520. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 18 Apr, 2026 Reviewers agreed at journal 15 Apr, 2026 Reviewers invited by journal 15 Apr, 2026 Editor invited by journal 14 Apr, 2026 Editor assigned by journal 14 Apr, 2026 Submission checks completed at journal 14 Apr, 2026 First submitted to journal 11 Apr, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9389538","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":625411349,"identity":"0319493e-5fa6-4ee6-ac04-b932097a2bda","order_by":0,"name":"Jin Li","email":"","orcid":"","institution":"Hospital of Stomatology, Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Jin","middleName":"","lastName":"Li","suffix":""},{"id":625411350,"identity":"f384a162-607d-4822-a4fa-e1d2bbc3db50","order_by":1,"name":"Yingzhu Bao","email":"","orcid":"","institution":"Hospital of Stomatology, Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Yingzhu","middleName":"","lastName":"Bao","suffix":""},{"id":625411351,"identity":"0610ad01-c452-4b93-b8b0-6cf2e7e2dd32","order_by":2,"name":"Haochen Zhang","email":"","orcid":"","institution":"Hospital of Stomatology, Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Haochen","middleName":"","lastName":"Zhang","suffix":""},{"id":625411352,"identity":"cdb5db83-4659-4ad9-ad6c-b40d324e7892","order_by":3,"name":"Keying Qi","email":"","orcid":"","institution":"Shenzhen Institutes of Advanced Technology","correspondingAuthor":false,"prefix":"","firstName":"Keying","middleName":"","lastName":"Qi","suffix":""},{"id":625411353,"identity":"fd49ed25-1a4a-405a-bd4b-d909bfbc88d7","order_by":4,"name":"Le Yang","email":"","orcid":"","institution":"Hospital of Stomatology, Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Le","middleName":"","lastName":"Yang","suffix":""},{"id":625411354,"identity":"59b87c76-17e5-4871-91e4-9e69719b94cb","order_by":5,"name":"Yufan Chen","email":"","orcid":"","institution":"Hospital of Stomatology, Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Yufan","middleName":"","lastName":"Chen","suffix":""},{"id":625411355,"identity":"4e106ca9-f27d-41f9-9754-cd78c4d4ad7a","order_by":6,"name":"Chen Zhou","email":"","orcid":"","institution":"Hospital of Stomatology, Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Chen","middleName":"","lastName":"Zhou","suffix":""},{"id":625411356,"identity":"dc62367b-191e-4276-90c2-67ab168ccaf9","order_by":7,"name":"Weicai Wang","email":"","orcid":"","institution":"Hospital of Stomatology, Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Weicai","middleName":"","lastName":"Wang","suffix":""},{"id":625411357,"identity":"167a4813-91a8-4ce4-abe8-88780b95cd1a","order_by":8,"name":"Baicheng Bao","email":"","orcid":"","institution":"Hospital of Stomatology, Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Baicheng","middleName":"","lastName":"Bao","suffix":""},{"id":625411358,"identity":"3dacd78d-b178-4908-9300-84df6c95c089","order_by":9,"name":"Guangsen Zheng","email":"","orcid":"","institution":"Hospital of Stomatology, Sun Yat-sen University","correspondingAuthor":false,"prefix":"","firstName":"Guangsen","middleName":"","lastName":"Zheng","suffix":""},{"id":625411359,"identity":"a485a9c3-8624-453c-bcfb-a9a68691ac3d","order_by":10,"name":"Xiaojing Long","email":"","orcid":"","institution":"Shenzhen Institutes of Advanced Technology","correspondingAuthor":false,"prefix":"","firstName":"Xiaojing","middleName":"","lastName":"Long","suffix":""},{"id":625411360,"identity":"f8c16f73-3699-4747-b081-b563753139c7","order_by":11,"name":"Xi Wang","email":"data:image/png;base64,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","orcid":"","institution":"Hospital of Stomatology, Sun Yat-sen University","correspondingAuthor":true,"prefix":"","firstName":"Xi","middleName":"","lastName":"Wang","suffix":""}],"badges":[],"createdAt":"2026-04-11 16:39:33","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9389538/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9389538/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107635870,"identity":"db19a7a4-d298-4512-82e2-35792c683e8e","added_by":"auto","created_at":"2026-04-23 12:36:36","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":790299,"visible":true,"origin":"","legend":"\u003cp\u003eThe flowchart of data selection\u003c/p\u003e","description":"","filename":"Fig.1.png","url":"https://assets-eu.researchsquare.com/files/rs-9389538/v1/d65a14152fd81f7bf5e730e7.png"},{"id":107869158,"identity":"f9c1a150-bfdd-403f-9e9f-b0466700dbad","added_by":"auto","created_at":"2026-04-27 07:36:15","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1672985,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eManual cropping process example\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Fig.2.png","url":"https://assets-eu.researchsquare.com/files/rs-9389538/v1/6d05106d6beb70892573b7fc.png"},{"id":107707196,"identity":"7b37faa8-d1d5-46a4-8cbe-29e416159105","added_by":"auto","created_at":"2026-04-24 09:19:46","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":768918,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFlowchart of the CNN models constructed based on cephalometric\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Fig.3.png","url":"https://assets-eu.researchsquare.com/files/rs-9389538/v1/8f628a7b09dd31df836c7dcd.png"},{"id":107635871,"identity":"0e9b6c01-5d70-4bb9-8ec4-940d51002008","added_by":"auto","created_at":"2026-04-23 12:36:36","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":231238,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDepiction of the means and standard deviations for the prediction accuracy of the four models\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Fig.4.png","url":"https://assets-eu.researchsquare.com/files/rs-9389538/v1/742f2c8fbc25ffe9b1eb0d5f.png"},{"id":107706590,"identity":"cd7b7ed6-c7e0-4f45-bf29-cfda0c20bd2d","added_by":"auto","created_at":"2026-04-24 09:18:25","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":787946,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eThe average performance of each model\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Fig.5.png","url":"https://assets-eu.researchsquare.com/files/rs-9389538/v1/150c7dd55b923ba289fbdefc.png"},{"id":107635873,"identity":"d2f97846-51c7-45e4-81cb-8523cd119500","added_by":"auto","created_at":"2026-04-23 12:36:37","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":161561,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eROC curve of deep learning models in internal test datasets\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Fig.6.png","url":"https://assets-eu.researchsquare.com/files/rs-9389538/v1/a0895210b6158a3c82cd80b1.png"},{"id":107635877,"identity":"0ac3bc6d-fedb-44e2-9247-bc395c42414a","added_by":"auto","created_at":"2026-04-23 12:36:37","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":1674554,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePlots for the performance metrics including accuracy values of each model at different epochs during training and validation phase\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Fig.7.png","url":"https://assets-eu.researchsquare.com/files/rs-9389538/v1/d914d1abbebe7cb9a3942e4b.png"},{"id":107707622,"identity":"34e27fe7-ab69-473b-93f1-1af4ac76ed56","added_by":"auto","created_at":"2026-04-24 09:20:46","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":575527,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eConfusion matrix showing the classification performance of the four models. The prediction labels are shown along the x-axis, whereas the true diagnosis labels are shown along the y-axis. Predictions matching the true diagnosis labels resulted in blue squares along the diagonal from top left to bottom right. The color mapping is determined based on the maximum and minimum values on the color bar to the right\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Fig.8.png","url":"https://assets-eu.researchsquare.com/files/rs-9389538/v1/e7304d9a1facc2930baf0df1.png"},{"id":107871712,"identity":"e5fec035-56df-4890-82fd-84467f57777c","added_by":"auto","created_at":"2026-04-27 07:53:47","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":7076175,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9389538/v1/259a55da-33d2-4bb2-8831-1c0daa4677be.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Construction of a CNN Model Based on Lateral Cephalometric Imaging Features and Its Application in Surgery-First Treatment Decision-Making for Skeletal Class III Malocclusion","fulltext":[{"header":"Introduction","content":"\u003cp\u003eIn the combined orthodontic\u0026ndash;orthognathic treatment of skeletal Class III malocclusion, the selection between surgery-first approach (SFA) and orthodontics-first approach(OFA) represents a complex clinical decision‑making process\u003csup\u003e[1][2][3]\u003c/sup\u003e.While SFA has gained significant attention for its potential to shorten total treatment duration and provide immediate facial esthetic improvement, its application requires careful patient selection\u003csup\u003e[4][5]\u003c/sup\u003e. Inappropriate treatment pathways may lead to increased postoperative orthodontic complexity and compromised long-term stability\u003csup\u003e[6]\u003c/sup\u003e. Currently, this decision relies primarily on comprehensive evaluation of lateral cephalometric parameters, dental model, clinician experience, etc. However, this process largely depends on manual analysis and lacks objective, standardized criteria\u003csup\u003e[11]\u003c/sup\u003e, which is necessary for consistent preoperative assessment across different institutions.\u003c/p\u003e \u003cp\u003eRecent advancements in digital orthodontics have introduced new possibilities for enhancing this decision-making process through deep learning, which excels at efficient feature extraction and automated prediction\u003csup\u003e[7][8]\u003c/sup\u003e. Nevertheless, the clinical adoption of these models is often hindered by their \"black-box\" nature and a perceived lack of interpretability\u003csup\u003e[9][10]\u003c/sup\u003e. In contrast, standard analysis offers established, theory-based criteria but is constrained by manual processing and a narrow range of predefined variables\u003csup\u003e[12]\u003c/sup\u003e. While several studies have attempted to integrate these two modalities\u0026mdash;utilizing CBCT-based nomograms or multi-modal data such as lateral cephalograms combined with 3D dental scans \u0026mdash;to assist SFA selection, certain considerations remain\u003csup\u003e[11][12]\u003c/sup\u003e. For instance, while CBCT provides a distinct three-dimensional advantage, its routine application may be tempered by concerns regarding higher radiation exposure\u003csup\u003e[13]\u003c/sup\u003e. In contrast, lateral cephalograms remain the most routine and available imaging modality, yet the overall decision-relevant morphological information they contain has not been fully explored. Furthermore, because clinical indications inherently vary across institutions, models trained exclusively on single-center data inevitably face limitations in generalizability. \u003csup\u003e[14]\u003c/sup\u003e. Nevertheless, while individual studies are often initially constrained to a single center, utilizing universally accessible imaging modalities allows for data pooling across multiple centers in the future\u003csup\u003e[15]\u003c/sup\u003e. This enables the gradual construction of expansive, diverse databases, progressively mitigating selection bias and approaching more generalizable outcomes\u003csup\u003e[16]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eAlthough previous studies have shown that candidates for SFA and OFA exhibit distinct maxillofacial, vertical skeletal, and dental compensation patterns, synthesizing these morphologic characteristics into systematic clinical guidelines remains challenging\u003csup\u003e[1][7][8]\u003c/sup\u003e. In traditional practice, surgical indications rely on implicit morphological criteria and individual clinical experience, making the decision-making process difficult to standardize\u003csup\u003e[17]\u003c/sup\u003e. In this context, the morphological decision-related information embedded within accessible 2D lateral cephalograms remains underexplored by modern deep learning approaches. Therefore, AI-driven imaging modeling is meaningful even within single-center settings, because it successfully translates unquantifiable clinical experience into an objective, mathematical framework\u003csup\u003e[18]\u003c/sup\u003e. By automatically extracting the core anatomical determinants from routine radiographs, the deep learning model explicitly clarifies and quantifies the shared morphological characteristics driving these complex surgical decisions\u003csup\u003e[19]\u003c/sup\u003e. This critical transformation of implicit clinical heuristics into a baseline provides evidence to support the development of more consistent and generalizable decision-making frameworks.\u003c/p\u003e \u003cp\u003eTo address these issues, we developed and validated a CNN-based decision-support framework for analyzing routinely acquired lateral cephalograms. By focusing on standard 2D imaging, we investigated whether the global morphological information it contains can provide reliable guidance for treatment selection without the need for 3D imaging or manual annotation. Among the evaluated models, MobileNetV2 was utilized for its balance between computational efficiency and classification performance. To enhance clinical interpretability, multivariate statistical analysis was incorporated to validate the independent predictors identified by the model, thereby linking automated efficiency with clinical insight. This study represents an independent evaluation of AI-driven decision-making in a new clinical setting and supports the development of more consistent and accessible treatment planning strategies.\u003c/p\u003e"},{"header":"METHODS","content":"\u003cp\u003e\u003cstrong\u003ePatients\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePatients with skeletal Class III malocclusion who underwent orthognathic surgery at the\u0026nbsp;\u003cstrong\u003e\u0026times;\u003c/strong\u003e between 2019 and 2024 were consecutively enrolled. A total of 530 patients were screened, and 511 were finally included, comprising 77 males and 114 females in the SFA group and 145 males and 175 females in the OFA group. Preoperative treatment plans were determined jointly by orthodontists and oral and maxillofacial surgeons, and patients were grouped according to the surgical approach. The study was approved by the Ethics Committee of the\u0026nbsp;\u003cstrong\u003e\u0026times;\u003c/strong\u003e(GHKQ-202401-L2), and all participants provided informed consent.\u0026nbsp;Details of data selection and grouping are shown in Figure 1 and Table 1.\u003c/p\u003e\n\u003cp\u003eThe exclusion criteria included: (1) a history of orthodontic treatment or orthognathic surgery; (2) dentition defects, defined as the absence of two or more teeth per quadrant, excluding third molars; (3) craniofacial developmental deficiencies or congenital deformities (e.g., cleft lip and palate), as well as a history of infection, trauma, or tumors; and (4) image quality issues that could compromise diagnostic validity, including motion artifacts, metal artifacts, image blurring, or insufficient resolution.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1 Distribution of imaging data\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 74px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width: 230px;\"\u003e\n \u003cp\u003eLateral cephalograms（N）\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 74px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 74px;\"\u003e\n \u003cp\u003eMan\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003eWoman\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 74px;\"\u003e\n \u003cp\u003eSFA\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 74px;\"\u003e\n \u003cp\u003e77\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003e114\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 74px;\"\u003e\n \u003cp\u003eOFA\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 74px;\"\u003e\n \u003cp\u003e145\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 156px;\"\u003e\n \u003cp\u003e175\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eOFA orthodontic-first approach, SFA surgery-first approach\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eVariables of structural parameters on lateral cephalograms\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe definitions or descriptions of the variable parameters used are provided in Table 2.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2. The definitions or descriptions of the variable measurements\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"633\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 205px;\"\u003e\n \u003cp\u003eVariable measurement\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 428px;\"\u003e\n \u003cp\u003eDescription\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 205px;\"\u003e\n \u003cp\u003eSNA(\u0026deg;)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 428px;\"\u003e\n \u003cp\u003eAngle formed between sella, nasion and point A\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 205px;\"\u003e\n \u003cp\u003eSNB(\u0026deg;)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 428px;\"\u003e\n \u003cp\u003eAngle formed between sella, nasion and point B\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 205px;\"\u003e\n \u003cp\u003eANB(\u0026deg;)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 428px;\"\u003e\n \u003cp\u003eAngle formed between point A, nasion and point B\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 205px;\"\u003e\n \u003cp\u003eSN-MP(\u0026deg;)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 428px;\"\u003e\n \u003cp\u003eSella\u0026ndash;Nasion to Mandibular Plane Angle\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 205px;\"\u003e\n \u003cp\u003eFH-MP(\u0026deg;)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 428px;\"\u003e\n \u003cp\u003eFrankfort Horizontal to Mandibular Plane Angle\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 205px;\"\u003e\n \u003cp\u003eU1-SN(\u0026deg;)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 428px;\"\u003e\n \u003cp\u003eAngle formed by SN-plane and the axis of the maxillary incisor (U1) in midsagittal plane\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 205px;\"\u003e\n \u003cp\u003eU1-NA(\u0026deg;)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 428px;\"\u003e\n \u003cp\u003eAngle between long axis of the maxillary central incisor and nasion\u0026ndash;point A line\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 205px;\"\u003e\n \u003cp\u003eL1-MP(\u0026deg;)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 428px;\"\u003e\n \u003cp\u003eMandibular Incisor to Mandibular Plane Angle\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 205px;\"\u003e\n \u003cp\u003eWits（mm）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 428px;\"\u003e\n \u003cp\u003eDrawn perpendiculars from points A and B onto the occlusal plane and measured the distance between these two points\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 205px;\"\u003e\n \u003cp\u003ePo-NB（mm）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 428px;\"\u003e\n \u003cp\u003ePogonion to NB Line Distance\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 205px;\"\u003e\n \u003cp\u003eCo-A（mm）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 428px;\"\u003e\n \u003cp\u003eDistance between condylion and point A\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 205px;\"\u003e\n \u003cp\u003eCo-Gn（mm）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 428px;\"\u003e\n \u003cp\u003eDistance between condylion and gonion\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eImage Preprocessing\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe dataset was randomly divided into training, validation, and test sets at a ratio of 7:1.5:1.5, ensuring independence to avoid data leakage. All images were resized to 224 \u0026times; 224 pixels and converted into tensor formats (Fig.2). To improve generalization, data augmentation was applied only to the training set, including random translation within \u0026plusmn;5 pixels, rotation of 0\u0026ndash;5\u0026deg;, and \u0026plusmn;10% adjustments in contrast and brightness. No augmentation was applied to the validation or test sets to ensure unbiased evaluation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCNN classification workflow based on lateral cephalograms\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFour pretrained models\u0026mdash;DenseNet121, InceptionV3, MobileNetV2, and VGG16\u0026mdash;were trained on lateral cephalograms. Four backbone architectures were trained using cross-entropy loss, ReLU activation, and the Adam optimizer. All models were evaluated using five‑fold cross‑validation with a batch size of 64. DenseNet121, InceptionV3, and MobileNetV2 were trained for 100 epochs, while VGG16 was trained for 300 epochs, with an initial learning rate of 0.001 and weight decay of 2\u0026nbsp;\u0026times;\u0026nbsp;10⁻⁴. Learning rate reduction and early stopping were applied according to validation performance. The workflow is illustrated in Fig. 3.\u003c/p\u003e\n\u003col\u003e\n \u003cli\u003eData cleaning and dataset construction: The dataset was first cleared of invalid, low-quality, or redundant images. Only qualified lateral cephalograms were retained for subsequent analysis.\u003c/li\u003e\n \u003cli\u003eImage preprocessing and data partitioning: The dataset was resized to 224\u0026nbsp;\u0026times;\u0026nbsp;224 pixels and converted into vector formats compatible with the deep learning framework, with grayscale normalization and other preprocessing steps applied.\u003c/li\u003e\n \u003cli\u003eData augmentation: Common augmentation techniques, such as rotation, scaling, and flipping, were applied to expand sample diversity and improve model generalization for the subsequent classification tasks.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eModel construction and optimization: The four models were selected as model backbones. Network parameters were tuned through learning rate adjustment, weight decay regularization, and data augmentation to enhance classification performance.\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eModel performance evaluation: The trained deep learning models processed the test set data, and metrics such as accuracy, recall, F1 score, and AUC evaluated their generalization and clinical applicability in classifying SFA versus OFA.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModel output evaluation and statistical methods\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll statistical analyses were performed using SPSS version 25.0. Descriptive statistics were used to summarize the baseline parameters of the SFA and OFA groups, followed by independent-samples t-tests for between-group comparisons of continuous variables. Univariate analysis was conducted to identify potential predictive factors, and multivariate logistic regression analysis was performed to determine independent predictors. All statistical tests were two-sided, with a P-value \u0026lt; 0.05 considered statistically significant. Intra-observer reliability of cephalometric measurements was assessed using the intraclass correlation coefficient (ICC) with a two-way random effects model and absolute agreement, based on repeated measurements of 511 patients by the same examiner at three time points with two-week intervals.\u003c/p\u003e\n\u003cp\u003eThe classification performance of the model for distinguishing SFA from OFA was evaluated by standard diagnostic metrics, including accuracy (ACC), sensitivity (Sen), specificity (Spec), precision, F1 score, positive predictive value (PPV), negative predictive value (NPV), and AUC; ROC curves and confusion matrices were generated, and 95% confidence intervals were estimated using 1,000 bootstrap resampling iterations.\u003c/p\u003e"},{"header":"RESULTS","content":"\u003cp\u003eThe means and standard deviations of the twelve cephalometric parameters for both groups from the 511 patients are presented in Table 3. The SFA group demonstrated smaller SN\u0026ndash;MP and FH\u0026ndash;MP angles, increased labial inclination and protrusion of the maxillary incisors, and more pronounced labial inclination of the mandibular incisors.\u0026nbsp;In addition, the intraclass correlation coefficients (ICCs) for the 12 variables ranged from 0.973 to 0.989, indicating excellent intra-observer reliability.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3.\u003c/strong\u003e \u003cstrong\u003eLateral cephalometric parameters in skeletal Class III patients under two treatment approaches\u003c/strong\u003e\u003c/p\u003e\n\u003cdiv align=\"center\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width: 132px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width: 156px;\"\u003e\n \u003cp\u003eOFA Group\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" style=\"width: 156px;\"\u003e\n \u003cp\u003eSFA Group\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" style=\"width: 109px;\"\u003e\n \u003cp\u003ep-value\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003eMean\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003eSD\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003eMean\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003eSD\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 132px;\"\u003e\n \u003cp\u003eSNA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e83.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e3.55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e84.05\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e3.74\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 109px;\"\u003e\n \u003cp\u003e0.430\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 132px;\"\u003e\n \u003cp\u003eSNB\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e86.77\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e4.35\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e87.02\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e4.38\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 109px;\"\u003e\n \u003cp\u003e0.521\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 132px;\"\u003e\n \u003cp\u003eANB\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e-2.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e3.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e-3.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e2.81\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 109px;\"\u003e\n \u003cp\u003e0.901\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 132px;\"\u003e\n \u003cp\u003eSN-MP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e33.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e7.54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e32.27\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e7.24\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 109px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.01\u003c/strong\u003e\u003cstrong\u003e3\u003c/strong\u003e\u003csup\u003e※\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 132px;\"\u003e\n \u003cp\u003eFH-MP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e27.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e6.79\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e26.09\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e6.53\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 109px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.01\u003c/strong\u003e\u003cstrong\u003e5\u003c/strong\u003e\u003csup\u003e※\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 132px;\"\u003e\n \u003cp\u003eU1-SN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e114.51\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e7.11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e117.55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e7.46\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 109px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e＜\u003c/strong\u003e\u003cstrong\u003e0.01\u003c/strong\u003e\u003csup\u003e※※\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 132px;\"\u003e\n \u003cp\u003eU1-NA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e30.59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e5.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e33.41\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e6.79\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 109px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e＜\u003c/strong\u003e\u003cstrong\u003e0.01\u003c/strong\u003e\u003csup\u003e※※\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 132px;\"\u003e\n \u003cp\u003eL1-MP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e80.13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e8.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e83.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e7.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 109px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e＜\u003c/strong\u003e\u003cstrong\u003e0.01\u003c/strong\u003e\u003csup\u003e※※\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 132px;\"\u003e\n \u003cp\u003e\u0026nbsp;Wits\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e-10.77\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e4.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e-9.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e5.25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 109px;\"\u003e\n \u003cp\u003e0.085\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 132px;\"\u003e\n \u003cp\u003ePo-NB\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e1.34\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e1.07\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e1.51\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e1.39\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 109px;\"\u003e\n \u003cp\u003e0.115\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 132px;\"\u003e\n \u003cp\u003eCo-A\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e81.95\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e6.37\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e82.38\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e8.66\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 109px;\"\u003e\n \u003cp\u003e0.518\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 132px;\"\u003e\n \u003cp\u003eCo-Gn\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e122.78\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e10.84\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e122.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 78px;\"\u003e\n \u003cp\u003e14.05\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 109px;\"\u003e\n \u003cp\u003e0.878\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e※p\u0026lt;0.05, ※※p\u0026lt;0.001\u003c/p\u003e\n\u003cp\u003eThe results of the univariate analysis are presented in Table 4, which identified five variables with statistical significance: SN\u0026ndash;MP, FH\u0026ndash;MP, U1\u0026ndash;SN, U1\u0026ndash;NA, and L1\u0026ndash;MP. Following these findings, multivariable analysis was conducted on the variables that showed significance in the univariate analysis. The results demonstrated L1\u0026ndash;MP as an independent predictive factor,\u0026nbsp;as detailed in Table 5.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 4. Results of univariate analysis in patients with skeletal Class III malocclusion\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"553\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 170px;\"\u003e\n \u003cp\u003eVariable\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e\u0026beta;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003eOR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 125px;\"\u003e\n \u003cp\u003e95%CI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 99px;\"\u003e\n \u003cp\u003eP\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eSNA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e-0.02\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 125px;\"\u003e\n \u003cp\u003e0.93 ~ 1.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 99px;\"\u003e\n \u003cp\u003e0.429\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eSNB\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e-0.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 125px;\"\u003e\n \u003cp\u003e0.95\u0026nbsp;~ 1.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 99px;\"\u003e\n \u003cp\u003e0.520\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eANB\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e0.004\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e1.004\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 125px;\"\u003e\n \u003cp\u003e0.94\u0026nbsp;~ 1.07\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 99px;\"\u003e\n \u003cp\u003e0.901\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eSN-MP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e0.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e1.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 125px;\"\u003e\n \u003cp\u003e1.01 ~ 1.05\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 99px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.0\u003c/strong\u003e\u003cstrong\u003e14\u003c/strong\u003e\u003csup\u003e※\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eFH-MP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e0.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e1.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 125px;\"\u003e\n \u003cp\u003e1.01 ~ 1.06\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 99px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.0\u003c/strong\u003e\u003cstrong\u003e15\u003c/strong\u003e\u003csup\u003e※\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eU1-SN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e-0.06\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e0.94\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 125px;\"\u003e\n \u003cp\u003e0.92\u0026nbsp;~ 0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 99px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e＜\u003c/strong\u003e\u003cstrong\u003e0.0\u003c/strong\u003e\u003cstrong\u003e0\u003c/strong\u003e\u003cstrong\u003e1\u003c/strong\u003e\u003csup\u003e※※\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eU1-NA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e-0.07\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 125px;\"\u003e\n \u003cp\u003e0.91\u0026nbsp;~ 0.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e＜\u003c/strong\u003e\u003cstrong\u003e0.0\u003c/strong\u003e\u003cstrong\u003e0\u003c/strong\u003e\u003cstrong\u003e1\u003c/strong\u003e\u003csup\u003e※※\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eL1-MP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e0.04\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e1.04\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 125px;\"\u003e\n \u003cp\u003e1.02 ~ 1.06\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e＜\u003c/strong\u003e\u003cstrong\u003e0.0\u003c/strong\u003e\u003cstrong\u003e0\u003c/strong\u003e\u003cstrong\u003e1\u003c/strong\u003e\u003csup\u003e※※\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eWits\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e-0.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e0.97\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 125px;\"\u003e\n \u003cp\u003e0.93 ~ 1.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 99px;\"\u003e\n \u003cp\u003e0.086\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003ePo-NB\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e-0.12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e0.89\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 125px;\"\u003e\n \u003cp\u003e0.77 ~ 1.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 99px;\"\u003e\n \u003cp\u003e0.116\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eCo-A\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e-0.08\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 125px;\"\u003e\n \u003cp\u003e0.97 ~ 1.02\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 99px;\"\u003e\n \u003cp\u003e0.520\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eCo-Gn\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e-0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e1.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 125px;\"\u003e\n \u003cp\u003e0.98~ 1.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 99px;\"\u003e\n \u003cp\u003e0.878\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e※p\u0026lt;0.05, ※※p\u0026lt;0.001\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 5. Results of multivariate analysis in patients with skeletal Class III malocclusion\u003c/strong\u003e\u003c/p\u003e\n\u003cdiv align=\"center\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"553\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 170px;\"\u003e\n \u003cp\u003eVariable\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e\u0026beta;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003eOR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 125px;\"\u003e\n \u003cp\u003e95%CI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 99px;\"\u003e\n \u003cp\u003eP\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eSN-MP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e-0.012\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e0.988\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 125px;\"\u003e\n \u003cp\u003e0.91\u0026nbsp;~ 1.07\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 99px;\"\u003e\n \u003cp\u003e0.768\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eFH-MP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e0.023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e1.023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 125px;\"\u003e\n \u003cp\u003e0.80 ~ 1.43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 99px;\"\u003e\n \u003cp\u003e0.661\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eU1-SN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e-0.023\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e0.978\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 125px;\"\u003e\n \u003cp\u003e0.92\u0026nbsp;~ 1.04\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 99px;\"\u003e\n \u003cp\u003e0.452\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eU1-NA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e-0.043\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 79px;\"\u003e\n \u003cp\u003e0.958\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 125px;\"\u003e\n \u003cp\u003e0.90\u0026nbsp;~ 1.02\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 99px;\"\u003e\n \u003cp\u003e0.176\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 170px;\"\u003e\n \u003cp\u003eL1-MP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 79px;\"\u003e\n \u003cp\u003e-0.033\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 79px;\"\u003e\n \u003cp\u003e0.968\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 125px;\"\u003e\n \u003cp\u003e0.95\u0026nbsp;~ 0.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 99px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.00\u003c/strong\u003e\u003cstrong\u003e5\u003c/strong\u003e\u003cstrong\u003e\u003csup\u003e※\u003c/sup\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e※p\u0026lt;0.05, ※※p\u0026lt;0.001\u003c/p\u003e\n\u003cp\u003eFollowing the standard statistical analyses,\u0026nbsp;we evaluated classification performance using deep learning to automatically distinguish between SFA and OFA. Five-fold cross-validation was used to evaluated the classification performance of DenseNet121, InceptionV3, MobileNetV2, and VGG16.\u0026nbsp;In the training set, the mean accuracies were 97.73% for DenseNet121, 99.56% for InceptionV3, 100% for MobileNetV2, and 90.25% for VGG16. In the test set, the corresponding accuracies were 87.01%, 87.27%, 89.35%, and 80.78%, respectively. In the validation set, the accuracies were 87.50%, 89.69%, 89.69%, and 82.50%, respectively. Taken together, the overall mean accuracies across all datasets were 90.74% for DenseNet121, 92.17% for InceptionV3, 93.01% for MobileNetV2, and 84.51% for VGG16.\u003c/p\u003e\n\u003cp\u003eIn the training set, the standard deviations of the models were 1.67% for DenseNet121, 0.28% for InceptionV3, 0.00% for MobileNetV2, and 2.08% for VGG16. In the test set, the standard deviations were 4.64% for DenseNet121, 3.22% for InceptionV3, 3.53% for MobileNetV2, and 2.08% for VGG16. In the validation set, the standard deviations were 2.96% for DenseNet121, 2.53% for InceptionV3, 3.22% for MobileNetV2, and 3.03% for VGG16. The overall standard deviations of the four models were 2.04%, 1.22%, 0.36%, and 1.22%, respectively.\u0026nbsp;All models demonstrated strong discriminative ability and high stability during the training process.\u003c/p\u003e\n\u003cp\u003eThe above results indicate that MobileNetV2 achieved consistently high accuracy across all datasets with relatively small standard deviations, suggesting stronger learning capability and greater stability. InceptionV3 and DenseNet121 also attained high accuracies with moderate standard deviations. In contrast, VGG16 showed lower accuracies and larger standard deviations across the datasets, indicating the need for further optimization and refinement (Fig. 4).\u003c/p\u003e\n\u003cp\u003eBuilding upon the evaluation of model stability, overall performance was analyzed across multiple classification metrics. As presented in Table 6 , Fig 5 and Fig 6, all major metrics exceeded 0.870 for the four deep learning models in distinguishing SFA from OFA. MobileNetV2 achieved the best performance (AUC = 0.949, accuracy = 0.953, sensitivity = 0.941, F1 = 0.914), outperforming the other models; its negative predictive value was 0.978, indicating a strong ability to exclude non-SFA cases.\u0026nbsp;\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" align=\"\" width=\"553\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width: 104px;\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"7\" style=\"width: 449px;\"\u003e\n \u003cp\u003eEvaluation metrics\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 12.3385%;\"\u003e\n \u003cp\u003eACC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11.5313%;\"\u003e\n \u003cp\u003eSen\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 69px;\"\u003e\n \u003cp\u003eSpec\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 63px;\"\u003e\n \u003cp\u003eF1\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 55px;\"\u003e\n \u003cp\u003ePPV\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003eNPV\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003eDenseNet121\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.3385%;\"\u003e\n \u003cp\u003e0.906\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11.5313%;\"\u003e\n \u003cp\u003e0.885\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 69px;\"\u003e\n \u003cp\u003e0.921\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e0.885\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 63px;\"\u003e\n \u003cp\u003e0.885\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e0.885\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e0.921\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003eInceptionV3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.3385%;\"\u003e\n \u003cp\u003e0.938\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11.5313%;\"\u003e\n \u003cp\u003e0.833\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 69px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.978\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.938\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 63px;\"\u003e\n \u003cp\u003e0.882\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.938\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e0.938\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003eMobileNetV2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.3385%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.953\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11.5313%;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.941\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 69px;\"\u003e\n \u003cp\u003e0.957\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e0.889\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 63px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.914\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e0.889\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.978\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 104px;\"\u003e\n \u003cp\u003eVGG16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12.3385%;\"\u003e\n \u003cp\u003e0.875\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 11.5313%;\"\u003e\n \u003cp\u003e0.810\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 69px;\"\u003e\n \u003cp\u003e0.907\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75px;\"\u003e\n \u003cp\u003e0.810\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 63px;\"\u003e\n \u003cp\u003e0.810\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e0.810\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 55px;\"\u003e\n \u003cp\u003e0.907\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003eTable 6 Prediction performance of each model\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eComputational parameters of each model were evaluated to further explain the differences in model performance, including network depth, parameter count, and inference time (Table 7). The results showed that MobileNetV2 achieved the best balance of lightweight design and efficiency, reducing hardware dependence and making it particularly suitable for small-dataset settings while helping to mitigate overfitting. VGG16, with its simpler architecture, achieved faster inference, which may partly explain its lower accuracy. In contrast, DenseNet121 and InceptionV3 had a larger number of parameters and longer inference times, reflecting the greater computational cost of dense connectivity and multi-branch structures. This more complex design allows them to strike a balance between performance and efficiency. Therefore, model selection should be guided by both computational resources and dataset size.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 7 Performance of each model\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"558\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 129px;\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 143px;\"\u003e\n \u003cp\u003eNetwork depth\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003eNumber of\u003c/p\u003e\n \u003cp\u003eparameters\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 151px;\"\u003e\n \u003cp\u003eInference time\u003c/p\u003e\n \u003cp\u003e(milliseconds)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 129px;\"\u003e\n \u003cp\u003eDenseNet-121\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 143px;\"\u003e\n \u003cp\u003e431\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e7563330\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 151px;\"\u003e\n \u003cp\u003e52\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 129px;\"\u003e\n \u003cp\u003eInceptionV3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 143px;\"\u003e\n \u003cp\u003e315\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e22852898\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 151px;\"\u003e\n \u003cp\u003e39\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 129px;\"\u003e\n \u003cp\u003eMobileNetV2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 143px;\"\u003e\n \u003cp\u003e158\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e2914882\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 151px;\"\u003e\n \u003cp\u003e31\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 129px;\"\u003e\n \u003cp\u003eVGG-16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 143px;\"\u003e\n \u003cp\u003e23\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 135px;\"\u003e\n \u003cp\u003e14978370\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 151px;\"\u003e\n \u003cp\u003e11\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eConsidering the influence of the aforementioned architectural characteristics on model learning, the convergence behavior of each model was analyzed. As presented in Fig 7, MobileNetV2 exhibited the fastest increase in training accuracy, with validation accuracy improving steadily and stabilizing around epoch 57, demonstrating rapid convergence and strong overall performance. DenseNet121 and InceptionV3 also performed well, reaching stability at epochs 76 and 84, respectively, though their validation accuracy gains gradually slowed, indicating mild overfitting. In contrast, VGG16 maintained high training accuracy but showed more variable and slower validation performance, consistent with the results in Table 4 and likely due to its deeper architecture. Overall, MobileNetV2 achieved high and stable accuracy in fewer epochs, demonstrating the best performance among the models.\u003c/p\u003e\n\u003cp\u003eFinally, we evaluated the practical performance of the models in the classification task by analyzing the confusion matrices derived from a single fold of data. As presented in Fig 8, the numbers of correctly classified SFA cases were 16 for DenseNet121, 15 for InceptionV3, 16 for MobileNetV2, and 17 for VGG16, whereas the numbers of correctly classified OFA cases were 42, 45, 45, and 39, respectively. MobileNetV2 demonstrated balanced identification of true positives (SFA) and true negatives (OFA), reflecting good discriminative performance.\u003c/p\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eIn the treatment of skeletal Class III malocclusion, the choice between SFA and OFA can substantially affect treatment duration and facial aesthetic outcomes. Although SFA offers advantages such as immediate correction and a shorter treatment course\u003csup\u003e[20]\u003c/sup\u003e, clear criteria for patient selection remain lacking in routine clinical practice. To address this issue, we developed and evaluated four deep learning models and, in parallel, performed multivariate statistical analysis, which identified L1\u0026ndash;MP as an independent predictor of SFA selection. Among the evaluated models, MobileNetV2 showed superior discriminative performance in distinguishing candidates for SFA and OFA, highlighting its potential as a valuable clinical decision-support tool.\u003c/p\u003e \u003cp\u003eIn recent years, the integration of artificial intelligence in orthodontic-orthognathic treatment planning has been extensively explored, with various models proposed to assist in selecting between OFA and SFA. These decision-support systems generally rely on high-dimensional clinical data, often combining traditional 2D cephalometric measurements with 3D intraoral scans, or utilizing advanced deep learning architectures on pure 3D imaging. For instance, in a previous work, several advanced deep learning architectures (including Simple-CNN, DenseNet121, MobileNetV2, EfficientNet-B0, ResNet10, and SEResNet50) were applied to classify OFA and SFA using 3D imaging data, achieving a high degree of performance with an accuracy of 0.909 and an AUC of 0.896\u003csup\u003e[11]\u003c/sup\u003e. Similarly, another representative study developed a multimodal ResNet34 model combining standard lateral cephalograms with 3D maxillary and mandibular scans, achieving an accuracy of 0.906\u003csup\u003e[12]\u003c/sup\u003e. While 3D maxillary models can achieve an accuracy of up to 0.970, the lack of mandibular information limits comprehensive evaluation of skeletal Class III malocclusion. Against the backdrop of increasingly complex, multi-dimensional methodologies, the present study sought to determine whether comparable or superior performance could be achieved using only routine 2D lateral cephalograms. Through architectural optimization with MobileNetV2, our 2D-based model achieved an accuracy of 0.953. This performance approaches or even exceeds that of the aforementioned 3D and multimodal models. Achieving high predictive robustness using standard 2D imaging\u0026mdash;which constitutes a foundational, low-radiation, and highly accessible diagnostic tool universally available across clinical settings\u0026mdash;highlights the immense clinical translatability of this approach. From a machine learning perspective, this may relate to the tendency of high-dimensional features to overfit under limited data, thereby constraining generalization\u003csup\u003e[21]\u003c/sup\u003e. This further indicates that lightweight architectural designs allow MobileNetV2 to efficiently extract essential discriminative features from 2D radiographs, preserving dependable generalization without the absolute necessity for resource-intensive 3D or multimodal data\u003csup\u003e[22]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eFurthermore, architectural variations, particularly network depth and parameter complexity, were associated with marked differences in model performance. MobileNetV2 demonstrated superior overall efficacy compared to heavier architectures like DenseNet121, InceptionV3, and VGG16. Theoretically, excessively deep networks may suffer from diminished generalization due to vanishing gradients, increased susceptibility to overfitting, and reduced backpropagation efficiency\u003csup\u003e[23][24]\u003c/sup\u003e. For example, although the dense connectivity of DenseNet121 and the multi-branch architecture of InceptionV3 facilitate feature fusion, their relatively high parameter counts (7.56M and 22.85M, respectively) under limited data conditions may lead to overfitting and reduced generalization performance\u003csup\u003e[25][26]\u003c/sup\u003e. Although the VGG16 model has fewer layers (16 layers) and a relatively simple architecture, its large number of parameters and complex fully connected layer design (14.98\u0026nbsp;million parameters) result in high computational cost and relatively weak resistance to overfitting\u003csup\u003e[27][28]\u003c/sup\u003e. In contrast, MobileNetV2 employs depthwise separable convolutions and linear bottleneck techniques to maintain network depth while compressing the parameter count to 2.91\u0026nbsp;million, thereby reducing computational complexity and memory consumption. This lightweight design enables stronger generalization capability in small-sample scenarios and is particularly well suited for applications with limited data availability\u003csup\u003e[29][30]\u003c/sup\u003e. At the feature extraction level, the linear bottleneck design of MobileNetV2 better preserves global information and captures skeletal features relevant to surgical decision-making. As a result, minimizing information distortion or loss may be particularly important for improving the accuracy of orthognathic surgery approach prediction\u003csup\u003e[31][32]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eA major barrier to the clinical adoption of deep learning in medicine is its \u0026ldquo;black-box\u0026rdquo; nature. To mitigate this, the present study positions deep learning as a decision-support tool rather than a replacement for clinical judgment. The consistency between statistical findings and model performance provides a degree of indirect interpretability. Multivariate analysis identified lower incisor compensation (L1\u0026ndash;MP) and vertical skeletal characteristics as key determinants. Meanwhile, the CNN achieved high classification accuracy using only raw lateral cephalograms, without explicit anatomical landmark annotation, suggesting that it implicitly learns relevant biomechanical features\u003csup\u003e[7]\u003c/sup\u003e. Moreover, MobileNetV2\u0026rsquo;s streamlined architecture promotes the progressive accumulation of broader spatial relationships rather than convoluted feature entanglements. Consequently, such lightweight networks minimize redundant feature assimilation, inherently reducing the likelihood of the model relying on erroneous or uncontrollable latent artifacts for its predictions\u003csup\u003e[33]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eDespite the promising findings, several limitations should be acknowledged. This study was a single-center retrospective analysis with a relatively limited sample size, and the generalizability of the results requires validation using large-scale, multicenter datasets. In addition, the analysis was based solely on lateral cephalograms, lacking multimodal data, which may limit comprehensive evaluation of three-dimensional structures. Future studies should incorporate multimodal imaging and explainable artificial intelligence techniques (e.g., Grad-CAM) to further enhance the model\u0026rsquo;s interpretability and clinical applicability.\u003c/p\u003e"},{"header":"CONCLUSIONS","content":"\u003cp\u003eA CNN-based framework utilizing routine lateral cephalograms demonstrated robust clinical potential for predicting the optimal treatment approach (SFA vs. OFA) in patients with skeletal Class III malocclusion. Among the evaluated architectures, MobileNetV2 exhibited consistently superior performance, achieving an accuracy of 0.953 and an AUC of 0.949, coupled with sF generalization across training and validation cohorts. By bridging computational efficiency with high predictive accuracy on standard 2D radiographs, this deep learning framework provides an accessible, evidence-based decision-support tool for orthodontic\u0026ndash;orthognathic treatment planning, offering a valuable methodological foundation for future translational investigations.\u003c/p\u003e"},{"header":"Declarations","content":" \u003cp\u003e \u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e \u003cp\u003e This study was approved by the Ethics Committee of Hospital of Stomatology, Sun Yat-sen University, with a reference number of (GHKQ-202401-L2), and the conduct of the study adheres to the Declaration of Helsinki and relevant ethical standards. Informed consent was obtained from all participants prior to their involvement in the study. Clinical trial number is not applicable.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eConsent for publication\u003c/strong\u003e \u003cp\u003eNot applicable.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eCompeting interests\u003c/strong\u003e \u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e \u003cp\u003eThis study was supported by IOF Young Research Grant (IOF2023Y09) and National Natural Science Foundation of China (82201060).\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eXW and XL was responsible for the study design, administration, drafting of the manuscript, critical revision, and final approval of the article. YB and JL contributedto the literature search, performed data acquisition and data analysis. HZ ,KQ and YCperformed data acquisition and data analysis. LY,CZ and WW were responsible for the drafting, data interpretation and manuscript editing. BB and GZ were responsible for critical revision. All authors read and approved the final manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgements\u003c/h2\u003e \u003cp\u003eNot applicable.\u003c/p\u003e\u003ch2\u003eAvailability of data and materials\u003c/h2\u003e \u003cp\u003eDe-identified data are available upon request to qualified researchers. Access requires (1) approval of a research proposal by the institutional ethics committee and (2) a signed data use agreement. Requests should be directed to corresponding author email.\u003c/p\u003e"},{"header":"Unsectioned Paragraphs","content":"\u003col\u003e\n \u003cli\u003eEslami S, Faber J, Fateh A, Sheikholaemmeh F, Grassia V, Jamilian A. Treatment decision in adult patients with class III malocclusion: surgery versus orthodontics. Prog Orthod. 2018 Aug 2;19(1):28. doi: 10.1186/s40510-018-0218-0. PMID: 30069814; PMCID: PMC6070451.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eAlhammadi MS, et al. Orthodontic camouflage versus orthodontic-orthognathic surgical treatment in borderline class III malocclusion: a systematic review. Clin Oral Investig. 2022;26:6443\u0026ndash;6455.\u003c/li\u003e\n \u003cli\u003eBeccuti ML, et al. \u0026ldquo;Surgery First\u0026rdquo; vs \u0026ldquo;Traditional Sequence\u0026rdquo; Surgery: A qualitative study of health experiences. J Maxillofac Oral Surg. 2022;21:1267\u0026ndash;1278.\u003c/li\u003e\n \u003cli\u003eSharma V, Yadav K, Tandon P. An overview of surgery-first approach: Recent advances in orthognathic surgery. J Orthod Sci. 2015;4(1).\u003c/li\u003e\n \u003cli\u003eKhalil AS, Alrehaili RS, Bajunaid M, Alhazmi M, Alshami A, Alharthy B, Fakhry O, Olfat Y, Taher A, Alotaibi R, Alrefai M, Barashid AA. Does Surgery-First Orthognathic Approach Improve Quality of Life of Orthodontic Patients With Skeletal Class III Malocclusion? A Systematic Review Following Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) Guidelines. Cureus. 2025 Mar 29;17(3):e81433. doi: 10.7759/cureus.81433. PMID: 40296934; PMCID: PMC12037207.\u003c/li\u003e\n \u003cli\u003eChen YC, et al. Dental occlusion characteristics for treatment decision-making regarding surgery-first approach. J Clin Med. 2023;12(18).\u003c/li\u003e\n \u003cli\u003eLo Giudice A, Ronsivalle V, Spampinato C, Leonardi R. Fully automatic segmentation of the mandible based on convolutional neural networks (CNNs). Orthod Craniofac Res. 2021 Dec;24 Suppl 2:100-107. doi: 10.1111/ocr.12536. Epub 2021 Dec 1. PMID: 34553817.\u003c/li\u003e\n \u003cli\u003eChoi HI, Jung SK, Baek SH, Lim WH, Ahn SJ, Yang IH, Kim TW. Artificial Intelligent Model With Neural Network Machine Learning for the Diagnosis of Orthognathic Surgery. J Craniofac Surg. 2019 Oct;30(7):1986-1989. doi: 10.1097/SCS.0000000000005650. Erratum in: J Craniofac Surg. 2020 Jun;31(4):1156. doi: 10.1097/SCS.0000000000006531. PMID: 31205280.\u003c/li\u003e\n \u003cli\u003eKhanagar SB, Al-Ehaideb A, Maganur PC, Vishwanathaiah S, Patil S, Baeshen HA, Sarode SC, Bhandi S. Developments, application, and performance of artificial intelligence in dentistry - A systematic review. J Dent Sci. 2021 Jan;16(1):508-522. doi: 10.1016/j.jds.2020.06.019. Epub 2020 Jun 30. PMID: 33384840; PMCID: PMC7770297.\u003c/li\u003e\n \u003cli\u003eBağ İ, Bilgir E, Bayrakdar İŞ, Baydar O, Atak FM, \u0026Ccedil;elik \u0026Ouml;, Orhan K. An artificial intelligence study: automatic description of anatomic landmarks on panoramic radiographs in the pediatric population. BMC Oral Health. 2023 Oct 17;23(1):764. doi: 10.1186/s12903-023-03532-8. PMID: 37848870; PMCID: PMC10583406.\u003c/li\u003e\n \u003cli\u003eBao Y, Qi K, Yang L, Chen Y, Zhou C, Wang W, Bao B, Long X, Zheng G, Wang X. CBCT radiomic features-based machine learning prediction models and nomogram for treatment decision-making regarding surgery-first approach in skeletal Class III malocclusion. BMC Oral Health. 2025 Sep 26;25(1):1464. doi: 10.1186/s12903-025-06702-y. PMID: 41013431; PMCID: PMC12465188.\u003c/li\u003e\n \u003cli\u003eChang JS, Ma CY, Ko EW. Prediction of surgery-first approach orthognathic surgery using deep learning models. Int J Oral Maxillofac Surg. 2024;53:942\u0026ndash;949.\u003c/li\u003e\n \u003cli\u003eLitjens G, et al. A survey on deep learning in medical image analysis. Med Image Anal. 2017;42:60\u0026ndash;88.\u003c/li\u003e\n \u003cli\u003eKrois J, Garcia Cantu A, Chaurasia A, Patil R, Chaudhari PK, Gaudin R, Gehrung S, Schwendicke F. Generalizability of deep learning models for dental image analysis. Sci Rep. 2021 Mar 17;11(1):6102. doi: 10.1038/s41598-021-85454-5. PMID: 33731732; PMCID: PMC7969919.\u003c/li\u003e\n \u003cli\u003eZaborowicz, K.; Zaborowicz, M.; Cieślińska, K.; Biedziak, B. Artificial Intelligence Methods in Cephalometric Image Analysis\u0026mdash;A Systematic Narrative Review. J. Clin. Med. 2026, 15, 1920. \u003ca href=\"https://doi.org/10.3390/jcm15051920\"\u003ehttps://doi.org/10.3390/jcm15051920\u003c/a\u003e\u003c/li\u003e\n \u003cli\u003eChang, Q., Bai, Y., Wang, S. et al. Automated orthodontic diagnosis via self-supervised learning and multi-attribute classification using lateral cephalograms. BioMed Eng OnLine 24, 9 (2025).\u0026nbsp;\u003ca href=\"https://doi.org/10.1186/s12938-025-01345-0\"\u003ehttps://doi.org/10.1186/s12938-025-01345-0\u003c/a\u003e\u003c/li\u003e\n \u003cli\u003eDur\u0026atilde;o AP, Morosolli A, Pittayapat P, Bolstad N, Ferreira AP, Jacobs R. Cephalometric landmark variability among orthodontists and dentomaxillofacial radiologists: a comparative study. Imaging Sci Dent. 2015 Dec;45(4):213-20. doi: 10.5624/isd.2015.45.4.213. Epub 2015 Dec 17. PMID: 26730368; PMCID: PMC4697005.\u003c/li\u003e\n \u003cli\u003eSchwendicke F, Samek W, Krois J. Artificial Intelligence in Dentistry: Chances and Challenges. J Dent Res. 2020 Jul;99(7):769-774. doi: 10.1177/0022034520915714. Epub 2020 Apr 21. PMID: 32315260; PMCID: PMC7309354.\u003c/li\u003e\n \u003cli\u003eSchwendicke, F., Chaurasia, A., Arsiwala, L. et al. Deep learning for cephalometric landmark detection: systematic review and meta-analysis. Clin Oral Invest 25, 4299\u0026ndash;4309 (2021). https://doi.org/10.1007/s00784-021-03990-w\u003c/li\u003e\n \u003cli\u003eUribe FA, Farrell B. Surgery-First Approach in the Orthognathic Patient. Oral Maxillofac Surg Clin North Am. 2020 Feb;32(1):89-103. doi: 10.1016/j.coms.2019.08.009. Epub 2019 Nov 1. PMID: 31685343.\u003c/li\u003e\n \u003cli\u003e]Parekh V, Jacobs MA. Radiomics: a new application from established techniques. Expert Review of Precision Medicine and Drug Development. 2016;1(2):207\u0026ndash;226.\u003c/li\u003e\n \u003cli\u003eVinayahalingam S, Kempers S, Limon L, et al. The Automatic Detection of Caries in Third Molars on Panoramic Radiographs Using Deep Learning: A Pilot Study[J].2021\u003c/li\u003e\n \u003cli\u003eHawkins DM. The problem of overfitting. J Chem Inf Comput Sci. 2004 Jan-Feb;44(1):1-12\u003c/li\u003e\n \u003cli\u003eYu XH. Can backpropagation error surface not have local minima. IEEE Trans Neural Netw. 1992;3(6):1019-21\u003c/li\u003e\n \u003cli\u003eLeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015 May 28;521(7553):436-44\u003c/li\u003e\n \u003cli\u003eZhang C, Bengio S, Hardt M,et al. Understanding deep learning requires rethinking generalization[J]. 2016\u003c/li\u003e\n \u003cli\u003eOmiotek Z, Kotyra A. Flame Image Processing and Classification Using a Pre-Trained VGG16 Model in Combustion Diagnosis. Sensors (Basel) 2021, 21(2)\u003c/li\u003e\n \u003cli\u003eKloster M, Langenk\u0026auml;mper D, Zurowietz M, Beszte et al. Deep learning-based diatom taxonomy on virtual slides. Sci Rep. 2020 Sep 2;10(1):14416\u003c/li\u003e\n \u003cli\u003eXie X, Liu Z, Wang Y, et al. Puppet Dynasty Recognition System Based on MobileNetV2. Entropy (Basel) 2024, 26(8)\u003c/li\u003e\n \u003cli\u003eYong L, Ma L, Sun D, et al. Application of MobileNetV2 to waste classification. PLoS One 2023, 18(3): e0282336\u003c/li\u003e\n \u003cli\u003eKim YH, Park JB, Chang MS, et al. Influence of the Depth of the Convolutional Neural Networks on an Artificial Intelligence Model for Diagnosis of Orthognathic Surgery. J Pers Med. 2021 Apr 29;11(5):356\u003c/li\u003e\n \u003cli\u003eKrizhevsky A, Sutskever I, Hinton G.E. ImageNet classification with deep convolutional neural networks. Commun. ACM. 2017; 60:84\u0026ndash;90\u003c/li\u003e\n \u003cli\u003eSandler M, Howard A, Zhu M, Zhmoginov A, Chen LC. MobileNetV2: Inverted residuals and linear bottlenecks. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2018:4510\u0026ndash;4520.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-oral-health","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ohea","sideBox":"Learn more about [BMC Oral Health](http://bmcoralhealth.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/ohea/default.aspx","title":"BMC Oral Health","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Skeletal Class III malocclusion, Surgery-first approach, Lateral cephalogram, Deep learning","lastPublishedDoi":"10.21203/rs.3.rs-9389538/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9389538/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA deep learning model using four convolutional neural networks (CNNs) was trained based on lateral cephalograms from patients with skeletal Class III malocclusion to assist in selecting between orthodontic-first and surgery-first approaches.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA total of 511 skeletal Class III patients undergoing orthodontic–orthognathic treatment were analyzed. Twelve cephalometric variables were assessed, and multivariate analysis identified independent predictors of treatment selection. Subsequently, the lateral cephalograms were split (7:1.5:1.5) into training, validation, and test sets. Four deep learning models were evaluated using five-fold cross-validation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eMultivariate analysis identified L1–MP was an independent predictor for the selection of SFA. All four deep learning models achieved accuracies above 0.850. MobileNetV2 performed best, with an AUC of 0.949, accuracy of 0.953, sensitivity of 0.941, specificity of 0.957, precision of 0.889, and an F1 score of 0.914.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eMobileNetV2 demonstrated high accuracy in classifying lateral cephalograms, providing significant decision support for the selection between orthodontic-first and surgery-first approaches in skeletal Class III malocclusion patients.\u003c/p\u003e","manuscriptTitle":"Construction of a CNN Model Based on Lateral Cephalometric Imaging Features and Its Application in Surgery-First Treatment Decision-Making for Skeletal Class III Malocclusion","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-23 12:36:32","doi":"10.21203/rs.3.rs-9389538/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-04-18T11:37:14+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"26601123321539499458049399078621960391","date":"2026-04-15T14:23:17+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-15T13:30:14+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-04-14T10:15:20+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-14T04:08:42+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-14T04:08:37+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Oral Health","date":"2026-04-11T16:31:12+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-oral-health","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ohea","sideBox":"Learn more about [BMC Oral Health](http://bmcoralhealth.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/ohea/default.aspx","title":"BMC Oral Health","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"80421b06-cb6e-4a2c-97b7-83f6ef037d7e","owner":[],"postedDate":"April 23rd, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-04-23T12:36:32+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-23 12:36:32","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9389538","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9389538","identity":"rs-9389538","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}