설명 가능한 인공지능을 이용한 파노라마 방사선 사진에서의 하악골 골절 분류 = Explainable Artificial Intelligence-Based Classification for Mandibular Fractures on Panoramic Radiographs

다국어 초록 (Multilingual Abstract)

Background and Objectives: Mandibular fractures are among the most prevalent maxillofacial injuries and require accurate diagnosis and classification to ensure optimal functional recovery and aesthetic outcomes. Although panoramic radiographs are widely used in routine clinical settings, they are inherently limited by structural superimposition, geometric distortion, and observer-dependent interpretive variability, which can undermine diagnostic reliability, particularly in complex or condylar- involved fractures. To overcome these challenges, this study developed an explainable deep learning–based automated classification framework and systematically evaluated its diagnostic performance and clinical interpretability on panoramic radiographs acquired in routine clinical practice. Methods: This retrospective study analyzed 800 panoramic radiographs acquired between 2014 and 2024. All images were annotated according to the clinically validated mandibular fracture classification described by Brown (2022), which comprises eight classes (Class I–V, IIc, IIIc, and IVc). Two convolutional neural network architectures, Inception-ResNet-v2 and NASNet-Large, were fine- tuned via transfer learning. Standardized preprocessing included CLAHE-based contrast enhancement and image normalization. A patient-level 5-fold cross-validation strategy was employed. To ensure transparent and clinically meaningful interpretation, multi-level explainability methods, including Grad-CAM, LIME, and SHAP, were implemented to provide global, local, and quantitative insights into the model decision-making process. Results: Inception-ResNet-v2 achieved a mean accuracy of 98.1% and an F1-score of 93.0%, while NASNet-Large achieved 97.8% accuracy and an F1-score of 92.9%. Both models demonstrated consistently strong performance even in low-frequency classes, including complex and condylar- involved fractures. Grad-CAM visualizations indicated that the models primarily attended to clinically relevant regions, such as cortical discontinuities and fracture line trajectories. LIME provided case-specific, pixel-level explanations that clarified the local features influencing predictions. SHAP analyses quantitatively illustrated consistent contribution patterns across the dataset, reinforcing the reliability and interpretive consistency of the models. Conclusion: The proposed explainable deep learning framework demonstrated high diagnostic accuracy and a transparent decision-making process for the automated classification of mandibular fractures on panoramic radiographs. By integrating multi-level interpretability into a robust diagnostic model, this study advances AI from a conventional black-box tool toward a clinically integrable decision-support system. These findings indicate that explainable artificial intelligence may enhance diagnostic consistency, strengthen clinician trust, and contribute to standardized, evidence-based workflows in maxillofacial trauma care. Keywords: Mandibular fracture; deep learning; artificial intelligence; explainable AI; panoramic radiography

목차 (Table of Contents)

• Abstract i
• Contents iii
• List of Tables and Figures v
• 1. Introduction 1
• 2. Materials and Methods 4
• 2.1 Dataset 5
• 2.2 Methods 6
• 2.2.1 Data Preprocessing and Labeling 6
• 2.2.1.1 Data Preprocessing 6
• 2.2.1.2 Data Labeling 6
• 2.2.2 Deep Learning Models 8
• 2.2.2.1 Model Selection and Rationale 8
• 2.2.2.2 Model Architecture and Objective Function 9
• 2.2.3 Training Configuration 9
• 2.2.3.1 Hardware and Software 9
• 2.2.3.2 Training Settings 9
• 2.2.3.3 Data Augmentation and Class Imbalance Handling 10
• 2.2.4 Validation Strategy 10
• 2.2.5 Performance Evaluation and Explainability 10
• 2.2.5.1 Performance Metrics 10
• 2.2.5.2 Explainability Methods 11
• 3. Results 13
• 3.1 Study Population and Imaging Data Distribution 13
• 3.2 Training Stability and Optimization 14
• 3.3 Model Performance Evaluation 15
• 3.4 Results of Explainability Analyses 18
• 3.4.1 Grad-CAM Analysis 18
• 3.4.2 LIME Analysis 21
• 3.4.3 SHAP Analysis 23
• 4. Discussion 25
• 5. Conclusion 28
• References 29
• Abstract (in Korean) 35