Deep Learning Framework for Classification, 3D Segmentation & Visualization of C-shaped Canals

Overview

The goal of this retrospective diagnostic accuracy study is to develop and validate a deep learning framework for the automated classification, three-dimensional (3D) segmentation, and visualization of C-shaped root canal anatomy using cone-beam computed tomography (CBCT) scans in adults with C-shaped root canals. The main questions it aims to answer are: Can a deep learning model accurately classify C-shaped root canal configurations from CBCT images? Can the model precisely segment the complex 3D anatomy of C-shaped root canals, including fins, webs, and isthmuses, with accuracy comparable to expert endodontists? Can the automated framework improve the efficiency and clinical utility of diagnosing and visualizing C-shaped root canal anatomy?

The framework is designed to identify C-shaped canal configurations and accurately segment their complex anatomical features, including fins, webs, and isthmuses. Index test: Deep Learning Model Design for Automated Classification and Segmentation Stage 1: Tooth Localization: * Objective: To identify and segment the target molar (primarily mandibular second molars) from the full CBCT volume. * Architecture: An Attention U-Net based architecture will be explored, known for its ability to focus on important regions and efficiently process dental descriptors. * Output: A cropped Region of Interest (ROI) containing the tooth of interest, reducing computational load for subsequent stages. Stage 2: C-shaped Root Canal Architecture Classification and Segmentation: * Objective: To precisely delineate the C-shaped root canal system, including the main canal lumen, fins, webs, and isthmuses, and to classify its specific type (e.g., C1, C2, C3, C4, C5) based on established criteria (e.g., Fan's classification). * Architecture: Advanced 3D U-Net variants will be explored, given their proven efficacy in medical image segmentation and ability to capture fine details. * Optimization: Models will be trained using robust optimizers (e.g., ADAM) with a managed learning rate schedule. Early stopping criteria will be implemented based on validation set performance to prevent overfitting. 3D Reconstruction and Advanced Visualization Pipeline 3D Model Generation: * Conversion: Segmented 3D masks will be converted into standard 3D file formats, such as Standard Triangle Language (STL), ensuring interoperability with various software and 3D printing platforms. Interactive Visualization Development: ● Software/Libraries: Open-source libraries like Open3D will be explored for interactive rendering and development of clinical utility features. Performance Evaluation and Validation Quantitative Metrics: * Segmentation: Dice Similarity Coefficient (DSC), Hausdorff Distance (HD) and Intersection over Union (IoU) will be used to assess spatial overlap and boundary agreement. * Classification: Accuracy, Precision, Recall, F1-score, and Area Under the Curve (AUC) will evaluate the model's ability to correctly categorize C-shaped canal types. Clinical Utility and Efficiency Assessment: * Qualitative Evaluation: Experienced endodontists will qualitatively assess the practical applicability and accuracy of the segmented 3D models for diagnosis, treatment planning, and identification of critical anatomical features. * Time Efficiency: The time efficiency of the automated framework will be measured and compared to manual segmentation processes. Reference standard: * Expert Annotation: Manual classification and segmentation will be performed by multiple experienced endodontists or dental-maxillofacial radiologists, establishing the "gold standard" ground truth for the dataset. Full manual 3D segmentation, including the intricate architectural features, will be meticulously performed using 3D Slicer software. For 2D annotations, such as those for initial classification tasks or specific cross-sectional views, Roboflow will be utilized. * Inter-observer Variability: Inter-observer variability among annotators will be assessed to ensure the consistency and quality of the ground truth.