ToothSeg: Robust Tooth Instance Segmentation and Numbering in CBCT Using Deep Learning and Self-Correction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oral and Maxillofacial Radiology

Background:

• Accurate interpretation of cone-beam computed tomography (CBCT) scans is crucial for dental diagnosis and treatment planning.
• Current automated tooth segmentation methods in CBCT struggle with imaging artifacts, anatomical variations, and often require manual corrections.
• These limitations hinder efficient clinical workflows and scalable research in oral health.

Purpose of the Study:

• To develop and evaluate ToothSeg, a fully automated deep learning approach for tooth instance segmentation and numbering in CBCT scans.
• To address the limitations of existing methods by incorporating self-correction for improved accuracy and robustness.
• To provide a tool that reduces manual workload and supports data-driven research in oral and craniofacial health.

Main Methods:

• ToothSeg utilizes a unified deep learning framework combining semantic and instance segmentation.
• A self-correction mechanism is integrated to resolve segmentation errors like merged or split teeth and optimize tooth numbering.
• The method was evaluated on a large in-house dataset (1282 scans) and the ToothFairy2 challenge dataset (480 scans), including comparisons with state-of-the-art techniques.

Main Results:

• ToothSeg significantly improved tooth segmentation accuracy (True Positive Dice: 93.6% to 94.3%) and tooth detection/numbering (multiclass instance F1: 94.2% to 95.5%) compared to an optimized semantic model.
• The approach outperformed existing methods on both datasets, showing superior performance especially in challenging cases (TP Dice: ≥ +0.4%, multiclass instance F1: ≥ +1.8%).
• Ablation studies confirmed the benefits of incorporating instance segmentation and self-correction.

Conclusions:

• ToothSeg offers a robust and accurate automated solution for tooth instance segmentation and numbering in CBCT.
• The method demonstrates significant potential for reducing manual effort in dental image analysis.
• This advancement facilitates scalable, data-driven research and clinical applications in oral and craniofacial health.