A Deep Learning Approach to Segment and Classify C-Shaped Canal Morphologies in Mandibular Second Molars Using...

Area of Science:

• Endodontics
• Dental Anatomy
• Artificial Intelligence in Dentistry

Background:

• C-shaped root canal anatomy presents diagnostic challenges in endodontic treatment planning.
• Accurate identification of C-shaped canals is crucial for successful clinical decision-making and treatment outcomes.
• Cone-beam computed tomography (CBCT) provides detailed 3D imaging for anatomical assessment.

Purpose of the Study:

• To develop and evaluate deep learning (DL) models for classifying C-shaped canal anatomy in mandibular second molars using CBCT data.
• To compare the performance of three distinct DL architectures: U-Net, residual U-Net, and Xception U-Net.
• To assess the efficacy of automated C-shaped anatomy detection and classification.

Main Methods:

• Three DL architectures (U-Net, residual U-Net, Xception U-Net) were employed for image segmentation and classification.
• Models were trained and validated on 100 CBCT volumes and tested on an independent set of 35 CBCT volumes.
• Performance was quantified using the Dice index for voxel-matching accuracy and mean sensitivity for subcategory prediction.

Main Results:

• Xception U-Net achieved the highest mean Dice coefficient (0.768 ± 0.0349), outperforming residual U-Net (0.736 ± 0.0297) and U-Net (0.660 ± 0.0354).
• Statistical analysis confirmed significant differences in performance among the architectures, with Xception U-Net and residual U-Net showing superior results compared to U-Net.
• The addition of contrast-limited adaptive histogram equalization improved overall architecture efficacy by 4.6%.

Conclusions:

• Deep learning models, particularly the Xception U-Net architecture, show significant potential for accurate detection and classification of C-shaped root canal anatomy.
• These AI-driven tools can aid clinicians in identifying complex root canal morphology, improving diagnostic accuracy.
• Further research can explore integrating these DL models into routine clinical workflows for enhanced endodontic diagnosis.