[Paper Review] Dental pathology detection in 3D cone-beam CT

Adel Zakirov, Matvey Ezhov+3 more|arXiv (Cornell University)|Oct 24, 2018

Dental Radiography and Imaging18 references4 citations

TL;DR

This paper presents a deep learning pipeline for automated detection of dental pathologies in 3D cone-beam CT (CBCT) scans, using a two-stage approach: first localizing teeth via a 3D U-Net variant (V-Net), then classifying six common conditions—fillings, crowns, implants, filled canals, impacted teeth, and missing teeth—using a 3D DenseNet. The system achieves 96.3% accuracy in tooth localization and an average AUROC of 0.94 across all pathology classes.

ABSTRACT

Cone-beam computed tomography (CBCT) is a valuable imaging method in dental diagnostics that provides information not available in traditional 2D imaging. However, interpretation of CBCT images is a time-consuming process that requires a physician to work with complicated software. In this work we propose an automated pipeline composed of several deep convolutional neural networks and algorithmic heuristics. Our task is two-fold: a) find locations of each present tooth inside a 3D image volume, and b) detect several common tooth conditions in each tooth. The proposed system achieves 96.3\% accuracy in tooth localization and an average of 0.94 AUROC for 6 common tooth conditions.

Motivation & Objective

• Address the time-consuming and complex interpretation of 3D cone-beam CT (CBCT) images in clinical dentistry.
• Develop an automated pipeline to detect common dental pathologies in 3D CBCT volumes, reducing reliance on manual interpretation.
• Improve diagnostic efficiency by enabling rapid, accurate localization of teeth and classification of pathological conditions.
• Provide a foundation for clinical decision support systems in dental imaging using deep learning.

Proposed method

• Train a 3D U-Net-based V-Net model for tooth localization, using sparse 2D bounding box annotations across axial slices and interpolating them into 3D volumetric masks.
• Apply a multi-view projection strategy (axial, frontal, sagittal) to identify the optimal region for tooth extraction.
• Extract a 3D sub-volume around each localized tooth, rescale it to 64³, and pass it to a 3D DenseNet classifier.
• Use a 3D DenseNet with 4 dense blocks of 6 layers each, a growth rate of 48, and compression factor 0.5, replacing 2D convolutions with 3D convolutions.
• Train the classification model with binary cross-entropy loss, weighted inversely by class frequency to address class imbalance.
• Optimize the localization pipeline by extending bounding boxes by 15 mm vertically and 8 mm horizontally to capture tooth context and improve robustness.

Experimental results

Research questions

• RQ1 Can a deep learning pipeline accurately localize all teeth in 3D CBCT volumes using sparse 2D annotations and 3D interpolation?
• RQ2 Can a 3D convolutional neural network achieve high performance in classifying multiple, non-mutually exclusive dental pathologies from localized tooth volumes?
• RQ3 How effective is weighted binary cross-entropy loss in mitigating class imbalance for rare dental conditions like implants and impacted teeth?
• RQ4 To what extent can the localization model support downstream tasks such as 2D panoramic reconstruction for clinical usability?

Key findings

• The V-Net-based localization model achieved a 96.3% accuracy in identifying the correct location of each tooth in 3D CBCT volumes.
• The 3D DenseNet classifier achieved an average AUROC of 0.94 across six dental pathology classes, with individual AUROC scores ranging from 0.892 (fillings) to 0.979 (implants).
• The most frequent condition in the dataset was fillings (21.5% of positive cases), while the rarest were implants and impacted teeth (1.5% and 1.8%, respectively).
• The classification model demonstrated strong performance on rare pathologies, with AUROC of 0.931 for impacted teeth and 0.979 for implants.
• The localization pipeline enables high-quality 2D panoramic reconstruction, offering a clinically familiar interface for dentists to review 3D CBCT data.
• The system’s performance, particularly in localization, is sufficient to support clinical integration as a preprocessing or decision-support tool.