Abstract: Convolutional neural networks (CNNs) are used in many areas of computer vision, such as object tracking and recognition, security, military, and biomedical image analysis. In this work, we describe the current methods, the architectures of deep convolutional neural networks used in CBCT. Literature from 2000-2020 from the PubMed database, Google Scholar, was analyzed. Account has been taken of publications in English that describe architectures of deep convolutional neural networks used in CBCT. The results of the reviewed studies indicate that deep learning methods employed in orthodontics can be far superior in comparison to other high-performing algorithms.
Introduction: The Challenge of CBCT Analysis
Orthodontic treatment planning relies heavily on understanding a patient’s unique craniofacial structure. Since 1998, Cone-Beam Computed Tomography (CBCT) has been the gold standard, providing high-resolution 3D visualizations of teeth, jawbone, and soft tissue. However, analyzing these scans is a significant bottleneck: practitioners must manually identify landmarks, segment individual teeth, and detect pathologies like root resorption or impacted teeth.
This manual process is time-consuming, subjective, and prone to human error. This literature review explores how modern machine learning architectures, specifically Convolutional Neural Networks (CNNs), can automate these tasks to make orthodontic planning more efficient and precise.
Deep Learning Architectures in Orthodontics
The transition from manual analysis to automated systems is driven by the explosion of deep learning. The literature categorizes these methods into two primary functional groups: classification and segmentation.
Classification and Segmentation
Architectures like AlexNet, VGG-16, and Inception-V3 are frequently used to identify tooth types or categorize malocclusions. However, segmentation represents the "heavy lifting" of the field. The paper highlights the transition from standard CNNs to encoder-decoder architectures:
- U-Net and V-Net: These are the most popular choices for volumetric (3D) segmentation of the mandible and individual teeth.
- Mask R-CNN: Employed for "instance segmentation," allowing the AI to identify and outline each individual tooth rather than treating them as a single mass.
- Coarse-to-Fine Approaches: A two-step method where a "coarse" model identifies the general region of interest, followed by a "fine" model for high-precision boundary detection.
Performance and Benchmarking
The quantitative leap in performance over the last five years is striking. Recent models have achieved high levels of accuracy, often rivaling or exceeding expert manual segmentation.
As shown in the literature summary, tooth segmentation models (e.g., Wu et al., 2020) have achieved a Dice Similarity Coefficient (DSC) of 0.962, where 1.0 represents a perfect match with an expert's manual segmentation. Mandible segmentation using U-Net architectures reported DSC scores between 0.917 and 0.94, proving that AI can isolate the jawbone with high reliability.
Key Insights
- The 3D Advantage: While 2D slices are easier to process, 3D (volumetric) models are essential for clinical accuracy. 3D-based architectures like V-Net consistently outperform 2D approaches in handling complex jaw geometry.
- Transfer Learning: Pre-trained models like Inception-V3 can be successfully "fine-tuned" for dental X-rays, even if originally trained on non-medical images.
- Overcoming Artifacts: Newer "mixed-scale" dense CNNs are being designed to ignore metal artifacts (streaking caused by braces or fillings) that typically hinder traditional algorithms.
Challenges and Future Directions
Despite high accuracy scores, several hurdles remain for clinical integration. The "annotation bottleneck" is a major barrier; training these models requires thousands of precisely labeled images, which is labor-intensive for experts. Furthermore, AI performs exceptionally well on standard cases but can struggle with rare genetic abnormalities or severe traumas where data is scarce.
Future work must focus on improving the morphological heterogeneity of training sets to ensure models generalize across diverse populations. The ultimate goal is to integrate these tools directly into surgical planning and "clear aligner" workflows, reducing the time from the initial scan to the start of treatment.
Summary updated