TL;DR
This paper introduces a novel deep learning model, ERF-Swin-OC, that automates the classification of dental occlusion from intraoral photographs. Achieving high accuracy in identifying molar, canine, and anterior overbite relationships, this model offers a promising tool to enhance efficiency and objectivity in orthodontic diagnosis and treatment planning.
What Problem Does This Paper Address?
Malocclusion, a common dental issue affecting oral health and aesthetics, has traditionally been diagnosed using physical exams or plaster casts, which are time-consuming and resource-intensive. While intraoral photography has become a standard for documentation, interpreting these images for occlusal classification still requires significant clinical expertise. Existing AI approaches have shown success in other dental imaging tasks but have been underexplored for occlusal classification from intraoral photos. This gap limits the potential for efficient, objective, and scalable orthodontic diagnosis and treatment planning.
Methods
The researchers developed a deep learning model called the Edge-Region Fusion Swin Transformer for Occlusion Relationship Classification (ERF-Swin-OC). This model utilizes a dual-branch network architecture based on the Swin Transformer.
- Tooth Segmentation: The process begins with a YOLOv11 network to perform instance segmentation of upper and lower teeth, extracting region information.
- Dual-Branch Processing:
- Region Branch (RSTB): This branch embeds the segmented tooth regions into the original image, using an offset operation to align embedded information with the original image. This helps the model understand the relative positioning of teeth and their occlusal relationships.
- Edge Branch (ESTB): This branch fuses the contour masks of the teeth with the original image, ensuring precise spatial alignment and enhancing the model's ability to perceive fine edge features.
- Fusion and Classification: Feature maps from both branches are fused and then passed to a classification head for the final occlusion classification.
- Data Augmentation: To improve robustness and generalization, techniques like random rotation (±10 degrees), color jittering, and Gaussian blur were applied to the training data.
- Loss Function: A weighted cross-entropy loss function was employed to address class imbalance, giving more emphasis to minority classes.
Datasets
The study utilized a large-scale dataset of intraoral photographs:
- Total Images: 6,200 intraoral photographs from 6,200 patients.
- Image Types:
- 5,000 45° lateral occlusal photographs (used for molar and canine classification).
- 2,200 90° lateral occlusal photographs (used for anterior overbite classification).
- Data Splitting: The dataset was split into a training and validation set (4,880 images) and a testing set (1,220 images) with an 8:2 ratio.
- Annotation: Images were classified by two senior dentists with over 10 years of experience, with a third senior dentist available for consultation. A Cohen's κ > 0.9 was achieved for inter-rater reliability.
- Classification Categories:
- Molar: Class I (M1), Class II (M2), Class III (M3).
- Canine: Class I (C1), Class II (C2), Class III (C3).
- Anterior Overbite: Normal overbite, deep overbite, edge-to-edge bite, open bite, anterior crossbite, single-tooth crossbite or segmental crossbite.
Key Insights and Ideas
- Fusion of Region and Edge Information: The ERF-Swin-OC model's success hinges on its ability to effectively combine both the spatial location of teeth (region) and their precise outlines (edge) for classification. This dual-branch approach appears to be more powerful than using either alone.
- Addressing Class Imbalance: The use of a weighted cross-entropy loss function was crucial for improving performance on less frequent occlusal subclasses, demonstrating the importance of tailored loss functions in medical imaging.
- Swin Transformer's Efficacy: The study highlights the Swin Transformer's capability in handling complex visual data like intraoral images, outperforming other established architectures like Vision Transformer and ConvNeXt in this specific task.
- Potential for Scalability: By automating classification, the model has the potential to significantly reduce the workload on clinicians and enable faster, more consistent diagnoses, especially in high-volume settings.
Key Results
The ERF-Swin-OC model demonstrated strong performance across all classification tasks:
- Molar Occlusal Relationships: Achieved a weighted average F1-score of 0.90. AUC values were 0.97 (M1), 0.97 (M2), and 0.99 (M3).
- Canine Occlusal Relationships: Achieved a weighted average F1-score of 0.87. AUC values were 0.96 (C1), 0.97 (C2), and 0.98 (C3).
- Anterior Overbite Relationships: Achieved a weighted average F1-score of 0.89. Subclass F1-scores ranged from 0.77 (normal overbite) to 0.94 (deep overbite).
- Comparison to Other Models: The ERF-Swin-OC model outperformed Vision Transformer, Swin Transformer, and ConvNeXt models in weighted average F1-scores for molar and canine classifications.
- Prediction Speed: The model achieved an average prediction time of approximately 15.4 ms per image on an RTX 4090 GPU.
Limitations and Future Work
- Subclass Granularity: The study acknowledges that the relatively small sample size for rarer subclasses (e.g., certain types of crossbites) limited the ability to subdivide occlusal relationships further, potentially affecting model training reliability for these specific cases.
- Image Quality Factors: Misclassifications were attributed to factors like camera angle distortion, incomplete tooth morphology, tooth stains, and variations in tooth position. This highlights the need for improved imaging techniques and preprocessing.
- Future Directions:
- Dataset Expansion: Increasing the dataset size, particularly with images captured via smartphones, is crucial for enhancing accuracy and model generalization.
- Smartphone Integration: The long-term goal is to leverage smartphone technology for easier image capture, enabling wider accessibility and public health awareness.
- 3D Integration: Exploring the incorporation of three-dimensional measurements from intraoral scans could further improve accuracy.
- Clinical Validation: Prospective clinical trials and long-term monitoring are needed to ensure reliability in practical settings.
- Complementary Tool: The model is intended to complement, not replace, traditional clinical examinations and expert judgment.
Summary updated