TL;DR
This paper introduces an AI model, specifically a ResNet-50 deep convolutional neural network, capable of accurately determining the positional relationship between mandibular third molars (M3) and the inferior alveolar nerve (IAN) from panoramic radiographs. The AI model significantly outperforms oral and maxillofacial surgery (OMFS) specialists, particularly in discerning the bucco-lingual position, offering a promising tool to aid in surgical planning and reduce the risk of nerve damage.
What Problem Does This Paper Address?
Extracting mandibular third molars (wisdom teeth) is a common surgical procedure. A critical challenge during these extractions is avoiding damage to the inferior alveolar nerve (IAN), which runs through the jawbone and can cause debilitating sensory impairments if injured. Panoramic radiography is the standard imaging technique for assessing the spatial relationship between the M3 and the IAN. However, interpreting these 2D images, especially when the structures are superimposed, can be difficult and prone to error, even for experienced surgeons. This ambiguity can lead to complications. While cone-beam computed tomography (CBCT) offers higher accuracy, it involves increased radiation exposure and cost. Therefore, there's a need for a more accurate and accessible method to precisely determine the M3-IAN positional relationship from standard panoramic radiographs.
Methods
The researchers developed and evaluated an AI model based on the ResNet-50 architecture, a well-established deep convolutional neural network for image classification. They trained the model to perform two distinct tasks:
- True Contact Position Determination: Identifying whether the M3 and IAN are in direct contact or merely intimate.
- Bucco-Lingual Position Determination: Ascertaining if the IAN is positioned buccally (towards the cheek) or lingually (towards the tongue) relative to the M3.
The dataset was randomly split into 75% for training/validation and 25% for testing. To enhance model robustness and mitigate overfitting, data augmentation techniques such as image rotation, horizontal flipping, and brightness adjustments were employed. The model's performance was evaluated using standard metrics like accuracy, precision, recall, F1 score, and Area Under the ROC Curve (AUC). Crucially, the AI's performance was directly compared against the assessments of six experienced OMFS specialists on the same test dataset.
Datasets
The study utilized a dataset of 571 panoramic images from 394 patients who underwent M3 extraction. These images were retrospectively collected from the Seoul National University Gwanak Dental Hospital. For ground truth labeling, CBCT scans were used to definitively classify the positional relationships. The dataset was categorized as follows:
- True contact: 202 images
- Intimate (non-contact): 246 images
- IAN buccal position: 61 images
- IAN lingual position: 62 images
This dataset, with its specific focus on superimposed M3 and IAN on panoramic radiographs and its ground truth derived from CBCT, serves as a benchmark for evaluating AI performance in this challenging diagnostic scenario.
Key Insights and Ideas
- AI excels where human eyes struggle: The study highlights that AI can overcome the inherent limitations of 2D panoramic radiography for complex spatial relationships, particularly when structures are superimposed.
- Bucco-lingual is AI's strong suit: The AI model demonstrated significantly better performance in determining the bucco-lingual position of the IAN relative to the M3 compared to its ability to detect true contact. This suggests that subtle directional cues are more readily learned by the AI than the precise boundary of contact.
- Specialist variability is high: The performance of OMFS specialists varied considerably, and their agreement (measured by Cohen's kappa) was often poor, especially for the bucco-lingual classification. This underscores the subjective nature of interpreting these radiographs for some aspects.
- Potential for pre-operative triage: The AI's superior accuracy suggests it could be used to identify high-risk cases from panoramic radiographs, potentially guiding which patients would benefit most from a more detailed CBCT scan, thereby optimizing resource allocation and patient exposure.
Key Results
The AI model achieved superior performance compared to OMFS specialists in both tasks:
- True Contact Position:
- AI Accuracy: 72.32%
- OMFS Specialist Accuracy: Ranged from 52.68% to 69.64%
- AI Cohen's Kappa: 0.41 (Moderate agreement)
- OMFS Specialist Cohen's Kappa: Ranged from Slight (0.05) to Fair (0.35) agreement.
- Bucco-Lingual Position:
- AI Accuracy: 80.65%
- OMFS Specialist Accuracy: Ranged from 32.26% to 48.39%
- AI Cohen's Kappa: 0.61 (Substantial agreement)
- OMFS Specialist Cohen's Kappa: Ranged from Poor (-0.15) to Poor (0.35) agreement.
The AI model's performance was notably better in the bucco-lingual determination (Experiment 2) with an average accuracy of 0.76, precision of 0.83, recall of 0.67, and F1 score of 0.73, compared to the true contact determination (Experiment 1) with an average accuracy of 0.63, precision of 0.62, recall of 0.63, and F1 score of 0.61.
Limitations and Future Work
The authors acknowledge several limitations:
- Small Dataset Size: The absolute number of images was relatively small, necessitating data augmentation. While effective, this can still lead to variability in model performance depending on the train-test split.
- Cropped Images: The images were cropped without specific domain knowledge about optimal size or resolution, which might have impacted performance.
- Lack of Interpretability: The study did not employ interpretable AI models (XAI), making it difficult to understand why the AI makes certain predictions and to identify specific weaknesses.
Future work could involve:
- Larger, Multi-center Datasets: Training with more extensive data from various institutions to improve model generalization and robustness.
- Incorporating Domain Knowledge: Using prior knowledge to construct datasets and potentially guide model architecture.
- Explainable AI (XAI): Implementing XAI techniques to understand model decision-making and identify areas for improvement.
- Ensemble Learning and Advanced Techniques: Exploring methods like ensemble learning, self-supervised learning, and contrastive learning to further boost performance, especially with limited data.
- Prospective Clinical Testing: Validating the model's performance in real-world clinical settings to assess its practical utility.
Summary updated