TL;DR
This paper introduces a deep learning model capable of diagnosing dens evaginatus (DE) from periapical radiographs with high accuracy, outperforming human specialists. This AI tool could significantly aid in early detection and management of DE, potentially improving long-term tooth prognosis, especially given the rarity of the condition and limited available data.
What Problem Does This Paper Address?
Dens evaginatus (DE) is a rare developmental anomaly where an extra cusp or tubercle forms on a tooth. This anomaly can lead to complications like fractures and pulp issues, making early diagnosis crucial. However, the rarity of DE poses a significant challenge for collecting sufficient data to train robust diagnostic models. Furthermore, as the extra cusp wears down over time, distinguishing DE from normal anatomy on periapical radiographs (PA) becomes increasingly difficult, even for experienced dentists. This study aims to bridge this gap by developing and evaluating an AI model for DE diagnosis on PAs, specifically addressing the challenge of limited data availability and comparing its performance against human experts.
Methods
The researchers employed deep learning techniques, specifically focusing on a pre-trained ResNet50 model due to the limited dataset size. They explored two approaches for image handling: using full periapical radiographs (F model) and using cropped images focusing on the tooth of interest (C model). To enhance model robustness with limited data, they utilized data augmentation techniques like rotation, flipping, and brightness adjustment. The models were trained and evaluated using a five-fold cross-validation strategy to mitigate overfitting. To understand the AI's decision-making process, a class activation map was used to visualize the regions of the image that influenced the AI's diagnosis. The performance of the AI model was then rigorously compared against the diagnoses of fourteen endodontic specialists.
Datasets
The study utilized a dataset of 402 periapical radiographs (PA) from patients at Wonkwang University Daejeon Dental Hospital. This dataset comprised 138 images showing dens evaginatus (DE) and 264 normal cases. The DE cases primarily involved mandibular first and second premolars, with a slight female predominance. The normal cases also predominantly featured mandibular second premolars. The dataset was split into training and testing sets (8:2 ratio for initial experiments, and five-fold cross-validation for the final evaluation).
Key Insights and Ideas
- AI's Superiority in Rare Conditions: The study demonstrates that AI can achieve superior diagnostic performance compared to human specialists, even for rare conditions like DE where data is scarce. This highlights AI's potential to overcome human limitations in recognizing subtle or infrequent anomalies.
- Image Cropping Not Always Beneficial for AI: Counterintuitively, the AI model's performance was not significantly different between using full PA images and cropped images. This suggests that the AI can effectively learn from the broader context of the full image, or that the cropping process didn't isolate the critical features as effectively as anticipated for the AI.
- AI Focuses on Coronal Pulp: The class activation map revealed that the AI model primarily focused on the coronal pulp area of the tooth when diagnosing DE. This is significant because the pulp extension within the abnormal cusp is a key indicator that can be visualized on PAs, even when the external morphology is worn.
- Substantial Agreement for AI: The AI model exhibited "substantial" agreement with itself (across different folds of cross-validation) and "fair" agreement with specialists, indicating a high level of internal consistency and a strong diagnostic capability.
Key Results
- The pre-trained ResNet50 model achieved the highest AUC of 0.878 among the initial DL models tested.
- The AI model, using five-fold cross-validation, achieved an average AUC of 0.895 for the F model (full images) and 0.901 for the C model (cropped images).
- There was no statistically significant difference in AUC between the AI's performance on full versus cropped images (p = 0.753).
- In contrast, specialists showed a statistically significant difference in AUC between full and cropped images (p = 0.040), with cropped images yielding a slightly higher average AUC.
- The AI model consistently outperformed endodontists across all performance metrics (accuracy, precision, recall, F1 score, and AUC). For instance, the AI's AUC was significantly higher than the average specialist AUC.
- Cohen's kappa scores indicated "substantial" agreement for the AI model (0.774 for F model, 0.684 for C model) and "fair" agreement for specialists (0.238 for F model, 0.359 for C model).
Limitations and Future Work
- Limited Dataset Size and Imbalance: The study acknowledges the small and imbalanced dataset due to the rarity of DE. While data augmentation and AUC as an evaluation metric were used, a larger, multicenter dataset would further enhance generalizability.
- Single Institution Data: The data was collected from a single institution, which might limit the model's applicability to different radiographic equipment and patient populations. Future work should involve multicenter studies.
- Inclusion of Periapical Lesions: The inclusion criteria for DE cases included periapical lesions. While this targeted more severe cases, it raises the possibility that the AI's determination might have been influenced by the presence of these lesions. Future research could investigate DE diagnosis without periapical pathology.
- Worn-Down DE: The current model might not be as effective for diagnosing worn-down DE, as the inclusion criteria focused on observable abnormal cusps. Developing models for panoramic radiography or those specifically trained on worn-down DE is suggested.
- AI as a Complementary Tool: The authors emphasize that AI should be viewed as a tool to support, not replace, clinical expertise. Future work should focus on integrating AI into clinical workflows and exploring its use with other imaging modalities or clinical parameters.
Summary updated