Artificial intelligent-driven decision-making for automating root fracture detection in periapical radiographs

Accurate and early detection of root fractures is crucial for effective treatment and prognosis. While periapical radiographs are the standard diagnostic tool, their interpretation can be subjective and prone to errors, especially among less experienced dentists. Three-dimensional imaging (CBCT) offers superior visualization but is limited by cost, radiation exposure, and image artifacts. Artificial intelligence (AI), particularly Convolutional Neural Networks (CNNs), has shown promise in analyzing medical images and assisting in diagnosis. This research explores the application of AI to improve the accuracy and efficiency of root fracture detection in periapical radiographs, aiming to reduce human error and improve patient care. The challenge lies in the subtle visual differences between fractured and unfractured roots on 2D radiographs, which often leads to misdiagnosis even by experienced clinicians. Integrating AI into dental diagnostics can provide a rapid and objective assessment, potentially leading to improved treatment outcomes and reduced diagnostic uncertainties for both experienced and less experienced practitioners.

Previous studies have investigated the use of AI, specifically neural networks, for root fracture detection. Kositbowornchai et al. validated a probabilistic neural network for detecting vertical root fractures in periapical radiographs, demonstrating its potential. Johari et al. compared neural network performance on CBCT and periapical radiographs, finding better results with CBCT. Hu et al. used AI algorithms (ResNet50, DenseNet169, VGG19) on CBCT images, showing ResNet50's superiority over other models and even outperforming experts. However, these studies often focused on specific AI models or used different imaging modalities. This study builds upon previous work by testing multiple state-of-the-art AI models and introducing a voting mechanism to improve the overall accuracy and robustness of fracture detection in 2D periapical radiographs.

Four hundred single-rooted anterior teeth (200 fractured, 200 intact) were used. Root fractures were artificially created horizontally. Digital periapical radiographs were taken and processed. Five pretrained CNN models (VGG16, VGG19, ResNet50, DenseNet121, DenseNet169) were used. Data were resized to 224x224x3 pixels. Training parameters included 50 epochs, 0.3 dropout, early stopping with patience of 15 epochs, and a learning rate reduction strategy. The Adam optimizer was used. A voting system aggregated the predictions of the five models. Model performance was evaluated using training/validation loss curves, sensitivity, specificity, positive predictive value (PPV), and the area under the receiver operating characteristic (ROC) curve (AUC). The ethical approval was obtained from the Institutional Review Board of Misr University for Science and Technology (MUST IRB) under approval number 2022/0097. Matplotlib was used for graphical representation of results.

VGG16 demonstrated excellent performance with low training and validation losses (0.09% and 0.18%, respectively), high specificity (93.6%), sensitivity (89.3%), and PPV (93.3% for fractured cases and 89.7% for unfractured cases). VGG19 showed signs of overfitting. ResNet50 exhibited bias toward unfractured cases. DenseNet121 effectively addressed overfitting and noise, achieving balanced metrics and high PPVs (0.933 for fractured and 0.898 for unfractured). DenseNet169, with increased depth, showed enhanced generalization. The voting system integrated the results from the five models to arrive at a final classification. The ROC curve analysis showed high AUC values for VGG16, DenseNet121, and DenseNet169 (0.99, 0.98, and 0.98 respectively), indicating their strong discriminative ability. The results are summarized in Table 1 of the paper.

The AI-based system using a voting mechanism demonstrated high accuracy in detecting root fractures in periapical radiographs. This approach has the potential to improve diagnostic accuracy and efficiency for both experienced and inexperienced dentists. The superior performance of VGG16, DenseNet121, and DenseNet169 highlights the potential of these CNN architectures for this specific task. The voting system mitigates the limitations of individual models, leading to a more robust and reliable diagnosis. The results are generally consistent with previous studies on AI-based fracture detection, although there are differences related to the type of fracture (horizontal vs. vertical), imaging modality (2D vs. 3D), and AI models used. The findings support the integration of AI in dental diagnostics to improve the accuracy and efficiency of root fracture detection.

This study demonstrates the feasibility and potential benefits of using an AI-based voting system for detecting root fractures in periapical radiographs. The high sensitivity and specificity achieved, particularly with VGG16, DenseNet121, and DenseNet169 models, indicate that this technology can significantly aid dentists in their diagnosis. Future research should focus on validating this approach in a clinical setting using a larger, more diverse dataset including multi-rooted teeth and different types of fractures. Further development could involve integrating this system into existing dental software for real-time diagnostic assistance.

This study used artificially created fractures on extracted teeth, which may not perfectly replicate the complexity and variability of fractures in clinical settings. The relatively small sample size could limit the generalizability of the findings. The study's focus on single-rooted anterior teeth limits the applicability to other tooth types. Future studies should address these limitations by using a larger dataset of clinical radiographs encompassing a wider range of tooth types and fracture characteristics.