Detection of vertical root fractures by cone-beam computed tomography based on deep learning

Area of Science:

• Dentistry
• Radiology
• Artificial Intelligence

Background:

• Vertical root fractures (VRF) are a common cause of tooth loss.
• Accurate detection of VRF is crucial for appropriate treatment planning.
• Cone-beam Computed Tomography (CBCT) is a valuable imaging modality for diagnosing dental conditions.

Purpose of the Study:

• To evaluate the performance of ResNet models for detecting vertical root fractures (VRF) in CBCT images.
• To compare the efficacy of different ResNet architectures (ResNet-18, ResNet-50, ResNet-101) in VRF detection.
• To assess the impact of incorporating in vitro VRF data on model performance.

Main Methods:

• Convolutional neural network (CNN) models, specifically ResNet architectures, were fine-tuned for VRF detection.
• Two CBCT datasets were used: one from patients (in vivo) and one from an in vitro VRF model.
• Performance metrics including sensitivity, specificity, accuracy, PPV, NPV, and AUC were calculated and compared.

Main Results:

• ResNet-50 achieved the highest Area Under the Curve (AUC) values for both patient data (0.929) and mixed data (0.936).
• The performance of ResNet models on patient data was comparable to that of oral and maxillofacial radiologists.
• Incorporating in vitro VRF data improved the AUC for ResNet-18 and ResNet-50 models.

Conclusions:

• Deep learning models, particularly ResNet-50, exhibit high accuracy in detecting VRF from CBCT images.
• The inclusion of in vitro VRF data enhances the training and performance of deep learning models for VRF detection.
• These findings suggest the potential of AI-powered tools for improving VRF diagnosis.