Deep Learning-Based Detection of Root Numbers in Maxillary Premolars

Area of Science:

• Dentistry
• Radiology
• Artificial Intelligence

Background:

• Determining the root number of maxillary premolars is crucial for endodontic diagnosis and treatment planning.
• Panoramic radiography is a common imaging modality, but root number detection can be challenging.
• Deep learning offers potential for automated analysis of radiographic images.

Purpose of the Study:

• To evaluate the efficacy of deep learning models in detecting the root number of maxillary premolars using panoramic radiographs.
• To compare the performance of different convolutional neural network (CNN) architectures.

Main Methods:

• A retrospective study utilized 925 maxillary premolars from 350 patients, with CBCT scans as the reference standard.
• Three CNN models (AlexNet, DenseNet-121, EfficientNet-B0) were trained using transfer learning on preprocessed panoramic images.
• Data augmentation, five-fold cross-validation, and an independent external validation set were employed to assess model performance and generalizability.

Main Results:

• The ensemble deep learning model achieved the highest accuracy (0.90) and AUC (0.94) in cross-validation.
• On external validation, the ensemble model also performed best (accuracy 0.87), outperforming an expert clinician (accuracy 0.82).
• The ensemble model demonstrated reduced variability and narrower confidence intervals, indicating robust performance.

Conclusions:

• Deep learning models, particularly the ensemble approach, show reliable performance in identifying maxillary premolar root numbers from panoramic radiographs.
• These AI systems hold promise as supportive tools to aid clinicians in decision-making.
• Further research can explore integration into routine dental practice.