Deep Learning Models for Evaluating the Anatomical Relationship Between Posterior Maxillary Teeth and Maxillary Sinus...

Area of Science:

• Radiology and Artificial Intelligence
• Dental Imaging Analysis
• Deep Learning in Healthcare

Background:

• Accurate assessment of the posterior maxillary teeth-sinus relationship is vital for dental procedures like orthodontics, surgery, and implantology.
• Panoramic radiography is a common imaging modality, but precise anatomical evaluation can be challenging.

Purpose of the Study:

• To evaluate the effectiveness of deep learning models in predicting the anatomical relationship between posterior maxillary teeth and the maxillary sinus using panoramic images.
• To compare the performance of different convolutional neural network architectures (VGG, ResNet, ResNeXt) for this task.

Main Methods:

• 300 panoramic images were processed into 1760 cropped slices (512x512 pixels).
• Three convolutional neural network models (VGG, ResNet, ResNeXt) were trained on 80% of the data, validated on 10%, and tested on 10%.
• Performance was assessed using accuracy, precision, recall, F1 score, and ROC-AUC.

Main Results:

• ResNet and ResNeXt models achieved superior performance with metrics of 0.88 and ROC-AUC values of 0.93-0.94.
• The VGG model achieved an accuracy of 0.84 and an ROC-AUC of 0.89.
• Overfitting was observed in ResNet and ResNeXt after 30-50 epochs; false positives were noted for second molars near the sinus.

Conclusions:

• Deep learning models, specifically ResNet and ResNeXt, reliably assess the maxillary teeth-sinus anatomical relationship from panoramic radiographs.
• These AI models can function as valuable diagnostic tools, enhancing clinical decision-making, especially when cone-beam CT (CBCT) is not accessible.