Advanced deep learning techniques for classifying dental conditions using panoramic X-ray images.

Area of Science:

• Artificial Intelligence in Dentistry
• Deep Learning for Medical Imaging
• Radiographic Dental Diagnostics

Background:

• Automated detection of dental conditions in panoramic radiographs is crucial for efficient diagnosis.
• Various deep learning models, including custom CNNs, hybrid approaches, and pre-trained architectures, have been explored for this task.
• Evaluating and comparing these diverse AI methodologies is essential to identify optimal solutions.

Purpose of the Study:

• To compare the performance of custom CNNs, hybrid CNN-machine learning models, and fine-tuned pre-trained architectures for detecting dental conditions.
• To assess the efficacy of these models in identifying fillings, cavities, implants, and impacted teeth from panoramic X-rays.
• To determine the most accurate and efficient deep learning approach for automated dental diagnostics.

Main Methods:

• A dataset of 1,512 panoramic X-rays with 11,137 annotations was used for AI-based classification.
• Class imbalance was managed using random downsampling, creating a balanced dataset.
• Five-fold cross-validation was employed to evaluate custom CNNs, hybrid models (CNN + SVM, Decision Tree, Random Forest), and fine-tuned VGG16, Xception, and ResNet50.

Main Results:

• The hybrid CNN-Random Forest model achieved the highest accuracy (85.4%) and macro-F1 score (0.843), outperforming the custom CNN by 11%.
• Among pre-trained architectures, VGG16 showed the best performance (82.3% accuracy, 0.817 macro-F1), followed by Xception and ResNet50.
• The CNN-Random Forest model excelled in fillings detection (F1: 0.860) but systematic misclassifications highlighted diagnostic challenges.

Conclusions:

• Hybrid CNN-based approaches with Random Forest classifiers offer superior discriminative power for detecting dental conditions in annotated regions.
• Computationally efficient hybrid models show potential as supplementary diagnostic tools.
• Observed misclassifications underscore the need for AI systems to augment, not replace, clinical judgment, necessitating prospective validation.