Deep learning with convolution neural network detecting mesiodens on panoramic radiographs: comparing four models.

Area of Science:

• Dentistry
• Radiology
• Artificial Intelligence

Background:

• Mesiodens, supernumerary teeth, can cause complications.
• Accurate detection on panoramic radiographs is essential for timely intervention.
• Current diagnostic methods may face challenges with unclear images.

Purpose of the Study:

• To develop an optimal, simple, and lightweight deep learning convolutional neural network (CNN) model.
• To evaluate the diagnostic performance of the developed CNN models for mesiodens detection.
• To utilize SHapley Additive exPlanations (SHAP) for model interpretability.

Main Methods:

• Trained and validated four modified CNN models using 628 panoramic radiographs.
• Evaluated model performance using accuracy, precision, recall, F1 scores, ROC curves, and AUC.
• Employed SHAP to visualize image features critical for classification.

Main Results:

• A binary_connect_mnist_LeNet model demonstrated the best diagnostic performance among the four deep learning models.
• The lightweight CNN model successfully detected mesiodens.
• SHAP analysis provided insights into the image features influencing the model's classifications.

Conclusions:

• A simple, lightweight deep learning model is capable of detecting mesiodens.
• AI-based diagnosis can be a valuable adjunct for unclear panoramic radiographs.
• Specialist re-evaluation is necessary due to the radiosensitivity of children.