Mandibular premolar identification system based on a deep learning model.

Area of Science:

• Dental diagnostics
• Machine learning in dentistry
• Forensic odontology

Background:

• Deep learning models are being developed for isolated tooth identification.
• Previous research showed potential but low accuracy in identifying mandibular premolars.
• Improving learning data is crucial for enhancing model performance.

Purpose of the Study:

• To modify learning data for isolated lower premolar identification.
• To increase the accuracy of a deep learning model for tooth identification.
• To enable reliable identification of individual lower premolars.

Main Methods:

• Utilized static occlusal images of premolars from dental plaster casts.
• Employed a convolutional neural network (AlexNet) with stochastic gradient descent.
• Constructed four learning models using modified and unmodified training data.

Main Results:

• Model accuracy improved when training data excluded adjacent teeth.
• The use of static occlusal images with adjacent teeth removed boosted identification accuracy.
• A perfectly accurate identification model was not achieved in this study.

Conclusions:

• Static occlusal images with adjacent teeth deleted are recommended for training and validation.
• Optimizing the ratio of training, validation, and test data is essential.
• Further data refinement is necessary to achieve perfect tooth identification accuracy.