The Improvement of Model Performance for Automated Tooth Identification

Automated tooth identification has been an emerging research topic in the field of dental research in recent years. It has a critical importance for dental health records, forensic dentistry and even age estimation. Instead of conventional methods, modern artificial intelligence techniques, like deep learning, have a significant potential for automatic recognition towards digital dentistry. This work aims to classify the type of teeth which are categorized into four different classes, incisor, canine, premolar and molar, using deep learning models. Cropped images of publicly available Tufts Dental Database, x-ray panoramic radiography image dataset, is used to train and test the models. It proposes modifications on the CNNs (Convolutional Neural Networks), namely ResNet152 and MobileNetV2 to get more robust and accurate results. The effect of neural network depth, convolution kernel size and residual blocks are investigated in the backbone of the network. Model performances are evaluated by confusion matrix, resulting in accuracy, precision and recall. MobileNetV2 yielded the highest accuracy of 84.25%. When the number of convolution layers is increased, it is observed that the classification performance improves more than the other two parameters. Automated tooth identification represents a significant advancement in the fields of dentistry and even forensics, streamlining the process of matching dental records and enhancing accuracy and efficiency in identifying individuals based on their dental characteristics.