Evaluation of mandibular impacted third molar teeth using panoramic radiographs with artificial intelligence

Artificial intelligence technologies have started to play significant roles in our lives with an increasing impact within the last few years. In this regard, new developments are also experienced rapidly in the fields of dentistry and maxillofacial radiology. In our thesis, it was aimed to automatically classify the mandibular third molar teeth, which is the most frequently impacted tooth, according to the Pell&Gregory and Winter classifications, with a model based on artificial intelligence and deep learning. Labeling of teeth on images was done using CranioCatch labeling software (CranioCatch, Eskişehir, Turkey). Separate models were developed for the assessment of position characteristics, impacted tooth detection and numbering. To evaluate the position characteristics; A total of 10 data set classes were created, 6 for the Pell & Gregory classification and 4 for the Winter classification, from the data of teeth 38 and 48, which were labeled on 3098 panoramic images. All available labeled data were used for impacted tooth detection and numbering studies. The achievements of the Inception v3 and Mask R CNN models developed for these tasks were calculated using the confusion matrix. In the Pell & Gregory classification of the Inception v3 model, the sensitivity, precision and F1 score values for Class I, consecutively; 0.7, 0.3684, 0.4827 for Class II; 0.7474, 0.9367, 0.8314 for Class III; 0.875, 0.6, 0.7118 for Position A; 0.8604, 0.7254, 0.7872 for Position B; 0.6267, 0.8245, 0.7121 for Position C; 0.7142, 0.4167, 0.5263 was found. Similarly, in the Winter classification of the Inception v3 model, the sensitivity, precision and F1 score values for the Vertical position are consecutively; 0.8333, 1, 0.9090, for Horizontal position; 0.7778, 0.7778, 0.7778 for Mesioangular position; 0.778, 0.7368, 0.7567 for Distoangular position; 0.944, 0.85, 0.8947 was found. The sensitivity, precision and F1 score values of the Mask R CNN model are consecutively; 0.9443, 0.9974, 0.9701 in impacted tooth numbering 0.9381, 1, 0.9680 was found. Deep learning models have been developed based on the results of our study; their performances in classifying impacted mandibular third molars according to their positions, identifying and numbering impacted teeth seem promising. Thanks to increasing number of the studies in this field, artificial intelligence applications based on deep learning can play a role in clinical practice and support physicians.