(12) Teeth identification and diagnosis of radiographic furcation involvement using machine learning.

Introduction: Convolutional neural networks (CNNs) are a subset of artificial intelligence that have been used in various fields of healthcare in recent years to develop deep machine learning models that automate pattern recognition, object localization, and object identification. Dentists can use such tools to assist with analyzing dental radiographic images to improve clinical efficiency, diagnostic accuracy, and training regimens. The objective of this study is to develop two deep machine learning models for analysis of full-mouth radiographs to (1) classify tooth type and number, and (2) identify furcation involvement.

Methods: The study consists of standardized full-mouth radiographs from 300 UTHealth School of Dentistry patients of record (HSC-DB-19-0994). Two calibrated periodontists classified the included radiographs based on the universal tooth numbering system and presence of radiographic furcation involvement or not.
The radiographs were then segmented and augmented into 6,298 standardized periapical images. 80% of the collected data was randomly distributed to train the models by five-fold cross validation and 20% was used to test the models. The Multi-ResNet 50 model was created for tooth type classification and numbering, and the U-net model was created for identification of proximal and buccal/lingual radiographic furcation involvement. Test accuracy, sensitivity, specificity, and positive and negative predictive values were calculated based on classification from calibrated periodontists.

Results: Multi-ResNet 50 successfully classified tooth types, including molars, premolars, canines, and incisors with respective accuracies of 99%, 98%, 95%, 99%. Mean teeth numbering accuracy was 94.9% (Figure 1).
U-net identified proximal furcation involvement with a sensitivity of 98% and specificity of 56% (Table 1) and buccal/lingual furcation involvement with a sensitivity of 91% and specificity of 82% (Table 2).

Conclusions: Machine learning on tooth type recognition and detection of furcation involvement showed potential in reducing clinical workload and benefiting dental education.

Lead Author's Contributions to the Research Project. : Lead author was responsible for the data collection, composition of the abstract.