IADR Abstract Archives
Data Augmentation Effect on the AI-Assisted Dental Disease Detection Model
Objectives: Artificial intelligence (AI) is a highly promising application to medicine and dentistry as a practical supplement tool. For both patient and clinician, discovering diseases at the incipient as well as the asymptomatic stage would be highly desired because it may reduce the amount of discomfort, time, and cost of the treatment. To mitigate these issues, this study developed the AI-assisted dental disease detection model. We utilized data augmentation techniques to improve the performance of the AI model.
Methods: We developed the AI model based on a convolutional neural network with 7 annotated classes: enamel caries, dentinal caries, impacted tooth, retained root tip, periapical radiolucency, widened PDL, and endo-perio lesions. The data augmentation was performed with an initial 399 periapical and bitewing radiographs. It generated a total of 1197 images by flipping horizontally and diagonally as well as adopting equalizing histogram techniques. The dataset was split into three subsets: 70% training, 15% testing, and 15% validation sets in model building. The performance of our AI model was measured with mean average precision (mAP50), and the comparison of the mAP50 values before and after data augmentation was obtained.
Results: Results showed that the mAP50 of the original dataset was 0.445 while that of the augmented dataset presented 0.826, i.e., an overall 85.6% performance improvement.
Conclusions: The model improvement resulting from data augmentation demonstrates the applicability and practicability of the AI model in dental clinics. Moreover, it may enhance dentists’ work efficiency as well as reduces the chance of missing clinical findings, misdiagnosis, and extra radiation exposure.
Methods: We developed the AI model based on a convolutional neural network with 7 annotated classes: enamel caries, dentinal caries, impacted tooth, retained root tip, periapical radiolucency, widened PDL, and endo-perio lesions. The data augmentation was performed with an initial 399 periapical and bitewing radiographs. It generated a total of 1197 images by flipping horizontally and diagonally as well as adopting equalizing histogram techniques. The dataset was split into three subsets: 70% training, 15% testing, and 15% validation sets in model building. The performance of our AI model was measured with mean average precision (mAP50), and the comparison of the mAP50 values before and after data augmentation was obtained.
Results: Results showed that the mAP50 of the original dataset was 0.445 while that of the augmented dataset presented 0.826, i.e., an overall 85.6% performance improvement.
Conclusions: The model improvement resulting from data augmentation demonstrates the applicability and practicability of the AI model in dental clinics. Moreover, it may enhance dentists’ work efficiency as well as reduces the chance of missing clinical findings, misdiagnosis, and extra radiation exposure.