Image Processing for Tooth Type Classification using Deep Learning

Objective: Tooth classification is a crucial aspect of dentistry, influencing effective diagnosis, treatment planning, and overall oral health. However, the subjectivity and variations in human judgment, coupled with the complexity of dental conditions, have led to disparities in tooth classification, particularly in cases involving multiple dentists' opinions, varying clinical expertise, and differing dental standards. Recent advances in technology and artificial intelligence have created new opportunities for innovative solutions in tooth classification. This paper aims to investigate the effect of image processing techniques on classification performance of teeth using deep learning. 4 classes - Incisor, Canine, Premolar, Molar- from panoramic radiographs are prepared. Methods: The state-of-the-art 6 deep learning classification models -Xception, GoogleNet, ResNet18, ShuffleNet, MobileNetV2, ResNext50- was implemented with transfer learning for model efficiency. Two models with the highest and lowest performance were chosen for further analysis related to image processing. 10 different image processing techniques (Gaussian Noise, Gaussian Blur, Wavelet Transform, Sharpness, Contrast Enhancement, Color Correction, Elastic Transform, Random Erasing, Local Binary Patterns, Local Max Min) were applied to these two models. Results: The Xception provided the highest accuracy of 90.25% while ResNet18 yielded the lowest accuracy of 74.86%. Additionally, findings indicated that certain image processing techniques can improve classification performance. Conclusion: The present work shows that image processing can enhance automated artificial intelligence-based solutions for more robust tooth classification, with the potential to improve dental diagnosis and treatment planning. Keywords: Dentistry, classification, deep learning, artificial intelligence, radiography