Human Tooth Crack Image Analysis with Multiple Deep Learning Approaches.

Abstract

Tooth cracks, one of the most common dental diseases, can result in the tooth falling apart without prompt treatment; dentists also have difficulty locating cracks, even with X-ray imaging. Indocyanine green (ICG) assisted near-infrared fluorescence (NIRF) dental imaging technique can solve this problem due to the deep penetration of NIR light and the excellent fluorescence characteristics of ICG. This study extracted 593 human cracked tooth images and 601 non-cracked tooth images from NIR imaging videos. Multiple imaging analysis methods such as classification, object detection, and super-resolution were applied to the dataset for cracked image analysis. Our results showed that machine learning methods could help analyze tooth crack efficiently: the tooth images with cracks and without cracks could be well classified with the pre-trained residual network and squeezenet1_1 models, with a classification accuracy of 88.2% and 94.25%, respectively; the single shot multi-box detector (SSD) was able to recognize cracks, even if the input image was at a different size from the original cracked image; the super-resolution (SR) model, SR-generative adversarial network demonstrated enhanced resolution of crack images using high-resolution concrete crack images as the training dataset. Overall, deep learning model-assisted human crack analysis improves crack identification; the combination of our NIR dental imaging system and deep learning models has the potential to assist dentists in crack diagnosis.