Caries detection with tooth surface segmentation on intraoral photographic images using deep learning

Abstract

Background Intraoral photographic images are helpful in the clinical diagnosis of caries. Moreover, the application of artificial intelligence to these images has been attempted consistently. This study aimed to evaluate a deep learning algorithm for caries detection through the segmentation of the tooth surface using these images. Methods In this prospective study, 2348 in-house intraoral photographic images were collected from 445 participants using a professional intraoral camera at a dental clinic in a university medical centre from October 2020 to December 2021. Images were randomly assigned to training (1638), validation (410), and test (300) datasets. For image segmentation of the tooth surface, classification, and localisation of caries, convolutional neural networks (CNN), namely U-Net, ResNet-18, and Faster R-CNN, were applied. Results For the classification algorithm for caries images, the accuracy and area under the receiver operating characteristic curve were improved to 0.813 and 0.837 from 0.758 to 0.731, respectively, through segmentation of the tooth surface using CNN. Localisation algorithm for carious lesions after segmentation of the tooth area also showed improved performance. For example, sensitivity and average precision improved from 0.890 to 0.889 to 0.865 and 0.868, respectively. Conclusion The deep learning model with segmentation of the tooth surface is promising for caries detection on photographic images from an intraoral camera. This may be an aided diagnostic method for caries with the advantages of being time and cost-saving.

Deep-learning-based automated detection and assistance in the diagnosis of early dental carious lesions

“…Among the available systems, the ICDAS, used in the present study, proves to be more e cient in the detection of early carious lesions because it includes several clinical aspects, from the very rst signs, and also because its diagnostic accuracy is akin to that of the histological analysis for detection of lesions on occlusal surfaces [22,23]. Previous studies that used dental images for the detection of dental carious lesions have not described the different stages of carious lesions [36, [39][40][41][42] and they have not assessed the effectiveness of the network in the distinction between early carious lesions and opacities caused by developmental defects of enamel [35].…”

“…CNNs are trainable algorithms capable of automating the detection of patterns and classifying changes by extracting data, such as shape, illumination, and color distribution, from images [30]. A large number of dental images can be utilized for CNN training, including radiographs [31,32] and infrared transillumination images [33,34], but intraoral photographs, regarded as the best diagnostic method for the detection of early carious lesions, closely resemble visual examination [35][36][37][38][39][40][41][42].…”

AI-Dentify: deep learning for proximal caries detection on bitewing x-ray - HUNT4 Oral Health Study

“…Data availability and reliable annotations [ 8 , 13 ] are the main bottlenecks in the development of machine learning (ML) methods in dentistry. A large portion of the published work uses a dataset of fewer than 300 images, only few studies have access to large datasets [ 8 ] with more than 1,000 images like [ 11 , 14 , 15 ]. Of these publications, the work presented in [ 4 – 6 , 11 , 15 ] focus on object detection, which is the scope of the present study.…”

“…A large portion of the published work uses a dataset of fewer than 300 images, only few studies have access to large datasets [ 8 ] with more than 1,000 images like [ 11 , 14 , 15 ]. Of these publications, the work presented in [ 4 – 6 , 11 , 15 ] focus on object detection, which is the scope of the present study. Object detection or object recognition refers to the task of localising and classifying objects in a picture [ 16 ].…”

“…Park et al. [ 15 ] proposed an ensemble of U-Net and Fast R-CNN for caries detection in colour image, trained with 2,348 RGB intraoral photographic images. Even though, the work done by Cantu et al.…”