Automated Teeth Extraction and Dental Caries Detection in Panoramic X-ray
MetadataShow full item record
Dental caries is one of the most chronic diseases that involves the majority of people at least once during their lifetime. This expensive disease accounts for 5-10% of the healthcare budget in developing countries. Caries lesions appear as the result of dental biofi lm metabolic activity, caused by bacteria (most prominently Streptococcus mutans) feeding on uncleaned sugars and starches in oral cavity. Also known as tooth decay, they are primarily diagnosed by general dentists solely based on clinical assessments. Since in many cases dental problems cannot be detected with simple observations, dental x-ray imaging is introduced as a standard tool for domain experts, i.e. dentists and radiologists, to distinguish dental diseases, such as proximal caries. Among different dental radiography methods, Panoramic or Orthopantomogram (OPG) images are commonly performed as the initial step toward assessment. OPG images are captured with a small dose of radiation and can depict the entire patient dentition in a single image. Dental caries can sometimes be hard to identify by general dentists relying only on their visual inspection using dental radiography. Tooth decays can easily be misinterpreted as shadows due to various reasons, such as low image quality. Besides, OPG images have poor quality and structures are not presented with strong edges due to low contrast, uneven exposure, etc. Thus, disease detection is a very challenging task using Panoramic radiography. With the recent development of Artificial Intelligence (AI) in dentistry, and with the introduction of Convolutional Neural Network (CNN) for image classification, developing medical decision support systems is becoming a topic of interest in both academia and industry. Providing more accurate decision support systems using CNNs to assist dentists can enhance their diagnosis performance, resulting in providing improved dental care assistance for patients. In the following thesis, the first automated teeth extraction system for Panoramic images, using evolutionary algorithms, is proposed. In contrast to other intraoral radiography methods, Panoramic is captured with x-ray film outside the patient mouth. Therefore, Panoramic x-rays contain regions outside of the jaw, which make teeth segmentation extremely difficult. Considering that we solely need an image of each tooth separately to build a caries detection model, segmentation of teeth from the OPG image is essential. Due to the absence of significant pixel intensity difference between different regions in OPG radiography, teeth segmentation becomes very hard to implement. Consequently, an automated system is introduced to get an OPG as input and gives images of single teeth as the output. Since only a few research studies are utilizing similar task for Panoramic radiography, there is room for improvement. A genetic algorithm is applied along with different image processing methods to perform teeth extraction by jaw extraction, jaw separation, and teeth-gap valley detection, respectively. The proposed system is compared to the state-of-the-art in teeth extraction on other image types. After teeth are segmented from each image, a model based on various untrained and pretrained CNN-based architectures is proposed to detect dental caries for each tooth. Autoencoder-based model along with famous CNN architectures are used for feature extraction, followed by capsule networks to perform classification. The dataset of Panoramic x-rays is prepared by the authors, with help from an expert radiologist to provide labels. The proposed model has demonstrated an acceptable detection rate of 86.05%, and an increase in caries detection speed. Considering the challenges of performing such task on low quality OPG images, this work is a step towards developing a fully automated efficient caries detection model to assist domain experts.
DegreeMaster of Science (M.Sc.)
CommitteeMcWalter, Emily; Zhang, Chris; Papagerakis, Petros
Copyright DateFebruary 2022
Convolutional neural networks