Delineation of Adolescent Alveolar Bone in Intraoral Ultrasonograph with Machine Learning
Introduction: Crowding and misalignment of teeth, known as malocclusion, and periodontal disease are two of the three most common dental anomalies and can cause psycho-social problems in children. Severe malocclusion can lead to oral function issues such as difficulty in jaw movement, chewing, speech, and high susceptibility to periodontal diseases which can cause adolescents’ tooth loss. Alveolar bone is one of the important periodontal structures to support and hold the teeth. Accurate assessment of alveolar bone level is essential for diagnosis of periodontal disease and orthodontic treatment planning. The evaluation of alveolar bone is typically performed using two-dimensional (2D) intraoral radiography on the mesial and distal surfaces, but not on the facial and lingual surfaces of the teeth. This limitation can be overcome by using cone-beam computed tomography (CBCT), which renders images without tissue superposition. Due to high radiation doses, there is a lack of evidence to support the use of CBCT as a routine and standard method of dental diagnosis, particularly for pediatric patients, who are more vulnerable to radiation. The use of intra-oral ultrasound has received great attention recently due to the advantages of being non-invasive, non-ionizing radiation, portable, and low-cost imaging solution for initial and continuing dental diagnosis. However, the interpretation of alveolar bone in ultrasound images depends on observer’s experience, which is a challenge for dental clinicians. This work aimed to automatically segment alveolar bone and locate the alveolar crest using machine learning (ML) approach for intra-oral ultrasound images.
Methods: Thirty adolescent patients (17 males and 13 females, aged 12-17 years), who underwent orthodontic treatment, were recruited and consented in Kaye Edmonton Dental Clinic. The data used for the present study were collected from a total of 51 mandibular and 59 maxillary central incisors. The scanning was performed at 20 MHz using a SonixTablet ultrasound scanner (Analogic, Vancouver, Canada). A subset of 10 unalike images for each tooth with the presence of alveolar bone was acquired and then the data was divided into three sets: training (700), validation (200), and testing (200). The training and validation sets were used to train and tune the ML model to automatically segment the alveolar bone. The alveolar bone crests were also identified on the labeled contours.
Results: Quantitative evaluations of over 200 images from the testing set compared to an expert clinician showed that the ML approach yielded an average Dice score of 85.3%, sensitivity of 88.5%, and specificity of 99.8%, respectively. Alveolar bone crests were also identified with a mean difference of 0.20 mm and excellent reliability (ICC ≥ 0.98) in less than a second.
Conclusion: This study demonstrated an application of ML to assist dental clinicians in the visualization of alveolar bone in ultrasound images. This can help improving the workflow efficiency and accuracy of diagnosis. Future study will involve different types of teeth of orthodontic, periodontal and healthy patients to allow for a more comprehensive assessment of the proposed ML algorithm.