Thesis
Using Artificial Intelligence to identify canine impaction in dental panoramic radiographs
University of Iowa
Master of Science (MS), University of Iowa
Spring 2023
DOI: 10.25820/etd.007096
Abstract
Introduction
In the field of dentistry, many developmental anomalies can be treated most effectively when identified and addressed early. When anomalies are discovered and treated early, treatment can be both less invasive as well as less time consuming for patients. Common among the type of anomalies just described are impacted maxillary canines. However, despite the ideal circumstances that would address this type of condition early, patients with impacted maxillary canines often first report to the orthodontist later than would be optimal. Consequences of this delayed diagnosis and treatment are frequently a more invasive approach, sometimes with uncertain prognosis for the affected teeth. Therefore, the dental profession, as well as the patient population to whom they serve, would greatly benefit from an automated way to identify and predict potential maxillary canine impactions. While our goal is to eventually predict maxillary canine impactions early (during primary or early mixed dentition), we first must achieve accurate automated identification of what an impacted maxillary canine is based on a dental panoramic radiograph.
Many fields within healthcare have successfully utilized the burgeoning technology of artificial intelligence. Artificial intelligence has proved invaluable to healthcare providers for the identification of trends or risk factors in their patients that would go unnoticed by the clinician. In the circumstance of potential maxillary canine impaction, the incorporation of artificial intelligence with dental records may successfully identify those at risk early enough to prevent future negative sequalae commonly associated with impacted maxillary canines. This study seeks to continue and to improve upon a previous study which utilized artificial intelligence on a subset of cases to identify and predict maxillary canine impaction based on dental panoramic radiographs.
Materials and Methods
Our study used a similar structure to the previous study performed by Welk et al. in 2021 as we utilized two separate image sets to train, validate, and test the chosen algorithm. One set of images was specific for impaction cases and a separate set of images for control cases. Our goal to improve upon the previous study included eliminating duplicate images and eliminating case and control images of questionable and poor quality. We also verified the maxillary canine diagnosis illustrated in each image and incorporated a larger impaction case sample size that would match our sample size of control images. A large difference between our study and the previous study was the incorporation of the image-recognition capabilities of ProtoPNet with the convolutional layers of DenseNet-121 and VGG-19.
Results
Convolutional Networks DenseNet-121 and VGG-19 were tested separately in conjunction with ProtoPNet to evaluate their ability to identify maxillary canine impaction based on a dental panoramic radiograph. Varying quantities of prototypes were used in the training process to refine and improve accuracy of the algorithm. Accuracy values obtained ranged from 62%-92% in correctly identifying impacted maxillary canine cases from non-impacted maxillary canine controls which is a substantial improvement from the 38-62% accuracy range obtained prior.
Conclusions
We utilized two different pre-trained, open-source convolutional neural networks in conjunction with the image-recognition model ProtoPNet to identify impacted maxillary canines versus normally erupting maxillary canines on dental panoramic radiographs. Using differing numbers of prototypes for identification showed an accuracy in identification that ranged from 62%-92%. This study shows that with high quality training and proper parameters, AI can be used to successfully identify impacted maxillary canines.
Details
- Title: Subtitle
- Using Artificial Intelligence to identify canine impaction in dental panoramic radiographs
- Creators
- Daniel Taylor
- Contributors
- Shankar Rengasamy Venugopalan (Advisor)Lina M Moreno-Uribe (Committee Member)Michael Callan (Committee Member)Shareef Dabdoub (Committee Member)
- Resource Type
- Thesis
- Degree Awarded
- Master of Science (MS), University of Iowa
- Degree in
- Orthodontics
- Date degree season
- Spring 2023
- DOI
- 10.25820/etd.007096
- Publisher
- University of Iowa
- Number of pages
- ix, 36 pages
- Copyright
- Copyright 2023 Daniel Taylor
- Language
- English
- Date submitted
- 04/07/2023
- Date approved
- 05/10/2023
- Description illustrations
- illustrations, tables, graphs
- Description bibliographic
- Includes bibliographical references (pages 33-36).
- Public Abstract (ETD)
- Within dentistry, maxillary impacted canines are a relatively common anomaly seen by dental clinicians. During the transitions that lead from primary to permanent dentition, the details shown in radiographs may suggest that maxillary canines are developing innocuously until it is suddenly obvious that intervention will be needed. However, if the potential for impaction is identified early enough, appropriate treatment modalities can successfully prevent the eventual impaction and facilitate normal eruption. Panoramic dental radiographs are routinely taken by general dentists to show the whole dentition, including developing teeth. The angulation of developing maxillary canines and the position of adjacent teeth may provide clues for diagnosis of maxillary canine impaction. Therefore, to mitigate the risk for negative sequelae associated with impacted maxillary canines, our study seeks to utilize a subset of artificial intelligence tools to identify impacted maxillary canines using standard panoramic dental radiographs. A previous study with the same goal achieved an accuracy that ranged from 38-62% in identifying whether a patient would develop impacted maxillary canines based on a panoramic radiograph. We used a similar open-sourced convolutional neural network (CNN), used in the previous study to train and test the radiographs. The CNN was used in conjunction with an image interpretability layer to better understand our results. Additionally, to improve results, we refined our sample of radiographs to increase the quality of training images for machine learning. We also increased our sample size of radiographs showing impacted maxillary canines to match the quantity of images in the control sample. Finally, we had three separate dental professionals (including two board-certified orthodontists) verify each image as either representing a clearly developing impacted maxillary canine or an obviously normal eruption. Our study sample used a case group of 723 dental panoramic images classified as impacted maxillary canine or obviously developing impacted maxillary canine and control group of 723 dental panoramic images showing no maxillary canine impaction or normal development. These images were used with a selected computerized neural network to train, validate, and test the ability to identify maxillary canine impaction based on dental panoramic radiographs. The results showed a significant improvement from the previous study in the identification of impacted maxillary canines. These results add to existing evidence to support that with appropriate training and parameter optimization, artificial intelligence can be used to assist dental professionals in the early diagnosis of developmental anomalies like impacted maxillary canines.
- Academic Unit
- Orthodontics; Craniofacial Anomalies Research Center
- Record Identifier
- 9984437259102771
Metrics
3 File views/ downloads
19 Record Views