Dental Crowding Categorization Network (DCC-Net): Explainable deep learning system for automatic categorization of dental crowding on intraoral photographs
2
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
Objective: Since categorization of dental crowding is a crucial parameter in orthodontic diagnosis and tooth-extraction decisions, we aimed to develop an automatic system to categorize crowding levels on intraoral photographs without space analysis. Methods: The Dental Crowding Categorization Network (DCC-Net), consisting of segmentation, extraction, and categorization modules, was proposed and optimized by extracting regions of interest and crown centroids. A multicenter dataset including 1,351 maxillary and 1,253 mandibular intraoral photographs was divided in an 8:2 ratio for model training and internal testing, and an additional 100 photographs were collected for external testing. The ground truth was obtained through measurements by experienced orthodontists using intraoral scanning data. The accuracy, precision, recall, and F1- score of the categorization module were calculated, and heatmaps were obtained for model interpretation. Furthermore, a clinical evaluation was performed to compare the diagnostic accuracy of junior orthodontists with and without the assistance of DCC-Net. Results: For the maxilla, the categorization accuracy, precision, recall, and F1-score were 0.7232, 0.7447, 0.6793, and 0.6962, respectively, whereas the corresponding values for categorization in the mandible were 0.7352, 0.7506, 0.6723, and 0.7019, respectively. The heatmaps indicated that DCC-Net could identify the dental arches and regions showing malocclusion. In the clinical evaluation, the diagnostic accuracy of junior orthodontists improved with DCC-Net’s assistance, increasing by 9.18% for the maxilla and 12.75% for the mandible. Conclusions: DCC-Net achieved accurate categorization of dental crowding on intraoral photographs. Its rapid predictions may offer insights for guiding tooth extraction in orthodontic treatment, providing valuable reference data for inexperienced orthodontists and improving doctor–patient communication.
Abstract
Objective: Since categorization of dental crowding is a crucial parameter in orthodontic diagnosis and tooth-extraction decisions, we aimed to develop an automatic system to categorize crowding levels on intraoral photographs without space analysis. Methods: The Dental Crowding Categorization Network (DCC-Net), consisting of segmentation, extraction, and categorization modules, was proposed and optimized by extracting regions of interest and crown centroids. A multicenter dataset including 1,351 maxillary and 1,253 mandibular intraoral photographs was divided in an 8:2 ratio for model training and internal testing, and an additional 100 photographs were collected for external testing. The ground truth was obtained through measurements by experienced orthodontists using intraoral scanning data. The accuracy, precision, recall, and F1- score of the categorization module were calculated, and heatmaps were obtained for model interpretation. Furthermore, a clinical evaluation was performed to compare the diagnostic accuracy of junior orthodontists with and without the assistance of DCC-Net. Results: For the maxilla, the categorization accuracy, precision, recall, and F1-score were 0.7232, 0.7447, 0.6793, and 0.6962, respectively, whereas the corresponding values for categorization in the mandible were 0.7352, 0.7506, 0.6723, and 0.7019, respectively. The heatmaps indicated that DCC-Net could identify the dental arches and regions showing malocclusion. In the clinical evaluation, the diagnostic accuracy of junior orthodontists improved with DCC-Net’s assistance, increasing by 9.18% for the maxilla and 12.75% for the mandible. Conclusions: DCC-Net achieved accurate categorization of dental crowding on intraoral photographs. Its rapid predictions may offer insights for guiding tooth extraction in orthodontic treatment, providing valuable reference data for inexperienced orthodontists and improving doctor–patient communication.