Fully automated film mounting in dental radiography: a deep learning model

Yu-Chun Lin; Meng-Chi Chen; Cheng-Hsueh Chen; Mu-Hsiung Chen; Kang-Yi Liu; Cheng-Chun Chang

doi:10.1186/s12880-023-01064-9

Fully automated film mounting in dental radiography: a deep learning model

BMC Med Imaging. 2023 Aug 18;23(1):109. doi: 10.1186/s12880-023-01064-9.

Authors

Yu-Chun Lin^{1

2}, Meng-Chi Chen³, Cheng-Hsueh Chen⁴, Mu-Hsiung Chen⁵, Kang-Yi Liu⁶, Cheng-Chun Chang⁷

Affiliations

¹ Department of Medical Imaging and Intervention, Chang Gung Memorial Hospital at Linkou, Taoyuan, Taiwan.
² Department of Medical Imaging and Radiological Sciences, Chang Gung University, Taoyuan, Taiwan.
³ Department of Dentistry, Chang Gung Memorial Hospital at Taipei, Taipei, Taiwan.
⁴ Department of Dentistry, National Taiwan University Hospital, Hsin-Chu Branch, Hsin-Chu, Taiwan.
⁵ Department of Dentistry, National Taiwan University Hospital, Taipei, Taiwan.
⁶ Department of Electrical Engineering, National Taipei University of Technology, 1, Sec. 3, Zhongxiao E. Rd, Taipei, 10608, Taiwan.
⁷ Department of Electrical Engineering, National Taipei University of Technology, 1, Sec. 3, Zhongxiao E. Rd, Taipei, 10608, Taiwan. ccchang@mail.ntut.edu.tw.

Abstract

Background: Dental film mounting is an essential but time-consuming task in dental radiography, with manual methods often prone to errors. This study aims to develop a deep learning (DL) model for accurate automated classification and mounting of both intraoral and extraoral dental radiography.

Method: The present study employed a total of 22,334 intraoral images and 1,035 extraoral images to train the model. The performance of the model was tested on an independent internal dataset and two external datasets from different institutes. Images were categorized into 32 tooth areas. The VGG-16, ResNet-18, and ResNet-101 architectures were used for pretraining, with the ResNet-101 ultimately being chosen as the final trained model. The model's performance was evaluated using metrics of accuracy, precision, recall, and F1 score. Additionally, we evaluated the influence of misalignment on the model's accuracy and time efficiency.

Results: The ResNet-101 model outperformed VGG-16 and ResNet-18 models, achieving the highest accuracy of 0.976, precision of 0.969, recall of 0.984, and F1-score of 0.977 (p < 0.05). For intraoral images, the overall accuracy remained consistent across both internal and external datasets, ranging from 0.963 to 0.972, without significant differences (p = 0.348). For extraoral images, the accuracy consistently achieved the highest value of 1 across all institutes. The model's accuracy decreased as the tilt angle of the X-ray film increased. The model achieved the highest accuracy of 0.981 with correctly aligned films, while the lowest accuracy of 0.937 was observed for films exhibiting severe misalignment of ± 15° (p < 0.001). The average time required for the tasks of image rotation and classification for each image was 0.17 s, which was significantly faster than that of the manual process, which required 1.2 s (p < 0.001).

Conclusion: This study demonstrated the potential of DL-based models in automating dental film mounting with high accuracy and efficiency. The proper alignment of X-ray films is crucial for accurate classification by the model.

Keywords: Deep learning; Dental; Radiography.

Publication types

Research Support, Non-U.S. Gov't

MeSH terms

Deep Learning*
Humans
Radiography, Dental