4.7 Article

Transformer based tooth classification from cone-beam computed tomography for dental charting

Journal

COMPUTERS IN BIOLOGY AND MEDICINE
Volume 148, Issue -, Pages -

Publisher

PERGAMON-ELSEVIER SCIENCE LTD
DOI: 10.1016/j.compbiomed.2022.105880

Keywords

Deep learning; Computer vision; Image classification; Medical imaging; Dental charting

Funding

  1. Natural Science Foundation of Shenzhen University General Hospital, China [SUGH2018QD029]
  2. NSFC-Youth [61902335]
  3. Key Area R&D Program of Guangdong Province, China [2018B030338001]
  4. National Key R&D Program of China [2018YFB1800800]
  5. Shenzhen Outstanding Talents Training Fund, China
  6. Guangdong Research Project [2017ZT07X152]
  7. Guangdong Regional Joint FundKey Projects, China [2019B1515120039]
  8. NSFC, China [81901058, 61931024, 81922046]
  9. helixon biotechnology company Fund, China
  10. CCFTencent Open Fund, China
  11. Research Fund for Overseas High-level Talents of Shenzhen, China [RC000336]

Ask authors/readers for more resources

In this paper, a deep neural network using a combination of CNN and Transformer structures is proposed for tooth classification. The method achieves improved accuracy in both clinical and publicly available datasets.
Dental charting is a useful tool in physical examination, dental surgery, and forensic identification. However, manual dental charting faces some difficulties, such as inaccuracy and psychiatric burden in forensic identification. As a critical step of dental charting, tooth classification can be completed on dental cone-beam computed tomography (CBCT) automatically to solve the above difficulties. In this paper, we build a deep neuron network which accepts a 3D CBCT image patch that contains the region of interest (ROI) of a tooth as input and then outputs the type of the tooth. Although Transformer-based neural networks outperform CNN-based neural networks in many natural image processing tasks, they are difficult to apply to 3D medical images. Therefore, we combine the advantages of CNN and Transformer structure to improve the existing methods and propose the Grouped Bottleneck Transformer to overcome the drawbacks of the Transformer, namely the requirement of large training dataset and high computational complexity. We conducted an experiment on a clinical data set containing 450 training samples and 104 testing samples. Experiments show that our network can achieve a classification accuracy of 91.3% and an AUC score of 99.7%. To further evaluate the effectiveness of our method, we tested our network on the publicly available medical image classification dataset MedMNIST3D. The result shows that our network outperforms other networks on 5 out of 6 3-dimensional medical image subsets.

Authors

I am an author on this paper
Click your name to claim this paper and add it to your profile.

Reviews

Primary Rating

4.7
Not enough ratings

Secondary Ratings

Novelty
-
Significance
-
Scientific rigor
-
Rate this paper

Recommended

No Data Available
No Data Available