Accuracy and efficiency of automatic

tooth segmentation in digital dental

models using deep learning

Joon Im¹, Ju-Yeong Kim², Hyung-SeogYu¹, Kee-Joon Lee¹,

Sung-Hwan Choi¹, Ji-Hoi Kim¹, Hee-KapAhn³ & Jung-Yul Cha¹

Digital Dentistry Mirai Nakayama

 Introduction

• The first step of successful orthodontic treatment

accurate diagnosis establishing an optimal treatment plan

• Several treatment options are typically available

depending on patients’ individual needs

2
 Introduction

intra and extra

clinical dental
radiographs oral clinical
examinations models
photographs

Orthodontist・・・
ｃｏｌｌｅｃｔｓ various types of patient data

prepares a problem list

sets treatment goals

establishes orthodontic treatment plans 3

 Introduction

• Manual setup using plaster models ・・・

・requires considerable time and effort
・difficult and cannot be used for all treatment cases

• Virtual setup with a digital modeｌｓ ・・・

・helps determine various treatment options
・no significant difference in reliability when compared
with a manual setup

1. Macchi, A., Carrafello, G., Cacciafesta, V. & Norcini, A. Tree-dimensional digital modeling and setup. Am. J. Orthod. Dentofac. Orthop. 4
129, 605–610. (2006).
 Purpose

• to compare and evaluate the accuracy and efficiency of

the DGCNN-based segmentation and classification of
digital dental models

• the null hypothesis

‘the three different tooth segmentation methods of the
digital tooth model do not significantly differ in the
success rate, time, and tooth size’

5
 Materials and Methods

• Patients treated by the Department of Orthodontics

at Yonsei University Dental Hospital

• between January 2010 and February 2019

• 546 digital dental models（training：516, evaluation：30）

6
 Criteria and exclusion criteria

【Criteria for Digital Models】

• Over 14-year-old orthodontic patients with second molars eruption
• No defects in the teeth or gums
• Mild and moderate crowding

【Exclusion Criteria for Digital Models】

• Congenital tooth deformity
• Severe dental caries and tooth wear
• Congenital or acquired missing tooth.
• Supernumerary tooth and Severe crowding 7
 Sample size and power calculation

• 28 segmented teeth were evaluated per digital dental model

→Total of 840 teeth were evaluated in four different methods

• Sample size : n=30 per group

（G*Power software version 3.1.9.2）

8
 Deep learning process

Wang, Y. et al. Dynamic graph cnn for learning on point clouds. ACM Trans. Graph. (TOG) 38, 1–12 (2019). 9
Tooth segmentation

10
Tooth segmentation

11
Tooth segmentation

12
Tooth segmentation

13
 Tooth segmentation

・precise MD point setting is required

・the tooth number and axis are determined

14
 Tooth segmentation

・approximate MD point setting is required

・the tooth number is designated

15
 Tooth segmentation

・fully automatic segmentation does not

require MD point setting

16
Tooth segmentation

LS DS AS

17
 Tooth segmentation

REF LS DS AS

18
Tooth segmentation

19
 Measurement of MD width and CCH

• a: Change tooth orientation

• b: Measurement of the MD width
• c: Measurement of the CCH
20
 Result：Reliability analysis

• Reliability analysis of tooth measurement

• ICC of MD width and CCH → 0.987∼0.997
• Very high reproducibility (ICC>0.7: excellent)

21
Result：success/failure of tooth segmentation

• Success and failure of segmentation

• Kappa values → 0.885∼1
• Very high evaluation reproducibility
22
 Result：Success rate among three methods

LS、AS > DS

• Success rates for automatic segmentation

23
 Result：Comparison of MD width, CCH, and
segmentation time

• The segmentation time → AS had the shortest

• Friedman test(Shapiro–Wilk’s test) *p<0.05.

24
 Result：Comparison of MD width and CCH

• Statistically significant differences in all tooth groups

25
Result：Main effects and first-order interactions

• The tooth size errors → statistically different

DS (-0.019) > AS (-0.052) > LS (-0.121)

26
 Discussion

• AS segmentation based on DGCNN → higher success rate

• Reduced need for manual intervention

• Decreased time and effort

• Increased user convenience

27
 考察

• This study：used digital models of permanent dentition

in good condition without teeth and gingival defects

Impossibility to determine tooth segmentation ability

Missing teeth Severe wearing Dental caries

Partial eruption Third molar

• Subsequent studies：It will be necessary to use digital

dental models of various conditions
28
 Conclusion

• Statistically significant differences in segmentation success rate,

time, and size of segmented teeth using three different
orthodontic CAD/CAM programs
→rejecting the null hypothesis

• Artificial intelligence including deep learning can be applied to

orthodontic extraction, Diagnosis and treatment planning for
orthognathic surgery

29
Thank you for your attention

30