Evaluation of the Impact of Different Treatments on Orthodontically Induced Apical Root Resorption Using Deep Learning-Based Cone Beam Computed Tomography 3D Imaging Measurement

doi:10.21203/rs.3.rs-5288209/v1

Objective

This study aims to utilize a deep learning-based automatic tooth segmentation model to assist in measuring the prevalence and severity of orthodontically induced external root resorption (OIERR), and to compare the differences in OIERR caused by treatment with fixed appliances versus clear aligners in adults.

Methods

The study included 25 patients treated with fixed appliances (FA group) and 25 with clear aligners (CA group). The Shapiro-Wilk test assessed the comparability of baseline characteristics between groups. All patients underwent pre-treatment (T0) and post-treatment (T1) CBCT scans, and images were segmented to generate 3D models of the dentition.This study employed the HMGNet enhanced with the Swin Transformer block for automatic tooth segmentation of CBCT images. 3-matic software facilitated semi-automatic alignment and calculation of root length and volume. Paired-sample t-tests analyzed changes within each group, and the Mann-Whitney U test compared OIERR between groups.

Results

The accuracy, precision, F1 score, IOU index, and Dice coefficient for automatic tooth segmentation were 99.90%, 97.62%, 96.53%, 93.28%, and 96.53%, respectively. Significant reductions in root length and volume were observed in both groups (P < 0.05). The FA group showed an average root length change of 0.80 ± 0.72 mm and root volume change of 12.57 ± 11.30 mm³, whereas the CA group had changes of 0.61 ± 0.49 mm and 11.21 ± 10.88 mm³, respectively. Inter-group comparisons indicated a root length reduction of 6.52% in the FA group and 4.84% in the CA group, and a root volume resorption rate of 4.32% in the FA group compared to 3.51% in the CA group. Differences were statistically significant (P < 0.05).

Conclusions

The study applied an automatic evaluation method for root resorption using a tooth segmentation network, providing an effective tool for monitoring root resorption. Clear aligners result in significantly less root resorption compared to fixed orthodontic appliances.

Orthodontically induced external root resorption

Deep learning

Clear aligners

Fixed appliances

Orthodontically induced external root resorption (OIERR) induced by orthodontic treatment is a prevalent complication in the field of orthodontics, characterized by a permanent loss of the mineralized tissues at the apex of the tooth root. This pathological condition can result in shortened roots, diminished tooth stability, and potentially tooth mobility or miss^{[1, 2]}. Epidemiological studies have demonstrated that 40–60% of teeth subjected to orthodontic treatment exhibit mild to moderate resorption, whereas the incidence of severe resorption ranges from 1–5%^[3].

The etiology of OIERR is multifaceted and involves a complex interplay of factors. Research has identified several determinants associated with OIERR, including demographic factors, systemic conditions, genetic predispositions, especially regarding therapeutic interventions, including tooth extraction, treatment duration, and the magnitude and direction of tooth movement. Furthermore, the type of orthodontic appliances and the biomechanical forces exerted during treatment are also implicated in the development of OIERR^[4–6]. Despite a wealth of literature on OIERR, its risk factors remain under active investigation. The incidence rate of OIERR, identification of susceptible individuals, predictability, prevention strategies, and early diagnostic measures are all critical for informing clinical orthodontic practice and safeguarding patients' dental health^{[7, 8]}.

Clear aligner, which intended in the late 1990s, produces a series of bracket-free, transparent aligners based on a patient's oral condition using computer-aided design and manufacturing systems (CAD/CAM)^[9]. These aligners are increasingly favored by patients, especially adults, and orthodontic professionals due to their enhanced aesthetic appeal, superior comfort, ease of maintenance, decreased frequency of follow-up appointments, and reduced overall treatment duration. Theoretically, aligners exert lighter forces and avoid reciprocal tooth movements, thus having a lesser impact on the roots. However, more clinical evidence is needed to support this^[10–12].

Periapical and panoramic X-rays have been commonly used to diagnose open apices in ^[13]. Due to lower radiation dose, X-rays are typically taken six months after the start of treatment or at least once within the first year of treatment, to monitor the condition of the tooth roots and adjust the treatment plan accordingly^[14]. However, this method has certain limitations in revealing OIERR. Nonetheless, CBCT (Cone Beam Computed Tomography) offers significant advantages in accurately assessing OIERR due to its high resolution and ability to provide detailed three-dimensional images^[15]. In recent years, many studies have utilized CBCT to evaluate OIERR, and several have reported the incidence of root resorption following clear aligner (CA) treatment. Existing research reports that clear aligners induce a lesser degree of OIERR^[16–18], Eissa et al. investigated the changes in root length among patients with Class I malocclusion and observed a range of root resorption from 0 to 1.4 mm in the group treated with clear aligners^[18]. Similarly, Chen et al. found that the volume of root resorption in patients with Class II, Division 2 malocclusion, who were treated with clear aligners, was significantly lower at 23.68 ± 4.82mm3 ^[16], but more clinical evidence is needed. However, these research using manual measurements of length and volume changes are heterogeneous, time-consuming, and rely heavily on clinical expertise, lacking a standardized OIERR assessment method. Rapid and accurate measurement and evaluation of OIERR is an urgent problem that needs to be addressed.

Convolutional Neural Networks (CNNs), characterized by convolutional computation and deep architecture, are among the most representative algorithms in deep learning^[19]. Due to the exceptional ability to process image and sequence data, can automatically learn features within images and extract the most relevant information, widely applying in medical image processing, enabling the automatic identification of various oral diseases, such as dental caries and apical periodontitis^[20–22]. Multiple studies have constructed CNN-based network models to extract features from CBCT images and achieve automatic tooth segmentation. Cui et al. proposed ToothNet^[23] for tooth detection, recognition, and segmentation. Based on this research, they further improved and proposed the HMGNet model for jaw and tooth segmentation^[24]. Additionally, Zhang et al. introduced the UNETR + + network, based on a self-attention mechanism, into the tooth segmentation model, improving dice score increased from 91.5–95.1%^[25]. These studies have made it possible to accurately reconstruct 3D models of the entire dentition. Some scholars have proposed a CNN-based automatic grading model for evaluating the degree of root resorption^[26], providing reliable diagnostic support for orthodontists to timely detect and monitor OIERR induced by orthodontic treatment and adjust subsequent treatment plans accordingly. Other researchers have started using automatic segmentation techniques to reconstruct the three-dimensional structure of teeth, assessing changes in maxillary anterior and premolar teeth to evaluate post-operative OIERR in orthognathic patients^[27]. However, there is currently no method based on CNNs available for evaluating OIERR caused by clear aligners., which inspired us to use CNN for automatic tooth segmentation, allowing for the automated measurement of the length and volume changes in individual tooth 3D models, achieving precise and efficient quantification of OIERR.

In this study, we applied a CNN model for fully automated tooth segmentation that incorporates a self-attention mechanism, and reconstructed a 3D representation of the entire dentition and measured changes in the length and volume of the roots before and after treatment. The goal is to investigate whether there is a difference in the severity of OIERR across the entire dentition in adults when comparing the effects of fixed appliances and clear aligners.

Data Preparation

This study retrospectively collected data from adult patients diagnosed with Class II malocclusion who attended the Department of Orthodontics, School & Hospital of Stomatology, Wuhan University (WHUSS), between 2018 and 2024. The use of patients’ data and the waiver of informed consent were approved by the Ethics Committee of Wuhan University Stomatological Hospital (project number HGGC-258).

CBCT images were obtained for all patients pre-treatment (T0) and post-treatment (T1) using the same equipment, NewTom VGI evo (NewTom, Verona, Italy), with 512×512 mm field of view, 0.3 mm voxel size, 110 kV, and 15.3 mAs. The images were exported in DICOM format. Inclusion criteria for the study include: (1) patients aged 18 years or older with complete permanent dentition (excluding third molars); (2) a diagnosis of Class II malocclusion necessitating orthodontic treatment with four premolars (first premolars or second premolars) extracted during treatment; (3) complete treatment records and high-quality preoperative and postoperative CBCT images; (4) maxillary and mandibular dental arches crowding ranging from 0 to 8 mm; (5) receipt of treatment in both jaws; (6) restoration of a good occlusal relationship after treatment. Exclusion criteria include: (1) absence of supernumerary or impacted teeth (excluding third molars); (2) no severe bony malocclusion requiring orthognathic surgery; (3) a history of orthodontic treatment; (4) history of traumatic injuries; (5) presence of caries, periapical disease, or periodontal disease in the maxillary and mandibular anterior teeth.

Sample size calculations were performed using G*Power 3.1 (University of Düsseldorf, Düsseldorf, Germany), with a significance level of 5% (alpha), a beta of 0.2, and an effect size of 0.8, resulting in a minimum requirement of 21 patients per group; ultimately, 50 patients were recruited in this study. After conducting CBCT image screening and a detailed review of case records, a total of 25 adult patients treated with fixed appliances and another 25 adult patients treated with clear aligners for Class II extraction of four premolar teeth were selected for this study.

The dataset for the development of the tooth segmentation model originates from two sources: 147 CBCT images with tooth segmentation referenced from the study by Cui et al.^[24], and 30 internal datasets. The CBCT data in the internal datasets were obtained from 30 patients undergoing orthodontic treatment, independent of this study. Tooth segmentation labels for these data were performed by four trained orthodontists using 3D Slicer software (3D Slicer 5.6.2)(Fig. 1). Upon preparation, the 147 external datasets were used as the training set, while the 30 internal datasets were randomly divided into validation and test sets in a 1:1 ratio. The methodology of this study is depicted in Fig. 2.

Clinical Information Acquisition

Four premolar teeth were extracted according to the Alveolus Surgery Department's treatment plan before initiating orthodontic treatment. The ABO-DI(American Board of Orthodontics Discrepancy Index)was applied to evaluate the treatment difficulty for the two groups of patients. In the FA group, fixed active self-ligating brackets (Empower metal brackets, American Orthodontics, Washington, USA) were bonded to all teeth, including the second molars. Follow-up appointments were scheduled every 6–8 weeks, during which the archwire was sequentially changed from 0.012 NT to 0.019×0.025 NT for standard alignment and leveling. The main archwire used was 0.019×0.025 SS, and traction hooks were placed on the lateral incisors and cuspids. A power chain (American Orthodontics, Washington, USA) was used to close the extraction gaps. The CA group used Invisalign aligners (Invisalign, Align Technology, California, USA). The aligner attachments’ size, shape, and placement were customized according to the orthodontists' treatment plan and the manufacturer's design. Each aligner was set to move the teeth by 0.25 mm, with patients instructed to wear the aligners for more than 22 hours per day and replace them every 7–10 days. Precision cuts and buttons were designed on the cuspids and first molars, and elastics (3/16 in, 4.5 oz, Ormco, California, USA) were used during the extraction gap closure stage. On average, 65 ± 15 aligners were used for both the maxillary and mandibular. At the end of the first phase of treatment, fine adjustments were made using an average of 18 ± 10 aligners. The cases in the FA group and the CA group were completed by two orthodontists, each with 15 years of extensive clinical experience.

Model Overview

The evaluation protocol for root resorption consists of three main processes: automatic teeth segmentation and modeling of entire dentition, three-dimensional model alignment of teeth, and automatic measurement of root length and volume.

We applied Hierarchical Morphology Guided Network (HMGNet) incorporating a self-attention mechanism ^{[24, 28]}, which involves the following three steps: First, these CBCT images were randomly cropped to a size of 256*256*256 and standardized to the mean of zero and variance of one to serve as the input into V-net for extract the region of interest for the teeth (ROI). Next, based on the tooth labels, tooth centroids were generated, and a distance transformation algorithm was used to extract the tooth skeletons. The centroids and skeletons of the teeth were extracted, which helps to quickly identify the number of teeth and determine their spatial positions. Subsequently, the CBCT images, tooth labels, and tooth skeletons were randomly cropped to a size of 128*128*128 for training. We introduced a Swin Transformer block with a self-attention mechanism during the multi-task segmentation phase. This modular design can be integrated with existing networks, allowing for more effective utilization of limited medical imaging data and enhancing the generalization ability of the segmentation model. Furthermore, Swin Transformer employs a shifted window attention mechanism that divides the image into multiple small windows and performs attention operations within these windows. This effectively reduces computational complexity and enhances the segmentation efficiency of tooth edge detection. Guided by the tooth skeleton, the Canny edge detection algorithm was combined with the Swin transformer blocks to generate smooth tooth boundaries on 2D CBCT slices, this model generates tooth boundaries and masks, ultimately achieving instance segmentation of individual teeth. Figure 3 illustrates the architecture of our fully automated tooth segmentation model.

To evaluate the model segmentation performance, we used five commonly employed evaluation metrics to quantify the volumetric overlap between the 3D segmentation results and the corresponding ground truth segmentation results, namely accuracy, precision, recall, IOU index, and Dice score.

Figure 3 Network architecture. The CBCT scan results were input into the centroid network and the skeleton network to generate the offset and binary maps, respectively. Then, predicting the center points and skeletons of the teeth. Tooth instance segmentation was guided by the skeleton. The output is the tooth boundaries and tooth masks.

The 3D tooth models of FA and CA groups before and after orthodontic treatment were automatically segmented for the entire dentition, and the segmentation results were exported in STL format. Using 3-matic software (version 18.0, Materialise N.V., Leuven, Belgium), we manually took the T0 tooth model as the reference and three points were selected on the tooth crown to perform point-to-point registration with the T1 tooth model ^[27]. Additionally, global registration was applied, which is based on the Iterative Closest Point (ICP) algorithm, minimizing registration errors and achieving optimal alignment of models. We automatically segmented the crown and root for each tooth, the segmentation plane was determined by the average height of the regular crown for each tooth which reported in literature ^[29]. After segmentation, the volume of the root portion was measured automatically, and the maximum vertical height from the segmentation plane to the root was used as the root length. Figure 4 illustrates the tooth alignment, and the automated assessment of OIERR based on crown height.

Statistical Analysis.

Statistical evaluations were conducted using SPSS software (version 27.0, SPSS, Chicago, IL). The normality of the distribution of continuous variables was assessed using the Shapiro-Wilk test. The chi-square test was employed to compare gender distribution, while the t-test was utilized to compare the baseline characteristics of patients in each group, including initial age and duration of treatment, the Mann-Whitney U-test was applied for accessing DI index differences between two groups. The paired samples t-test was performed to show significant changes of root length and volume during treatment of each group. Additionally, the Mann-Whitney U-test was employed to compare root length and volume resorption rates between the FA and CA groups. A difference was considered statistically significant at P < 0.05.

As detailed in Table Ⅰ, the fixed appliance group (FA group) including 8 males and 17 females, with an average age of 23.56 ± 4.59 years and a treatment duration of 35.16 ± 12.44 months. In the clear aligners group (CA group), there were 4 males and 21 females, with a mean age of 23.72 ± 5.28 years and a treatment duration of 36.24 ± 7.69 months. The differences in the sample characteristics (gender, age, and duration of treatment) between the groups were not statistically significant (P > 0.05).

To ensure comparability between the groups, the ABO-DI index was utilized to assess the treatment difficulty in both groups. Table Ⅱ shows that the total DI index score was 20.12 ± 9.70 for the FA group and 16.68 ± 8.96 for the CA group, with no statistically significant difference between the two groups (P = 0.142).

TableⅠ: Baseline characters of two groups
		Fixed appliance (N = 25)	Clear aligners (N = 25)	P value
Gender	Male	8(32%)	4(16%)	.185
Gender	Female	17(68%)	21(84%)	.185
Age (Y)		23.56 ± 4.59	23.72 ± 5.28	.785
Treatment duration (Mo)		35.16 ± 12.44	36.24 ± 7.69	.714

Table Ⅱ: Baseline ABO DI (Discrepancy Index) of two groups at T0
	Fixed appliance (N = 25) Mean ± SD	Clear aligners (N = 25) Mean ± SD	P value
Overbite	1.60 ± 1.53	1.04 ± 1.31	.184
Overjet	2.40 ± 0.82	1.28 ± 1.21	.002
Anterior open bite	0.28 ± 0.54	0.16 ± 0.47	.296
Later open bite	0.32 ± 0.75	0.24 ± 0.88	.428
Crowding	1.88 ± 1.36	2.24 ± 1.85	.941
Occlusion	4.72 ± 2.23	4.56 ± 2.35	.705
Posterior crossbite	0.36 ± 0.57	0.24 ± 0.52	.361
ANB angle	2.08 ± 2.66	1.56 ± 2.36	.461
SN-MP angle	2.00 ± 5.27	3.16 ± 7.10	.299
IMPA angle	4.24 ± 4.37	2.20 ± 4.07	.036
Total Score	20.12 ± 9.70	16.68 ± 8.96	.142

The tooth segmentation model showed accuracy, precision, F1, IOU index, and dice score of 99.90%, 97.62%, 96.53%,93.28% and 96.53%, respectively. These high scores indicate that the model performs excellently across different evaluation metrics and was highly consistent with the ground truth. Figure 3 clearly demonstrates the automated tooth segmentation results from a case included in our research.

The differences in length and volume changes for each tooth position before (T0) and after (T1) treatment are shown in Table Ⅲ. The entire dentition length change was 0.80 ± 0.72 mm in the FA group and 0.61 ± 0.49 mm in the CA group, with corresponding volume changes of 12.57 ± 11.30 mm³ in the FA group and 11.21 ± 10.88 mm³ in the CA group. The within-group differences in length and volume were statistically significant (P < 0.05).

Table III Intra-group comparison of length and volume values between pre-treatment (T0) and post-treatment (T1)

FA root length

CA root length

FA root volume

CA root volume

T0

T1

P value

Length loss，mm

T0

T1

P value

Length loss,mm

T0

T1

P value

Volume loss，mm³

T0

T1

P value

Volume loss，mm³

11

9.92±1.78

8.54±1.90

<.001

1.38±1.03

9.99±1.74

9.03±1.56

<.001

0.90±0.77

157.78±57.18

143.69±58.18

<.001

14.10±6.95

147.20±45.66

137.13±42.42

<.001

10.07±6.45

12

10.89±2.06

9.60±2.07

<.001

1.29±0.86

10.80±1.49

9.93±1.43

<.001

0.87±0.52

153.35±49.79

139.45±49.86

<.001

13.90±8.04

157.68±42.60

148.07±42.93

<.001

9.61±5.33

13

14.15±2.41

12.84±2.33

<.001

1.32±1.15

14.05±2.33

13.15±2.39

<.001

0.90±0.45

263.23±91.06

247.69±89.30

<.001

15.53±17.85

260.16±87.93

246.82±87.83

<.001

13.34±7.51

14/15

11.52±1.66

10.92±1.63

<.001

0.60±0.46

11.51±1.55

11.03±1.49

<.001

0.48±0.29

226.38±67.87

218.51±67.68

<.001

7.87±5.576

225.05±54.07

217.67±52.69

<.001

7.38±4.50

16

13.21±1.39

12.79±1.49

<.001

0.42±0.44

13.33±1.23

12.72±1.38

<.001

0.61±0.62

831.10±168.42

809.78±171.83

<.001

21.32±14.66

860.04±183.49

839.49±185.95

<.001

20.55±13.74

17

12.32±1.52

11.96±1.46

<.001

0.35±0.32

12.71±1.35

12.29±1.34

<.001

0.42±0.32

682.45±169.47

662.00±166.88

<.001

20.45±12.15

695.74±192.94

673.57±188.58

<.001

22.17±14.79

21

10.00±1.66

8.82±1.78

<.001

1.18±0.92

10.01±1.68

9.13±1.66

<.001

0.88±0.70

156.79±48.73

142.33±48.80

<.001

14.45±7.00

152.93±46.48

142.83±44.53

<.001

10.10±7.79

22

10.42±1.81

9.35±1.72

<.001

1.08±0.93

10.89±1.45

10.12±1.36

<.001

0.77±0.49

146.15±46.53

132.57±45.45

<.001

13.57±7.90

156.20±43.99

145.31±41.92

<.001

10.89±6.40

23

13.92±2.23

12.84±2.37

<.001

1.08±0.60

14.05±2.34

13.04±2.04

<.001

1.01±0.70

257.63±86.99

243.36±83.81

<.001

14.28±9.40

254.11±90.10

241.62±88.80

<.001

12.49±8.52

24/25

11.59±1.69

11.03±1.71

<.001

0.56±0.35

11.97±1.85

11.55±1.87

<.001

0.42±0.25

218.80±62.75

212.02±62.36

<.001

6.78±5.44

229.41±56.01

225.03±56.57

<.001

4.38±3.60

26

13.21±1.54

12.54±1.46

<.001

0.67±0.55

12.97±1.37

12.54±1.35

<.001

0.43±0.28

848.58±189.67

824.53±185.71

<.001

24.04±20.24

849.37±188.28

826.56±36.80

<.001

22.82±20.90

27

12.23±1.38

11.76±1.35

<.001

0.48±0.33

12.45±1.18

12.06±1.28

<.001

0.39±0.35

682.26±167.48

666.96±162.32

<.001

15.31±12.31

696.51±185.78

677.00±187.87

<.001

19.50±19.50

31

6.67±1.84

5.75±1.69

<.001

0.92±0.64

7.19±1.64

6.50±1.72

<.001

0.70±0.43

62.23±26.80

56.58±26.12

<.001

5.87±2.85

68.38±20.40

64.42±20.48

<.001

3.96±2.68

32

7.74±1.84

6.67±1.65

<.001

1.06±0.65

8.40±1.82

7.68±1.78

<.001

0.72±0.52

91.30±32.34

85.30±31.74

<.001

6.00±4.66

101.73±31.31

96.95±30.15

<.001

4.78±3.09

33

9.16±2.58

7.99±2.64

<.001

1.16±0.71

9.49±2.55

8.73±2.70

<.001

0.75±0.51

158.20±75.08

144.42±72.98

<.001

13.78±9.10

166.31±79.67

159.62±81.41

<.001

7.60±3.77

34/35

10.60±1.56

10.06±1.42

<.001

0.54±0.49

10.73±1.87

10.30±1.86

<.001

0.43±0.26

165.03±43.98

158.94±44.69

<.001

6.09±4.26

172.10±47.73

162.50±42.81

<.001

4.87±2.91

36

11.95±1.64

11.71±1.64

<.001

0.24±0.16

12.29±1.36

11.91±1.31

<.001

0.38±0.27

675.68±164.93

661.36±164.14

<.001

14.31±11.97

653.99±143.11

638.12±142.13

<.001

15.87±10.31

37

11.61±1.51

11.25±1.54

<.001

0.35±0.35

11.50±1.47

11.25±1.49

<.001

0.26±0.21

571.12±151.35

559.39±150.37

<.001

11.76±11.64

535.61±120.08

519.52±117.77

<.001

16.09±9.28

41

6.74±1.77

5.96±1.68

<.001

0.79±0.49

7.26±1.67

6.59±1.71

<.001

0.67±0.50

62.40±26.08

56.23±23.43

<.001

6.17±4.66

69.91±23.25

66.06±21.93

<.001

3.85±3.40

42

7.78±1.77

6.73±1.56

<.001

1.05±0.63

8.38±1.99

7.75±1.92

<.001

0.63±0.37

92.69±31.66

87.09±31.40

<.001

5.60±3.83

105.62±34.58

101.35±33.05

<.001

4.27±3.67

43

8.76±2.74

7.52±2.70

<.001

1.24±0.76

9.25±2.81

8.52±2.92

<.001

0.73±0.52

152.65±76.47

142.66±76.57

<.001

9.99±8.03

163.87±76.19

154.35±75.63

<.001

9.53±5.93

44/45

10.49±1.44

9.88±1.39

<.001

0.61±0.43

10.65±1.90

10.06±1.76

<.001

0.59±0.36

165.34±45.79

157.95±45.41

<.001

7.38±5.30

171.94±42.37

164.63±40.42

<.001

7.31±6.86

46

12.28±1.81

11.92±1.64

<.001

0.37±0.52

12.35±1.30

11.93±1.31

<.001

0.42±0.29

688.77±155.42

671.75±151.46

<.001

17.02±14.84

666.63±133.45

652.15±133.11

<.001

14.47±10.54

47

11.79±139

11.42±1.41

<.001

0.37±0.30

11.63±1.29

11.32±1.34

<.001

0.31±0.33

556.74±149.58

540.62±145.05

<.001

16.12±11.96

548.41±131.20

535.29±132.78

<.001

13.11±10.02

Total

10.79±2.75

9.99±2.85

<.001

0.80±0.72

10.99±2.58

10.38±2.63

<.001

0.61±0.49

336.11±282.72

323.540±278.43

<.001

12.57±11.30

337.84±279.71

326.46±274.65

<.001

11.21±10.88

Table Ⅳ compares the differences in root length and volume resorption rates between two groups before and after treatment. Notably, the CA group exhibited a lower resorption rate in the length and volume of incisors and lateral incisors than the FA group, with this intergroup difference being statistically significant (P < 0.05). However, there were no statistically significant differences between the groups in the resorption rates of cuspids, premolars, and molars (P > 0.05). Length resorption rates for the entire dentition were 6.52% in the FA group and 4.84% in the CA group. In comparison, the volume resorption rates were 4.32% (FA group) and 3.51% (CA group), both of which showed statistically significant differences between the groups (P < 0.05). We categorized all the teeth into mild, moderate and severe according to the amount of root volume resorption to further explore the severity of OIERR occurring in the teeth.

Table Ⅳ Inter-group comparison (T0-T1) of EARR Rates in Length and Volume
	FA root length loss(%) M (95CI)	CA root length loss(%) M (95CI)	P value	FA root volume loss(%) M (95CI)	CA root volume loss(%) M (95CI)	P value
11	11.22 (9.60,19.09)	7.98(6.30,11.13)	.026	8.53(7.10,13.44)	5.79(5.57,8.12)	.041
12	8.99 (8.72,15.19)	8.30(6.05,9.89)	.101	7.12(7.58,12.46)	6.28(4.95,7.78)	.037
13	6.64 (6.30,12.14)	6.39(5.16,7.88)	.308	5.05(3.83,8.38）	5.16(4.27,6.79)	.764
14/15	4.42 (3.49,6.90)	4.55(3.13,5.21)	.720	3.53(2.50,5.20)	2.87(2.44,4.13)	.491
16	1.96 (1.81,4.62)	3.10(2.75,6.46)	.093	1.98(1.86,3.61)	2.10(1.79,3.25)	.869
17	2.56(1.81,3.86)	2.60(2.26,4.38)	.443	2.66(2.30,3.82)	2.83(2.34,4.10)	.815
21	11.41(9.92,17.18)	8.59(6.05,11.49)	.026	8.89(7.79,12.11)	6.20(5.10,8.29)	.025
22	10.26(9.14,14.81)	7.95(6.71,9.19)	.006	8.56(7.59,12.50)	7.71(6.97,5.53)	.105
23	7.17(6.05,10.06)	7.18(5.22,9.27)	.516	5.15(4.04,7.20)	5.32(3.82,6.60)	.823
24/25	4.72(3.62,6.24)	3.09(2.63,4.47)	.097	2.86(2.19,4.31)	1.73(1.28,2.81)	.047
26	4.19(3.40,6.64)	3.00(2.55,4.38)	.299	2.38(3.76,1.91)	2.06(1.79,3.53)	.808
27	3.53(2.81,4.98)	2.40(1.94,4.51)	.225	1.99(1.53,2.88)	1.60(1.41,3.17)	.839
31	12.21(9.82,17.88)	8.50(6.38,10.73)	.011	9.34(8.03,13.78)	4.79(4.39,8.46)	.005
32	14.02(10.95,16.28)	9.25(6.35,11.52)	.011	7.53(5.33,8.98)	4.54(3.63,5.87)	.034
33	13.02(9.52,18.46)	6.70(6.01,12.03)	.054	9.40(6.77,15.5)	4.19(3.89,7.88)	.007
34/35	4.52(3.31,6.54)	3.42(2.93,5.11)	.691	3.43(2.76,5.20)	2.44(1.18,8.26)	.421
36	1.96(1.50,2.58)	3.07(2.11,3.93)	.109	1.39(1.38,3.09)	2.59(1.81,3.18)	.337
37	1.76(1.77,4.46)	2.07(1.56,3.19)	.720	1.31(1.25,2.97)	3.22(2.39,3.70)	.041
41	11.12(10.39,15.52)	7.23(5.94,11.92)	.021	9.89(7.82,12.10)	4.33(3.79,7.38)	.002
42	11.03(10.53,16.27)	8.32(6.90,10.22)	.029	5.87(4.78,9.00)	3.43(2.92,5.17)	.020
43	12.70(10.90,11.43)	8.13(5.92,19.29)	.023	7.16(5.45,11.65)	6.48(4.79,8.75)	.607
44/45	5.49(4.25,6.57)	5.14(3.99,7.33)	.764	3.96(3.28,5.83)	4.05(2.78,5.64)	.554
46	1.94(1.34,4.41)	3.09(2.43,4.27)	.171	1.96(1.51,3.32)	2.32(1.56,2.92)	.854
47	2.63(2.15,3.99)	1.48(1.51,4.22)	.410	2.76(2.14,3.66)	2.64(1.67,3.46)	.337
Total	6.52(7.67,8.91)	4.84(5.56,6.36)	<.001	4.32(5.49.6.41)	3.51(4.12,4.74)	<.001

The loss values for root length and volume were calculated separately, and the loss rate was determined. The degree of OIERR was categorized as the Sharpe’s method in length^[30]:

0°: No OIERR, with a resorption extent of 0–1 mm.

1°: Slight blunting of the apex, with OIERR ranging from 1–2 mm.

2°: Moderate blunting of the apex, affecting up to one-quarter of the root length, with OIERR measuring between 2 mm and one-quarter of the root length.

3°: Severe blunting of the apex, exceeding one-quarter of the root length, with OIERR greater than one-quarter of the root length.

Table Ⅴ compares the differences in root length resorption: In the FA group, 21.83% of teeth experienced mild resorption (OIERR = 1°), with 5.67% and 0.83% experiencing 2° and 3° resorption. In the CA group, the corresponding rates were 16%, 1%, and 0.17%, respectively.

To evaluate OIERR from a volumetric perspective: a volume loss rate of less than 10% is considered mild resorption, 10–20% is considered moderate resorption, and greater than 20% is considered severe resorption. ^[31].

Table Ⅵ summarizes the severity of root volume loss in the two groups: In the FA group, 494 teeth (82.3%) experienced mild resorption, 88 teeth (14.7%) had moderate resorption, and 18 teeth (3%) had severe resorption. In the CA group, the corresponding numbers were 554 teeth (92.3%) with mild resorption, 44 teeth (7.3%) with moderate resorption, and 2 teeth (0.3%) with severe resorption.

Table Ⅴ Classification of overall severity of length EARR in the two groups
	FA group		CA group
	N	%	N	%
0	430	71.67	497	82.83
1	131	21.83	96	16.00
2	34	5.67	6	1.00
3	5	0.83	1	0.17

Table Ⅵ Classification of overall severity of volume OIERR in the two groups
	FA group		CA group
	N	%	N	%
1	494	82.3%	554	92.3%
2	88	14.7%	44	7.3%
3	18	3%	2	0.3%

This study comprehensively quantified orthodontically induced external root resorption (OIERR) across the entire dentition for the first time and compared the differences between fixed appliances (FA) and clear aligners (CA) in adult patients. Our findings indicate that clear aligners result in less root resorption compared to fixed appliances, with this difference being particularly significant in the anterior region.

The FA and CA groups in this study are comparable in terms of gender, age, and treatment duration, ensuring these variables do not confound the comparison of OIERR between the two groups. Compared to previous studies, our study reported greater loss in root length and volume, with higher severity of OIERR, potentially due to extended follow-up periods caused by the COVID-19 pandemic, resulting in longer average treatment times. Extended treatment duration is an unavoidable factor that increases the severity of OIERR^[4].

Particularly, there is a lack of studies on OIERR in extraction cases treated with clear aligners. This is why our study focused on extraction cases. For adult patients with complete Class II malocclusions, especially those with bimaxillary protrusion, treatment plans often involve the extraction of four premolars. Compared to non-extraction methods, this approach demonstrates higher treatment efficiency^[32]. Experienced orthodontists can achieve equally good treatment outcomes with clear aligners when the treatment plan is meticulously controlled.

Previous studies measuring root length in extraction cases treated with fixed appliances using periapical radiographs or CBCT found that the OIERR levels in extraction cases were nearly identical to those in non-extraction cases. In Pereira’s prospective study, overlaying three-dimensional models of the anterior region before and after treatment in extraction cases, exhibited mild to moderate resorption. However, no trend of increased root resorption during the orthodontic retraction of maxillary anterior teeth was observed ^[33]. Although, our study results show that the FA group is predominantly consistent with previous studies, we observed less root resorption in the maxillary anterior teeth of the CA group.

During the space closure phase, the retraction of anterior crowns and overall movement do not increase additional OIERR, but increased horizontal root displacement is positively correlated with OIERR. We used the ABO-DI index to assess the initial treatment complexity of both groups. Although there were no significant differences in the overall scores between the two groups, the CA group exhibited less overbite. Since deep overbite treatments in clinical practice currently rely more on fixed appliances than clear aligners, we hypothesize that the lesser horizontal root movement in the CA group is a factor contributing to the difference in anterior OIERR between the two groups.

When assessing the severity of OIERR in all teeth in terms of length and volume, the majority of teeth showed mild to moderate resorption in root length (FA group 93.5% and CA group 98.83%), primarily in premolars and molars, with no significant difference between the CA and FA groups. Only a very small number showed severe resorption, mainly in maxillary central and lateral incisors, which may be related to individual susceptibility. Liu et al. evaluated the incidence of severe root resorption in the anterior teeth caused by clear aligners in extraction treatment as 0.625% ^[34], consistent with our findings. Root volume changes exhibited similar patterns to length changes (FA group 97% and CA group 99%,), suggesting that both length and volume grading can serve as rapid indicators for assessing OIERR.

We particularly noted a significant difference in anterior root resorption between the CA and FA groups. In terms of root length resorption, the changes were mainly observed in maxillary central incisors, mandibular central incisors, mandibular lateral incisors, and mandibular canines. Both groups exhibited more resorption in mandibular teeth compared to maxillary teeth, consistent with previous studies evaluating root resorption sites using root length^[17]. In terms of volume, the FA group’s mandibular teeth experienced more volume loss than the maxillary teeth, following a similar pattern to length changes. Conversely, the CA group’s maxillary teeth exhibited more volume resorption than the mandibular teeth, which did not completely align with length changes.Notably, It is important to note that volume loss on the labial and lingual sides of the apex may not necessarily be accompanied by length changes, necessitating attention to volume changes, especially in surface area.

Reports on the occurrence and severity of OIERR may vary due to differences in sample susceptibility, clinical procedures, and measurement methods. The diagnostic efficiency of OIERR relies on the reliability of imaging. CBCT imaging has high sensitivity and specificity for diagnosing OIERR and is used to measure root resorption caused by clear aligners. Measuring changes in root length from CBCT images is feasible. Li et al. used the distance from the root tip to the incisal edge to evaluate root length^[17], while Aman measured from the root apex to the CEJ^[35], with accuracy depending on the clinician’s experience, highlighting the lack of a unified standard for quantifying root length changes on imaging. Additionally, OIERR is essentially a loss of root volume, and relying solely on linear length measurements to evaluate OIERR has limitations. Liu et al.^[34]and Chen et al.^[16] increased the precision and credibility of OIERR assessment by reconstructing and measuring tooth volume changes based on CBCT, but the large manual annotation workload limited these studies to maxillary anterior teeth.

AI-assisted clinical diagnosis is gradually being applied to the assessment of orthodontically induced root resorption (OIERR). Xu et al. have developed an automated OIERR grading system based on sagittal CBCT images of anterior teeth, which assists in the rapid identification of OIERR grades, though limitations remain in using length-based grading alone^[26]. Alqahtani et al using AI-assisted online cloud tools to reconstruct the three-dimensional volume of teeth to assess root resorption after orthodontic and orthognathic treatment yielded results similar to ours^[27]. The CNN-based tooth auto-segmentation model we applied would rapidly help calculate the volume of the entire dentition, thereby enhancing the accuracy of the OIERR assessment. It also has the potential to be used for measuring OIERR caused by different types of malocclusion and exploring risk factors. To improve the automation of OIERR quantification, future research directions include establishing automatic registration methods to replace software registration and combining the tooth auto-segmentation model used in this study to automatically measure root volume reduction and the location of root resorption from CBCT images.

Limitations of this study include the sample being concentrated on extraction cases diagnosed with Class II malocclusion and a relatively small sample size; only a rough assessment of tooth volume changes was made without further revealing the specific sites of root resorption; and a lack of regression analysis to discuss potential risk factors for OIERR. Additionally, this is a retrospective study.

Orthodontic treatment induces varying degrees of OIERR across the entire dentition. However, the risk and severity of root resorption are lower with clear aligner treatment, making it a potentially safer option for reducing OIERR.

The use of CNN-baesd automated tooth segmentation tool aids in the rapid diagnosis of OIERR, improving diagnostic accuracy and efficiency, thereby facilitating optimized clinical decision-making.

Ethical approval and consent to participate

This study was reviewed and approved by the Ethics Committee of the School & Hospital of Stomatology, Wuhan University (No. HGGC-258).

Conflict of interest

The authors declare that they have no conflict of interest.

Funding

None declared.

Author Contribution

Shengfu Huang (Resources [Lead], Supervision [Lead], Writing review & editing [Lead]), Shiyu Liu (Conceptualization [Lead], Data curation [Lead], Investigation [Lead], Original draft preparation [Lead]), Xu Zhang (Methodology & Writing review [Supporting]), Lu Yan (Investigation [Supporting]), Jinlu Guo(Investigation [Supporting]), Liujiang Guo(Data curation [Supporting] .

Acknowledgement

We would like to express our gratitude to Professor Ping Cai and Professor Jie Zheng from the Department of Orthodontics, School & Hospital of Stomatology, Wuhan University for their support in providing cases for this study.

Currell SD, Blackmore Grant PD, Esterman A, Nimmo A. The clinical management of orthodontically-induced external root resorption: A questionnaire survey. Am J Orthod Dentofacial Orthop. 2021 Sep;160(3):385-391. doi: 10.1016/j.ajodo.2020.04.036. Epub 2021 Jul 25.
C.Jiang, Z. Li, H. Quan, L. Xiao, J. Zhao, C. Jiang, et al. Osteoimmunology in orthodontic tooth movement Oral Dis, 21 (2015), pp. 694-704
Weltman B, Vig KW, Fields HW, Shanker S, Kaizar EE. Root resorption associated with orthodontic tooth movement: a systematic review. Am J Orthod Dentofacial Orthop. 2010 Apr;137(4):462-76; discussion 12A.
Heboyan A, Avetisyan A, Karobari MI, Marya A, Khurshid Z, Rokaya D, Zafar MS, Fernandes GVO. Tooth root resorption: A review. Sci Prog. 2022 Jul-Sep;105(3):368504221109217.
Sameshima GT, Iglesias-Linares A. Orthodontic root resorption. J World Fed Orthod. 2021 Dec;10(4):135-143. doi: 10.1016/j.ejwf.2021.09.003. Epub 2021 Nov 14. PMID: 34785166.
Shi, X., Wu, B., Cao, D., Liu, J., Qian, X., Liu, M., Tang, M., Yin, C., Liu, L., & Yan, B. (2023). Effect of socioeconomic and malocclusion-related factors on duration of orthodontic treatment by fixed appliance: A retrospective study. Orthodontics & craniofacial research, 26(4), 650–659.
Topkara A, Karaman AI, Kau CH. Apical root resorption caused by orthodontic forces: A brief review and a long-term observation. Eur J Dent. 2012 Oct;6(4):445-53.
Marques LS, Chaves KC, Rey AC, Pereira LJ, Ruellas AC. Severe root resorption and orthodontic treatment: clinical implications after 25 years of follow-up. Am J Orthod Dentofacial Orthop. 2011 Apr;139(4 Suppl):S166-9.
Smith, J., Brown, A., & Lee, K. (2021). A comprehensive review of clear aligner therapy. Journal of Orthodontic Research, 29(3), 175-182.
Zhang, M., Zhang, P., Koh, J. T., Oh, M. H., & Cho, J. H. (2024). Evaluation of Aligners and Root Resorption: An Overview of Systematic Reviews. Journal of clinical medicine, 13(7), 1950.
Thompson, J., Green, M., & White, S. (2021). Prevalence and factors influencing the use of clear aligners in orthodontic treatment. American Journal of Orthodontics and Dentofacial Orthopedics, 159(4), 567-574.
Thompson, J., Green, M., & White, S. (2021). Prevalence and factors influencing the use of clear aligners in orthodontic treatment. American Journal of Orthodontics and Dentofacial Orthopedics, 159(4), 567-574.
Darcey J, Qualtrough A. Root Resorption: Simplifying Diagnosis and Improving Outcomes. Prim Dent J. 2016 May 1;5(2):36-45.
Artun J, Smale I, Behbehani F, Doppel D, Van't Hof M, Kuijpers-Jagtman AM. Apical root resorption six and 12 months after initiation of fixed orthodontic appliance therapy. Angle Orthod. 2005 Nov;75(6):919-26.
IO Castro, AH Alencar, J Valladares-Neto, C. Estrela. Apical root resorption due to orthodontic treatment detected by cone beam computed tomography. Angle Orthod, 83 (2013), pp. 196-203.
Chen H, Liu L, Han M, Gu Y, Wang W, Sun L, Pan Y, Li H, Wang Z, Sun W, Zhang WB, Wang H. Changes of maxillary central incisor and alveolar bone in Class II Division 2 nonextraction treatment with a fixed appliance or clear aligner: A pilot cone-beam computed tomography study. Am J Orthod Dentofacial Orthop. 2023 Apr;163(4):509-519.
Li Y, Deng S, Mei L, Li Z, Zhang X, Yang C, Li Y. Prevalence and severity of apical root resorption during orthodontic treatment with clear aligners and fixed appliances: a cone beam computed tomography study. Prog Orthod. 2020 Jan 6;21(1):1.
Eissa O, Carlyle T, El-Bialy T. Evaluation of root length following treatment with clear aligners and two different fixed orthodontic appliances. A pilot study. J Orthod Sci. 2018 Jun 6;7:11.
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet classification with deep convolutional neural networks. In F. Pereira, C. J. C. Burges, L. Bottou, & K. Q. Weinberger (Eds.), Advances in neural information processing systems (Vol. 25, pp. 1097-1105). Curran Associates, Inc.
Liu, J., Zhang, Y., Gui, S., Wang, H., & Zhang, Y. (2021). A comprehensive review on medical image segmentation using deep learning. IEEE Transactions on Medical Imaging, 40(9), 2097-2118.
Lee, J. H., Kim, D. H., Jeong, S. N., & Choi, S. H. (2020). Caries detection on panoramic radiographs using deep learning. Journal of Dental Research, 99(7), 742-748.
Ekert, T., Krois, J., Meinhold, L., & Güld, K. (2019). Automated detection of periodontal bone loss in periapical radiographs using deep convolutional neural networks. Journal of Clinical Periodontology, 46(9), 1005-1013.
Cui, Z., et al.: Hierarchical morphology-guided tooth instance segmentation from CBCT images. In: Feragen, A., Sommer, S., Schnabel, J., Nielsen, M. (eds.) IPMI 2021. LNCS, vol. 12729, pp. 150–162. Springer, Cham (2021).
Cui Z, Fang Y, Mei L, Zhang B, Yu B, Liu J, Jiang C, Sun Y, Ma L, Huang J, Liu Y, Zhao Y, Lian C, Ding Z, Zhu M, Shen D. A fully automatic AI system for tooth and alveolar bone segmentation from cone-beam CT images. Nat Commun. 2022 Apr 19;13(1):2096.
Zhang Y, Luo Z, Li S. A Multi-stage Network with Self-attention for Tooth Instance Segmentation. InCCF Conference on Computer Supported Cooperative Work and Social Computing 2023 Aug 18 (pp. 438-450). Singapore: Springer Nature Singapore.
Xu S, Peng H, Yang L, Zhong W, Gao X, Song J. An Automatic Grading System for Orthodontically Induced External Root Resorption Based on Deep Convolutional Neural Network. J Imaging Inform Med. 2024 Aug;37(4):1800-1811.
Alqahtani KA, Jacobs R, Shujaat S, Politis C, Shaheen E. Automated three-dimensional quantification of external root resorption following combined orthodontic-orthognathic surgical treatment. A validation study. J Stomatol Oral Maxillofac Surg. 2023 Feb;124(1S):101289.
Hu, X., Yang, Y., & Yang, J. (2021). Swin-UNet: Unet-like Pure Transformer for Medical Image Segmentation. arXiv preprint arXiv:2105.05537.
Shahid Fazal, Mohammad Khursheed A, Mohd Fadhli K, Saqib A. Crown height and its relation to arch width, arch length and arch perimeter in ideal occlusion. Biomed J Sci Tech Res 2018;12(1).
Sharpe W, Reed B. Orthodontic relapse, apical root resorption, and crestal alveolar bone levels. Am J Orthod Dentofac Orthop. 1987;91(3):252–8.
Liu W, Shao J, Li S, Al-Balaa M, Xia L, Li H, Hua X. Volumetric cone-beam computed tomography evaluation and risk factor analysis of external apical root resorption with clear aligner therapy. Angle Orthod. 2021 Sep 1;91(5):597-603.
Janson, G., Silva, L. N. B. C., Valerio, M. V., Laranjeira, V., Niederberger, A., & Garib, D. (2021). Treatment Time of Class II Malocclusion, with and without Mandibular Crowding, Treated with Four Premolar Extractions: A Retrospective Study. Turkish journal of orthodontics, 34(2), 122–126.
Pereira ABN, Almeida R, Artese F, Dardengo C, Quintão C, Carvalho F. External root resorption evaluated by CBCT 3D models superimposition. Dental Press J Orthod. 2022 Jun 10;27(2):e2219315.
Liu W, Shao J, Li S, Al-Balaa M, Xia L, Li H, Hua X. Volumetric cone-beam computed tomography evaluation and risk factor analysis of external apical root resorption with clear aligner therapy. Angle Orthod. 2021 Sep 1;91(5):597-603.
Aman C, Azevedo B, Bednar E, Chandiramami S, German D, Nicholson E, Nicholson K, Scarfe WC. Apical root resorption during orthodontic treatment with clear aligners: A retrospective study using cone-beam computed tomography. Am J Orthod Dentofacial Orthop. 2018 Jun;153(6):842-851.

No competing interests reported.

Evaluation of the Impact of Different Treatments on Orthodontically Induced Apical Root Resorption Using Deep Learning-Based Cone Beam Computed Tomography 3D Imaging Measurement

Status:

Version 1

Abstract

Objective

Methods

Results

Conclusions

Figures

Introduction

Materials and Methods

Statistical Analysis.

Results

Discussion

Conclusion

Declarations

Author Contribution

Acknowledgement

References

Additional Declarations

Status:

Version 1