Introduction

External cervical resorption (ECR) is defined as an irreversible loss of dental structures (i.e., cementum and dentin) followed by the action of odontoclastic cells1. Such resorption resulting from mechanical or chemical factors can be associated with orthodontic treatment, ectopic eruption, tooth impaction, parafunctional occlusion, periodontal/periapical infections, restorations, internal bleaching, trauma, and some systemic conditions such as Paget’s disease1,2,3,4,5,6. The process of resorption initiates at some point below the epithelial attachment and can gradually progress into the entire cervical region7. ECR is mainly asymptomatic and would incidentally be detected through radiographic examinations unless it involves the pulp or periodontal tissues. Since effective treatments for ECR are needed to be timely, precise identification of their location and magnitude is crucial4,8,9,10.

Conventional radiographs are routinely used as the first modality for ECR detection10. However, because of the limitations of two-dimensional radiographs such as superimposition of the anatomical structures, uneven magnification, and distortion; panoramic and periapical radiographs could result in misdiagnosis by having false-positive and false-negative results4,10,11,12. Cone beam computed tomography (CBCT) has proved to be an efficient method for diagnosis of ECR12,13. According to the previous studies, CBCT scans using a 0.3 mm voxel size would demonstrate root resorption with high accuracy14. Nevertheless, streaking artifacts generated from high-density materials such as intracanal posts and metallic restorations could corrupt the CBCT data by disfiguring the anatomical structures15,16,17,18,19,20,21,22,23. Such artifacts could jeopardize the diagnosis of lesions such as ECR especially when they are small and therefore postpone their treatment. Hence, developing techniques for enhancing the visibility of ECR lesions in the presence of metallic artifacts is necessary.

Canny edge-detection algorithm has been introduced as a technique for identifying the outlines of an object as well as sharp intensity changes in the images24,25. It has the capability of detecting the outer borders of an object precisely, without being affected by exposure parameters or external conditions17. This algorithm has been applied for quantification of metal artifacts in CBCT, diagnosis of osteoporosis, and detection of dental implant fractures15,17,26. Nevertheless, no studies so far have used the Canny algorithm to determine whether it facilitates the diagnosis of cervical resorption in the presence of metallic artifacts. Following the STARD guidelines27, for the first time we assessed the performance of this algorithm in the diagnosis of ECR defects when metal artifacts generated from intracanal posts worsens the condition.

Methods

Study design and sample preparation

This experimental in vitro study was approved by the Research Ethics Committee of Guilan University of Medical Sciences (Approval ID: IR.GUMS.REC.1401.039). All experiments were performed in accordance with relevant guidelines and regulations of the Declaration of Helsinki.

60 human premolar teeth that were extracted for orthodontic reasons were selected. Extraction of the teeth was performed as part of the patients’ orthodontic treatments and informed consent was obtained from all subjects to use their extracted teeth as the samples of the present study. All selected teeth had a single root. Teeth with caries, restorations, or anatomical concavities in the cervical region as well as any evidence of root resorption or crack were excluded. Root canal treatment was performed for all of the premolar teeth and the canals were obturated with gutta-percha points (Meta Biomed; Cheongju-si, Korea). Afterwards, post spaces with 8 mm length were prepared using Peeso reamers no. 2, 3, and 4. A prefabricated size 2 titanium post (Nordin; Montreux, Switzerland) was passively inserted in the root canals and self-cured luting glass ionomer cement (Fuji I; GC, Tokyo, Japan) was applied. The teeth were then divided into 3 groups randomly: 1- teeth with no cervical resorption (control group), 2- teeth with 0.5 mm cervical resorption, and 3- teeth with 1 mm cervical resorption. The resorption cavities were created at the CEJ of the teeth with two different depths (0.5 and 1 mm) using a round diamond bur. Following that, the margins were beveled with a flame diamond bur to create more natural shapes for the resorption defects. Depth of the cavities was controlled and measured with the use of a periodontal probe. In order to simulate the shape and density of the surrounding soft tissues, each tooth was inserted in a condensational type impression material (Speedex; Coltene, Altstatten, Switzerland) which was placed within a wax arch formed in the shape of the mandible (Fig. 1).

Figure 1
figure 1

Wax arch containing a pair of teeth for each CBCT exposure.

Full size image

Radiographic examinations and image processing

The wax arch contained 2 teeth that were randomly selected for each scanning procedure. The arch was placed in the central position of a CBCT unit (Pax-i 3D; Vatech, Hwaseong-si, Korea) and exposure parameters included 95 kV, 5.2 mA, FOV of 120 mm × 90 mm, and voxel size of 0.2 mm. Ez3D-i software version 1.1 (Ez3D-i; Vatech, Hwaseong-si, Korea, Available at: https://www.vatech.com/software_3d/417) was used to reconstruct the CBCT data. Three successive axial slices that referred to the cervical region of each tooth were selected. The selected images were saved in their original formats and at the same time exported to MATLAB version 2018a (MathWorks; Natick, MA, USA) to create the Canny-type images. In this way, 6 images including 3 original and 3 canny types were prepared from the CBCT scan of each tooth. In applying the Canny edge detection algorithm, firstly, image noise is removed by a Gaussian filter. Then, using a gradient magnitude thresholding technique, the edges of the image are identified. These steps are sequentially performed by MATLAB version 2018a software (MathWorks; Natick, MA, USA) in an automatic manner. Three parameters including standard deviation of the Gaussian filter, high, and low thresholds are defined by the operator. In the current study, standard deviation of the Gaussian filter was set to 1.2 and high and low thresholds were selected at 0.07 and 0.028, respectively17. All images were randomly numbered to be presented to the observers. Random numbering of the original and Canny images was performed to ensure no bias occurs during evaluation of the images. The observers included 3 experienced oral and maxillofacial radiologists who were asked to determine whether ECR existed in the images or not. Figure 2 refers to the original and Canny-type images of two teeth with different depths of cervical resorption.

Figure 2
figure 2

Comparison of the original (left) and Canny (right) images in displaying the resorption defects of 0.5 mm (A), 1 mm (B), and control group (C). Arrows on the Canny images point to the resorption areas that are unidentifiable on the original images.

Full size image

Analysis

Diagnostic accuracy of the index tests (original and Canny images) against the reference standard (direct observation of the teeth) is reported following the STARD guidelines27. Calculation of the index tests sensitivity and specificity as well as receiver operating characteristic (ROC) analysis was performed by the MedCalc software version 20.026 (MedCalc; Ostend, Belguim). The standard error (SE) was calculated with the DeLong test and confidence interval (CI) for the area under the ROC curve (AUC) was defined by the Binominal exact method. Pairwise comparisons were also performed using the DeLong test. P values less than 0.05 were considered to be statistically significant. The inter-rater reliability was defined by kappa coefficient which was over 0.7 between each pair of the three observers. The kappa values were measured using SPSS software version 28 (IBM; Armonk, NY, USA).

Results

Table 1 presents the sensitivity, specificity, and AUC values for the original and Canny images in detecting different depths of cervical resorption. It is noteworthy to mention that the significance values in Table 1 refer to the comparison of each image type with the random technique which is a technique with 50% chance of correct diagnosis. P values in Table 1 suggest that both image types are significantly more accurate than the random technique except for the original images in detecting 0.5 mm resorption cavities. ROC curves for the different image types are presented in Fig. 3.

Table 1 Diagnostic accuracy of the original and Canny images for the detection of ECR.
Full size table
Figure 3
figure 3

ROC curves of the original and Canny images in the different resorption depths: (A) original images in 0.5 mm depth, (B) Canny images in 0.5 mm depth, (C) original images in 1 mm depth, (D) Canny images in 1 mm depth.

Full size image

Comparison of the original and Canny images for the diagnosis of different depths of cervical resorption showed that in both 0.5 and 1 mm cavities, Canny images were superior to the original ones in terms of revealing the resorption defects. Table 2 refers to the comparison of the two image types in identifying the resorption cavities in the teeth.

Table 2 Comparison of the diagnostic accuracy of original and Canny images in detecting ECR.
Full size table

Considering the depth of resorption cavities, accuracy of the original images was significantly different between the two resorption depths (p = 0.009). The accuracy was higher in the detection of larger cavities which were 1 mm deep. Nevertheless, no significant difference existed in the diagnosis of 0.5 and 1 mm cavities by the Canny images (p = 0.168). Figure 4 graphically shows the comparison of AUC values among different image types and resorption depths. The highest accuracy belonged to Canny images for 1 mm defects, followed by Canny images for 0.5 mm defects, original images for 1 mm defects, and original images for 0.5 mm defects, respectively.

Figure 4
figure 4

Comparison of AUC among the original and Canny images in 0.5 mm and 1 mm resorption cavities.

Full size image

Discussion

Root resorption results in loss of dental hard tissue leading to reduction of crown-root ratio or loss of the whole tooth in severe cases28,29. ECR usually initiates more apically than the epithelial attachment and can progress in any direction7. This pathologic condition occurs in response to trauma, failure of periodontal or root canal treatments, undesirable position of the nearby impacted teeth, and ectopic eruption2,3,4. Since ECR is usually asymptomatic, better prognosis can be expected with an early diagnosis30. Having more accuracy than 2D plain radiographies, CBCT images are more reliably used for the diagnosis of root resorption1,31,32. Moreover, the influence of parameters such as voxel size and FOV dimensions on the diagnostic accuracy of root resorption has been investigated previously33,34,35,36. Nevertheless, treated teeth with root canal filling material and intracanal metallic posts produce interfering artifacts that make the diagnosis of root resorption difficult31. A number of previous studies have assessed the influence of artifacts caused from endodontically treated teeth on the accuracy of CBCT images10,17,37; however, no studies so far have investigated the accuracy of CBCT images for the diagnosis of ECR when intracanal metallic posts are present.

In the present study, CBCT images of three tooth groups based on the presence and depth of ECR (0, 0.5, and 1 mm) were examined. We created the resorption defects at the cervical region of the teeth since the intensity of artifacts from metallic posts is greatest at this region. According to the previous studies, an increased cavity size of ECR improves its detection32,38,39. Neves et al. concluded that better sensitivity, accuracy, positive and negative predictive values were obtained with larger defects38,39. Similarly, in our research, ECR defects with 1 mm depth were more accurately diagnosed than those with 0.5 mm depth. Nikneshan et al.35 reported that although increasing the size of defects increases the diagnosis accuracy, the difference is not statistically significant. Lack of a significant difference between the depth groups could be attributed to the absence of root filling material, metal restorations, or intracanal metallic posts and hence, lack of metallic artifacts. Neves et al. reported that an intracanal metallic post significantly decreases the accuracy of CBCT images for the diagnosis of vertical root fractures39. Thus, developing techniques that are capable of compensating for the adverse effects of metal artifacts is of utmost importance in diagnostic radiology.

In the present study, we used the Canny edge-detection technique as a novel method to determine whether it could improve the CBCT images in a way that the diagnosis of ECR in the vicinity of metallic posts is facilitated. We compared the detectability of ECR among original and Canny images derived from CBCT scans. The Canny edge-detection technique uses an algorithm which provides an accurate instrument for defining the object's outlines and sudden intensity changes in an image. Three parameters of the Canny algorithm including standard deviation of the Gaussian filter, high and low sensitivity thresholds can be set manually by the operator. Previous studies have suggested certain values for these three parameters to achieve the best image quality15,17. In this study, two experienced oral and maxillofacial radiologists who were not among the observers viewed the images to confirm that the suggested values for the aforementioned parameters provided the finest image details.

In order to assess the accuracy of different image types, original and Canny images were first compared with the random technique. Subsequently, the two image types (original and Canny) were compared with each other. The only image type that showed no significant difference with the random technique was original images in 0.5 mm depth of resorption (p > 0.999). This means that CBCT images in their original format could not be helpful for the diagnosis of small ECR defects when metallic artifacts exist. In other words, the diagnostic accuracy of CBCT images without using the Canny algorithm is adversely affected by the depth of the defect.

Comparison of the original and Canny images revealed significant differences in the diagnosis of ECR. Canny images were more accurate in either the 0.5 mm (p < 0.001) or 1 mm (p = 0.009) defect depths. This finding confirms the positive effect of applying the Canny algorithm for the diagnosis of ECR. Another important result that we obtained was that the diagnostic accuracy of Canny images did not differ significantly between the two ECR depths (p = 0.168) while the original images performed differently in the depth groups (p = 0.009). Hence, it could be concluded that use of the Canny edge-detection technique improves the CBCT images to the point that different ECR sizes are diagnosed with almost similar accuracy.

There are also limitations noteworthy to mention for this in vitro study. First, although resorption cavities were beveled to make them more natural-looking, clinical conditions are frequently associated with more bizarre-shaped defects that make the diagnosis more challenging. Hence, further studies in clinical situations are required to determine the accuracy of Canny edge-detection. Second, different voxel sizes have to be examined with the CBCT scans to assess the accuracy of Canny edge-detection in various resolutions.

Conclusion

Application of the Canny edge-detection algorithm on CBCT images with interfering metallic artifacts greatly improves the diagnosis of different depths of cervical resorption in endodontically treated teeth.