Detection of vertical root fractures by cone-beam computed tomography based on deep learning

Abstract

Objectives:

This study aims to evaluate the performance of ResNet models in the detection of in vitro and in vivo vertical root fractures (VRF) in Cone-beam Computed Tomography (CBCT) images.

Methods:

A CBCT image dataset consisting of 28 teeth (14 intact and 14 teeth with VRF, 1641 slices) from 14 patients, and another dataset containing 60 teeth (30 intact and 30 teeth with VRF, 3665 slices) from an in vitro model were used for the establishment of VRFconvolutional neural network (CNN) models. The most popular CNN architecture ResNet with different layers was fine-tuned for the detection of VRF. Sensitivity, specificity, accuracy, PPV (positive predictive value), NPV (negative predictive value), and AUC (the area under the receiver operating characteristic curve) of the VRF slices classified by the CNN in the test set were compared. Two oral and maxillofacial radiologists independently reviewed all the CBCT images of the test set, and intraclass correlation coefficients (ICCs) were calculated to assess the interobserver agreement for the oral maxillofacial radiologists.

Results:

The AUC of the models on the patient data were: 0.827(ResNet-18), 0.929(ResNet-50), and 0.882(ResNet-101). The AUC of the models on the mixed data get improved as:0.927(ResNet-18), 0.936(ResNet-50), and 0.893(ResNet-101). The maximum AUC were: 0.929 (0.908–0.950, 95% CI) and 0.936 (0.924–0.948, 95% CI) for the patient data and mixed data from ResNet-50, which is comparable to the AUC (0.937 and 0.950) for patient data and (0.915 and 0.935) for the mixed data obtained from the two oral and maxillofacial radiologists, respectively.

Conclusions:

Deep-learning models showed high accuracy in the detection of VRF using CBCT images. The data obtained from the in vitro VRF model increases the data scale, which is beneficial to the training of deep-learning models.

Introduction

Vertical root fracture (VRF) is a completeor incomplete fracture initiating from the root at any level, usually directs buccolingually.¹ Tamse et al. found that the VRF, whether incomplete or complete, extends to the periodontal ligament, and bacterial elements are constantly released into the area, leading to a hopeless prognosis.² In Tsesis’s³ study, whether to extract a VRF tooth or to choose a more conservative treatment is a multifactorial clinical decision-making process. Although some clinicians have tried to retain teeth with VRF, the outcomes are not predictive.^4,5 Therefore, early diagnosis of VRF is essential to avoid excessive alveolar bone loss.

However, early diagnosis of VRF is difficult due to the absence of pathognomonic signs and symptoms.⁶ Rud et al.⁷evaluated 468 teeth with vertical or oblique root fractures in periapical radiographs. Their results showed that only one-third of the fractured roots could be visualized directly in the radiograph and that the fracture line or plane could be identified only if the central X-ray beam is parallel to it or within 4°. Although periapical radiograph was widely used, the intrinsic limitations of anatomical superimposition and geometric distortion interfere with the diagnosis of VRF.^8,9 In the late 1990s, cone beam computed tomography (CBCT) was introduced into the field of dentistry.¹⁰ It provides a three-dimensional visualization of the region of interest without superimposing adjacent structures.^11,12 Talwar et al.¹² included four studies for a meta-analysis and concluded that CBCT images showed better sensitivity and specificity in the detection of VRFs in non-root filled teeth. The high diagnostic accuracy of CBCT for root fractures was also admitted in Ma’s meta-analysis.¹³ CBCT plays an important role in early diagnosis of VRF;¹⁴ however, the diagnosis of VRF from CBCT images is time-consuming, even for an experienced dental radiologist.

Recently, deep learning has been widely used for medical imaging tasks. Deep learning is a subset of artificial intelligence (AI) techniques. Deep learning can establish an end-to-end model to fit data by adjusting the interconnection relationship between a large number of internal nodes. A convolution neural network (CNN) is a subset of deep-learning algorithms that employ the convolution operation in neural network layers. CNN shows a significant performance improvement in many medical imaging analysis fields.^15,16 CNN was also introduced to many dental fields.^17–19 Kwon et al developed a CNN model for detecting and classifying odontogenic cysts and tumors in1,282 panoramic radiographs. The CNN model showed high performance, though the number of panoramic images was limited.¹⁷ Lee et al used CNN for the detection and diagnosis of dental caries in 3, 000 periapical radiographs and the area under the receiver operating characteristic (ROC) curve (AUC) of the premolar model was 0.917.¹⁸ CNN was also applied to the prediction of periodontally compromised teeth¹⁹ and the classification of mandibular third molar.²⁰ Liu et al.²⁰ proposed a CNN approach based on ResNet 34 to classify the relationship between the mandibular third molar and mandibular canal on CBCT images. The mean accuracy was 93.3% in their study and on par with the dental radiology residents. As for the diagnosis of VRF by deep learning, there have been several studies on periapical radiographs (PAs) and panoramic images.^21–23 In addition, a probabilistic neural network (PNN) has been used to diagnose VRF in periapical radiographs and CBCT images of intact and endodontically treated teeth. Hu et al²⁴ evaluated the diagnostic efficiency of deep residual neural network (ResNet) 50, VGG19 and DensenNet169 on the CBCT of VRF teeth, ResNet50 presented the highest accuracy, the AUC was 0.99. Since the ResNet is one of the well-performed CNNs,²⁵ this study aims to evaluate the performance of ResNet models in the detection of in vitro and in vivo VRFs in CBCT images.

Methods and material

Data collection

One CBCT image dataset (dataset 1) consisting of 1,641 slices from 14 patients was collected between 2020 and 2021 at the Department of Radiology, Peking University School and Hospital of Stomatology. Twenty-eight (14 intact and 14 teeth with VRF) teeth were extracted for the dataset 1. These images were acquired by CBCT unit NewTom VG (Quantitative Radiology, Verona, Italy). The acquisition parameters used were a field of view (FOV) of 15 cm (diameter)×15 cm (height), a voxel size of 0.25 mm, 110kVp, 1.60 ~ 4.30 mA, and scan time 24 s. The volume data were exported as Digital Imaging and Communications in Medicine (DICOM) format. The inclusion criteria included: (1) at least one tooth with VRF and one tooth without VRF were clearly displayed without obvious movement artefact in the region of interest (ROI) of the CBCT images; (2) the diagnosis of VRF was confirmed radiographically, intraoperative, or post-extraction. The exclusion criteria included: (1) teeth with VRF were endodontically treated; (2) the presence of metal or other radiopaque restorative materials in adjacent teeth or teeth in the opposite arch; (3) obvious movement artefact was observed in the CBCT images.

Another CBCT image dataset (dataset 2) consisting of 3665 slices from 60 teeth (30 artificially induced VRF teeth and 30 non-VRF teeth) from a previous study in dry human mandibles was used.²⁶ The simulation procedures are addressed briefly here. The teeth were preprocessed with 3% sodium hypochlorite solution for 15–30 min for the removal of residual hard and soft tissues and then stored in normal saline. All teeth were evaluated with a magnifying glass (2×) for the exclusion of root fractures. When there was any doubt about the existence of root fractures, a 3D laser scanning microscope (VK-X100/X200; Keyence, Japan) was used for confirmation. Periapical radiographs were used to exclude abnormal teeth such as calcified root canal (s), root resorption, and endodontically treated teeth. Thirty teeth were randomly selected for the induction of VRF. The 30 teeth were fixed with their longitudinal axes of roots parallel to the platform, and a 0.25-mm-diameter diamond-coated wire was used to make the VRF from the apex to the crown. The cyanoacrylate adhesive (32543564, Shenzhen, China) was used to glue the two halves from the same tooth.The fracture widths of teeth for VRF were scanned with the Inveon Micro-CT (Siemens, Munich, Germany). The fracture widths were from 110 to 170 um. The mean width of fractures was 140 ± 26.8 um. All 60 teeth were numbered and allocated randomly into suitable sockets of dry human mandibles. The gap between the root and the socket walls was filled with paraffin wax to simulate the periodontal space and to keep the teeth stable. Mandibles with teeth were then placed into a cylindrical 20-mm-thick water phantom to simulate soft tissues.

The CBCT images were acquired by a NewTom VGi (Quantitative Radiology, Verona, Italy) CBCT unit. The acquisition parameters used were a field of view (FOV) of 6 × 6 cm, a voxel size of 0.125 mm, 110 kVp, 1 ~ 3.65 mA, a scan time of 36 s. The CBCT slices were all exported as digital imaging and communications in medicine (DICOM) format.

This study was approved by the Institutional Review Board of Peking University School and Hospital of Stomatology (PKUSSIRB-202281141).

Data preparation

1641 CBCT slices of 28 teeth from patients (dataset 1) were first randomly divided into three parts: training set (16 teeth, 57.1%), validation set (four teeth, 14.3%), and test set (eight teeth, 28.6%). Then, the CBCT slices acquired from the 60 in vitro teeth in the dataset two were combined with the dataset 1, giving 88 teeth in total (dataset 3). Then, the dataset three was randomly divided and allocated into three subsets: the training set (52 teeth, 59.1%), the validation set (10 teeth, 11.4%), and the test set (26 teeth, 29.5%), respectively.

Workflow of the deep learning approach

The workflow of the deep learning approach is displayed in Figure 1. The region of interest in the 1641 CBCT slices from the dataset one was first cropped for data preprocessing. The chosen ROIs were then resized 224 × 224 for training and validation purposes. The pre-trained ResNet with different layers (18,50,121) in the ImageNet dataset²⁷ were fine-tuned to accelerate the convergence of the model training. The data augmentation, weight decay, and early-stopping were applied to avoid overfitting. The Softmax cross-entropy loss was adjusted as loss function for classification model training. The training epochs, initial learning rate, and batch size were set as 50, 0.0002, 32, respectively. The Adam optimizer was used for training. After that, the shuffled mixed CBCT slices for VRF in the in vitro teeth and patients were trained in the same way. The trained CNN models in the mixed data were then applied for comparison. To understand the mechanism by which the CNN models work, class activation mapping (CAM)²⁸ was visualized and heat maps were generated. The heat maps could show the area to which the CNN models pay more attention when classifying the teeth with VRF.The CBCT slices were first classified. If a series of consecutive CBCT slices (more than 50%) were diagnosed as VRF, the tooth would be classified as a VRF tooth.

Figure 1.

Classification workflow

To evaluate the CNN models’ performance in the detection of VRF, two oral and maxillofacial radiologists (OMRs) with more than 7 years’ experience in the interpretation of CBCT images were invited to independently review the in vivo (dataset 1) and mixed (dataset 3) CBCT images. The radiologists evaluated all the CBCT slices in the test sets from the dataset one and the dataset three by the dedicated software of the CBCT unit. A five-point confidence rating scale was used to classify all the teeth in the CBCT images for the presence or absence of VRF: (1) definitely absent, (2) probably absent, (3) unsure, (4) probably present, and (5) definitely present. Intraclass correlation coefficients (ICCs) were calculated to assess the interobserver agreement.

Statistical analysis

The sensitivity, specificity, accuracy, PPV (positive predictive value), and NPV (negative predictive value) for the detection of VRF in the CBCT datasets were classified by the CNN models. The ROC and AUC of different models were compared. All statistical analyses were performed using SPSS Statistics 24.0 (SPSS Inc., Chicago, IL, USA).

Intraclass correlation coefficients (ICCs) were calculated to assess the interobserver agreement. ICC estimates and their 95% confidence intervals were calculated based on a single-rating (mean-rating (interagreement)), absolute-agreement, 2-way mixed-effects model.The ICC measures were divided into four levels: poor (below 0.50), moderate (between 0.50 and 0.75), good (between 0.75 and 0.90), and excellent (above 0.90). The chi-square test was used to compare the diagnosis results from the CNN model and oral and maxillofacial radiologists.

Result

The sensitivity, specificity, accuracy, PPV and NPV obtained fromdifferent models and OMRs were shown in Table 1. The sensitivity values of three models training with patients’ data (dataset 1) were: 86.8% (ResNet18), 94.5% (ResNet50), and 89.4% (ResNet101). The specificity values were: 64.3% (ResNet18), 73.2% (ResNet50), and 67.7% (ResNet101). The accuracy values were: 75.6% (ResNet18), 83.9% (ResNet50), and 78.6% (ResNet101).The positive predictive values were: 71.17% (ResNet18), 78.18% (ResNet50), and 73.72% (ResNet101). The negative predictive values were: 82.78% (ResNet18), 92.92% (ResNet50), and 86.26% (ResNet101).The models trained with mixed data (dataset 3) showed better performance than those solely trained with patient data (dataset 1).The sensitivity values were: 94.6% (ResNet18), 95.9% (ResNet50), and 85.9% (ResNet101). The specificity values were: 64.5% (ResNet18), 74.1% (ResNet50), and 76.1% (ResNet101). The accuracy values were:78.0% (ResNet18), 84.0% (ResNet50), and 80.5% (ResNet101).The positive predictive values were: 68.61% (ResNet18), 75.30% (ResNet50), and 74.72% (ResNet101). The negative predictive values were: 93.59% (ResNet18), 95.66% (ResNet50), and 86.77% (ResNet101). Overall, the performance of ResNet50 is the best.

Table 1.

Accuracy of different models and OMRs for the detection of vertical root fracture using cone beam computed tomography

Accuracy	CNN models						Meanoftwo OMRs
	ResNet-18		ResNet-50		ResNet-101
	Patient Data	Mixed Data	Patient Data	Mixed Data	Patient Data	Mixed Data	Patient Data	Mixed Data
Sensitivity (%)	86.8	94.6	94.5	95.9	89.4	85.9	90.9	93.15
Specificity (%)	64.3	64.5	73.2	74.1	67.7	76.1	86.35	84.05
Accuracy (%)	75.6	78.0	83.9	84.0	78.6	80.5	88.65	88.65
PPV	71.17	68.61	78.18	75.30	73.72	74.72	87.12	81.11
NPV	82.78	93.59	92.92	95.66	86.26	86.77	90.45	90.89
AUC	0.827	0.927	0.929	0.936	0.882	0.893	0.944	0.925
(95% CI)	(0.793-0.861)	(0.914–0.940)	(0.908-0.950)	(0.924-0.948)	(0.855-0.909)	(0.878-0.909)

The ROC curves of different models are plotted in Figures 2 and 3 and Figure 3. The maximum AUC were obtained from the ResNet 50, giving 0.929 (0.908–0.950, 95% confidence interval, CI) for the patient data and 0.936 (0.924–0.948, 95% CI) for the mixed data.The mean sensitivity, specificity, and accuracy values of OMRs were: 90.9%, 86.35%, and 88.65% on the patient data and 93.15%, 84.05%, and 88.65% on the mixed data.

Figure 2.

ROC curve of different models on the patient data

Figure 3.

ROC curve of different models on the mixed data

The essential features were visualized in the heat maps, that is, the CNN models paid more attention to the fracture line on the CBCT images, as shown in Figure 4.

Figure 4.

CAM results for visualization of the attention regions for VRF

The ICC value for the interobserver consistencyof two radiologists was 0.917. The diagnosis consistency was excellent. The AUCs from the two radiologists were 0.937 and 0.950 for patient data and 0.915 and 0.935 for the mixed data, respectively. The chi-square tests showed that there was no significant difference between the AUCs from the CNN model and oral and maxillofacial radiologists (radiologists1: p = 0.67>0.05, and radiologists 2: p = 0.83>0.05).

Discussion

Since there are no specific clinical symptoms and signs for VRF, early and accurate diagnoses are challenging for VRF teeth. Progression of VRF represents not only tooth extraction but also bone resorption, resulting in poor conditions for the tooth implant. Deep CNN models have high performance in 2D dental image detection, classification, and segmentation.^17–20 Deep-learning algorithms usually require balanced and large-scale data to optimize the parameters. Compared to other dental diseases such as caries, the number of VRF images is limited. Transfer learning and fine-tune techniques of pre-trained models on large-scale data generate a specific result for similar tasks.²⁹ On the other hand, the image features of the fracture line are usually continuous in VRF teeth, which makes it possible to train the classification models using CBCT slices. In Liu et al.’s²⁰ study, 2D CBCT slices were trained for classification of the relationship between mandibular third molar and mandibular canal and they get high performance in the test set. Therefore, 2D CBCT slices were adopted in this study. The class activation mapping (Figure 4) shows that image features learned by CNN models focus on the most obvious fracture line, which is consistent with human experts’ experience. Furthermore, the AUC of ResNet-50 is 0.936 in the mixed data, which is comparable to the performance from the two oral maxillofacial radiologists (0.915 and 0.935).

In the studies of VRF CBCT diagnosis, the in vitro models are often used to simulate actual clinical situations. Jakobson et al assessed the influence of metallic posts in the CBCT images of VRF using in vitro teeth model and Mohamed et al. compared the diagnostic accuracy of CBCT images in 80 extracted teeth.^30,31 Guo et al investigated the effect of voxel size and fracture width using CBCT images on the in vitro models.²⁶ VRF in clinical patients usually occurs with significant displacement or root resorption. The collection of early VRF teeth for research purpose is difficult. Thus, the in vitro mandibular models²⁶ were also applied in this study. The mixed data show a better performance for classification than sole use of patients’ data (Table 1), and this means that it is a viable approach by increasing the amount of data through in vitro models for identification of VRF.

Fukuda et al²¹ evaluated the CNN for detecting VRF on panoramic radiography, the performance was displayed as recall (0.75), precision (0.93), and F measure (0.83). Johari et al²² and Kositbowornchai et al²³ evaluated the performance of the artificial neural networks using Matlab on in vitro VRF teeth’ periapical radiographs and the 2D reconstructed CBCT images. The performance was reported as sensitivity: 98%, 93.3%, specificity: 90.5%, 100%, and accuracy: 95.7%, 96.6% for PAs and CBCT images, respectively. The angle of the fracture line and X-ray beam was not mentioned in their studies, which is a key factor for diagnosis in periapical radiographs as explained in the introduction.

The diagnosis of VRF in CBCT images usually requires an experienced radiologist and endodontist.Pradeep Kumar et al.reviewed eight articles for meta-analysis to determine the diagnostic accuracy of CBCT in detecting VRF. The pooled sensitivity, specificity and accuracy values were: 0.78 (95% CI, 0.64–0.88), 0.80 (95% CI, 0.63–0.91), and 0.86 (95% CI, 0.83–0.89),respectively.³² InTalwar’s review, four studies were considered for meta-analysis, and the pooled sensitivity, specificity, and the diagnostic odds ratio of CBCT in filled and unfilled teeth were: CBCT (filled): 0.752, 0.652, and 5.527; CBCT(unfilled): 0.776, 0.946, and 94.26, respectively.¹² In the present study, ResNet-50 on the mixed data performed a better sensitivity (95.9%) and AUC (93.6%), which means a low rate of missed diagnosis. However, the lower specificity (74.1%) of ResNet-50 means a high rate of misdiagnosis.The cutoff value of ROC needs adjustment to optimize the sensitivity and specificity. CNN shows potential for computer-aided diagnosis. In general, CNN provided a new method for automated classification toimprove the diagnostic accuracy and efficiency of VRF.

Although CNN models show high performance in VRF classification, there also are some limitations in this study. First, 2D CBCT slices were used instead of 3D images can provide more information for classification. Secondly, this study collected CBCT of VRF teeth without filling, but VRF usually occurs in the teeth after root canal treatment.Thirdly, the width of the VRF model in the present study was not a real simulation of clinical situations. Detection of smaller fractures remains a challenge. CBCT images with filled teeth still need to be collected for further analysis. Finally, the model training only adopted the ResNet architecture, and the algorithm has room for improvement.

Conclusion

Deep-learning models showed high accuracy in VRF detection which can be used for automated diagnosis of VRF for time-saving and compensating for the lack of oral maxillofacial radiologists. VRF in vitro model CBCT can increase data scale, which is beneficial to the training of deep-learning models.