An artificial intelligence system for detecting the root canal morphology of the mandibular first premolar on panoramic radiographs

doi:10.21203/rs.3.rs-3201787/v1

Background:

The internal root canal morphology of the mandibular first premolar can be quite diversified, which pose great challenge and can directly affect the outcome of endodontic therapy. The aim of this paper was to examine the diagnostic performance of a deep learning system to determine whether the mandibular first premolar has single root canal or not. Dental cone beam CT was used as the gold standard.

Methods:

CBCT and panoramic radiographs of 1630 mandibular first premolar from 815 patients were analyzed. The CBCT images of the roots were examined to determine whether the mandibular first premolar has single root canal or not. Image patches of the mandibular first premolars were segmented from panoramic radiographs. Then deep-learning network approaches were applied to the patched images. The performance of the deep-learning networks and the dentists were compared in terms of Precision (PR), Recall (RE), Accuracy (ACC), F1-score and Area Under Curve (AUC).

Results:

Using the ResNet-101(CBAM) system showed an accuracy of 85%, a positive predictive value of 81.4%%, and a negative predictive value of 82.35% in the determination of whether the mandibular first premolar has single root canal or not. There was a significant difference between the results of experienced dentists and those of less experienced dentists. There was no difference between the results of experienced dentists and those of the ResNet-101(CBAM). And the results of the ResNet-101(CBAM) are much better than the average of the dentists.

Conclusions:

The deep learning systems showed high levels of performance in differentiating whether the mandibular first premolar has single root canal or not based on panoramic radiographs. It is a promising artificial intelligence approach that may assist dentists in preoperative diagnosis.

artificial intelligence

the mandibular first premolar

root canal morphology

A comprehensive understanding of internal anatomy of human teeth is essential for root canal treatment[1]. Precise identification of root canal anatomy followed by perfect root canal preparation and obturation guarantee the success of root canal treatment. Variations in root canal morphology represent a great challenge in identification and management during the treatment of endodontic cases, as missing, untreated parts of the root canal were responsible for up to 42% of case’s failure which compromise the outcome[2].

In particular, the mandibular first premolars (MFPs) are complicated teeth to endodontically treat due to the anatomical variations that can present. Some studies have reported anatomical variations of this dental group, such as the presence of two or more roots, more than one canal and the presence of bifurcations of the root canal system. [3, 4]Other studies have reported unusual anatomy in the mandibular premolars described as radicular grooves that cause a C-shaped morphology when these teeth are analyzed in their cross sections[5]

Generally, periapical radiographs, panoramic radiographs, and cone beam computed tomography (CBCT) imaging are the most widely used diagnostic strategies for determining appropriate treatments. CBCT imaging has been demonstrated to provide higher accuracy than radiography for assessing root and root canal configurations. However, the risk of excessive radiation exposure and costs limits its use in routine clinical practice. Alternatively, panoramic radiographs are commonly employed in diagnosis and treatment[6], and if all changes in tooth and root canal morphology are successfully identified, this imaging method can greatly facilitate initial adjunctive diagnoses in clinic.

Convolutional neural network-based deep learning artificial intelligence (AI) models have driven considerable advances in processing imaging data for a variety of medical fields, consequently improving diagnostic and treatment outcomes. Deep learning neural network models can learn explicit and implicit features from large amounts of training data, substantially reducing the manual workload. In image recognition, deep learning models can make full use of multi-dimensional features through a flexible architecture that enables broad image recognition, while an end-to-end learning framework makes tuning of the entire system feasible.

AI has been used in endodontics primarily for detection of periapical lesions[7, 8], crown and root fractures[9, 10]; working length determination; and morphological characterization of root architecture. For example, Hiraiwa et al tested a deep learning system for image analysis of panoramic radiographs and found that it provided high accuracy in the differential diagnosis of single or multiple roots in the distal roots of mandibular first molars[11]. However, the MFP root canal has been largely overlooked in previous imaging studies.

In this study, we establish a dataset suitable for deep learning-based morphological classification of the root canal and applied a regional segmentation model based on Transformer to achieve segmentation of each tooth position. The target classification model was then used to determine whether or not the MFP has a single root canal in a set of patched images.

2.1 Clinical data collection and group classification

Panoramic radiographs acquired between Jan 2020 and Mar 2022 were randomly selected from the picture and communication system (PASC) of the Stomatology Hospital, School of Stomatology, Zhejiang University School of Medicine. All panoramic radiographs were captured with a Dentsply Sirona (Bensheim, Germany) dental X-ray instrument. The inclusion criteria for panoramic radiographs were as follows: 1) patients age >18 years old; 2) at least one MFP with fully developed dental roots must be present; 3) no prior root canal treatment of the MFP; 4) panoramic radiographs and CBCT scans were both available for the MFP. Exclusion criteria were as follows: 1) poor quality panoramic radiographs; 2) detectable external resorption, internal resorption, or apical foramen destruction in the MFP; 3) significant buccolingual inclination of the MFP. The study protocol was approved by the Ethics Committee of the Stomatology Hospital, School of Medicine, Zhejiang University, China, and all procedures were performed in accordance with the Helsinki Declaration.

The panoramic radiographs were screened by three independent dentists, and MFP root canal morphology was then examined in CBCT images. The MFP images were divided into single or non-single groups, the former group including MFPs with only one root canal（Figure1.A） and the latter group containing MFP images with multiple root canals and more complex root canal systems(Figure1.B-F). The single group contained 459 MFPs, while 366 MFPs were classified to the non-single group. The results of root canal on CBCT were used as the gold standard.

Prior to the observation, the intraclass correlation for the evaluation of root canal was examined. The three dentists evaluated this for 50 teeth. The resulting intraclass correlation coefficient was 1.000.

2.2 Automatic Root Canal Classification

The workflow of automatic root canal classification could be divided into following steps. In the first step, all panoramic radiographs were resized to 1440 ×2976 pixels. When the MFP had only one root canal on CBCT images, MFPs were divided into contact group, the MFPs were labeled “one” with the open-source software Labellmg. In total, 1630 MFPs were included, with 1304 for training, 163 for validation and 163 for testing.

In the second step, the U-NET deep learning model is applied for segmentation of teeth[12], which builds a U-shaped configuration from convolutional network layers with skip connection, which is widely used in biomedical image segmentation to prevent loss of information from layers. The symmetrical encoder-decoder architecture includes down-sampling blocks for extracting features and up-sampling blocks to infer the segmentation at the same resolution. The network hierarchically extracts low-level features, then recombines them into higher-level features in the encoder. The skip connection between the falling and the rising part of the U-shape, it expands the contextual information and performs element-wise classification from multiple features in the decoder. Details of the model architecture are shown in Figure 2. After U-Net segmentation of individual teeth, the MFPs can be located and the ROI is cropped for classification of canal architecture.

In the third step, two convolutional neural network (CNN) model, the ResNet-101 model and ResNet-152 model were applied in root canal classification. we also introduced the convolutional block attention module (CBAM; Figure 3) [13] to force the network to select class-specific features through a combination of channel and spatial attention strategies which we named ResNet-101(CBAM). A channel attention module was included to perform average and max pooling operations in order to extract more high-level features. To account for the influence of different MFP locations, the spatial attention module was employed to examine different places within the same channel, enhancing the representation of key features.

The three models were trained and tested with PyTorch on a GeForce RTX 2080Ti running on the Ubuntu (version 18.04) operating system with the CUDA10.1 toolkit and CUDNN8.0.5 library. A stochastic gradient descent (SGD) optimization strategy was used during model training. The initial learning rate was 0.001, epochs were 100, batch size was 10, and the learning rate optimization strategy was ReduceLROnPlateau.

2.3 Diagnostic performance analysis

To assess the performance of the proposed framework for classification, we used Precision (PR), Recall (RE), Accuracy (ACC), F1-score, and Area Under Curve (AUC) of Receiver Operating Characteristic curves, defined below:

Results:

in which TP represents true positives, TN represents true negatives, FP represents false positives, and FN represents false negatives.

Panorama radiograph results for MFPs were interpreted by a panel of three dentists with <3 years experiences and two dentists with >10 years. Metrics based on the results of MFP detection were used to estimate the performance of the deep learning network for comparison with the assessments of the dentists.

2.4 Statistical analysis

Diagnostic performance was compared using the Mann-Whitney U Test. AUC values were compared using McNemar’s chi-square analysis. Statistical analyses were performed with IBM SPSS Statistics 24.0, The level of significance was set at p < 0.05.

3.1 Accuracy of the deep learning networks

Measurements of the deep learning system’s diagnostic performance in determining whether an MFP has a single root canal or not are shown in Table 1. Using the ResNet-101(CBAM) system resulted in 85% accuracy, with a positive predictive value of 81.4%, and a negative predictive value of 82.35%. Similarly, the ResNet-101 system without CBAM showed an accuracy of 82%, a positive predictive value of 72.09%, and a negative predictive value of 77.50%. By contrast, the ResNet-152 system showed a relatively lower accuracy of 80%, with a positive predictive value of 62.79%, and a negative predictive value of 72.97%. These results indicated that the ResNet-101 with CBAM could provide higher accuracy in classifying MPF root canal architecture than other models.

In addition, we generated a normalized confusion matrix of the accuracy of the proposed model (Figure 4) showing the ratio of predicted results to observed results of classification by ResNet-101(CBAM), with a value of 1 indicating 100% accuracy.

Table 1. Accuracy metrics for deep learning model classification of MFP root canal image data.

methods	results（%）
methods	Accuracy	AUC	Precision	Recall	F1-score
ResNet-101(CBAM)	85.00	87.21	83.33	81.40	82.35
ResNet-101	82.00	85.01	83.78	72.09	77.50
ResNet-152	80.00	83.29	87.10	62.79	72.97

3.2 Comparison of manual and automatic classifications

Comparison of manual (dentist) and automatic (AI) classification of single or not single MFP root canal number indicated that deep learning-based classification of radiography imaging data could provide significantly higher accuracy than human predictions, which was showed in Table 2. However, it should be noted that average accuracy of inexperienced dentists was significantly lower than that of experienced dentists (73.0% vs 85.88%; inexperienced vs experienced; p<0.05), and the predictions of experienced dentists did not significantly differ from those of ResNet-101(CBAM).

Table 2. Comparison of deep learning model and dentists’ accuracy in classifying MFPs as single or not single root canals by radiograph

Parameter	inexperienced	experienced	average of all	ResNet-101(CBAM)
Accuracy(%)	73.00	85.88	78.16	84.04
AUC(%)	74.29	86.33	80.31	87.21
Precision(%)	75.12	86.45	79.66	82.00
Recall(%)	84.44	90.00	86.67	91.11
F1-score(%)	77.37	88.09	81.67	86.31

Knowledge of the variations in root canal morphology is a clinically important factor for proper endodontic treatment. In clinical practice, dentists often rely on experience to evaluate the root canal number of the mandibular first premolar with the naked eye based on panoramic radiographs. It is difficult for dentist to recognize the root using the two-dimensional panoramic radiographs especially when there is buccal lingual side bifurcate in the root canal. the cone-beam CT should be used as a first hand supplement when the roots seem irregular in 2D X-ray films. In our study, the accuracy of the dentists with low experience is only 73%, it is much lower than the accuracy of the dentists with ten years working experience. And the ICC of five dentists was only 0.32 in our experiment, indicating that the five dentists' judgments regarding the real root canal number based on the same panoramic radiographs were in poor agreement. Inexperienced doctors often treat non-single root canals as single root canals, the precision is 75.12% which will lead to root canal leftover and greatly reduce the success rate of root canal treatment. If a single root canal is misjudged as a non-single root canal, and the patient is required to undergo CBCT examination for confirmation, although the occurrence of root canal omission is effectively avoided, it also increases the medical expenses of the patient and receives too much radiation.

In our experiment, with the purpose of suppressing the interference in irrelevant regions, we applied U-Net to detect specific tooth to obtain key regions of interest. Subsequently, the pre-trained ResNet-101 model is used for the classification of ROI, which is proven to be effective and improves the classification precision and recall of the neural network. Compared to DenseNet model, by increasing the depth, ResNet can find the optimal shallow network and ensure that the deep network will not degrade due to the increase of the number of network layers. [14]Compared with the serial structure of ordinary networks, ResNet-101 uses residual connections that can leverage feature information in shallow layers, allowing the network to preserve global information while extracting deep features. By increasing depth, ResNet can identify the optimal shallow network to ensure that the deep network does not degenerate as the number of network layers increases[15]. Since the features of single canal and non-single canal are less distinguishable, we introduce an attention module to enhance the extraction ability of crucial features by jointly using channel and spatial attention strategy (CBAM; Figure 2) to force the network to select class-specific features through a combination of channel and spatial attention strategies[13]. Similar to SE-Net, a channel attention module was included to perform average and max pooling operations in order to extract more high-level features. To account for the influence of different MFP locations, the spatial attention module was employed to examine different places within the same channel, enhancing the representation of key features. This improvement has increased the accuracy of the model from 82.00% to 85.00%, and AUC is increased from 85.01% to 87.21%. The experimental results show that the two-stage approach produces the most accurate results. Therefore, our method could be clinically meaningful and have promising prospects in practical clinical examination.

However, there are still many deficiencies in this study. There are not enough training set data for the deep learning model and the number of doctors participating in panoramic reading is also insufficient; moreover, the age, gender and race of the patients were not taken into account, which would have an impact on the results, and further experimental verification is needed.

The deep learning was performed with panoramic radiographs that are easy to obtain from healthy patients. This study is one of a series of studies aimed at using deep learning artificial intelligence to automate diagnoses on panoramic radiography. In the future, we will use this model to study the root canal morphology of all tooth positions of complete teeth. At the same time, apical slices were used in the experiment to improve the accuracy of the results.

The deep learning systems showed high levels of performance in differentiating whether the mandibular first premolar has single root canal or not based on panoramic radiographs. It is a promising artificial intelligence approach that may assist dentists in preoperative diagnosis.

Acknowledgments

The authors want to thank the Radiology Department and Electronic Information Department for their cooperation and help.

Author contributions

All authors conceived of the study and have read and agreed to the published version of the manuscript. Conceptualization, HZ, FZ and FW; methodology, HZ, FZ and FW; software, DF, KS.; formal analysis, DF; investigation, DF; resources, DF; data curation, DF, KS; writing—original draft preparation, DF; writing—review and editing, DF, KS, HZ, FZ.and FW; supervision, HZ, FZ.and FW; project administration, H.Z., F.Z. and F.W.; funding acquisition, HZ, FZ and FW.

Funding

This work was supported by Natural Science Foundation of Zhejiang Province (TGY23H140016), Exploration and R&D project of the Affiliated Stomatology Hospital of Zhejiang University School of Medicine (RD2022YFYB01),

Availability of data and materials

The data are not publicly available due to privacy. The data presented in this study are available on request from the corresponding author.

Ethics approval and consent to participate

The protocol was reviewed and approved by Ethnics Committee of Stomatology Hospital, School of Medicine, Zhejiang University, China. (Ethics Approval No. 2022-033). The study procedure was performed in accordance with Helsinki Declaration. All methods were performed in accordance with relevant guidelines and regulations. All the patients provided informed consent for their data to be used for research purposes.

Consent for publication

Not applicable.

Competing interests

The authors declare no conflict of interest. The funder had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.

Aung NA-O, Myint KA-O: Diagnostic Accuracy of CBCT for Detection of Second Canal of Permanent Teeth: A Systematic Review and Meta-Analysis. (1687-8728 (Print)).
Wolf TG, Anderegg AL, Wierichs RJ, Campus G: Root canal morphology of the mandibular second premolar: a systematic review and meta-analysis. BMC Oral Health 2021, 21(1):309.
Bürklein S, Heck R, Schäfer E: Evaluation of the Root Canal Anatomy of Maxillary and Mandibular Premolars in a Selected German Population Using Cone-beam Computed Tomographic Data. (1878-3554 (Electronic)).
Liu N, Li X Fau - Liu N, Liu N Fau - Ye L, Ye L Fau - An J, An J Fau - Nie X, Nie X Fau - Liu L, Liu L Fau - Deng M, Deng M: A micro-computed tomography study of the root canal morphology of the mandibular first premolar in a population from southwestern China. (1436-3771 (Electronic)).
Brea G, Gomez F, Gomez-Sosa JF: Cone-beam computed tomography evaluation of C-shaped root and canal morphology of mandibular premolars. BMC Oral Health 2021, 21(1):236.
Nardi C, Calistri L, Grazzini G, Desideri I, Lorini C, Occhipinti M, Mungai F, Colagrande S: Is Panoramic Radiography an Accurate Imaging Technique for the Detection of Endodontically Treated Asymptomatic Apical Periodontitis? (1878-3554 (Electronic)).
Ying S, Wang B, Zhu H, Liu W, Huang F: Caries segmentation on tooth X-ray images with a deep network. J Dent 2022, 119:104076.
Lian L, Zhu T, Zhu F, Zhu H: Deep Learning for Caries Detection and Classification. Diagnostics (Basel) 2021, 11(9).
Takahashi T, Nozaki K, Gonda T, Mameno T, Ikebe K: Deep learning-based detection of dental prostheses and restorations. Sci Rep 2021, 11(1):1960.
Fukuda M, Inamoto K, Shibata N, Ariji Y, Yanashita Y, Kutsuna S, Nakata K, Katsumata A, Fujita H, Ariji E: Evaluation of an artificial intelligence system for detecting vertical root fracture on panoramic radiography. Oral Radiology 2019, 36(4):337-343.
Hiraiwa T, Ariji Y, Fukuda M, Kise Y, Nakata K, Katsumata A, Fujita H, Ariji E: A deep-learning artificial intelligence system for assessment of root morphology of the mandibular first molar on panoramic radiography. Dentomaxillofac Radiol 2019, 48(3):20180218.
Ronneberger O, Fischer P, Brox T: U-Net: Convolutional Networks for Biomedical Image Segmentation. In: Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015: 2015// 2015; Cham: Springer International Publishing; 2015: 234-241.
Woo S, Park J, Lee J-Y, Kweon IS: CBAM: Convolutional Block Attention Module. In: Computer Vision – ECCV 2018: 2018// 2018; Cham: Springer International Publishing; 2018: 3-19.
Warin K, Limprasert W, Suebnukarn S, Jinaporntham S, Jantana P: Performance of deep convolutional neural network for classification and detection of oral potentially malignant disorders in photographic images. (1399-0020 (Electronic)).
Park WA-O, Schwendicke FA-O, Krois JA-O, Huh JA-O, Lee JA-O: Identification of Dental Implant Systems Using a Large-Scale Multicenter Data Set. (1544-0591 (Electronic)).

No competing interests reported.

An artificial intelligence system for detecting the root canal morphology of the mandibular first premolar on panoramic radiographs

Status:

Version 1

Abstract

Figures

1. Background

2. Methods

3. Results

4. Discussion

5. Conclusions

Declarations

References

Additional Declarations

Status:

Version 1