Applying deep learning techniques to identify tonsilloliths in panoramic radiography

Abstract

Tonsilloliths can be seen on panoramic radiographs (PRs) as deposits located on the middle portion of the ramus of the mandible. Although tonsilloliths are clinically harmless, the high risk of misdiagnosis leads to unnecessary advanced examinations and interventions, thus jeopardizing patient safety and increasing unnecessary resource use in the healthcare system. Therefore, this study aims to meet an important clinical need by providing accurate and rapid diagnostic support. The dataset consisted of a total of 275 PRs, with 125 PRs lacking tonsillolith and 150 PRs having tonsillolith. ResNet and EfficientNet CNN models were assessed during the model selection process. An evaluation was conducted to analyze the learning capacity, intricacy, and compatibility of each model with the problem at hand. The effectiveness of the models was evaluated using accuracy, recall, precision, and F1 score measures following the training phase. Both the ResNet18 and EfficientNetB0 models were able to differentiate between tonsillolith-present and tonsillolith-absent conditions with an average accuracy of 89%. ResNet101 demonstrated underperformance when contrasted with other models. EfficientNetB1 exhibits satisfactory accuracy in both categories. The EfficientNetB0 model exhibits a 93% precision, 87% recall, 90% F1 score, and 89% accuracy. This study indicates that implementing AI-powered deep learning techniques would significantly improve the clinical diagnosis of tonsilloliths.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-10489-x.

Introduction

Soft tissue calcifications (STCs) are radiopaque formations that occur as a result of the deposition of calcium salts. They are typically asymptomatic and can be observed on panoramic radiographs (PRs) acquired during routine dental examinations¹. Calcified lymph nodes, tonsilloliths, cysticercosis, and vascular calcifications are examples of dystrophic calcifications that appear radiopaque in the head and neck region.

These calcifications are generally asymptomatic². Tonsiloliths, often called tonsil stones, are calcified deposits that typically develop in the tonsil crypts of older persons, either on one side or both sides. In PRs, the presence of several radiopacities, resembling rice grains or irregular clusters, can be observed in the central sections of the ramus. It seldom attains significant dimensions, and its density is similar to that of compact bone. It is common for hard and soft tissue features to overlap on radiographic pictures in this anatomical location, which makes interpretation difficult. Dentists have the duty to identify tonsilloliths when they are overlapping with anatomical structures and to distinguish them from normal anatomical formations that appear radiopaque, dental abnormalities, bone abnormalities, foreign objects, artifacts, and other non-specific findings^3,4.

PRs are utilized in dental practice due to their numerous benefits, including low radiation exposure, simple application, affordability, and convenient availability. They continue to hold their position as the predominant imaging technique utilized in the diagnosis and treatment planning of numerous oral and maxillofacial diseases⁵. Tonsilloliths should be identified during regular PRs and distinguished from clinical and anatomical abnormalities in order to establish the necessity for therapy.

Currently, artificial intelligence (AI) technologies are quickly gaining popularity in the field of dentistry. The efficient utilization of vast amounts of data is contingent upon the implementation of machine learning. As a subfield of machine learning, deep learning employed artificial neural networks. The transmission of data representations layer by layer in these networks enables the development of greater abstraction and complexity⁶. Object detection, image classification, and segmentation are the most prevalent applications of Convolutional Neural Networks (CNNs) in this field. Convolutions are employed by these networks to derive features from images, such as colors, textures, edges, geometric shapes, and macroscopic structures. CNNs have been demonstrated to be extremely effective in both two-dimensional and three-dimensional radiographs in the field of dentistry, according to research⁷.This approach helps prevent inexperienced or careless physicians from overlooking asymptomatic lesions, leading to more accurate radiography interpretations. As a result, clinical diagnosis and treatment planning benefit from improved recall and reduced errors. Furthermore, its objective is to streamline the tasks of physicians, enhance patient care, and optimize treatment outcomes by actively supporting the education of physicians during their training^8,9.

The objective of this study is to enhance clinical decision-making and optimize patient health management by utilizing AI-based deep learning techniques. This powered by AI approach will enhance the precision of tonsillolith diagnosis and reduce errors in clinical assessments.

Methods

Study design

A retrospective analysis was conducted on the panoramic images of patients who underwent panoramic radiography at Dicle University Faculty of Dentistry for a variety of reasons between January 2018 and December 2022. Patients with at least one tooth in the molar region were included in the investigation. The following were excluded from the study: edentulous patients, PRs that contained foreign objects other than dental restorations, phantom images, and superpositions caused by patient positioning.

Tonsilloliths often manifest as numerous, tiny, indistinct radiopaque areas on PRs. Syphilis, enostosis, mycosis, or calcifying granulomatous disease should be taken into account during the differential diagnosis. The diagnostic criteria for the tonsilloliths are as follows: The image of single or multiple radiopacities obscures the midportion of the mandibular ramus in the area where the dorsal surface of the tongue intersects the ramus¹⁰. In order to construct a comprehensive image dataset, two maxillofacial radiologists (EK and SCB), each with eleven years of experience, collected PRs with and without tonsilloliths, reaching consensus on each image. Out of the 172 films that were suspected of having tonsillolith, the study included 150 PRs that both radiologists accepted as fulfilling the diagnostic and inclusion criteria in the study. The dataset comprised an overall of 275 PRs, of which 125 lacked tonsillolith and 150 had tonsillolith (Fig. 1). All PRs were taken using the standard exposure mode (66 kV, 6 mA, 16 s) on a Planmeca ProMax (Planmeca ProMax, Helsinki, Finland) device.

Fig. 1.

a An image showing a tonsillolith located on the right ramus of the mandible (white arrow). b Panoramic image without tonsillolith.

Convolutional neural network

In the study, a deep learning-based classification system was proposed as a way to automatically identify tonsilloliths in PRs. After evaluating the presence or absence of tonsilloliths, individuals were classified and each image was divided into two main classes: tonsillolith-present and tonsillolith-absent. In order to prepare the images appropriately for model training, a series of preprocessing methods were applied, including illumination adjustments and resizing. Lighting adjustment was performed to reduce the contrast and brightness differences that may occur in different radiographic shots. This process was performed by adaptive histogram equalization using the OpenCV library. The pixel values ​​of the images were normalized between 0 and 1 to provide a more stable learning process while the model was being trained. The generalization ability of the model was increased by using data augmentation techniques. In this step, random horizontal flip, random rotation up to ± 5° and slight brightness and contrast changes were applied to the training data. In the model selection process, ResNet and EfficientNet CNN models were evaluated because they provide a deep network model and are frequently preferred in image processing applications. The image dimensions were adjusted according to the input size accepted by each model. While the images were resized to 224*224 for ResNet, the resizing for the EfficientNet model was performed according to all models between B0-B7. The learning ability, complexity and problem compatibility of each model were evaluated. Hyperparameter optimization was used to determine the optimal learning rate, sample size and number of epochs for the selected models to reduce the probability of overfitting. In the model training phase, training curves and loss functions representing the training progress were obtained. The effectiveness of the models was evaluated using accuracy, recall, precision and F1 score measurements after the training phase. The perspectives of dentists who are authorities in their fields were taken into account throughout the development of the model. The results obtained during the development of the model were validated with the insights provided by these authorities. Figure 2 illustrates this methodical approach.

Fig. 2.

Designed System Flow.

After the completion of the data preprocessing stage, the model selection was performed according to the characteristics of the dataset and the investigated problem. In deep learning-based image processing applications, the model selection stage is critical because it requires considering the characteristics of the dataset and the investigated problem. During the model selection procedure, models widely used in the literature on image processing in healthcare were analyzed. For this research, ResNet-based models ResNet18 and ResNet101 were preferred over EfficientNet-based models EfficientNetB0 and EfficientNetB1. After model selection, the dataset was divided into two subsets as training and testing. In this process, 80% of the dataset was used for training and 20% for testing. In the data splitting process, it was aimed to represent the “tonsillolith-present” and “tonsillolith-absent” classes in a balanced way in the training and testing data. For this purpose, the distribution of the classes was reviewed, and a proportional separation was applied. Thus, it was aimed to evaluate the fair representation between the classes in both the training process and the testing phase of the model. The aim of this applied approach is to prevent biases that may arise from class imbalance and to ensure that the model has sufficient information in learning both classes. By adhering to this division, all training and evaluation processes were carried out over this structure.

The ResNet18 and ResNet101 models address the difficulties associated with training deep neural networks by utilizing the residual learning framework. This framework increases the efficiency of the learning process in order to circumvent the vanishing gradient issue that can arise during deep network training. Due to its reduced number of layers, ResNet18 demands an expedited training procedure and decreased computational resources. Despite having a limited number of parameters, this model is capable of achieving high levels of accuracy. In contrast, ResNet101 exhibits a more extensive architecture, facilitating the acquisition of more intricate features and enhancing its performance on sizable datasets¹¹. An architecture of CNN known as EfficientNet, which was created in 2019 by Quoc V. Le and Mingxing Tan¹²efficiently classifies images. EfficientNet optimizes computational expenses and enhances model performance through the implementation of its compound scaling methodology, which involves proportional scaling of the model’s depth, breadth, and resolution. Model sizes, complexity, and capacities spanning from “B0” to “B7” were incorporated into the design of EfficientNet, which performs exceptionally well on extensive benchmarks like ImageNet. On account of these characteristics, EfficientNet is regarded as a significant advancement in the domain of deep learning¹².

Optimization of hyperparameters to maximize the performance of the model determines the classification and generalization capacity of deep learning models. In this research, a methodology was used where certain hyperparameters were kept constant to increase the generalizability and learning efficiency of the model. The batch size for each model was configured as 8, while the number of epochs was fixed as 50. Considering that the learning rate is a parameter that varies between models, it was calculated independently and autonomously for each model. The plots of the calculated learning rates are presented in Fig. 3. This technique allows the model to dynamically adapt the learning process throughout its iterations. The aim of incorporating the modified hyperparameters into the model training procedures is to optimize the performance of each model.

Fig. 3.

Model based learning rate graphs.

An analysis of the diagrams presented in Fig. 3 reveals that the learning rates for the ResNet18, ResNet101, EfficientNetB0, and EfficientB1 models are, respectively, 1.2e-2, 1.2e-2, 1.2e-2, and 1.3e-3. After selecting the hyperparameters, we conducted model training by partitioning the data set into 80% training data and 20% training data. Model training was performed on the Google Colab platform. Among the GPU types provided in the Colab environment, T4 GPU was used. Hardware assignments are performed dynamically by Colab. The training process was carried out with Python 3.10 and PyTorch framework. OpenCV, Pillow and torchvision libraries were used for image processing operations, and matplotlib libraries were used for graphical analysis. All CNN architectures on ImageNet were initialized with pre-trained weights. This transfer learning approach increased the generalization success of the model by providing better learning with a small dataset.

Following the completion of model training for each model, performance evaluation metrics including accuracy, precision, recall, and F1 score were employed to assess their performance. The mathematical expressions for these metrics are presented in Table 1.

Table 1.

Performance evaluation metrics.

Metrics	Mathematical Expressions
Accuracy
Precision
Recall
F1 Score

True Positives (TP): Number of positive instances correctly identified by the model, True Negatives (TN): Number of negative instances correctly identified by the model, False Positives (FP): Number of negative instances incorrectly identified as positive by the model, False Negatives (FN): Number of positive instances incorrectly identified as negative by the model

Results

This study involved the development of a diagnostic support system utilizing deep learning techniques on PRs, specifically focusing on cases with and without tonsil stones. The data preprocessing stage involved training images using CNN models, namely ResNet18, ResNet101, EfficientNetB0, and EfficientNetB1, with hyperparameters that were optimized. The CNN models were comparatively evaluated using metrics such as accuracy, precision, recall, and F1 score, which are displayed in Table 2.

Table 2.

Performance evaluation metrics after the classification process.

Class model		Precision (%)	Recall(%)	F1 Score(%)	Accuracy (%)
Tonsillolith-absent	ResNet18	85	92	88	89
	ResNet101	69	72	71	73
	EfficientNetB0	85	92	88	89
	EfficientNetB1	75	96	84	84
Tonsillolith-present	ResNet18	93	87	90	89
	ResNet101	76	73	75	73
	EfficientNetB0	93	87	90	89
	EfficientNetB1	96	73	83	84

Table 2 results show that ResNet18 and EfficientNetB0 models show high classification performance for both classes. For the “tonsillolith-absent” class, both models achieved 85% precision, 92% recall and 88% F1-score values, while for the “tonsillolith-present” class, 93% precision, 87% recall and 90% F1-score values ​​were obtained. These measurements show that, beyond just their accuracy rates, the models are both accurate in their detection (recall) and reliable in their predictions. This success shows that ResNet18 and EfficientNetB0 can effectively differentiate between people with “tonsillolith-present” and “tonsillolith-absent”. In both models, balanced rates of precision, recall, F1 score, and accuracy are observed. The models demonstrated remarkable precision and recall in distinguishing individuals with and without tonsillolith cases. The reduced likelihood of case misses in the tonsillolith-absent class is supported by the EfficientNetB1 model’s high recall for this class. Although the model’s recall is comparatively lower for the tonsillolith class, this suggests that it may erroneously classify certain cases that are in fact positive as negative. This implies that enhancing the dataset or implementing supplementary modifications could be required to optimize the model’s performance in this specific category. The comparative underperformance of the ResNet101 model in relation to alternative models suggests that the model might not be well-suited for the specific classification task at hand. This is subject to variation dependent upon the efficacy of deep learning models across different tasks and the meticulous tuning of model architecture and hyperparameter configurations. There are multiple potential strategies that could be implemented to enhance the performance of ResNet101. These include augmenting the diversity of the dataset utilized during model training or altering the depth of the model to better correspond with the task requirements. Figure 4 illustrates the confusion matrices utilized to derive the performance evaluation metrics presented in Table 2.

Fig. 4.

Confusion matrices of the models after the classification process.

In addition to these results, a series of comparative analyses were used to determine whether the performance differences between the models were statistically significant rather than evaluating them only on the basis of accuracy rates. In the assessments made, Cohen’s d effect size values ​​and corresponding p-values ​​between the models were found. These analyses are of critical importance in terms of showing how different the models are from each other. The comparison between EfficientNetB0 and ResNet101 shows that EfficientNetB0 model has a better classification ability for this task (Cohen’s d = 0.88) and statistical significance (p = 0.044).

Although the Cohen’s d values ​​found in the other model matches are generally moderate, the fact that the p values ​​remain above the significance level indicates that there is no statistically significant difference, but there may be a significant performance change at a practical level. Relevant statistical comparisons are presented in Table 3.

Table 3.

Statistical comparison of model performances (Cohen’s d and p-values).

Model Comparison		Cohen’s d	p-value
EfficientNetB0 vs. ResNet18		0.00	1.000
EfficientNetB0 vs. ResNet101		0.88	0.044
EfficientNetB0 vs. EfficientNetB1		0.47	0.118
ResNet18 vs. ResNet101		0.60	0.092
ResNet18 vs. EfficientNetB1		0.50	0.108

The exceptional classification performance of the ResNet18 model in both the tonsillolith-absent and tonsillolith-present classes is illustrated by the confusion matrices in Fig. 4. It gains attention due to its ability to detect tonsillolith cases with a reduced number of false-negative (FN) outcomes in comparison to alternative models. While ResNet101 exhibits superior accuracy in tonsillolith case classification, it also demonstrates a considerable false-positive (FP) rate. It seems to exhibit diminished performance in the classification of tonsillolith-absent data as a consequence. When EfficientNet models are analysed, it is seen that EfficientNetB1 model achieves a high accuracy rate in both classes and provides a balanced result, while EfficientNetB0 model shows a high level of accuracy in tonsillolith case classification. When the overall performance is analysed, it is seen that the EfficientNetB0 model shows a robust and balanced performance in both classes. In contrast, the ResNet models and the EfficientNetB1 model display discernible advantages and disadvantages in specific classes.

In addition to confusion matrices, we monitored the error levels of the models and assessed the efficacy of the training procedure via training loss graphs. The error rates as well as the efficacy of the models on the training set are depicted in these graphs. Furthermore, the stability of the learning algorithm and its propensity for overfitting are illustrated by these curves. The training loss graphs analyzed in this investigation are depicted in Fig. 5.

Fig. 5.

Training loss graphs of classification models.

Training loss is a metric that indicates how models’ efficacy on the training data set evolves with time. Analyzing the training loss graphs presented in Fig. 5 indicates that during the initial stages, training loss decreases precipitously for all models. As the process progressed, the error rates gradually decreased. This signifies that a certain level of saturation has been attained by the learning function. It is anticipated that ideal learning processes will exhibit a gradual reduction in training loss and eventually reach a value near zero. Training losses for every model exhibit a progressive decline, culminating in consistently low loss values. This indicates that the models are capable of generalizing the data set without being overfit.

Discussion

The current study examined a variety of CNN models to evaluate their efficacy, learning capacity, and complexity in the context of tonsillolith identification. To the best of our knowledge, there is no research on the identification of tonsilloliths in the literature. We are of the opinion that this investigation will be the first to address this subject.

Deep learning models have been employed in numerous published research projects to explore different aspects of dentistry practice. AI systems are being utilized to research the identification of dental anomalies, such as taurodontism¹³. Jae-Hong¹⁴ showed that CNN architectures may be utilized for identifying and categorizing dental implant systems. However, the number of studies that have employed deep learning techniques to identify calcifications in soft tissues of the head and neck is relatively small^15–18. Ishibashi et al.¹⁹ employed deep learning techniques to identify sialoliths in panoramic radiography. They developed three distinct models and assessed the accuracy of these models in detecting sialoliths by organizing the datasets in a differing manner. They attained the highest detection rates (recall: 85%, precision: 100%, and F score of 91.9%) for sialoliths in panoramic radiographs with two models. Tonsilloliths were identified by EfficientNetB0 in our investigation with a recall rate of 87%, precision of 93%, and F score of 90%. Despite the fact that our dataset was smaller, the recall values for ResNet18 and EfficientNetB0 in our study were 87%, which was slightly higher than those in the study conducted by Ishibashi et al.¹⁹. In contrast, the highest precision value achieved in our analysis was 90% (ResNet18 and EfficientNetB0), which was lower than that of the previous study. The different number of epochs employed in the investigations may be the cause of this situation. Vinayahalingam et al.²⁰ investigated the detection of carotid artery calcifications on panoramic radiographs through the application of deep learning techniques. The results indicate that the proposed method, which was designed using Faster R-CNN and Swin Transformer, achieved an F1 score of 89%, a recall of 88%, and a rate of precision of 89%. It is possible to conclude that the study results are consistent when observed in comparison to the results of our study, as similar values were obtained.

By enabling visualization of the maxillary and mandibular arches and their adjacent structures on a single film, panoramic radiography is an extensively utilized technique across all domains of dentistry^10,21. Soft tissue calcifications, including tonsilloliths, sialoliths, lymph node calcifications, and calcified atheroma plaque, may also be detected via panoramic imaging. However, it is probable that inexperienced physicians will fail to notice it¹⁵. According to studies utilizing panoramic imaging, tonsillitis prevalence ranged from 7.3 to 13.4%^22–25. Studies using computed tomography have shown a higher occurrence rate. Fauroux et al.²⁶ reported a prevalence rate of 24.6%, whereas Oda et al.²² reported a higher prevalence rate of 46.1%. The identification of this relatively common soft tissue calcification is accomplished through the examination of its spatial distribution on the panoramic film. They are capable of being superimposed on the mandible as well as the adjacent soft tissue. Distinguishing palatine tonsilloliths from parotid sialoliths, submandibular salivary ducts, and phleboliths, which are calcified structures that should not be present, is a critical medical judgment for physicians²⁴.

The application of AI techniques can provide physicians with support in making decisions pertaining to patient care²⁷. Song et al.¹⁵ conducted a study comparing the reading speed and detection success of two oral and maxillofacial radiologists and two general dentists when analyzing panoramic films with soft tissue calcification (sialolith, lymph node calcification and carotid artery calcification) using Fast R-CNN. They reported a recall of 77.4%, and a specificity of 71.4% for carotid artery calcification (CAC). The study also showed that expert radiologists took less time to analyze films during the second reading, as compared to the first reading conducted without the use of AI. Additionally, general dentists exhibited improved ability to detect STCs. Kats et al.¹⁶ employed the Fast R-CNN technique on a limited dataset to ascertain the presence of CAC on PRs. Using a limited set of 65 panoramic photos, they were able to identify CACs with a recall of 75%, a specificity of 80%, and an accuracy of 83%. Amitay et al.¹⁷ used 500 PRs in their study to automatically identify CACs using a CNN method implemented using a transfer learning technique. The algorithm employs deep learning CNN with a transfer learning approach. It successfully obtained accurate labels for each corner and achieved a recall of 82% and a specificity of 97% for individual arteries. Additionally, it achieved a recall of 87% and a specificity of 97% for individual patients. Ajami et al.¹⁸ favored the use of CBCT images for the automated identification and localization of CACs. Their analysis yielded a recall of 94.2% and a specificity of 96.5%. The ResNet18 model exhibited outstanding classification accuracy for both the tonsillolith-absent and tonsillolith-present classes in our investigation. The precision for both tonsillolith-absent and tonsillolith-present classes ranged from 85 to 93%. The recall ranged from 92 to 87%. The F1 score ranged from 88 to 90%. The accuracy for both classes was 89%. Furthermore, it has a lower FN ratio when compared to other models. Although the ResNet101 model shows impressive accuracy in identifying tonsillolith, it also has a notable FP rate. The EfficientNetB1 model has a notable level of precision in both the tonsillolith-absent and tonsillolith-present groups, achieving an accuracy of 84%. Conversely, the EfficientNetB0 model has demonstrated exceptional accuracy, particularly in detecting tonsillolith instances. When evaluating the overall performance, the EfficientNetB0 model exhibits a comprehensive and strong performance with a precision of 93%, a recall of 87%, a F1 score of 90%, and an accuracy of 89%. However, the ResNet and EfficientNetB1 models exhibit distinct strengths and weaknesses. The high performance of the EfficientNetB0 model in the limited size and unbalanced class distribution dataset used in the study can be related to some of the technical features of its architectural structure. The EfficientNet architecture uses a compound scaling approach in which network dimensions such as depth, width, and resolution are scaled together and evenly. In this way, the model can maintain high accuracy with less computational cost while maintaining parameter efficiency. In addition, due to its smaller and optimized architecture, EfficientNetB0 can be more resistant to overfitting and show a more stable generalization performance on small-to-medium-sized data sets. The relatively small number of parameters makes it easier for the model to learn more robustly, especially in health data sets with a limited number of samples. EfficientNetB0 has been able to achieve high F1 scores and accuracy values in both tonsillolith-present and -absent cases as a result of these technical advantages.

The implementation of our proposed model in the dental clinic will assist inexperienced dentists in the identification of tonsilloliths during radiological examinations, thereby preventing the need for unnecessary radiological investigations. In the literature, accuracy rates of 85% and higher in assistive decision support systems are seen as encouraging^28,29. Since our study has an average accuracy of 89%, it serves as a significant and useful starting point for situations like diagnosing tonsilloliths. However, we would like to emphasize that this performance should be positioned as a supportive tool rather than completely taking over the physician’s decision-making processes. In this direction, it is aimed to test the model in larger and multi-center data sets and to increase its clinical validity by using it together with expert evaluations in future studies.

Our study has some limitations. We utilized panoramic imaging to identify tonsilloliths. Panoramic radiography is a frequently utilized technology in dentistry that enables the visualization of both the upper and lower dental arches, as well as the surrounding structures, on a single image. Nevertheless, it is imperative that we should not disregard any alterations that may arise as a result of the patient’s location³⁰. The PRs utilized in our investigation were exclusively obtained from a single center. We included cases in which the presence of tonsilloliths was accepted by two oral and maxillofacial radiologists, as we did not re-evaluate the cases with three-dimensional imaging. Consequently, our dataset was restricted. Additionally, inter-observer variability analysis was not conducted due to the concurrent nature of the evaluation. The probability of obtaining successful results will be increased by incorporating multicenter studies into the dataset. Additionally, all PRs were acquired using the same device. It is uncertain how the model’s performance would be impacted, as images obtained from various devices were not utilized. The diagnosis of tonsilloliths was limited to cases in which two dentomaxillofacial radiologists reached a consensus on panoramic images in this study. There is a possibility that certain cases were misidentified, as verification was not conducted using three-dimensional imaging. Multidetector computed tomography or cone-beam computed tomography imaging may be required to ascertain the location of a suspected tonsillolith¹⁰. Nevertheless, this may lead to an elevated level of radiation exposure for the patient. We incorporated tonsillolith cases that were inadvertently identified on panoramic films that were obtained from patients for the purpose of investigation in our study. We claim that three-dimensional imaging has the capability to retrieve a more extensive collection of data as a result of its greater prevalence in computed tomography scans. Therefore, it is intended to conduct future research on CBCT images that utilize a more extensive dataset. Furthermore, it is intended to improve the interpretability of deep learning models by employing explanatory visualization techniques, including Grad-CAM + + and comparable methodologies. It is anticipated that these methods will significantly enhance the assessment of the reliability of model decisions, particularly in the field of medical imaging. In the same vein, semi-supervised learning approaches are also being investigated for future research, as the labeling process necessitates both time and expertise. It is anticipated that these methods will enhance the efficacy of systems that operate with restricted data and lower the annotation burden by enabling the more efficient utilization of unlabeled data.

Conclusion

It is possible to train a dependable deep learning model to identify tonsilloliths on PRs, which would be a beneficial resource for oral healthcare professionals who are not familiar with STCs in the head and neck region. It is crucial to radiologically identify tonsilloliths in order to differentiate them from other abnormalities that necessitate treatment and to prevent the need for superfluous radiological evaluations.

Our study has limitations such as being conducted with a limited dataset and only two-dimensional imaging. Additional research is required to verify the clinical applicability and efficacy of this model. Future studies using data from multiple centers and a larger dataset size are important for making the model more reliable and applicable, especially when using three-dimensional imaging for testing.

Nevertheless, the results indicate that deep learning methods supported by AI have significant potential to enhance the precision of clinical diagnosis of tonsilloliths and minimize errors.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

STC

Soft tissue calcification

PR

Panoramic radiography

AI

Artificial intelligence

CNN

Convolutional Neural Network

FN

False negative

FP

False positive

CBCT

Cone-beam computed tomography

CAC

Carotid artery calcification

CT

Computed tomography

Author contributions

E.K. and S.C.B. were responsible for designing the research, collecting and interpreting image data, processing the image data, and writing the report. Ç.D. and S.A.T. utilized deep learning techniques and performed statistical analysis. The final manuscript was read and approved by all writers.

Funding

The authors certify that they did not receive any specific funding from public, private, or non-profit organizations for this research.

Data availability

The data that support the findings of this study are available from Dicle University Faculty of Dentistry, but restrictions apply to the availability of these data, which were used under license for the current study and so are not publicly available. Data are, however, available from the corresponding author upon reasonable request and with permission from the Dicle University Faculty of Dentistry.

Declarations

Competing interests

The authors declare no competing interests.