1. Introduction
Dental diseases such as periodontal disease, dental caries, and fluorosis affect a large population worldwide. However, their prevalence varies among different geographical distributions. Based on the research conducted by World Health Organization, dental diseases affect 621 million children, and almost 2.40 billion adults are affected by dental caries [
1]. Furthermore, results obtained from the National Health and nutrition examination survey show that almost 41% of children between the age of 2–11, 42% of adolescents from 6–19 years of age, and 90% of adults over the age of 20 are affected by dental caries [
2].
Periapical radiographs, a kind of intraoral radiograph, are widely used in endodontics. Periapical diseases are substantially inflammatory lesions that lead to dental caries causing detrimental injury to the teeth. These diseases are classified as apical cysts, abscesses, and granulomas affecting the dental pulp protected by enamel, cementum, and dentin. If left untreated, chronic damage to the pulp chamber can lead to inflammation, eventually turning into pulp necrosis and periradicular pathosis. Scenarios triggering periapical radiolucency include trauma, tooth wear, or caries. However, it is possible to prevent their spread through non-surgical endodontic treatment. Most periapical lesions heal through meticulous nonsurgical endodontic treatments. A timeframe of 6 to 12 months is required for assessing the healing potential, after which root canal treatment should be considered [
3,
4]. Additionally, complete healing of the lesion might take up to four years if not diagnosed in time. This requires frequent visits and interactions with the dentist and associated staff, which has become difficult in the post-COVID-19 environment. Postponing treatment can increase the risk of tooth fracture [
5]. Therefore, it is essential to diagnose teeth lesions in a timely way through routine oral examination of the teeth and gums as well as the soft tissues in and around the mouth. Conventional periapical radiographs are obtained by exposing X-ray radiation that is then processed chemically to produce images. This film-based conventional method has certain drawbacks, and digital radiography has been introduced to overcome these drawbacks. This process involves acquiring images digitally which are then manipulated using a computer. There are several ways of obtaining such images, including intraoral sensors, charge-coupled devices, and scanning of radiographs [
6]. Due to the emergence and spread of low-cost connectivity in underdeveloped and developing countries, the Internet of Things (IoT) has enabled dental services to be more broadly leveraged, providing better dental health in developing countries and decreasing the overall demand for dental care. Looking at the successful adoption of the IoT in different sectors [
7,
8], it can serve as an attractive area of adoption in dentistry. IoT-enabled devices can obtain information, including pathological parameters, to predict oral health and make decisions regarding the treatment of the disease at earlier stages. With the growth in IoT-based healthcare services, it has become possible to achieve significant enhancement in dental healthcare, including early detection and prediction [
9].
Recent advancements in the microcontroller-based wearable and implantable smart electronic devices has given rise to touchless technologies, which are gaining momentum and are regarded as the future of medical science. Such wearables are data transmission devices worn on the human body, such as the eyes, ears, knees, feet, fingers, and hands. These devices can detect, process, analyze, receive, and transmit related vital body signals and information such as the pulse rate, heart rate, body temperature, and other ambient data which do not require human intervention and may allow immediate biofeedback to the wearer [
10,
11]. They are an essential part of an IoT scenario in which they allow the exchange of high-quality data through the global internet via processor and relay agents to the operator, data collector, company, and/or manufacturer.
As a result, effective biomedical monitoring systems can be designed for medical diagnosis, physiological and psychological health monitoring, and evaluation through continuous measurement of critical biomarkers from various human body parts. A possible range of wearable devices in the form of basic clothing, common accessories, attachments, and body implants is shown in
Figure 1.
The information available from these sensors can be transmitted through modern communication links, as shown in
Figure 2, for intelligent processing and storage in an IoT environment. The IoT infrastructure makes possible the monitoring and recording of real-time temporal characteristics of chronic/acute patients, allowing for both immediate diagnosis and further research around finding and tracking disease trends.
Wearable technology is expanding into newer applications, and has gained significant space in consumer electronics through smartwatches, activity trackers, advanced e-textiles, healthcare products, navigation systems, and geofencing applications. Within the healthcare domain, parameters such as blood pressure, exercise time, steps walked/ran, running speed, calories burned, heart rate, body temperature, pulse rate, seizures, physical stress/strain, and level of certain body chemicals can be measured and used to evaluate users’ health. The issue of power sourcing for wearables has been suitably addressed by various methods as well. Smart wristbands can be used to generate their own electrical signals through thermoelectric generation technology (TEG) [
12]. Power can be generated using the temperature differential between human skin and the environmental temperature. These implants have been used to monitor the diet and nutrition content of meals, and play a great role in improving diet plans, leading to a more healthy and fit life [
13]. These devices collect information about salt, glucose, and even alcohol consumption, can share these data with other smart devices that a person uses, and are capable of updating physicians about diet routines and health conditions.
Additionally, IoT infrastructure allows monitoring and recording of the real-time temporal characteristics of chronic/acute patients for immediate diagnosis and tracing of disease trends. For diagnosing dental caries, digital radiographs are employed by dentists to assess and evaluate the extent of caries and to determine the need for treatment. Among the common intraoral types of radiographs (periapical, bitewing, and panoramic), periapical radiographs are the more common, as they provide localized information related to the length and adequacy of caries and periodontal ligament space [
14]. Furthermore, challenges faced in interpreting conventional radiographs create the potential for inconsistencies among dentists due to a number of factors, including contrast variation, magnification, and lack of experience. Nevertheless, due to the clinical reliability of periapical radiographs these have become a popular choice for diagnosing dental caries. Additionally, the above-mentioned challenges pave the way for utilizing automated solutions for improving diagnosis and standardization of care.
Recently, deep learning-based techniques have demonstrated excellent performance in computer vision, including object detection, tracking, and recognition, improving the ability to build software for automated analysis and evaluation of images. Different deep learning techniques have demonstrated the potential for automated identification of radiological and pathological features. Furthermore, different image processing and recognition procedures have been adopted for medical segmentation, with high accuracy and efficiency observed in the classification of different diseases using deep learning-based models, including cystic lesions [
15], skin cancer [
16], COVID-19 [
17], and thoracic diseases [
18], demonstrating improved accuracy and efficiency.
On the contrary, few studies have been observed based on deep learning, specifically deep convolutional neural networks (CNNs), in the dental field, limiting research investigating the diagnosis and detection of dental caries. Additionally, limited attempts at automatic dental radiograph analysis using deep learning have been observed from previous studies [
19,
20,
21]. Other than this, variable accuracy is observed in the methods currently utilized, highlighting the need for further research in this field. Approaches involving CNN and transfer learning have been used in dental diseases based on dental X-rays [
22,
23].
However, due to ease of classification, most approaches consider specific issues such as periodontitis, periapical lesions, and dental caries. Other teeth lesions, such as endo-perio and perio-endo lesions, are more challenging to diagnose due to their closely matching attributes. Therefore, they have been grouped into five significant lesions: primary endodontic with or without secondary periodontal involvement, primary periodontal with or without secondary endodontic involvement, and true combined lesion [
24,
25].
There are several limitations and challenges in the existing approaches. First, most of the work relies on traditional machine learning techniques, which may not be effective in detecting complex lesions due to the closely matching attributes of dental lesions, making them difficult to diagnose using traditional methods. Second, data insufficiency is a common challenge, as there may not be enough data available to train accurate models.
Keeping in view the limitations of previous studies, the present study aims to predict five types of endo-perio and perio-endo lesions in periapical radiographs by adopting a hybrid approach using transfer learning and a fine-tuning method after acquiring radiographs from a wearable device through an IoT infrastructure. Additionally, the presented work compares the diagnosis efficacy of different machine learning-based techniques, including support vector machine (SVM) and K-nearest neighbors (KNN), on features extracted using a pretrained AlexNet. Finally, the objective is to evaluate the efficacy of CNN algorithms and transfer learning to detect and classify dental caries in periapical radiographs and to overcome data insufficiency using data augmentation techniques. The main contributions of this research include:
Effective detection and classification of dental caries in periapical radiographs by leveraging deep learning techniques such as transfer learning and image augmentation;
Integrating IoT technology through an IoT-enabled device to capture tooth lesions from radiographs, allowing for remote monitoring and management of dental health, which is significant in situations where in-person visits to dental clinics might be limited.
The rest of the paper is structured in five sections, with
Section 2 presenting a literature review on wearables in the healthcare sector and machine, deep, and transfer learning concepts and techniques used for diagnosis of teeth lesions in dental X-rays,
Section 3 explaining the methodology of the proposed framework,
Section 4 discussing the results and providing a comparative analysis, and
Section 5 concluding the paper.
2. Related Work
Recently, new technologies have been evolving rapidly, and the need for these technologies has increased due to the COVID-19 pandemic. The IoT, specifically the Internet of Dental Things (IoDT), is a novel strategy that aids in managing and preventing dental caries as well as periodontal and other disorders [
9,
26]. Diagnostic screening and visualization incur significant costs, as dental diseases are very common. With the development of the Internet of Things (IoT), internet-based systems have shown great potential in healthcare, particularly dental healthcare. Smart dental IoT-based systems based on intelligent hardware and deep learning can allow for better identification and monitoring, improving the preventive care process.
Over the past decade, various retention and restoration methods have been proposed for detecting dental caries. However, there remains room for improvement in diagnosing dental caries due to various teeth morphologies and restoration shapes [
27]. Furthermore, detecting lesions in their earlier stages is challenging using these methods. The final diagnosis ultimately depends on empirical evidence, even though different dental radiographs (periapical, panoramic, and bitewing) are widely used. A range of options are available to detect dental lesions, including real-time ultrasound imaging [
28], contrast media, Papanicolaou smears, and cone beam computed tomography (CBCT), with the latter showing the highest discriminatory ability. However, it is limited in general dental practice due to its cost and high radiation dose [
29]. Panoramic radiographs allow all teeth to be assessed simultaneously; however, these methods have proven to be less accurate due to limited datasets. Therefore, using other imaging systems, such as periapical radiographs, which is the current standard of endodontic radiography, can enhance the chance of more accurate preoperative diagnosis.
Regardless of their discriminatory ability, the reliability of the radiographs depends on different factors, including the examiner’s reliability and experience. To overcome this, automated systems for dental radiographs using machine learning and deep learning techniques have demonstrated high performance. Patil and fellow authors [
30] proposed an attractive model for caries detection that detects dental cavities using MPCA-based feature extraction and an NN classifier trained using an adaptive DA algorithm to classify extracted features. Nonlinear programming optimization is used to maximize the distance between the feature output. This approach performs better than methods such as PCA, LDA, MPCA, and ICA, reaching 95% accuracy on three developed test cases. In addition, the proposed model works better than machine learning techniques such as KNN, SVM, and Naıve Bayes for dental lesion detection.
Kühnisch et al. proposed a deep learning approach using a convolutional neural network (CNN) trained using image augmentation and transfer learning to detect, categorize, and compare the diagnostic performance. The dataset comprised 2417 anonymized photographs of permanent teeth with 1317 occlusal and 1100 smooth surfaces. The images were organized into three main categorie: non-cavitated caries lesion, caries-free, and caries-related cavitation. The proposed model was able to classify caries in 92.5% of all images and 93.3% of all tooth surfaces, achieving an accuracy of 90.6% for caries-free surfaces, 85.2% for non-cavitated caries lesions, and 79.5% for cavitated caries lesions [
19].
Privado et al. [
31] reviewed different studies investigating caries detection using neural networks with different dental images. Various neural networks for detection and diagnosis of dental caries were discussed, and a database was constructed containing each neural network’s parameters. In 2019, Casalegno et al. [
32] presented a deep learning model for the automatic detection and localization of dental lesions in near-infrared TI images based on a convolutional neural network (CNN). Their dataset consisted of 217 grayscale images of upper and lower molars and premolars obtained using the DIAGNOcam system. Over 185 training samples were used to train the model. The model’s performance was tested based on pixel segmentation into classes and binary labeling of the region of interest. In addition, the model was tested using Monte Carlo cross-validation. However, the proposed model showed physically unrealistic labeling artifacts, especially in underexposed and overexposed regions. Geetha et al. [
33] proposed a dental lesion classification model based on an Artificial Neural Network (ANN) and performed ten-fold cross-validation, achieving an accuracy of 97.1%.
Duong et al. formulated a computational algorithm to automate and recognize carious lesions on tooth occlusal surfaces using smartphone images. The images were evaluated for caries diagnosis using International Caries Detection and Assessment System (ICDAS) II codes. Support vector machine (SVM) was used for classification; the proposed model was able to diagnose caries with a highest accuracy of 92.37%. Schwendicke and Paris discussed the cost-effectiveness of using artificial intelligence in caries detection. It was found that in most cases using AI is more effective and cost-efficient [
34]. Gracia Cantu proposes a model involving training a CNN; the model was validated and tested on 3686 collected bitewing radiographs. However, this research involved proximal caries lesions on permanent teeth. To measure cost-effectiveness, a Markov simulation model was used. Several studies have used AI, more specifically, deep learning techniques, to analyze dental images. These studies have mainly focused on developing and evaluating models that are more accurate and efficient [
20].
A deep convolutional neural network to detect caries lesions in near-infrared light transillumination images was proposed by Schwendicke and Rossi. The model involves two CNNs, namely, Resnet 18 and Resnext50, which were pre-trained on the ImageNet dataset. The dataset was processed digitally before being fed into the CNN. Ten-fold cross-validation was used to validate the model, which was optimized with respect to recall, F1 score, and precision. The model obtained an accuracy of 0.73 (0.67/0.80) using Resnet18 and an accuracy of 0.74 (0.66/0.82) using Resnext50, with a mean 95% confidence interval (CI). However, there are certain limitations of deep neural networks. One is that they are opaque prediction models and have a complex and nonlinear structure that makes it difficult to make decisions based on their results. Furthermore, limited augmentation and optimization processes are performed, and higher accuracies can be achieved by combining this approach with a larger dataset and in different settings with different diagnostic standards [
21].
Another interesting model was formulated by Saleh et al. [
35], combining a deep convolutional neural network (CNN) and optical coherence tomography (OCT) imaging modality to classify human oral tissues for earlier detection of dental caries. The CNN utilized two convolutional and pooling layers for feature extraction, and used the probabilities of the SoftMax classification layer to classify each patch. The sensitivity and specificity for oral tissues were found to be 98% and 100%, respectively. Thus, this model can help to classify oral tissues with various densities for earlier dental caries detection [
10].
Prajapati and Nagaraj [
22] proposed a model that combines a deep convolutional neural network (CNN) and optical coherence tomography (OCT) imaging modality to classify human oral tissues for earlier detection of dental caries. Their CNN utilized two convolutional and pooling layers for feature extraction and used the probabilities of the SoftMax classification layer to classify each patch. The sensitivity and specificity for the oral tissues were found to be 98% and 100%, respectively. Thus, this model can help to classify oral tissues with various densities for earlier dental caries detection. An automatic lesion detection model to analyze and locate lesions in panoramic radiographs using different image processing techniques was proposed by Bridal et al. in [
36]. Their system was capable of root localization, tooth segmentation, jaw separation, and detection of periapical lesions. First, the input image is enhanced by observing the smooth variations between intensities of neighboring pixels, then a Gaussian filter is used to smooth the photos. The jaws are separated by feeding discrete wavelet transformation into polynomial regression, then tooth segmentation and apex localization are performed. Their model achieved a specificity and specificity of 89% and 70%, respectively.
Ghaedi et al. [
37] proposed a method for examining dental caries in which dental caries are examined using optical images and the histogram equalization method. Segmentation is achieved in two steps; first, the circular Hough transform and region growth is used to the segment the tooth surface. Second, the morphology method is applied to identify unstable regions within the tooth boundaries; the authors extracted 77 features from these unstable regions when using a suitable window size. Feature space is reduced using a heuristic approach based on the information gain ratio method. Gawad et al. [
38] formulated a caries status detection and classification model based on a low-powered and less hazardous 635 nm He-Ne laser–tissue interaction mechanism to characterize human teeth into regular, moderate, and severe caries degree status.
Similarly, another study provided a framework for diagnosing periodontal, periapical, and dental caries using the CNN and transfer learning approaches [
23]. This method employed a CNN consisting of five convolution layers, four fully connected layers, and two max-pooling layers. The transfer learning technique was used in two different ways. First, a pre-trained VGG16 (a CNN model used for classification and detection) was combined with another CNN trained earlier. Eight convolutions, four zero padding, five max pooling, and eight fully connected layers were used. Second, the pretrained VGG16 model was fine-tuned using a dataset and the results were analyzed based on accuracy.
Table 1 summarizes the related research performed by different authors for the detection of teeth lesions using deep learning approaches. Although several studies have incorporated artificial intelligence, and more specifically deep learning techniques, for dental lesion detection, there remains room for further improvement in analyzing and classifying dental caries.