Nowadays Machine Learning is widely used for Object Recognition, Pattern Recognition, Natural Language Processing, and image processing tasks [
4]. In the field of medical image processing, several works have been done. Deep learning, a subfield of machine learning, has shown great potential in various image recognition and classification tasks, including medical image analysis [
5]. However, it has gained significant popularity in recent years due to its ability to achieve amazing results, even at human-level performance [
6]. In [
7], a deep learning model was developed for the classification of COVID-19 based on CT images. Convolutional neural network (CNN) is one of the most popular architectures of Deep Learning networks [
8]. The main advantage of CNN compared to its predecessors is that it automatically detects significant features without any human supervision which makes it the most used [
9]. Recent advancements in artificial intelligence have made it possible to diagnose dental caries via machine learning techniques, with a particular focus on neural networks and deep learning. This development is highly significant; as traditional diagnosis methods can often result in dentists misidentifying healthy teeth as carious (false positives) or decayed teeth as healthy (false negatives). Additionally, the availability of dentists to diagnose caries rapidly may be limited, underscoring the importance of leveraging AI in this domain. In this article, we will delve into how artificial neural networks and deep learning can be leveraged for the accurate diagnosis of tooth caries from radiographic images of teeth. Lian et al., [
10], used to detect caries lesions, classify different radiographic extensions on panoramic films, and compare the classification results with those of expert dentists. Experts evaluated 1160 dental panoramic films to detect and classify caries lesions based on depth. The study used no new net (nnU-Net) for segmentation and DenseNet121 for classification. Results showed high accuracy and recall rates for both techniques, and one of the positive points of this study was that they followed both segmentation and classification techniques. A study by Faria et al. [
11] has introduced a method that uses artificial intelligence neural network to detect and predict regular caries or radiation-related caries (RRC) in head and neck cancer patients undergoing radiotherapy. The study analyzed 420 teeth retrospectively from 15 HNC patients using PyRadiomics, and an artificial neural network classifier (ANN) was utilized to analyze the data. The proposed method demonstrated a sensitivity of 98.8% in detecting RRC and predicted an RRC with 99.2% accuracy. This study presented a new perspective on dental caries, however, its smaller sample size compared to other studies may limit its impact. Lee et al. [
12] used 3000 periapical radiographic images to train a pre-trained GoogLeNet Inception v3 CNN network for processing and transfer learning. The diagnostic accuracies for premolar, molar, and both were found to be 89.0, 88.0, and 82.0%, respectively. The premolar model achieved an AUC of 0.917, the molar model achieved an AUC of 0.890, and both premolar and molar models achieved an AUC of 0.845 using the deep CNN algorithm. This study utilized a large dataset and applied its models to various teeth. Sornam et al., [
13], a different approach was used. They used Linearly Adaptive Particle Swarm Optimization [LA-PSO] and Back Propagation Neural Network for the classification of dental caries based on the features that have been extracted from the Panoramic X-ray images. They achieved a 99% accuracy. Their method was novel, however it is hard to this model in other studies. In this study, the combination of statistics and neural networks can be seen which is an important feature. A combination of CNN and LSTM networks known as CNN-LSTM was suggested by Singh et al. [
14] . The aim was to classify dental caries according to the G.V. black classes. The optimal CNN-LSTM model proposed achieved a 96% accuracy rate. Mao et al. [
3], used CNNs to classify restorations and caries. They implemented transfer learning CNNs on Bitewing films by dint of Gaussian high-pass filter and Otsu’s threshold image enhancement technology. In the study, four networks were evaluated for their effectiveness in restoration and caries diagnosis. AlexNet achieved an accuracy of 95.56%, making it a valuable tool for computer-aided diagnosis in dentistry. Moran et al. [
15], utilized Inception and ResNet networks with three different learning rates (0.1, 0.01, 0.001), and after 2000 iterations, the Inception model with a 0.001 learning rate achieved the best results. The accuracy on the test set was 73.3%. This study detected both caries and restorations which was worth noticing. Their accuracy was not too high which can be challenging. Mertens et al. [
16], found that an AI-powered diagnostic-support software for detecting proximal caries in bitewing radiographs helped dentists increase their sensitivity and mean area under the Receiver-Operating- Characteristics curve. Their method was new and it compared the AI methods with dentists diagnosis power. In a study by Bayraktar et al. [
17], they assessed the effectiveness of using CNNs to diagnose interproximal caries lesions in digital bitewing radiographs. They analyzed 1000 bitewing images and found 11,521 approximal surfaces through augmentations. The YOLO algorithm was used for detection. The CNN model showed an overall accuracy of 94.59%, with a sensitivity of 72.26%, specificity of 98.19%, PPV of 86.58%, and NPV of 95.64%. Using YOLO was an important feature of this study, however, there was a gap between sensitivity value and other criteria reported in their study. Bayrakdar et al., [
18], an AI system called CranioCatch was used to detect and segment dental caries on 621 bitewing radiographs using VGG16 and U-net models. The results showed high rates of sensitivity, precision, and F-measure for caries detection and segmentation. The AI models outperformed 5 experienced observers on an external dataset which is an important achievement. A study was conducted by Canas et al. [
19] to evaluate the reliability of a web-based AI program for detecting interproximal caries in bitewing radiographs. They analyzed 300 images using a convolutional neural network and calculated various metrics such as accuracy rate, sensitivity, specificity, positive and negative predictive value, positive and negative likelihood ratio, and areas under the receiver operating characteristic curves. Imak et al. [
20] have proposed a new method for detecting dental caries using a multi-input deep convolutional neural network ensemble model. The approach involves pre-processing, deep convolutional neural network, and score-based fusion phases. The team used pre-learned weights of the AlexNet architecture and a transfer learning approach to adapt this architecture in dental caries detection. The study analyzed 340 periapical images from 310 patients and achieved an impressive accuracy rate of 99.13%. The study by Oztekin et al. [
21] used heat maps to explain deep learning-based models to physicians. The maps were created using the Grad-CAM method and applied to dental images. The dataset used was composed of 562 subjects labeled as caries and non-carious. The study employed data augmentation methods and chose Adam optimization, cross-entropy loss, 16 batch size, and a learning rate of 0.001. Two CNN-based models were used, with the ResNet model performing the best, achieving an accuracy of 92.00%, a sensitivity of 87.33%, and an F1-score of 91.61%.
The research undertaken encompasses all the necessary steps from data preparation and sorting to pre-processing. The data has been meticulously sourced and no ready-made data has been employed. The study employs four networks, with comparisons drawn among them. Additionally, two of the networks were trained from scratch while the other two were trained through the use of transfer learning, a novel combination for network training. The outcomes of the present study can serve as a valuable tool for dental practitioners to expediently and remotely diagnose caries. Additionally, this study can establish a correlation between artificial intelligence and dental science, leading to a more effective utilization of artificial intelligence in the field of dentistry. By utilizing the findings of this study and its subsequent enhancements, a diagnostic support system can be developed to identify dental caries.