Research on Cyst of Jaw Detection Algorithm Based on Alex Net Deep Learning Model

Research on Cyst of Jaw Detection Algorithm Based on Alex Net Deep Learning Model | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Research on Cyst of Jaw Detection Algorithm Based on Alex Net Deep Learning Model Wang Guangyan, Jia Yanan, Gulibstan Aihemaiti, Wang Kexin, Qiao Feng, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-3856379/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract In clinical medicine, jawbone cysts are a common dental disease, and their symptoms are similar to other dental diseases, making diagnosis difficult. To address this issue, this paper proposes a cyst detection system based on Alex Net to achieve the detection of cysts on dental radiographs using a CBCT dataset. The system can detect potential cyst lesions and locations in a timely manner to assist doctors in diagnosis. The improved model achieves an average accuracy of 83.5% and a maximum accuracy of 99.9%, achieving a high cyst recognition rate on existing datasets. In addition, the extensive image enhancement techniques introduced in the Alex Net model also improve the performance of the model. The experimental results show that compared to Res Net and VGG Net, both networks are not ideal for the classification of jawbone cysts, and may not be able to effectively extract key features from medical images, resulting in low classification accuracy. Therefore, it is important to choose a suitable deep learning model for the diagnosis of specific dental diseases. In future research, it is possible to further explore how to combine multiple deep learning models to improve the accuracy of diagnosis of dental diseases such as jaw cysts. In addition, improving data preprocessing and enhancing techniques can further improve the generalization ability of the model. In summary, by combining deep learning and clinical medicine concepts and methods, more effective auxiliary diagnostic systems can be developed to improve the accuracy and efficiency of dental disease diagnosis. Medical imaging diagnosis Dental cysts detection Convolutional neural network Deep learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Figure 13 Figure 14 Figure 15 Figure 16 Figure 17 Figure 18 Figure 19 Figure 20 Figure 21 1 Introduction There are various types of dental diseases, such as dental calculus and gingivitis, which can be observed directly by the naked eye, as well as jaw cysts, which can be diagnosed through dental images. Jaw cyst is a fluid-filled cyst found within the jaw bone. It is generally benign, with no signs of infiltration or growth, and does not metastasize to other sites. However, as it grows slowly over time and expands in size, it may ultimately squeeze the surrounding tissue, resulting in bone resorption, facial swelling, loose teeth, and tooth displacement in patients. This type of disease can be diagnosed through CBCT image. Cyst usually appear as circular or oval low-density areas with clear borders and smooth sharp edges. Depending on the type and size of the cyst, its internal structure and relationship with surrounding tissues may also vary. Medical images have high grayscale quality, the influence of noise and metal artifacts, and there are certain limitations [ 1 ] , which brings some difficulties to diagnosis. However, doctors have the problem of being unable to objectively and accurately diagnose images, and medical images are susceptible to external factors. The accuracy of shooting angles or operations can lead to images deviating from the true situation. Misdiagnosis caused by doctors' negligence in judgment can miss the best time for treatment, leading to the worst outcome of oral treatment. Introducing artificial intelligence to assist doctors in making judgments about images greatly improves the accuracy, efficiency and objectivity of diagnosis, which is also the purpose and significance of this study. This article mainly focuses on a series of studies on dental imaging. At present, the main type of medical image used in the dental industry is CBCT, also known as Cone beam Computer Tomography (CBCT) [ 2 ] . CBCT belongs to low-dose CT image, and the biggest difference compared to volumetric CT is that it can obtain high-definition three-dimensional images and reduce the impact of metal artifacts caused by two-dimensional slice stacking [ 3 – 5 ] . CBCT type dental images can be used for three-dimensional observation, allowing for better observation of teeth; The dental film can be scanned in a sitting position, which is more user-friendly; Due to low radiation dose, shorter acquisition time, higher X-ray utilization rate, and simpler and easier equipment operation; It also has extremely high isotropic spatial resolution, making it easy to observe the root canal structure of teeth. Therefore, this article chooses CBCT as the main research object. With the continuous development and progress of computer technology, deep learning has made a breakthrough in the field of clinical Medical imaging auxiliary diagnosis. In recent years, deep learning has also been successfully introduced into the field of Oral medicine [ 6 ] , especially the classic models represented by Convolutional neural network show strong robustness and universality, that is, to promote Oral medicine to become digital, intelligent and automated, so as to achieve data cooperation and sharing. Deep learning has more successful cases in the auxiliary diagnosis of common oral diseases, the identification, positioning and segmentation of anatomical structures in Oral medicine images, and guiding the fine clinical operation of dentists. Compared to previous traditional manual operations and physicians' diagnostic accuracy has been improved. Therefore, dental clinicians can make more accurate decision analysis of oral imaging data under the guidance of auxiliary diagnostic systems, reduce the probability of missed diagnosis and misdiagnosis, improve the accuracy and homogeneity of oral diagnosis, and thus improve the level of oral health services in China. This article aims to design a smart jaw cyst assistant detection system, which is based on the Alex Net model of deep learning network. After the model has been effectively learned and trained, the patient's dental image is input into the system, and the network will determine whether there is a cyst in the patient's jaw bone based on the training results. Clinical practice is an important basis for testing whether artificial intelligence can assist doctors in diagnosis and treatment, but there may be deviations from reality and a lack of feasibility. Digital healthcare should also be appropriately adjusted based on the facilities and medical resources in different places, fundamentally improving the medical gap and playing a role [ 7 ] . 2 Related research methods 2.1 Basic principle of Convolutional neural network Deep learning is applied to various occasions in image processing, and also widely developed in the Medical imaging industry. In 1962, scholars Hubel and Wiesel first proposed the concept of CNN in biological neurology. Although the CNN [ 8 ] model only emerged in 2012, Lo [ 9 ] and others proposed the use of artificial neural networks for chest image detection in 1995, and CNN was used as a classifier for the final detection of lung diseases. Until 2012, the Alex Net model had better performance compared to traditional methods in the large-scale visual recognition challenge of ImageNet. Since then, CNN technology has emerged, and researchers have begun to widely use and research it. The full name of CNN is Convolutional neural network, which mainly consists of three modules, the convolutional layer (CONV), pooling layer (POOL), and full connection layer (FC). Figure 1shows the architecture of CNN, and the most important functional layer of Convolutional neural network is the convolutional layer. After receiving the signal in the input layer of CNN, it is processed in the early stage and sent to the convolutional working layer for operation. After the feature extraction is completed, it is sent to the fully connected layer for output. CNN convolution layer is mainly composed of convolution kernel and filter, and it is not only composed of a simple layer, but can be formed by unlimited stacking and complex combination. With continuous stacking, the desired effect can be constantly approached, which is one of the reasons why many scholars continue to develop and study Convolutional neural network. After the convolution operation is completed, it will enter the pooling layer step. Due to the continuous convolution operation and transformation of the input image signal in the convolution layer, the filter may perform a "dimensionality increase", and the pooling layer is also the process of "dimensionality reduction". The process of pooling is the step of filtering the matrix eigenvalues obtained in the convolutional layer. Displaying the maximum detected eigenvalues is called max-pooling [ 10 – 12 ] , and displaying the mean eigenvalues is called average-pooling [ 13 , 14 ] . The max-pooling method is relatively commonly used in convolutional networks. The assembly process after feature extraction is the fully connected layer, and finally fitting is performed to connect the extracted features, reducing the loss of feature information and outputting. The training process of Convolutional neural network is divided into two parts, forward propagation and backward propagation. The essence of forward propagation is to use the output of the previous layer to calculate the output of the next layer, and process the input image. It is the process of learning the training set. After processing, backward propagation detects the sensitivity of the entire network, pushes back the possible errors in the previous layers, checks the algorithm, and then adjusts the parameters of each layer based on the received errors. Both the pooling layer and the convolutional layer have backpropagation, so errors can be detected and adjusted in a timely manner. Compared to traditional neural networks such as linear regression [ 15 – 17 ] and linear to nonlinear transformations [ 18 ] , CNN have two major advantages: local connections and weight sharing [ 19 , 20 ] . In traditional neural networks, all neurons in the network are connected, creating a high degree of interdependence between the input signal and all output results. Any errors in the intermediate steps can lead to significant deviations in the final output. In contrast, CNN process images through independent and independent outputs for different regions of the image. This allows for smaller impact of errors and reduces the overall loss in image processing. Weight sharing in CNN refers to the ability to share the parameters of a convolutional kernel across different parts of an image, regardless of their spatial locations. This eliminates the need to repeatedly add a convolutional kernel wherever it is needed and reduces the complexity of the entire network, as well as the number of parameters required for convolutional kernels. By enabling weight sharing, the same convolutional kernel can be used across the entire image, significantly reducing the number of parameters required for each kernel. 2.2 Alex Net model based on CNN network The original convolutional network LeNet-5 from 1980 [ 21 ] was relatively outdated and had not yet proposed the use of stride and padding for optimization when dealing with images. It also used average-pooling, which resulted in a decrease in both the height and width of the images. In 2012, the Alex Net network [ 22 ] emerged, introducing the ReLU activation function [ 23 ] and switching to max-pooling, which improved the training speed of the network and reduced its complexity. In 2014, Simonyan proposed the VGG [ 24 ] network, which better displayed the convolutional layer, increased the depth of the model, and truly simplified the neural network structure. In 2015, the Res-Net residual network [ 25 ] was a deeper neural network that introduced cross-layer transmission combined with convolutional layers, greatly improving accuracy. It is also the most advanced model of convolutional neural networks to date, and these models are frequently used in analyzing dental images. Alex Net also uses multi-GPU training technology, which enables faster training of the model and can take advantage of parallel computing to accelerate the training process, making it more suitable for processing large data sets. The advantage of deep learning in the medical field lies in its applicability to large amounts of data. Due to the variety of perspectives and angles in medical image acquisition, the data volume is extremely large, resulting in high requirements for image processing systems. In recent years, the U-Net network model, CNN convolutional neural network, and Alex Net network architecture have been popular research directions, while other deep networks are also continuously being studied in the medical field. This article investigates a cyst of jaw detection system based on the Alex Net framework using CNN convolutional neural networks, as shown in Fig. 2 . The innovation of this model lies in the use of the ReLU activation function, which provides better performance than traditional Sigmoid and Tanh activation functions. It also implements LRN local response normalization and uses Droupout random neural deactivation in the first two fully connected layers to reduce overfitting. From the graph, it can be seen that the network has deepened, consisting of 5 convolutional layers, 3 pooling layers, 2 fully connected layers, and finally 1 softmax layer. In deep learning neural networks, activation functions map the output of neurons in the network through nonlinear functions, allowing the network to approximate any function arbitrarily well. The weighted input from a node is transformed into an activation or output of that input node. ReLU (rectified linear unit) is a piecewise linear function that extends the learning capability of neural networks and makes the network more stable. It has become the default activation function for many types of neural networks because models using it are easier to train and typically perform better. The ReLU function is linear for positive inputs, has fast convergence and calculation speed, fully passes the gradient, and does not suffer from gradient vanishing problems (saturated gradients). 2.3 Compared to other CNN network models This article will compare two network models, Res Net and VGG Net, train the established dataset, and analyze the classification effect. The Res Net network model solves the gradient problem by connecting residual units, allowing for the construction of deeper neural networks, and focuses on solving the degradation problem caused by network deepening. As shown in Fig. 3 , the network architecture of Res Net50 includes 50 convolutional layers, which can extract more features. In order to reduce the error caused by backpropagation, the residual structure is added to complete image classification. The VGG Net network structure is more concise, and uses small convolutional kernels instead of large convolutional layers, resulting in a reduction in parameters and a simpler overall structure that is easier to understand. As shown in Fig. 4 , the VGG Net network architecture includes 13 convolutional layers, 5 max pooling layers, 3 fully connected layers, and finally outputs the prediction result through soft max. Subsequent experiments will compare three deep learning network models to find the algorithm with the best classification performance. 2.4 Related research literature based on deep learning Machine learning algorithms for dental imaging use a combination of multiple algorithms or improvements. Deep learning is also constantly evolving, and various algorithms are constantly being improved and optimized, resulting in better results in medical image processing. Research based on the U-Net network model accounts for a high proportion. A.Fariza [ 26 ] proposed an improved U-Net network model for automatic segmentation of teeth and background based on X-ray images of teeth, removing the influence of tooth overlap; Kirnbauer [ 27 ] improved the U-Net architecture for binary segmentation of tooth apical lesions based on CBCT images, achieving a sensitivity of 97.1% for detecting lesion locations; Y. Rao [ 28 ] implemented single tooth recognition and segmentation based on panoramic dental images, achieving an accuracy of 97.93% through training a U-Net model; Estai M [ 29 ] developed a convolutional neural network-based system for detecting and classifying permanent teeth on surface tomographic images. The results showed sensitivities and precisions of 0.99 for the tooth detection module, and sensitivities, precisions, and F1 scores of 0.98 for the tooth numbering module. Secondly, some research based on convolutional neural network models, Yang [ 30 ] proposed a method to determine the location of dental pulp based on CNN, due to the blurred edges of teeth in CBCT images. They described the shape and size of teeth using mathematical methods and prior information. The experimental results demonstrated the feasibility and effectiveness of the model; Chung [ 31 ] reduced the influence of metal artifacts on the segmentation process during dental implant simulation. They extracted patient alignment information using a pose regression neural network to obtain a volume of interest (VOI) region and realign the input image, reducing the mutual overlap between tooth boundary boxes. Then, they converted the pixel-wise labeling task into a distance regression task using a CNN network to segment individual teeth, which increased the accuracy by 30.3%; M.P Muresan [ 32 ] trained a CNN using labeled data to obtain semantic segmentation information. They performed multiple image processing operations to segment and refine the boundary boxes corresponding to tooth detection; Ying [ 33 ] proposed a method for caries segmentation based on dental X-ray images, using visual transformation, extended convolution and feature pyramid fusion to enhance multi-scale and global feature extraction capabilities. The Dice coefficient reached 74.87%. Secondly, there are studies on the classification of dental images. Oktay [ 34 ] proposed a RCNN-based method for detecting, segmenting, and classifying teeth in panoramic X-ray images; Chen [ 35 ] used a TensorFlow-based R-CNN network to classify and number the periapical periodontal membranes, achieving an accuracy and recall rate of over 90%. The performance of the entire network system almost reached the level of a primary dentist; Shamim [ 36 ] used a Vgg19-based DCNN model to classify benign lesions and precancerous lesions of the tongue, achieving an accuracy of 98%. ResNet50 can achieve an accuracy of 97%, almost reaching the "human-like" classification level of doctors. The extremely high accuracy provides possibilities for screening oral cancer. Deep learning methods are fast and efficient in processing medical imaging data. Their excellent performance enables rapid classification, localization, and other operations in oral analysis and processing, improving the efficiency of dentists' diagnosis and treatment. However, research in the field of oral medicine is still in its infancy and has bright prospects to be explored. Currently, although convolutional neural networks (CNN) are a very effective method in recent years, the construction of their networks and the adjustment of their parameters cannot be fully explained, and it still requires continuous attempts to build suitable models. From traditional algorithms to higher-level algorithms in deep learning, each has its own advantages. In the experimental process, multiple algorithms are used for comparison and analysis of experimental results. The introduction of machine learning and deep learning in medical imaging reduces potential medical risks and can also appropriately share the repetitive and complex work of doctors, thereby improving work efficiency and reducing work intensity. 3 Research on jaw cyst detection algorithm based on deep learning network model 3.1 Basic scheme design This article will design a CNN-based cysts detection system that can predict whether a patient contains cysts based on its input. The system will be developed using Windows 10 as the operating environment and compiled in PyCharm using the Python language. As shown in Fig. 5 , the design process begins with collecting and organizing input data, followed by image preprocessing including de-noising and filtering, as well as data set expansion for limited data sets. The preprocessed data is then input into a pre-built CNN model for training, and the optimal model is stored. The system then validates the results using a test set and a validation set. Finally, the experimental results are presented through loss functions and accuracy rates of the validation set, test set, and training set. The specific design process of the plan is as follows: The dental image dataset used in this study primarily consists of CBCT, which are currently the most widely used type of dental image. Prior to conducting the experiment, the collected tooth images were divided into two categories according to the dentist's annotation: Healthy teeth and teeth with jaw cyst disease. The data set was then allocated according to a 7:2:1 ratio. As this study used a self-collected dataset, the sample size and type of lesion were limited. In order to prevent interference from extraneous factors, the teeth portion was cropped separately and filled with 0-pixels. Due to the small size of the dataset, techniques such as adjusting image contrast and brightness were used to augment the dataset and improve the accuracy of the training model. Based on the Alex Net framework, a model was built considering the high requirements of the deep learning runtime environment to accelerate training time. The entire Alex Net model was scaled down proportionally, with the input image size set to 65 x 65 pixels. The preprocessed dataset was then input into the model for training, resulting in three model training outcome graphs: the training set loss graph, the validation set loss graph, and the validation set accuracy graph. The trained model will generate three graphs showing the results of the training, validation, and testing sets, as well as the predicted outcome of the image. First, the unlabeled test set of tooth images will be used as input into the system, undergoing preprocessing and cysts labeling before entering the classification system. The system will display the predicted results for the test set, and finally classify and store the predicted tooth images. During training, network models will be optimized by continuously adjusting parameters. 3.2 Dataset establishment The dataset used in this article is derived from actual cases from Tianjin Medical University Stomatology Hospital, but the data quantity is relatively small, which poses certain limitations. The collected oral CT dataset is divided into three parts, divided according to the ratio of 7:2:1. the first part is used to update model parameters and improve performance, with the largest amount of data, consisting of 350 images. The second part is the validation set, responsible for adjusting the model's hyper parameters, the training set for 100 images. The third part is the test set, used to evaluate the accuracy of the model, primarily assessing the difference between the CNN model's estimated predictions and actual predictions, with the smallest amount of data, consisting of 50 images. The dataset distribution is presented in Table 1 . Please note that despite the relatively small size of the dataset, efforts have been made to ensure its representativeness and accuracy through various methods mentioned above. This experiment plans to run 50 and 100 rounds of comparison of three algorithms, using the same dataset to obtain more accurate classification results, and observed and analyzed the influence of the number of operations. Table 1 Data set allocation Data set allocation Name Training set Validation set Testing set Data volume 350 100 50 3.2.1 Dataset evaluation criteria CBCT dental images are based on X-rays. Because teeth block X-rays, the enamel appears as a high-brightness white color in the panoramic film, while the root canals in the middle of the teeth appear as a lighter gray color. When there is a cyst in the jawbone, a large amount of fluid accumulates in the tooth tissue, as shown in Fig. 6 , which appears as a black shadow on the dental image, and the size of the shadow indicates the amount of fluid accumulation. From this, it can be designed to preprocess all raw images and then put them into the Alex Net model for learning and training. The system will diagnose jaw cyst disease based on the presence of shadows on CBCT images. In addition, CBCT can clearly show the positional relationship between the cyst and surrounding teeth. If the cyst is connected to the tooth, residual roots or teeth may be visible, or there may be a space within the cyst, and the continuity of the bone cortex may be interrupted. For larger jaw cysts, swelling in the surrounding soft tissues may be seen. 3.3 Algorithm optimization and experimental results 3.3.1 Model optimization For the prediction of disease based on higher resolution and more precise oral and maxillofacial cone-beam CT images, the following improvements are proposed with a scheme design shown in Fig. 7 : Improve the CNN model: modify the optimizer and optimizer parameters to compare the changes in training and validation losses and accuracy of various training results, and analyze and select appropriate improvement methods to maximize the similarity of the model simulation results to the ideal state. Attempt to identify the parts of the teeth that contain cysts: theoretically, the cysts site can be separately marked through image segmentation, and then image overlap can be used to overlap the marked portion with the original image, thereby achieving a clear display of the cysts part on the dental images. After labeling, the size and location of the cyst can be visually detected. Implement batch prediction of images: if the number of images input into the system for prediction is too large, the efficiency of single-image prediction will decrease with the increase of data volume. To solve this problem, a function that can switch between single-image prediction and batch prediction based on the number of input images can effectively address the issue of large volumes of images to be predicted. Additionally, if the batch-predicted images can be stored and organized according to their prediction results, it will further facilitate secondary judgments of the predicted images by doctors. The following is a scheme design for simulation verification based on four steps: 3.3.2 Dental cysts markers CBCT images are a collection of cross-sectional views of teeth, with a vertical perspective, so a watershed algorithm can be used for image segmentation. This algorithm divides the image based on different gray values, resulting in a vertical perspective of the image with different levels of division. The cyst of jaw inside the dental tissue presents a dark image with large shadows, which is different from normal dental tissue. This image will be processed by the Alex Net model to threshold and convert it into a binary image. The distance transformation result is adjusted to determine the background (healthy tooth tissue), and the thresholding result is marked. Then, the watershed algorithm is used to segment the image, and the image markers are set as green. The background of the labeled decayed portion is processed as transparent and overlaid on the original CBCT image, resulting in the marked results of cysts shown in Fig. 8 . 3.3.3 Batch prediction function When dealing with large amounts of data, using a single image for prediction can be inefficient. As shown in Fig. 9 , this is the result of predicting with a single image. To address this issue, we design a function that can perform batch recognition. The main parameters are the model address, the folder address where the images need to be detected, and the result address where the predictions will be saved. The predicted results will be stored in the corresponding folders. Finally, each image's corresponding prediction results will be output as shown in Fig. 10 , and the labeled folders completed according to the results will be displayed in the result saving address as shown in Fig. 11 . 4 Expand function implementation 4.1 Change optimizer Optimizers play a crucial role in machine learning, as they update the model parameters to minimize the training error based on mathematical operations involving gradients and loss functions. After each sample passes through the model, it produces an estimated value, but due to the model's imperfections, this estimated value often differs from the actual value, and the difference is referred to as the error or loss. The smaller the loss, the more successful the model. The function introduced to calculate this loss is called the loss function. Gradient descent and adaptive moment estimation are widely used optimization methods. The choice of optimizer can lead to different experimental results. (1)Stochastic Gradient Descent (SGD) Stochastic Gradient Descent [ 37 , 38 ] does not use the loss function on the entire training dataset, but in each iteration, it randomly optimizes the loss function on a single training example. Therefore, the parameter update speed is faster in each round. SGD is commonly used for training on large-scale datasets because it can effectively reduce training time. However, since it only considers the information from a randomly selected sample, which may not represent the trend of the entire dataset, it can easily fall into local optimal situations. Therefore, in experimental processes, it is possible to reduce the learning rate or increase the number of gradient descent iterations to enable the algorithm to better converge to the optimal solution. (2)Adaptive Moment Estimation (ADAM) Adaptive Moment Estimation [ 39 ] is an optimization algorithm proposed in 2014, which combines the advantages of adaptive learning rate gradient descent algorithms and momentum gradient descent algorithms. It can adapt to sparse gradients (as in natural language and computer vision problems) and is well-suited for large-scale data and parameter scenarios. Although its implementation is simple, it requires little memory, and is very convenient to use, it is prone to oscillation around the optimal value, so its computational results are slightly inferior to stochastic gradient descent. In the experiment, it was found that changing the optimizer to ADAM significantly increased the training speed and the loss generally showed a decreasing trend. Figure 12 shows the training results when the optimizer was changed to ADAM, and it can be seen that the training set loss showed a decreasing trend until a small fluctuation at around 36 iterations. However, the validation set loss experienced a significant increase around 36 iterations, and the validation set accuracy fluctuated widely. Therefore, when using ADAM as the optimizer, there are large fluctuations around the optimal value, and after 40 iterations, the validation set loss still shows an upward trend, indicating that the overfitting problem has not been resolved. Moreover, ADAM performs worse than SGD in terms of performance. Therefore, it is recommended to use SGD as the optimizer. As shown in the Fig. 13 , the experimental results of 100 rounds of running are shown. It can be seen that there is still overfitting in the loss rates of the training set and the validation set, but the more times of training, the better the effect. 4.2 Optimizer parameter change The optimizer selects the SGD gradient descent method optimizer, where the parameters lr are learning rate, momentum is momentum, and weight_ Decay is the weight attenuation, and the parameter changes are as follows: (1)Change learning rate When using SGD as the optimizer, changing the learning rate: Generally, the learning rate of deep learning is between 0.01 and 0.001. In this design, the learning rate was initially set to the maximum value of 0.01 and the decay initial value was set to 0.0005 to shorten training time. However, it was found that overfitting occurred between 20 and 40 rounds of training, as shown in Fig. 14 . Therefore, the learning rate was reduced to 0.005, resulting in the training results shown in Fig. 15 . It can be seen that the training loss function and the validation loss function both show an ideal decreasing trend, and the training loss is lower than the validation loss, solving the problem of overfitting. In the later stages of training, the validation accuracy is high and stable, solving the problem of overfitting. Therefore, reducing the learning rate is very effective in improving CNN. (2)Reduce attenuation value Next, with SGD as the optimizer and a learning rate of 0.005, the learning rate decay value was changed from 0.0005 to 0.0003, resulting in the training results shown in Fig. 16 . It can be seen that the loss suddenly rises and fluctuates greatly in the later stages of training, and the validation loss is still smaller than the training loss. The accuracy has a large fluctuation and a relatively high value. Therefore, it can be concluded that the training results with a lower learning rate decay value are not as good as the training results with a decay value of 0.0005. (3)Increase attenuation value Because the training results with a lower learning rate decay value were not ideal, attempts were made to increase the decay value and obtain the result shown in Fig. 17 with a decay value of 0.0008. It can be seen that the training loss presents a downward curve, and the validation loss graph has fluctuations from 21 to 30 rounds, but gradually tends to a stable descending state. The validation loss value is significantly lower than the training loss value, and the whole model is in an ideal state. Moreover, the validation accuracy tends to be stable after 30 rounds and increases as the number of training rounds increases. Therefore, the improved method of increasing the learning rate decay value is selected, so the learning rate is set to 0.05 and the decay value is set to 0.0008 as the final parameters. As shown in the Fig. 18 , the results of training 100 rounds with the final adjusted parameters show that there is still overfitting in the first 20 rounds, but it gradually stabilizes later. Therefore, the more training rounds, the more accurate the classification effect of the model. The experiment improved the model training results by changing the dataset, reducing the learning rate of the optimizer, and other methods. The average accuracy was increased to 75%, and the highest accuracy was increased to 99.3%, solving the problem of validation loss being lower than data loss and overfitting. The average accuracy of the model reached 83.5%, and the highest accuracy reached 99.9%, achieving a relatively high cyst of jaw identification rate on the existing dataset. 4.3 Compare with the other two algorithms 4.3.1Analysis of classification results of Res Net algorithm The experiment compared the Res Net algorithm to classify the same dataset, running for 100 rounds and using the SGD optimizer. As shown in Fig. 19 , the loss rates of the training and validation sets fluctuated greatly, and overfitting remained a serious problem after 100 rounds of training, with the longest running time. Figure 20 shows the accuracy rates of the training and validation sets, which also fluctuated erratically and were far inferior to the classification performance of the Alex Net network. 4.3.2 Analysis of classification results of VGG Net algorithm VGG Net has the worst classification performance, with a high loss rate in the training set and severe overfitting, as shown in Fig. 21 . The classification accuracy is only 50%, and the classification performance is poor and takes a long time. According to the comparison of the effects of the above three algorithms, Alex Net performs the best on existing datasets, with high classification accuracy and fast processing speed; Res Net algorithm has high accuracy but suffers from overfitting and large fluctuations, and due to the large network depth and relatively large number of parameters, it results in long processing time; VGG Net algorithm has the lowest accuracy, and due to the relatively simple network structure, it suffers from overfitting and needs to be improved and upgraded. 5 Conclusion 5.1 Experimental Summary The above experimental results indicate that taking X-ray dental images is an effective diagnostic method for the timely treatment and prevention of worsening oral diseases such as Jawbone cysts. Oral health is an increasingly concerned medical issue in society, and analyzing dental images as a prerequisite for diagnosis is crucial for the diagnosis and treatment of oral diseases. In the field of computer vision, convolutional neural networks have prominent advantages, and the Deep Learning-based Alex Net network model based on PyTorch can be effectively applied to oral medical image analysis, providing powerful support for dentists in disease recognition. However, despite the achievements of our experiment, there are still some areas that need to be improved and perfected: (1)In terms of dataset Firstly, the size and quality of the dataset have a certain impact on the learning effect of deep learning models. Due to the limited protection of patient privacy data, the collection efficiency is not high, and the amount of data used in this article is not sufficient to fully support the fine-grained learning of the model. In addition, this experiment only targeted young and middle-aged people, and did not cover data from different age groups, such as children, the elderly, and denture wearers, which limited the scope of application of the model. Therefore, it is hoped that future research can address the limitations of dataset recognition and expand the application scenarios of models. (2)Display of data results Secondly, the display of data results still needs to be optimized. Although the experiment proposed in this article can determine whether there is a cyst in an image, it cannot visualize the location and extent of the cyst. In addition, there is an unstable state during model training, and there are fluctuations in loss value and accuracy. Follow-up studies can optimize the experiment, such as labeling the cyst range and using large data sets to train the model to improve accuracy. The final judgment result of the experiment is based on visual inspection, which lacks reliability. Future work can include medical evaluation indicators in the experiment code, and make the final conclusion based on the most accurate and objective numerical data. 5.2 Prospect Medical images are often stored in the DICOM electronic format, different software applications may require different export formats for processing, and there are also issues with accessing specialized medical imaging analysis and processing software such as 3D-slicer [ 40 ] , which may be inaccessible to non-professionals and difficult to operate without guidance. The open-source database for dental image analysis needs to be further improved. It is hoped that a more comprehensive database can be established to support broader applications in research, in terms of quantity, type, and quality. Researchers in the field of information processing lack professional medical knowledge, and it is important to establish a comprehensive evaluation standard system in the future. The lack of understanding of professional medical jargon by information technology personnel undoubtedly increases the difficulty of the experimental process. A new evaluation standard would be more accessible to non-professionals and would facilitate their understanding and judgment. Non-medical professionals should also have a deeper understanding of the structure and morphology of tooth tissue to better analyze the causes of dental diseases. Although traditional methods are still being optimized continuously, deep learning has enormous potential in fields such as image segmentation. Therefore, future research directions can include optimizing the training methods of deep learning models, such as using cross-validation and oversampling to improve model training effectiveness and generalization ability. Additionally, the training of deep learning models requires a strong database as support, which is also one of the serious challenges faced by digital healthcare. How to enhance the practicality of data and quickly, efficiently, and accurately analyze larger datasets is also a key direction for future research. models. The medical industry is an area with low error tolerance, and the training process of digital algorithms can be repeatedly tested and trained in other industries, but cannot be achieved in the medical industry. The unknown, scarcity, and practicality of medical data are greatly limited. Deep learning network models require powerful databases as support, which is also an extremely serious challenge for digital healthcare. There is still a lot of room for improvement in the application of image segmentation technology in medicine. How to enhance the practicality of data and quickly, efficiently, and accurately analyze larger datasets is also the main focus of future research. The medical industry still has a long process of effort from proposing new ideas to applying them to practical situations, and auxiliary diagnosis and treatment should also start from the foundation and continuously make improvements based on the actual situation. Declarations Funding No funding was received for conducting this study. Conflict of interest/Competing interests (check journal-specific guidelines for which heading to use) The authors have no relevant financial or non-financial interests to disclose. Ethics approval Attachment submission Consent to participate All authors agree Consent for publication All authors agree Availability of data and materials Not applicable Code availability Not applicable Authors’ contributions W and J wrote the main content of the manuscript, G and W were responsible for the preliminary investigation and preparation, Q provided data, and G was responsible for reviewing and revising the entire research content. W: Wang Guangyan J: Jia Yanan G: Gulibstan Aihemaiti W: Wang Kexin Q: Qiao Feng G: Geng Duyan References LI, P., LIU, Y., CUI, Z., et al.: Semantic Graph Attention With Explicit Anatomical Association Modeling for Tooth Segmentation From CBCT Images [J]. IEEE Trans. Med. Imaging. 41 (11), 3116–3127 (2022) HAO, J., ZHANG, L., LI, L., et al.: Nov. An improved non-local means regularized iterative reconstruction method for low-dose dental CBCT; proceedings of the 2012 IEEE Nuclear Science Symposium and Medical Imaging Conference Record (NSS/MIC), F 27 Oct.-3 2012 [C]. (2012) THAKUR, A., PARGAIN, V., SINGH, P., et al.: An efficient fuzzy and morphology based approach to metal artifact reduction from dental CBCT image; proceedings of the 2017 International Conference on Computing, Communication and Automation (ICCCA), F 5–6 May 2017, 2017 [C] AHN, J., NGUYEN T P, KIM, Y-J, et al.: Automated analysis of three-dimensional CBCT images taken in natural head position that combines facial profile processing and multiple deep-learning models [J], vol. 226, p. 107123. Computer Methods and Programs in Biomedicine (2022) LIUGANG, G., JIANFENG, S.: Metal Artifact Reduction Method Based on Noncoplanar Scanning in CBCT Imaging [J]. IEEE Access. 8 , 7236–7243 (2020) YANG, J., XIE, Y., LIU, L., et al.: Automated Dental Image Analysis by Deep Learning on Small Dataset; proceedings of the 2018 IEEE 42nd Annual Computer Software and Applications Conference (COMPSAC), F] KHERA, R., BUTTE A J, BERKWITS, M., et al.: AI in Medicine—JAMA’s Focus on Clinical Outcomes, Patient-Centered Care, Quality, and Equity [J]. JAMA. 330 (9), 818–820 (2023) MAHMUD B U, HONG G Y:. Semantic Image Segmentation using CNN (Convolutional Neural Network) based Technique; proceedings of the 2022 IEEE World Conference on Applied Intelligence and Computing (AIC), F 17–19 June 2022, 2022 [C] LO, S-C B, CHAN H-P, LIN, J-S, et al.: Artificial convolution neural network for medical image pattern recognition [J]. Neural Netw. 8 (7), 1201–1214 (1995) MURRAY N, PERRONNIN F. Generalized Max Pooling; proceedings of the 2014 IEEE Conference on Computer Vision and Pattern Recognition, F 23–28: 2014 [C]. (2014) MASCI, J., GIUSTI, A., CIRESAN, D., et al.: Sept. A fast learning algorithm for image segmentation with max-pooling convolutional networks; proceedings of the 2013 IEEE International Conference on Image Processing, F 15–18 2013 [C]. (2013) CHRISTLEIN, V., SPRANGER, L., SEURET, M., et al.: Deep Generalized Max Pooling; proceedings of the 2019 International Conference on Document Analysis and Recognition (ICDAR), F 20–25 Sept. 2019, 2019 [C] YANG, J., XIE, F., FAN, H., et al.: Classification for Dermoscopy Images Using Convolutional Neural Networks Based on Region Average Pooling [J]. IEEE Access. 6 , 65130–65138 (2018) WANG, S., JIANG, Y., HOU, X., et al.: Cerebral Micro-Bleed Detection Based on the Convolution Neural Network With Rank Based Average Pooling [J]. IEEE Access. 5 , 16576–16583 (2017) SHAKOURI, H., NADIMI G R, GHADERI, F.: Dec. Fuzzy linear regression models with absolute errors and optimum uncertainty; proceedings of the 2007 IEEE International Conference on Industrial Engineering and Engineering Management, F 2–4 2007 [C]. (2007) KAVITHA, S., VARUNA, S.: RAMYA R. A comparative analysis on linear regression and support vector regression; proceedings of the 2016 Online International Conference on Green Engineering and Technologies (IC-GET), F 19–19 Nov. 2016, 2016 [C] HIROSE, H., SOEJIMA, Y.: HIROSE K. NNRMLR: A Combined Method of Nearest Neighbor Regression and Multiple Linear Regression; proceedings of the 2012 IIAI International Conference on Advanced Applied Informatics, F 20–22 Sept. 2012 [C]. (2012) SONTAG E. From linear to nonlinear: some complexity comparisons; proceedings of the Proceedings of 1995 34th IEEE Conference on Decision and Control, F 13–15: 1995 [C]. (1995) ZHANG, H., HE, J.: KO S B. Improved Hybrid Memory Cube for Weight-Sharing Deep Convolutional Neural Networks; proceedings of the 2019 IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS), F 18–20 March 2019, 2019 [C] MOHAZZEBI, M., EFTEKHARI, M.: SHARIATZADEH M. How weight-sharing mechanisms affect the performance of deep Siamese networks; proceedings of the 2022 9th Iranian Joint Congress on Fuzzy and Intelligent Systems (CFIS), F 2–4 March 2022, 2022 [C] ISLAM, M.R.: MATIN A. Detection of COVID 19 from CT Image by The Novel LeNet-5 CNN Architecture; proceedings of the 2020 23rd International Conference on Computer and Information Technology (ICCIT), F 19–21 Dec. 2020 [C]. (2020) RANI, S., KUMAR, S., GHAI, D., et al.: Automatic Detection of Brain Tumor from CT and MRI Images using Wireframe model and 3D Alex-Net; proceedings of the 2022 International Conference on Decision Aid Sciences and Applications (DASA), F 23–25 March 2022, 2022 [C] STURSA, D.: DOLEZEL P. Comparison of ReLU and linear saturated activation functions in neural network for universal approximation; proceedings of the 2019 22nd International Conference on Process Control (PC19), F 11–14 June 2019, 2019 [C] BAGASKARA, A.: SURYANEGARA M. Evaluation of VGG-16 and VGG-19 Deep Learning Architecture for Classifying Dementia People; proceedings of the 2021 4th International Conference of Computer and Informatics Engineering (IC2IE), F 14–15 Sept. 2021, 2021 [C] ZHOU, J., JIN, Y.I.S.: F. Super resolution reconstruction of single image based on Res-Net and Sub-Pixel; proceedings of the 2022 IEEE 2nd International Conference on Electronic Technology, Communication and Information (ICETCI), F 27–29 May 2022, 2022 [C] FARIZA, A., ARIFIN A Z, A.S.T.U.T.I.E.R.: Automatic Tooth and Background Segmentation in Dental X-ray Using U-Net Convolution Network; proceedings of the 2020 6th International Conference on Science in Information Technology (ICSITech), F 21–22 Oct. 2020 [C]. (2020) KIRNBAUER, B., HADZIC, A., JAKSE, N., et al.: Automatic Detection of Periapical Osteolytic Lesions on Cone-beam Computed Tomography Using Deep Convolutional Neuronal Networks [J]. Journal of Endodontics (2022) RAO, Y., NARTEY O T, Z.E.N.G.S., et al.: LeFUNet: UNet with Learnable Feature Connections for Teeth Identification and Segmentation in Dental Panoramic X-ray Images; proceedings of the 2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), F 6–8 Dec. 2022, 2022 [C] ESTAI, M., TENNANT, M.: Deep learning for automated detection and numbering of permanent teeth on panoramic images [J]. Dento Maxillo Fac. Radiol. 51 (2), 20210296 (2022) YANG, Y., XIE, R., JIA, W., et al.: Accurate and automatic tooth image segmentation model with deep convolutional neural networks and level set method [J]. Neurocomputing. 419 , 108–125 (2021) CHUNG, M., LEE, M., HONG, J., et al.: Pose-aware instance segmentation framework from cone beam CT images for tooth segmentation [J]. Comput. Biol. Med. 120 , 103720 (2020) MURESAN M P, BARBURA A R, N.E.D.E.V.S.C.H.I.S.: Sept. Teeth Detection and Dental Problem Classification in Panoramic X-Ray Images using Deep Learning and Image Processing Techniques; proceedings of the 2020 IEEE 16th International Conference on Intelligent Computer Communication and Processing (ICCP), F 3–5 2020 [C]. (2020) YING, S., WANG, B., ZHU, H., et al.: Caries segmentation on tooth X-ray images with a deep network [J]. J. Dent. 119 , 104076 (2022) OKTAY A B, G.U.R.S.E.S.A.: Chapter 4 - Detection, segmentation, and numbering of teeth in dental panoramic images with mask regions with convolutional neural network features [M]//EL-BAZ A S, SURI J S. State of the Art in Neural Networks and their Applications. Academic Press. : 73–90. (2021) CHEN, H., ZHANG, K., LYU, P., et al.: A deep learning approach to automatic teeth detection and numbering based on object detection in dental periapical films [J]. 9 (1). (2019) SHAMIM M Z M, SYED, S., SHIBLEE, M., et al.: Automated detection of oral pre-cancerous tongue lesions using deep learning for early diagnosis of oral cavity cancer [J]. (2019) VINISKI A D, BARDDAL J P, DE SOUZA BRITTO JR, A, et al.: Incremental specialized and specialized-generalized matrix factorization models based on adaptive learning rate optimizers [J]. Neurocomputing. 552 , 126515 (2023) HABIB, G.: Optimization and acceleration of convolutional neural networks: A survey [J]. J. King Saud Univ. - Comput. Inform. Sci. 34 (7), 4244–4268 (2022) NA G S: Efficient learning rate adaptation based on hierarchical optimization approach [J]. Neural Netw. 150 , 326–335 (2022) BAHKALI I M, S.E.M.W.A.L.S.K. Multiple Ways for Medical Data Visualization Using 3D Slicer; proceedings of the 2020 International Conference on Computational Science and, Intelligence, C.: (CSCI), F 16–18 Dec. 2020 [C]. (2020) Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-3856379","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":266824564,"identity":"2c1a1d94-e1f2-4fb5-b60f-a728b0828d84","order_by":0,"name":"Wang Guangyan","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA00lEQVRIiWNgGAWjYPACGwbGBhDNRryWNAmStRyWgNDEaDE4fvbwy69t5+uY288YMHwoO8zAP7uBgJYzeWnWsm23JRh7cgwYZ5w7zCBx5wB+LWYHcsyMJUFaZvAYMPO2HWYwkEggoOX8G5CWcxAtf4nSciPH+OHHtgMQLYzEaLG/8caMmeFcsmRjT1rBwZ5z6TwSNwhokezPMf74o8yO37D98MYHP8qs5fhnENACBGzSvMDoMGxgYDgA5PEQVA8EzB9//GFgkCdG6SgYBaNgFIxMAACb7kOuXC5RtgAAAABJRU5ErkJggg==","orcid":"","institution":"Tianjin University of Commerce","correspondingAuthor":true,"prefix":"","firstName":"Wang","middleName":"","lastName":"Guangyan","suffix":""},{"id":266824565,"identity":"d8ec21e2-a326-4e05-856e-f989829e9204","order_by":1,"name":"Jia Yanan","email":"","orcid":"","institution":"Tianjin University of Commerce","correspondingAuthor":false,"prefix":"","firstName":"Jia","middleName":"","lastName":"Yanan","suffix":""},{"id":266824566,"identity":"2c0d9df8-a12e-43e7-9225-b9c50a52f929","order_by":2,"name":"Gulibstan Aihemaiti","email":"","orcid":"","institution":"Tianjin University of Commerce","correspondingAuthor":false,"prefix":"","firstName":"Gulibstan","middleName":"","lastName":"Aihemaiti","suffix":""},{"id":266824567,"identity":"7ad8db33-9b05-4ee4-adcd-a7d42890b9ca","order_by":3,"name":"Wang Kexin","email":"","orcid":"","institution":"Tianjin University of Commerce","correspondingAuthor":false,"prefix":"","firstName":"Wang","middleName":"","lastName":"Kexin","suffix":""},{"id":266824568,"identity":"8bb38e5a-642e-4983-b648-95f0c9c636f5","order_by":4,"name":"Qiao Feng","email":"","orcid":"","institution":"Tianjin Medical University","correspondingAuthor":false,"prefix":"","firstName":"Qiao","middleName":"","lastName":"Feng","suffix":""},{"id":266824569,"identity":"e7f3a583-e76b-448f-99fd-eddaa3fbccaa","order_by":5,"name":"Geng Duyan","email":"","orcid":"","institution":"Hebei University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Geng","middleName":"","lastName":"Duyan","suffix":""}],"badges":[],"createdAt":"2024-01-12 09:14:49","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-3856379/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-3856379/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":49709595,"identity":"e5168c85-aa27-4e28-9da9-a19ca5061e8f","added_by":"auto","created_at":"2024-01-16 19:37:56","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":48676,"visible":true,"origin":"","legend":"\u003cp\u003eCNN architecture diagram\u003c/p\u003e","description":"","filename":"fig1.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/6e9eeaaa6d25d3bb70439e25.png"},{"id":49710754,"identity":"4e77844b-54e4-424e-a326-c242334156bb","added_by":"auto","created_at":"2024-01-16 19:45:57","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":125641,"visible":true,"origin":"","legend":"\u003cp\u003eAlex Net Network Architecture\u003c/p\u003e","description":"","filename":"fig2.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/5108e023cacf2162b8de1fc3.png"},{"id":49711398,"identity":"f7c1f0ef-c9c3-441c-8002-de7fd4b08e1e","added_by":"auto","created_at":"2024-01-16 20:01:57","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":163578,"visible":true,"origin":"","legend":"\u003cp\u003eRes Net Network Architecture\u003c/p\u003e","description":"","filename":"fig3.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/8a4ca38a043f2628bc49f939.png"},{"id":49711669,"identity":"1b1127c5-c6a5-40c3-a64a-8d75f77bd2d6","added_by":"auto","created_at":"2024-01-16 20:09:57","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":98153,"visible":true,"origin":"","legend":"\u003cp\u003eVGG Net Network Architecture\u003c/p\u003e","description":"","filename":"fig4.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/851ed35933ac93bf8156a3fc.png"},{"id":49711123,"identity":"c9e895e1-7800-41f3-9160-264d963e384b","added_by":"auto","created_at":"2024-01-16 19:53:57","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":174986,"visible":true,"origin":"","legend":"\u003cp\u003eOverall design diagram of the plan\u003c/p\u003e","description":"","filename":"fig5.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/0d59314132c05b1a1c702825.png"},{"id":49711400,"identity":"c714d77b-b7fb-4ebf-a413-70ad5a787e79","added_by":"auto","created_at":"2024-01-16 20:01:57","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":262119,"visible":true,"origin":"","legend":"\u003cp\u003eCyst of jaw\u003c/p\u003e","description":"","filename":"fig6.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/57d8589686d2f407af4b8d14.png"},{"id":49709602,"identity":"9a10bbe5-4c15-42b4-87e1-6269a8e03958","added_by":"auto","created_at":"2024-01-16 19:37:57","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":99538,"visible":true,"origin":"","legend":"\u003cp\u003eExpanding Function Flowchart\u003c/p\u003e","description":"","filename":"fig7.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/8f6a90ce20c1e2768e955bb3.png"},{"id":49709597,"identity":"ece6edec-e31b-46cf-9bab-c9265cb4ae9b","added_by":"auto","created_at":"2024-01-16 19:37:57","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":60941,"visible":true,"origin":"","legend":"\u003cp\u003eOriginal image cysts labeling\u003c/p\u003e","description":"","filename":"fig8.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/951e2b9a56ffb9f95b7c999a.png"},{"id":49710753,"identity":"e7bc4a11-8042-4a62-8f72-69a3a92dc76f","added_by":"auto","created_at":"2024-01-16 19:45:56","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":27501,"visible":true,"origin":"","legend":"\u003cp\u003eSingle image prediction results\u003c/p\u003e","description":"","filename":"fig9.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/eee2b91bce58cb9d50d74062.png"},{"id":49709609,"identity":"88e621ca-7f01-4f6e-8472-38beafc74f03","added_by":"auto","created_at":"2024-01-16 19:37:57","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":199121,"visible":true,"origin":"","legend":"\u003cp\u003eBatch prediction of image results\u003c/p\u003e","description":"","filename":"fig10.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/be438e8f5ea277445ea04294.png"},{"id":49710758,"identity":"452dd24a-59c4-47c0-8694-ed273ebdf67f","added_by":"auto","created_at":"2024-01-16 19:45:57","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":44531,"visible":true,"origin":"","legend":"\u003cp\u003eClassified images\u003c/p\u003e","description":"","filename":"fig11.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/3b865580840bc89c868ef7c7.png"},{"id":49711126,"identity":"4602e2fc-2daf-4b24-adc7-f64a3327f510","added_by":"auto","created_at":"2024-01-16 19:53:57","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":57183,"visible":true,"origin":"","legend":"\u003cp\u003eADAM optimizer experimental results\u003c/p\u003e","description":"","filename":"fig12.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/e41d96d69ba8a5fc9eec4c34.png"},{"id":49710765,"identity":"fe458cca-a18a-47f2-8b34-83dcaccef38a","added_by":"auto","created_at":"2024-01-16 19:45:57","extension":"png","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":51570,"visible":true,"origin":"","legend":"\u003cp\u003eExperimental results of ADAM optimizer running for 100 rounds\u003c/p\u003e","description":"","filename":"fig13.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/95e5357a27f2c942b876b695.png"},{"id":49710762,"identity":"6fe81b7a-5c64-4546-af49-f6b66cedea10","added_by":"auto","created_at":"2024-01-16 19:45:57","extension":"png","order_by":14,"title":"Figure 14","display":"","copyAsset":false,"role":"figure","size":71441,"visible":true,"origin":"","legend":"\u003cp\u003eExperimental results with a learning rate of 0.01\u003c/p\u003e","description":"","filename":"fig14.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/bf1122aef5f998d59efd1580.png"},{"id":49710763,"identity":"a297c8f3-f66e-47e9-bc1e-4b5883b8616b","added_by":"auto","created_at":"2024-01-16 19:45:57","extension":"png","order_by":15,"title":"Figure 15","display":"","copyAsset":false,"role":"figure","size":71802,"visible":true,"origin":"","legend":"\u003cp\u003eExperimental results with a learning rate of 0.005\u003c/p\u003e","description":"","filename":"fig15.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/5a0275b7edcf394470b499ba.png"},{"id":49709615,"identity":"025581dd-0b79-4fd5-af85-2cf2f62bb040","added_by":"auto","created_at":"2024-01-16 19:37:57","extension":"png","order_by":16,"title":"Figure 16","display":"","copyAsset":false,"role":"figure","size":60069,"visible":true,"origin":"","legend":"\u003cp\u003eExperimental results with attenuation value of 0.0003\u003c/p\u003e","description":"","filename":"fig16.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/ea192071d122e77b9c6fb8cd.png"},{"id":49709613,"identity":"29e1ae87-bd27-4e93-8d32-6243ed1f883d","added_by":"auto","created_at":"2024-01-16 19:37:57","extension":"png","order_by":17,"title":"Figure 17","display":"","copyAsset":false,"role":"figure","size":61389,"visible":true,"origin":"","legend":"\u003cp\u003eExperimental results with attenuation value of 0.0008\u003c/p\u003e","description":"","filename":"fig17.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/4f8e8482579d140232cbd375.png"},{"id":49711129,"identity":"8789c528-6fe2-4c19-be94-b6893b5bf61c","added_by":"auto","created_at":"2024-01-16 19:53:57","extension":"png","order_by":18,"title":"Figure 18","display":"","copyAsset":false,"role":"figure","size":58650,"visible":true,"origin":"","legend":"\u003cp\u003eThe result of 100 rounds of training\u003c/p\u003e","description":"","filename":"fig18.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/9446f9c69ffc037bde9316ba.png"},{"id":49709614,"identity":"cc8f217b-53fa-49a7-96e0-ad162a3092ee","added_by":"auto","created_at":"2024-01-16 19:37:57","extension":"png","order_by":19,"title":"Figure 19","display":"","copyAsset":false,"role":"figure","size":38044,"visible":true,"origin":"","legend":"\u003cp\u003eLoss rate comparison\u003c/p\u003e","description":"","filename":"fig19.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/1374e52f93557dbd843d01bb.png"},{"id":49711670,"identity":"806f1d99-6c73-4567-bff1-f103d56ac5ea","added_by":"auto","created_at":"2024-01-16 20:09:57","extension":"png","order_by":20,"title":"Figure 20","display":"","copyAsset":false,"role":"figure","size":37855,"visible":true,"origin":"","legend":"\u003cp\u003eAccuracy comparison\u003c/p\u003e","description":"","filename":"fig20.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/d5d4df2c6cac02768311ce1f.png"},{"id":49709610,"identity":"47a9f1e1-97f1-46ea-bee4-953c15aa5699","added_by":"auto","created_at":"2024-01-16 19:37:57","extension":"png","order_by":21,"title":"Figure 21","display":"","copyAsset":false,"role":"figure","size":22321,"visible":true,"origin":"","legend":"\u003cp\u003eLoss rate of VGG Net training set\u003c/p\u003e","description":"","filename":"fig21.png","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/47d34924d44f4460cade7e5d.png"},{"id":50463750,"identity":"7de2d763-9083-4d87-bb3f-a285b151c538","added_by":"auto","created_at":"2024-01-31 22:37:16","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2197164,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-3856379/v1/b67d372a-927a-4ef7-be73-89cff19327d8.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Research on Cyst of Jaw Detection Algorithm Based on Alex Net Deep Learning Model","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eThere are various types of dental diseases, such as dental calculus and gingivitis, which can be observed directly by the naked eye, as well as jaw cysts, which can be diagnosed through dental images. Jaw cyst is a fluid-filled cyst found within the jaw bone. It is generally benign, with no signs of infiltration or growth, and does not metastasize to other sites. However, as it grows slowly over time and expands in size, it may ultimately squeeze the surrounding tissue, resulting in bone resorption, facial swelling, loose teeth, and tooth displacement in patients. This type of disease can be diagnosed through CBCT image. Cyst usually appear as circular or oval low-density areas with clear borders and smooth sharp edges. Depending on the type and size of the cyst, its internal structure and relationship with surrounding tissues may also vary. Medical images have high grayscale quality, the influence of noise and metal artifacts, and there are certain limitations\u003csup\u003e[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]\u003c/sup\u003e, which brings some difficulties to diagnosis. However, doctors have the problem of being unable to objectively and accurately diagnose images, and medical images are susceptible to external factors. The accuracy of shooting angles or operations can lead to images deviating from the true situation. Misdiagnosis caused by doctors' negligence in judgment can miss the best time for treatment, leading to the worst outcome of oral treatment. Introducing artificial intelligence to assist doctors in making judgments about images greatly improves the accuracy, efficiency and objectivity of diagnosis, which is also the purpose and significance of this study.\u003c/p\u003e \u003cp\u003eThis article mainly focuses on a series of studies on dental imaging. At present, the main type of medical image used in the dental industry is CBCT, also known as Cone beam Computer Tomography (CBCT)\u003csup\u003e[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]\u003c/sup\u003e. CBCT belongs to low-dose CT image, and the biggest difference compared to volumetric CT is that it can obtain high-definition three-dimensional images and reduce the impact of metal artifacts caused by two-dimensional slice stacking\u003csup\u003e[\u003cspan additionalcitationids=\"CR4\" citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]\u003c/sup\u003e. CBCT type dental images can be used for three-dimensional observation, allowing for better observation of teeth; The dental film can be scanned in a sitting position, which is more user-friendly; Due to low radiation dose, shorter acquisition time, higher X-ray utilization rate, and simpler and easier equipment operation; It also has extremely high isotropic spatial resolution, making it easy to observe the root canal structure of teeth. Therefore, this article chooses CBCT as the main research object.\u003c/p\u003e \u003cp\u003eWith the continuous development and progress of computer technology, deep learning has made a breakthrough in the field of clinical Medical imaging auxiliary diagnosis. In recent years, deep learning has also been successfully introduced into the field of Oral medicine\u003csup\u003e[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u003c/sup\u003e, especially the classic models represented by Convolutional neural network show strong robustness and universality, that is, to promote Oral medicine to become digital, intelligent and automated, so as to achieve data cooperation and sharing. Deep learning has more successful cases in the auxiliary diagnosis of common oral diseases, the identification, positioning and segmentation of anatomical structures in Oral medicine images, and guiding the fine clinical operation of dentists. Compared to previous traditional manual operations and physicians' diagnostic accuracy has been improved. Therefore, dental clinicians can make more accurate decision analysis of oral imaging data under the guidance of auxiliary diagnostic systems, reduce the probability of missed diagnosis and misdiagnosis, improve the accuracy and homogeneity of oral diagnosis, and thus improve the level of oral health services in China. This article aims to design a smart jaw cyst assistant detection system, which is based on the Alex Net model of deep learning network. After the model has been effectively learned and trained, the patient's dental image is input into the system, and the network will determine whether there is a cyst in the patient's jaw bone based on the training results. Clinical practice is an important basis for testing whether artificial intelligence can assist doctors in diagnosis and treatment, but there may be deviations from reality and a lack of feasibility. Digital healthcare should also be appropriately adjusted based on the facilities and medical resources in different places, fundamentally improving the medical gap and playing a role\u003csup\u003e[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]\u003c/sup\u003e.\u003c/p\u003e"},{"header":"2 Related research methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Basic principle of Convolutional neural network\u003c/h2\u003e \u003cp\u003eDeep learning is applied to various occasions in image processing, and also widely developed in the Medical imaging industry. In 1962, scholars Hubel and Wiesel first proposed the concept of CNN in biological neurology. Although the CNN\u003csup\u003e[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]\u003c/sup\u003e model only emerged in 2012, Lo\u003csup\u003e[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]\u003c/sup\u003e and others proposed the use of artificial neural networks for chest image detection in 1995, and CNN was used as a classifier for the final detection of lung diseases. Until 2012, the Alex Net model had better performance compared to traditional methods in the large-scale visual recognition challenge of ImageNet. Since then, CNN technology has emerged, and researchers have begun to widely use and research it. The full name of CNN is Convolutional neural network, which mainly consists of three modules, the convolutional layer (CONV), pooling layer (POOL), and full connection layer (FC). Figure\u0026nbsp;1shows the architecture of CNN, and the most important functional layer of Convolutional neural network is the convolutional layer. After receiving the signal in the input layer of CNN, it is processed in the early stage and sent to the convolutional working layer for operation. After the feature extraction is completed, it is sent to the fully connected layer for output.\u003c/p\u003e \u003cp\u003eCNN convolution layer is mainly composed of convolution kernel and filter, and it is not only composed of a simple layer, but can be formed by unlimited stacking and complex combination. With continuous stacking, the desired effect can be constantly approached, which is one of the reasons why many scholars continue to develop and study Convolutional neural network.\u003c/p\u003e \u003cp\u003eAfter the convolution operation is completed, it will enter the pooling layer step. Due to the continuous convolution operation and transformation of the input image signal in the convolution layer, the filter may perform a \"dimensionality increase\", and the pooling layer is also the process of \"dimensionality reduction\". The process of pooling is the step of filtering the matrix eigenvalues obtained in the convolutional layer. Displaying the maximum detected eigenvalues is called max-pooling \u003csup\u003e[\u003cspan additionalcitationids=\"CR11\" citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]\u003c/sup\u003e, and displaying the mean eigenvalues is called average-pooling\u003csup\u003e[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]\u003c/sup\u003e. The max-pooling method is relatively commonly used in convolutional networks. The assembly process after feature extraction is the fully connected layer, and finally fitting is performed to connect the extracted features, reducing the loss of feature information and outputting.\u003c/p\u003e \u003cp\u003eThe training process of Convolutional neural network is divided into two parts, forward propagation and backward propagation. The essence of forward propagation is to use the output of the previous layer to calculate the output of the next layer, and process the input image. It is the process of learning the training set. After processing, backward propagation detects the sensitivity of the entire network, pushes back the possible errors in the previous layers, checks the algorithm, and then adjusts the parameters of each layer based on the received errors. Both the pooling layer and the convolutional layer have backpropagation, so errors can be detected and adjusted in a timely manner.\u003c/p\u003e \u003cp\u003eCompared to traditional neural networks such as linear regression\u003csup\u003e[\u003cspan additionalcitationids=\"CR16\" citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]\u003c/sup\u003e and linear to nonlinear transformations\u003csup\u003e[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]\u003c/sup\u003e, CNN have two major advantages: local connections and weight sharing\u003csup\u003e[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]\u003c/sup\u003e. In traditional neural networks, all neurons in the network are connected, creating a high degree of interdependence between the input signal and all output results. Any errors in the intermediate steps can lead to significant deviations in the final output. In contrast, CNN process images through independent and independent outputs for different regions of the image. This allows for smaller impact of errors and reduces the overall loss in image processing. Weight sharing in CNN refers to the ability to share the parameters of a convolutional kernel across different parts of an image, regardless of their spatial locations. This eliminates the need to repeatedly add a convolutional kernel wherever it is needed and reduces the complexity of the entire network, as well as the number of parameters required for convolutional kernels. By enabling weight sharing, the same convolutional kernel can be used across the entire image, significantly reducing the number of parameters required for each kernel.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Alex Net model based on CNN network\u003c/h2\u003e \u003cp\u003eThe original convolutional network LeNet-5 from 1980\u003csup\u003e[\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]\u003c/sup\u003e was relatively outdated and had not yet proposed the use of stride and padding for optimization when dealing with images. It also used average-pooling, which resulted in a decrease in both the height and width of the images. In 2012, the Alex Net network\u003csup\u003e[\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]\u003c/sup\u003e emerged, introducing the ReLU activation function\u003csup\u003e[\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]\u003c/sup\u003e and switching to max-pooling, which improved the training speed of the network and reduced its complexity. In 2014, Simonyan proposed the VGG\u003csup\u003e[\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]\u003c/sup\u003e network, which better displayed the convolutional layer, increased the depth of the model, and truly simplified the neural network structure. In 2015, the Res-Net residual network\u003csup\u003e[\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]\u003c/sup\u003e was a deeper neural network that introduced cross-layer transmission combined with convolutional layers, greatly improving accuracy. It is also the most advanced model of convolutional neural networks to date, and these models are frequently used in analyzing dental images. Alex Net also uses multi-GPU training technology, which enables faster training of the model and can take advantage of parallel computing to accelerate the training process, making it more suitable for processing large data sets.\u003c/p\u003e \u003cp\u003eThe advantage of deep learning in the medical field lies in its applicability to large amounts of data. Due to the variety of perspectives and angles in medical image acquisition, the data volume is extremely large, resulting in high requirements for image processing systems. In recent years, the U-Net network model, CNN convolutional neural network, and Alex Net network architecture have been popular research directions, while other deep networks are also continuously being studied in the medical field. This article investigates a cyst of jaw detection system based on the Alex Net framework using CNN convolutional neural networks, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. The innovation of this model lies in the use of the ReLU activation function, which provides better performance than traditional Sigmoid and Tanh activation functions. It also implements LRN local response normalization and uses Droupout random neural deactivation in the first two fully connected layers to reduce overfitting. From the graph, it can be seen that the network has deepened, consisting of 5 convolutional layers, 3 pooling layers, 2 fully connected layers, and finally 1 softmax layer.\u003c/p\u003e \u003cp\u003eIn deep learning neural networks, activation functions map the output of neurons in the network through nonlinear functions, allowing the network to approximate any function arbitrarily well. The weighted input from a node is transformed into an activation or output of that input node. ReLU (rectified linear unit) is a piecewise linear function that extends the learning capability of neural networks and makes the network more stable. It has become the default activation function for many types of neural networks because models using it are easier to train and typically perform better. The ReLU function is linear for positive inputs, has fast convergence and calculation speed, fully passes the gradient, and does not suffer from gradient vanishing problems (saturated gradients).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Compared to other CNN network models\u003c/h2\u003e \u003cp\u003eThis article will compare two network models, Res Net and VGG Net, train the established dataset, and analyze the classification effect. The Res Net network model solves the gradient problem by connecting residual units, allowing for the construction of deeper neural networks, and focuses on solving the degradation problem caused by network deepening. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, the network architecture of Res Net50 includes 50 convolutional layers, which can extract more features. In order to reduce the error caused by backpropagation, the residual structure is added to complete image classification.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe VGG Net network structure is more concise, and uses small convolutional kernels instead of large convolutional layers, resulting in a reduction in parameters and a simpler overall structure that is easier to understand. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, the VGG Net network architecture includes 13 convolutional layers, 5 max pooling layers, 3 fully connected layers, and finally outputs the prediction result through soft max.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eSubsequent experiments will compare three deep learning network models to find the algorithm with the best classification performance.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Related research literature based on deep learning\u003c/h2\u003e \u003cp\u003eMachine learning algorithms for dental imaging use a combination of multiple algorithms or improvements. Deep learning is also constantly evolving, and various algorithms are constantly being improved and optimized, resulting in better results in medical image processing. Research based on the U-Net network model accounts for a high proportion. A.Fariza\u003csup\u003e[\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]\u003c/sup\u003e proposed an improved U-Net network model for automatic segmentation of teeth and background based on X-ray images of teeth, removing the influence of tooth overlap; Kirnbauer\u003csup\u003e[\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]\u003c/sup\u003e improved the U-Net architecture for binary segmentation of tooth apical lesions based on CBCT images, achieving a sensitivity of 97.1% for detecting lesion locations; Y. Rao\u003csup\u003e[\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]\u003c/sup\u003e implemented single tooth recognition and segmentation based on panoramic dental images, achieving an accuracy of 97.93% through training a U-Net model; Estai M\u003csup\u003e[\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]\u003c/sup\u003e developed a convolutional neural network-based system for detecting and classifying permanent teeth on surface tomographic images. The results showed sensitivities and precisions of 0.99 for the tooth detection module, and sensitivities, precisions, and F1 scores of 0.98 for the tooth numbering module.\u003c/p\u003e \u003cp\u003eSecondly, some research based on convolutional neural network models, Yang\u003csup\u003e[\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]\u003c/sup\u003e proposed a method to determine the location of dental pulp based on CNN, due to the blurred edges of teeth in CBCT images. They described the shape and size of teeth using mathematical methods and prior information. The experimental results demonstrated the feasibility and effectiveness of the model; Chung\u003csup\u003e[\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]\u003c/sup\u003e reduced the influence of metal artifacts on the segmentation process during dental implant simulation. They extracted patient alignment information using a pose regression neural network to obtain a volume of interest (VOI) region and realign the input image, reducing the mutual overlap between tooth boundary boxes. Then, they converted the pixel-wise labeling task into a distance regression task using a CNN network to segment individual teeth, which increased the accuracy by 30.3%; M.P Muresan\u003csup\u003e[\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e]\u003c/sup\u003e trained a CNN using labeled data to obtain semantic segmentation information. They performed multiple image processing operations to segment and refine the boundary boxes corresponding to tooth detection; Ying\u003csup\u003e[\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]\u003c/sup\u003e proposed a method for caries segmentation based on dental X-ray images, using visual transformation, extended convolution and feature pyramid fusion to enhance multi-scale and global feature extraction capabilities. The Dice coefficient reached 74.87%.\u003c/p\u003e \u003cp\u003eSecondly, there are studies on the classification of dental images. Oktay \u003csup\u003e[\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]\u003c/sup\u003eproposed a RCNN-based method for detecting, segmenting, and classifying teeth in panoramic X-ray images; Chen\u003csup\u003e[\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]\u003c/sup\u003e used a TensorFlow-based R-CNN network to classify and number the periapical periodontal membranes, achieving an accuracy and recall rate of over 90%. The performance of the entire network system almost reached the level of a primary dentist; Shamim\u003csup\u003e[\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]\u003c/sup\u003e used a Vgg19-based DCNN model to classify benign lesions and precancerous lesions of the tongue, achieving an accuracy of 98%. ResNet50 can achieve an accuracy of 97%, almost reaching the \"human-like\" classification level of doctors. The extremely high accuracy provides possibilities for screening oral cancer.\u003c/p\u003e \u003cp\u003eDeep learning methods are fast and efficient in processing medical imaging data. Their excellent performance enables rapid classification, localization, and other operations in oral analysis and processing, improving the efficiency of dentists' diagnosis and treatment. However, research in the field of oral medicine is still in its infancy and has bright prospects to be explored. Currently, although convolutional neural networks (CNN) are a very effective method in recent years, the construction of their networks and the adjustment of their parameters cannot be fully explained, and it still requires continuous attempts to build suitable models. From traditional algorithms to higher-level algorithms in deep learning, each has its own advantages. In the experimental process, multiple algorithms are used for comparison and analysis of experimental results. The introduction of machine learning and deep learning in medical imaging reduces potential medical risks and can also appropriately share the repetitive and complex work of doctors, thereby improving work efficiency and reducing work intensity.\u003c/p\u003e \u003c/div\u003e"},{"header":"3 Research on jaw cyst detection algorithm based on deep learning network model","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Basic scheme design\u003c/h2\u003e \u003cp\u003eThis article will design a CNN-based cysts detection system that can predict whether a patient contains cysts based on its input. The system will be developed using Windows 10 as the operating environment and compiled in PyCharm using the Python language. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, the design process begins with collecting and organizing input data, followed by image preprocessing including de-noising and filtering, as well as data set expansion for limited data sets. The preprocessed data is then input into a pre-built CNN model for training, and the optimal model is stored. The system then validates the results using a test set and a validation set. Finally, the experimental results are presented through loss functions and accuracy rates of the validation set, test set, and training set.\u003c/p\u003e \u003cp\u003eThe specific design process of the plan is as follows:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cli\u003e \u003cp\u003eThe dental image dataset used in this study primarily consists of CBCT, which are currently the most widely used type of dental image. Prior to conducting the experiment, the collected tooth images were divided into two categories according to the dentist's annotation: Healthy teeth and teeth with jaw cyst disease. The data set was then allocated according to a 7:2:1 ratio. As this study used a self-collected dataset, the sample size and type of lesion were limited.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eIn order to prevent interference from extraneous factors, the teeth portion was cropped separately and filled with 0-pixels. Due to the small size of the dataset, techniques such as adjusting image contrast and brightness were used to augment the dataset and improve the accuracy of the training model.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eBased on the Alex Net framework, a model was built considering the high requirements of the deep learning runtime environment to accelerate training time. The entire Alex Net model was scaled down proportionally, with the input image size set to 65 x 65 pixels. The preprocessed dataset was then input into the model for training, resulting in three model training outcome graphs: the training set loss graph, the validation set loss graph, and the validation set accuracy graph.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThe trained model will generate three graphs showing the results of the training, validation, and testing sets, as well as the predicted outcome of the image. First, the unlabeled test set of tooth images will be used as input into the system, undergoing preprocessing and cysts labeling before entering the classification system. The system will display the predicted results for the test set, and finally classify and store the predicted tooth images. During training, network models will be optimized by continuously adjusting parameters.\u003c/p\u003e \u003c/li\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Dataset establishment\u003c/h2\u003e \u003cp\u003eThe dataset used in this article is derived from actual cases from Tianjin Medical University Stomatology Hospital, but the data quantity is relatively small, which poses certain limitations. The collected oral CT dataset is divided into three parts, divided according to the ratio of 7:2:1. the first part is used to update model parameters and improve performance, with the largest amount of data, consisting of 350 images. The second part is the validation set, responsible for adjusting the model's hyper parameters, the training set for 100 images. The third part is the test set, used to evaluate the accuracy of the model, primarily assessing the difference between the CNN model's estimated predictions and actual predictions, with the smallest amount of data, consisting of 50 images. The dataset distribution is presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Please note that despite the relatively small size of the dataset, efforts have been made to ensure its representativeness and accuracy through various methods mentioned above. This experiment plans to run 50 and 100 rounds of comparison of three algorithms, using the same dataset to obtain more accurate classification results, and observed and analyzed the influence of the number of operations.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eData set allocation\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003eData set allocation\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eName\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTraining set\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eValidation set\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTesting set\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eData volume\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e350\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e100\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e50\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cdiv id=\"Sec10\" class=\"Section3\"\u003e \u003ch2\u003e3.2.1 Dataset evaluation criteria\u003c/h2\u003e \u003cp\u003eCBCT dental images are based on X-rays. Because teeth block X-rays, the enamel appears as a high-brightness white color in the panoramic film, while the root canals in the middle of the teeth appear as a lighter gray color. When there is a cyst in the jawbone, a large amount of fluid accumulates in the tooth tissue, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e, which appears as a black shadow on the dental image, and the size of the shadow indicates the amount of fluid accumulation. From this, it can be designed to preprocess all raw images and then put them into the Alex Net model for learning and training. The system will diagnose jaw cyst disease based on the presence of shadows on CBCT images. In addition, CBCT can clearly show the positional relationship between the cyst and surrounding teeth. If the cyst is connected to the tooth, residual roots or teeth may be visible, or there may be a space within the cyst, and the continuity of the bone cortex may be interrupted. For larger jaw cysts, swelling in the surrounding soft tissues may be seen.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Algorithm optimization and experimental results\u003c/h2\u003e \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e \u003ch2\u003e3.3.1 Model optimization\u003c/h2\u003e \u003cp\u003eFor the prediction of disease based on higher resolution and more precise oral and maxillofacial cone-beam CT images, the following improvements are proposed with a scheme design shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cli\u003e \u003cp\u003eImprove the CNN model: modify the optimizer and optimizer parameters to compare the changes in training and validation losses and accuracy of various training results, and analyze and select appropriate improvement methods to maximize the similarity of the model simulation results to the ideal state.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eAttempt to identify the parts of the teeth that contain cysts: theoretically, the cysts site can be separately marked through image segmentation, and then image overlap can be used to overlap the marked portion with the original image, thereby achieving a clear display of the cysts part on the dental images. After labeling, the size and location of the cyst can be visually detected.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eImplement batch prediction of images: if the number of images input into the system for prediction is too large, the efficiency of single-image prediction will decrease with the increase of data volume. To solve this problem, a function that can switch between single-image prediction and batch prediction based on the number of input images can effectively address the issue of large volumes of images to be predicted. Additionally, if the batch-predicted images can be stored and organized according to their prediction results, it will further facilitate secondary judgments of the predicted images by doctors.\u003c/p\u003e \u003c/li\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eThe following is a scheme design for simulation verification based on four steps:\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section3\"\u003e \u003ch2\u003e3.3.2 Dental cysts markers\u003c/h2\u003e \u003cp\u003eCBCT images are a collection of cross-sectional views of teeth, with a vertical perspective, so a watershed algorithm can be used for image segmentation. This algorithm divides the image based on different gray values, resulting in a vertical perspective of the image with different levels of division. The cyst of jaw inside the dental tissue presents a dark image with large shadows, which is different from normal dental tissue. This image will be processed by the Alex Net model to threshold and convert it into a binary image. The distance transformation result is adjusted to determine the background (healthy tooth tissue), and the thresholding result is marked. Then, the watershed algorithm is used to segment the image, and the image markers are set as green. The background of the labeled decayed portion is processed as transparent and overlaid on the original CBCT image, resulting in the marked results of cysts shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section3\"\u003e \u003ch2\u003e3.3.3 Batch prediction function\u003c/h2\u003e \u003cp\u003eWhen dealing with large amounts of data, using a single image for prediction can be inefficient. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e, this is the result of predicting with a single image. To address this issue, we design a function that can perform batch recognition. The main parameters are the model address, the folder address where the images need to be detected, and the result address where the predictions will be saved. The predicted results will be stored in the corresponding folders. Finally, each image's corresponding prediction results will be output as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e, and the labeled folders completed according to the results will be displayed in the result saving address as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"4 Expand function implementation","content":"\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Change optimizer\u003c/h2\u003e \u003cp\u003eOptimizers play a crucial role in machine learning, as they update the model parameters to minimize the training error based on mathematical operations involving gradients and loss functions. After each sample passes through the model, it produces an estimated value, but due to the model's imperfections, this estimated value often differs from the actual value, and the difference is referred to as the error or loss. The smaller the loss, the more successful the model. The function introduced to calculate this loss is called the loss function.\u003c/p\u003e \u003cp\u003eGradient descent and adaptive moment estimation are widely used optimization methods. The choice of optimizer can lead to different experimental results.\u003c/p\u003e \u003cp\u003e(1)Stochastic Gradient Descent (SGD)\u003c/p\u003e \u003cp\u003eStochastic Gradient Descent\u003csup\u003e[\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e]\u003c/sup\u003e does not use the loss function on the entire training dataset, but in each iteration, it randomly optimizes the loss function on a single training example. Therefore, the parameter update speed is faster in each round. SGD is commonly used for training on large-scale datasets because it can effectively reduce training time. However, since it only considers the information from a randomly selected sample, which may not represent the trend of the entire dataset, it can easily fall into local optimal situations. Therefore, in experimental processes, it is possible to reduce the learning rate or increase the number of gradient descent iterations to enable the algorithm to better converge to the optimal solution.\u003c/p\u003e \u003cp\u003e(2)Adaptive Moment Estimation (ADAM)\u003c/p\u003e \u003cp\u003eAdaptive Moment Estimation\u003csup\u003e[\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]\u003c/sup\u003e is an optimization algorithm proposed in 2014, which combines the advantages of adaptive learning rate gradient descent algorithms and momentum gradient descent algorithms. It can adapt to sparse gradients (as in natural language and computer vision problems) and is well-suited for large-scale data and parameter scenarios. Although its implementation is simple, it requires little memory, and is very convenient to use, it is prone to oscillation around the optimal value, so its computational results are slightly inferior to stochastic gradient descent.\u003c/p\u003e \u003cp\u003eIn the experiment, it was found that changing the optimizer to ADAM significantly increased the training speed and the loss generally showed a decreasing trend. Figure\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e shows the training results when the optimizer was changed to ADAM, and it can be seen that the training set loss showed a decreasing trend until a small fluctuation at around 36 iterations. However, the validation set loss experienced a significant increase around 36 iterations, and the validation set accuracy fluctuated widely. Therefore, when using ADAM as the optimizer, there are large fluctuations around the optimal value, and after 40 iterations, the validation set loss still shows an upward trend, indicating that the overfitting problem has not been resolved. Moreover, ADAM performs worse than SGD in terms of performance. Therefore, it is recommended to use SGD as the optimizer. As shown in the Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e, the experimental results of 100 rounds of running are shown. It can be seen that there is still overfitting in the loss rates of the training set and the validation set, but the more times of training, the better the effect.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Optimizer parameter change\u003c/h2\u003e \u003cp\u003eThe optimizer selects the SGD gradient descent method optimizer, where the parameters lr are learning rate, momentum is momentum, and weight_ Decay is the weight attenuation, and the parameter changes are as follows:\u003c/p\u003e \u003cp\u003e(1)Change learning rate\u003c/p\u003e \u003cp\u003eWhen using SGD as the optimizer, changing the learning rate: Generally, the learning rate of deep learning is between 0.01 and 0.001. In this design, the learning rate was initially set to the maximum value of 0.01 and the decay initial value was set to 0.0005 to shorten training time. However, it was found that overfitting occurred between 20 and 40 rounds of training, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e14\u003c/span\u003e. Therefore, the learning rate was reduced to 0.005, resulting in the training results shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig15\" class=\"InternalRef\"\u003e15\u003c/span\u003e. It can be seen that the training loss function and the validation loss function both show an ideal decreasing trend, and the training loss is lower than the validation loss, solving the problem of overfitting. In the later stages of training, the validation accuracy is high and stable, solving the problem of overfitting. Therefore, reducing the learning rate is very effective in improving CNN.\u003c/p\u003e \u003cp\u003e(2)Reduce attenuation value\u003c/p\u003e \u003cp\u003eNext, with SGD as the optimizer and a learning rate of 0.005, the learning rate decay value was changed from 0.0005 to 0.0003, resulting in the training results shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig16\" class=\"InternalRef\"\u003e16\u003c/span\u003e. It can be seen that the loss suddenly rises and fluctuates greatly in the later stages of training, and the validation loss is still smaller than the training loss. The accuracy has a large fluctuation and a relatively high value. Therefore, it can be concluded that the training results with a lower learning rate decay value are not as good as the training results with a decay value of 0.0005.\u003c/p\u003e \u003cp\u003e(3)Increase attenuation value\u003c/p\u003e \u003cp\u003eBecause the training results with a lower learning rate decay value were not ideal, attempts were made to increase the decay value and obtain the result shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig17\" class=\"InternalRef\"\u003e17\u003c/span\u003e with a decay value of 0.0008. It can be seen that the training loss presents a downward curve, and the validation loss graph has fluctuations from 21 to 30 rounds, but gradually tends to a stable descending state. The validation loss value is significantly lower than the training loss value, and the whole model is in an ideal state. Moreover, the validation accuracy tends to be stable after 30 rounds and increases as the number of training rounds increases. Therefore, the improved method of increasing the learning rate decay value is selected, so the learning rate is set to 0.05 and the decay value is set to 0.0008 as the final parameters. As shown in the Fig.\u0026nbsp;\u003cspan refid=\"Fig18\" class=\"InternalRef\"\u003e18\u003c/span\u003e, the results of training 100 rounds with the final adjusted parameters show that there is still overfitting in the first 20 rounds, but it gradually stabilizes later. Therefore, the more training rounds, the more accurate the classification effect of the model.\u003c/p\u003e\u003cp\u003eThe experiment improved the model training results by changing the dataset, reducing the learning rate of the optimizer, and other methods. The average accuracy was increased to 75%, and the highest accuracy was increased to 99.3%, solving the problem of validation loss being lower than data loss and overfitting. The average accuracy of the model reached 83.5%, and the highest accuracy reached 99.9%, achieving a relatively high cyst of jaw identification rate on the existing dataset.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Compare with the other two algorithms\u003c/h2\u003e \u003cdiv id=\"Sec19\" class=\"Section3\"\u003e \u003ch2\u003e4.3.1Analysis of classification results of Res Net algorithm\u003c/h2\u003e \u003cp\u003eThe experiment compared the Res Net algorithm to classify the same dataset, running for 100 rounds and using the SGD optimizer. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig19\" class=\"InternalRef\"\u003e19\u003c/span\u003e, the loss rates of the training and validation sets fluctuated greatly, and overfitting remained a serious problem after 100 rounds of training, with the longest running time. Figure\u0026nbsp;\u003cspan refid=\"Fig20\" class=\"InternalRef\"\u003e20\u003c/span\u003e shows the accuracy rates of the training and validation sets, which also fluctuated erratically and were far inferior to the classification performance of the Alex Net network.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section3\"\u003e \u003ch2\u003e4.3.2 Analysis of classification results of VGG Net algorithm\u003c/h2\u003e \u003cp\u003eVGG Net has the worst classification performance, with a high loss rate in the training set and severe overfitting, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig21\" class=\"InternalRef\"\u003e21\u003c/span\u003e. The classification accuracy is only 50%, and the classification performance is poor and takes a long time.\u003c/p\u003e \u003cp\u003eAccording to the comparison of the effects of the above three algorithms, Alex Net performs the best on existing datasets, with high classification accuracy and fast processing speed; Res Net algorithm has high accuracy but suffers from overfitting and large fluctuations, and due to the large network depth and relatively large number of parameters, it results in long processing time; VGG Net algorithm has the lowest accuracy, and due to the relatively simple network structure, it suffers from overfitting and needs to be improved and upgraded.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"5 Conclusion","content":"\u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003e5.1 Experimental Summary\u003c/h2\u003e \u003cp\u003eThe above experimental results indicate that taking X-ray dental images is an effective diagnostic method for the timely treatment and prevention of worsening oral diseases such as Jawbone cysts. Oral health is an increasingly concerned medical issue in society, and analyzing dental images as a prerequisite for diagnosis is crucial for the diagnosis and treatment of oral diseases. In the field of computer vision, convolutional neural networks have prominent advantages, and the Deep Learning-based Alex Net network model based on PyTorch can be effectively applied to oral medical image analysis, providing powerful support for dentists in disease recognition. However, despite the achievements of our experiment, there are still some areas that need to be improved and perfected:\u003c/p\u003e \u003cp\u003e(1)In terms of dataset\u003c/p\u003e \u003cp\u003eFirstly, the size and quality of the dataset have a certain impact on the learning effect of deep learning models. Due to the limited protection of patient privacy data, the collection efficiency is not high, and the amount of data used in this article is not sufficient to fully support the fine-grained learning of the model. In addition, this experiment only targeted young and middle-aged people, and did not cover data from different age groups, such as children, the elderly, and denture wearers, which limited the scope of application of the model. Therefore, it is hoped that future research can address the limitations of dataset recognition and expand the application scenarios of models.\u003c/p\u003e \u003cp\u003e(2)Display of data results\u003c/p\u003e \u003cp\u003eSecondly, the display of data results still needs to be optimized. Although the experiment proposed in this article can determine whether there is a cyst in an image, it cannot visualize the location and extent of the cyst. In addition, there is an unstable state during model training, and there are fluctuations in loss value and accuracy. Follow-up studies can optimize the experiment, such as labeling the cyst range and using large data sets to train the model to improve accuracy.\u003c/p\u003e \u003cp\u003eThe final judgment result of the experiment is based on visual inspection, which lacks reliability. Future work can include medical evaluation indicators in the experiment code, and make the final conclusion based on the most accurate and objective numerical data.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section2\"\u003e \u003ch2\u003e5.2 Prospect\u003c/h2\u003e \u003cp\u003e \u003col\u003e \u003cli\u003e \u003cp\u003eMedical images are often stored in the DICOM electronic format, different software applications may require different export formats for processing, and there are also issues with accessing specialized medical imaging analysis and processing software such as 3D-slicer\u003csup\u003e[\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e]\u003c/sup\u003e, which may be inaccessible to non-professionals and difficult to operate without guidance. The open-source database for dental image analysis needs to be further improved. It is hoped that a more comprehensive database can be established to support broader applications in research, in terms of quantity, type, and quality.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eResearchers in the field of information processing lack professional medical knowledge, and it is important to establish a comprehensive evaluation standard system in the future. The lack of understanding of professional medical jargon by information technology personnel undoubtedly increases the difficulty of the experimental process. A new evaluation standard would be more accessible to non-professionals and would facilitate their understanding and judgment. Non-medical professionals should also have a deeper understanding of the structure and morphology of tooth tissue to better analyze the causes of dental diseases.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eAlthough traditional methods are still being optimized continuously, deep learning has enormous potential in fields such as image segmentation. Therefore, future research directions can include optimizing the training methods of deep learning models, such as using cross-validation and oversampling to improve model training effectiveness and generalization ability. Additionally, the training of deep learning models requires a strong database as support, which is also one of the serious challenges faced by digital healthcare. How to enhance the practicality of data and quickly, efficiently, and accurately analyze larger datasets is also a key direction for future research. models.\u003c/p\u003e \u003c/li\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eThe medical industry is an area with low error tolerance, and the training process of digital algorithms can be repeatedly tested and trained in other industries, but cannot be achieved in the medical industry. The unknown, scarcity, and practicality of medical data are greatly limited. Deep learning network models require powerful databases as support, which is also an extremely serious challenge for digital healthcare. There is still a lot of room for improvement in the application of image segmentation technology in medicine. How to enhance the practicality of data and quickly, efficiently, and accurately analyze larger datasets is also the main focus of future research. The medical industry still has a long process of effort from proposing new ideas to applying them to practical situations, and auxiliary diagnosis and treatment should also start from the foundation and continuously make improvements based on the actual situation.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003eFunding\u003c/p\u003e\n\u003cp\u003eNo funding was received for conducting this study.\u003c/p\u003e\n\u003cp\u003eConflict of interest/Competing interests (check journal-specific guidelines for which heading to use)\u003c/p\u003e\n\u003cp\u003eThe authors have no relevant financial or non-financial interests to disclose.\u003c/p\u003e\n\u003cp\u003eEthics approval\u003c/p\u003e\n\u003cp\u003eAttachment submission\u003c/p\u003e\n\u003cp\u003eConsent to participate\u003c/p\u003e\n\u003cp\u003eAll authors agree\u003c/p\u003e\n\u003cp\u003eConsent for publication\u003c/p\u003e\n\u003cp\u003eAll authors agree\u003c/p\u003e\n\u003cp\u003eAvailability of data and materials\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003eCode availability\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003eAuthors\u0026rsquo; contributions\u003c/p\u003e\n\u003cp\u003eW and J wrote the main content of the manuscript, G and W were responsible for the preliminary investigation and preparation, Q provided data, and G was responsible for reviewing and revising the entire research content.\u003c/p\u003e\n\u003cp\u003eW: Wang Guangyan\u003c/p\u003e\n\u003cp\u003eJ: Jia Yanan\u003c/p\u003e\n\u003cp\u003eG: Gulibstan Aihemaiti\u003c/p\u003e\n\u003cp\u003eW: Wang Kexin\u003c/p\u003e\n\u003cp\u003eQ: Qiao Feng\u003c/p\u003e\n\u003cp\u003eG: Geng Duyan\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eLI, P., LIU, Y., CUI, Z., et al.: Semantic Graph Attention With Explicit Anatomical Association Modeling for Tooth Segmentation From CBCT Images [J]. IEEE Trans. Med. Imaging. \u003cb\u003e41\u003c/b\u003e(11), 3116\u0026ndash;3127 (2022)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHAO, J., ZHANG, L., LI, L., et al.: Nov. An improved non-local means regularized iterative reconstruction method for low-dose dental CBCT; proceedings of the 2012 IEEE Nuclear Science Symposium and Medical Imaging Conference Record (NSS/MIC), F 27 Oct.-3 2012 [C]. (2012)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTHAKUR, A., PARGAIN, V., SINGH, P., et al.: An efficient fuzzy and morphology based approach to metal artifact reduction from dental CBCT image; proceedings of the 2017 International Conference on Computing, Communication and Automation (ICCCA), F 5\u0026ndash;6 May 2017, 2017 [C]\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAHN, J., NGUYEN T P, KIM, Y-J, et al.: Automated analysis of three-dimensional CBCT images taken in natural head position that combines facial profile processing and multiple deep-learning models [J], vol. 226, p. 107123. Computer Methods and Programs in Biomedicine (2022)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLIUGANG, G., JIANFENG, S.: Metal Artifact Reduction Method Based on Noncoplanar Scanning in CBCT Imaging [J]. IEEE Access. \u003cb\u003e8\u003c/b\u003e, 7236\u0026ndash;7243 (2020)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYANG, J., XIE, Y., LIU, L., et al.: Automated Dental Image Analysis by Deep Learning on Small Dataset; proceedings of the 2018 IEEE 42nd Annual Computer Software and Applications Conference (COMPSAC), F]\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKHERA, R., BUTTE A J, BERKWITS, M., et al.: AI in Medicine\u0026mdash;JAMA\u0026rsquo;s Focus on Clinical Outcomes, Patient-Centered Care, Quality, and Equity [J]. JAMA. \u003cb\u003e330\u003c/b\u003e(9), 818\u0026ndash;820 (2023)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMAHMUD B U, HONG G Y:. Semantic Image Segmentation using CNN (Convolutional Neural Network) based Technique; proceedings of the 2022 IEEE World Conference on Applied Intelligence and Computing (AIC), F 17\u0026ndash;19 June 2022, 2022 [C]\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLO, S-C B, CHAN H-P, LIN, J-S, et al.: Artificial convolution neural network for medical image pattern recognition [J]. Neural Netw. \u003cb\u003e8\u003c/b\u003e(7), 1201\u0026ndash;1214 (1995)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMURRAY N, PERRONNIN F. Generalized Max Pooling; proceedings of the 2014 IEEE Conference on Computer Vision and Pattern Recognition, F 23\u0026ndash;28: 2014 [C]. (2014)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMASCI, J., GIUSTI, A., CIRESAN, D., et al.: Sept. A fast learning algorithm for image segmentation with max-pooling convolutional networks; proceedings of the 2013 IEEE International Conference on Image Processing, F 15\u0026ndash;18 2013 [C]. (2013)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCHRISTLEIN, V., SPRANGER, L., SEURET, M., et al.: Deep Generalized Max Pooling; proceedings of the 2019 International Conference on Document Analysis and Recognition (ICDAR), F 20\u0026ndash;25 Sept. 2019, 2019 [C]\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYANG, J., XIE, F., FAN, H., et al.: Classification for Dermoscopy Images Using Convolutional Neural Networks Based on Region Average Pooling [J]. IEEE Access. \u003cb\u003e6\u003c/b\u003e, 65130\u0026ndash;65138 (2018)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWANG, S., JIANG, Y., HOU, X., et al.: Cerebral Micro-Bleed Detection Based on the Convolution Neural Network With Rank Based Average Pooling [J]. IEEE Access. \u003cb\u003e5\u003c/b\u003e, 16576\u0026ndash;16583 (2017)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSHAKOURI, H., NADIMI G R, GHADERI, F.: Dec. Fuzzy linear regression models with absolute errors and optimum uncertainty; proceedings of the 2007 IEEE International Conference on Industrial Engineering and Engineering Management, F 2\u0026ndash;4 2007 [C]. (2007)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKAVITHA, S., VARUNA, S.: RAMYA R. A comparative analysis on linear regression and support vector regression; proceedings of the 2016 Online International Conference on Green Engineering and Technologies (IC-GET), F 19\u0026ndash;19 Nov. 2016, 2016 [C]\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHIROSE, H., SOEJIMA, Y.: HIROSE K. NNRMLR: A Combined Method of Nearest Neighbor Regression and Multiple Linear Regression; proceedings of the 2012 IIAI International Conference on Advanced Applied Informatics, F 20\u0026ndash;22 Sept. 2012 [C]. (2012)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSONTAG E. From linear to nonlinear: some complexity comparisons; proceedings of the Proceedings of 1995 34th IEEE Conference on Decision and Control, F 13\u0026ndash;15: 1995 [C]. (1995)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZHANG, H., HE, J.: KO S B. Improved Hybrid Memory Cube for Weight-Sharing Deep Convolutional Neural Networks; proceedings of the 2019 IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS), F 18\u0026ndash;20 March 2019, 2019 [C]\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMOHAZZEBI, M., EFTEKHARI, M.: SHARIATZADEH M. How weight-sharing mechanisms affect the performance of deep Siamese networks; proceedings of the 2022 9th Iranian Joint Congress on Fuzzy and Intelligent Systems (CFIS), F 2\u0026ndash;4 March 2022, 2022 [C]\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eISLAM, M.R.: MATIN A. Detection of COVID 19 from CT Image by The Novel LeNet-5 CNN Architecture; proceedings of the 2020 23rd International Conference on Computer and Information Technology (ICCIT), F 19\u0026ndash;21 Dec. 2020 [C]. (2020)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRANI, S., KUMAR, S., GHAI, D., et al.: Automatic Detection of Brain Tumor from CT and MRI Images using Wireframe model and 3D Alex-Net; proceedings of the 2022 International Conference on Decision Aid Sciences and Applications (DASA), F 23\u0026ndash;25 March 2022, 2022 [C]\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSTURSA, D.: DOLEZEL P. Comparison of ReLU and linear saturated activation functions in neural network for universal approximation; proceedings of the 2019 22nd International Conference on Process Control (PC19), F 11\u0026ndash;14 June 2019, 2019 [C]\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBAGASKARA, A.: SURYANEGARA M. Evaluation of VGG-16 and VGG-19 Deep Learning Architecture for Classifying Dementia People; proceedings of the 2021 4th International Conference of Computer and Informatics Engineering (IC2IE), F 14\u0026ndash;15 Sept. 2021, 2021 [C]\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZHOU, J., JIN, Y.I.S.: F. Super resolution reconstruction of single image based on Res-Net and Sub-Pixel; proceedings of the 2022 IEEE 2nd International Conference on Electronic Technology, Communication and Information (ICETCI), F 27\u0026ndash;29 May 2022, 2022 [C]\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFARIZA, A., ARIFIN A Z, A.S.T.U.T.I.E.R.: Automatic Tooth and Background Segmentation in Dental X-ray Using U-Net Convolution Network; proceedings of the 2020 6th International Conference on Science in Information Technology (ICSITech), F 21\u0026ndash;22 Oct. 2020 [C]. (2020)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKIRNBAUER, B., HADZIC, A., JAKSE, N., et al.: Automatic Detection of Periapical Osteolytic Lesions on Cone-beam Computed Tomography Using Deep Convolutional Neuronal Networks [J]. Journal of Endodontics (2022)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRAO, Y., NARTEY O T, Z.E.N.G.S., et al.: LeFUNet: UNet with Learnable Feature Connections for Teeth Identification and Segmentation in Dental Panoramic X-ray Images; proceedings of the 2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), F 6\u0026ndash;8 Dec. 2022, 2022 [C]\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eESTAI, M., TENNANT, M.: Deep learning for automated detection and numbering of permanent teeth on panoramic images [J]. Dento Maxillo Fac. Radiol. \u003cb\u003e51\u003c/b\u003e(2), 20210296 (2022)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYANG, Y., XIE, R., JIA, W., et al.: Accurate and automatic tooth image segmentation model with deep convolutional neural networks and level set method [J]. Neurocomputing. \u003cb\u003e419\u003c/b\u003e, 108\u0026ndash;125 (2021)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCHUNG, M., LEE, M., HONG, J., et al.: Pose-aware instance segmentation framework from cone beam CT images for tooth segmentation [J]. Comput. Biol. Med. \u003cb\u003e120\u003c/b\u003e, 103720 (2020)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMURESAN M P, BARBURA A R, N.E.D.E.V.S.C.H.I.S.: Sept. Teeth Detection and Dental Problem Classification in Panoramic X-Ray Images using Deep Learning and Image Processing Techniques; proceedings of the 2020 IEEE 16th International Conference on Intelligent Computer Communication and Processing (ICCP), F 3\u0026ndash;5 2020 [C]. (2020)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYING, S., WANG, B., ZHU, H., et al.: Caries segmentation on tooth X-ray images with a deep network [J]. J. Dent. \u003cb\u003e119\u003c/b\u003e, 104076 (2022)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOKTAY A B, G.U.R.S.E.S.A.: Chapter 4 - Detection, segmentation, and numbering of teeth in dental panoramic images with mask regions with convolutional neural network features [M]//EL-BAZ A S, SURI J S. State of the Art in Neural Networks and their Applications. Academic Press. : 73\u0026ndash;90. (2021)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCHEN, H., ZHANG, K., LYU, P., et al.: A deep learning approach to automatic teeth detection and numbering based on object detection in dental periapical films [J]. \u003cb\u003e9\u003c/b\u003e(1). (2019)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSHAMIM M Z M, SYED, S., SHIBLEE, M., et al.: Automated detection of oral pre-cancerous tongue lesions using deep learning for early diagnosis of oral cavity cancer [J]. (2019)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVINISKI A D, BARDDAL J P, DE SOUZA BRITTO JR, A, et al.: Incremental specialized and specialized-generalized matrix factorization models based on adaptive learning rate optimizers [J]. Neurocomputing. \u003cb\u003e552\u003c/b\u003e, 126515 (2023)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHABIB, G.: Optimization and acceleration of convolutional neural networks: A survey [J]. J. King Saud Univ. - Comput. Inform. Sci. \u003cb\u003e34\u003c/b\u003e(7), 4244\u0026ndash;4268 (2022)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNA G S: Efficient learning rate adaptation based on hierarchical optimization approach [J]. Neural Netw. \u003cb\u003e150\u003c/b\u003e, 326\u0026ndash;335 (2022)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBAHKALI I M, S.E.M.W.A.L.S.K. Multiple Ways for Medical Data Visualization Using 3D Slicer; proceedings of the 2020 International Conference on Computational Science and, Intelligence, C.: (CSCI), F 16\u0026ndash;18 Dec. 2020 [C]. (2020)\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Medical imaging diagnosis, Dental cysts detection, Convolutional neural network, Deep learning","lastPublishedDoi":"10.21203/rs.3.rs-3856379/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-3856379/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eIn clinical medicine, jawbone cysts are a common dental disease, and their symptoms are similar to other dental diseases, making diagnosis difficult. To address this issue, this paper proposes a cyst detection system based on Alex Net to achieve the detection of cysts on dental radiographs using a CBCT dataset. The system can detect potential cyst lesions and locations in a timely manner to assist doctors in diagnosis. The improved model achieves an average accuracy of 83.5% and a maximum accuracy of 99.9%, achieving a high cyst recognition rate on existing datasets. In addition, the extensive image enhancement techniques introduced in the Alex Net model also improve the performance of the model. The experimental results show that compared to Res Net and VGG Net, both networks are not ideal for the classification of jawbone cysts, and may not be able to effectively extract key features from medical images, resulting in low classification accuracy. Therefore, it is important to choose a suitable deep learning model for the diagnosis of specific dental diseases. In future research, it is possible to further explore how to combine multiple deep learning models to improve the accuracy of diagnosis of dental diseases such as jaw cysts. In addition, improving data preprocessing and enhancing techniques can further improve the generalization ability of the model. In summary, by combining deep learning and clinical medicine concepts and methods, more effective auxiliary diagnostic systems can be developed to improve the accuracy and efficiency of dental disease diagnosis.\u003c/p\u003e","manuscriptTitle":"Research on Cyst of Jaw Detection Algorithm Based on Alex Net Deep Learning Model","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-01-16 19:37:51","doi":"10.21203/rs.3.rs-3856379/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"b7c67df8-77c8-4912-9618-7ccef81a127a","owner":[],"postedDate":"January 16th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2024-10-29T15:53:53+00:00","versionOfRecord":[],"versionCreatedAt":"2024-01-16 19:37:51","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-3856379","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-3856379","identity":"rs-3856379","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}