6. Technology and services

(1) Relevant definitions

1. Artificial intelligence

Artificial intelligence was officially born in 1956 and was known at the time as the science and engineering of making intelligent machines. This term is applied to a wide range of areas of medicine, such as robotics, medical diagnostics, medical statistics, and human biology, up to today's "omics". It has two main branches: virtual and physical. The virtual branch includes informatics approaches ranging from deep learning information management to health management system controls, including electronic health records, and active guidance by physicians in treatment decisions.

2. Machine learning

Machine learning is a multidisciplinary interdisciplinary discipline, covering probability theory, linear algebra, statistics, numerical computing, information theory, approximation theory, convex analysis, algorithm complexity theory and other fields. How to conduct in-depth analysis of complex and diverse data based on machine learning and make more efficient use of information has become the main direction of machine learning research in the current big data environment. Therefore, machine learning is increasingly developing in the direction of intelligent data analysis, and has become an important source of intelligent data analysis technology. In addition, in the era of big data, with the continuous acceleration of data generation, the amount of data has increased unprecedentedly, and new types of data that need to be analyzed are also emerging, such as text comprehension, text sentiment analysis, image recognition, graphics and network data analysis, etc. Therefore, in the era of big data, intelligent computing technologies such as machine learning and data mining play an extremely important role in intelligent analysis and processing applications.

3. Convolutional Neural Networks (CNN).

Convolutional neural network is a kind of feedforward neural network with convolutional computation and deep structure, which is one of the representative algorithms of deep learning. Convolutional neural networks have the ability to represent learning and can translate and classify input information according to their hierarchical structure, so they are also called "translational invariant artificial neural networks."

The research on convolutional neural networks began in the 80s and 90s of the twentieth century, and the time delay network and LeNet-5 were the earliest convolutional neural networks. After the 21st century, with the proposal of deep learning theory and the improvement of numerical computing equipment, convolutional neural networks have developed rapidly and have been applied to computer vision, natural language processing and other fields.

4. Deep learning

Deep learning is the internal rules and representation levels of learning sample data, and the information obtained in the learning process is of great help to the interpretation of data such as text, images, and sounds. The ultimate goal is for machines to be able to learn analytically like humans, and to recognize data such as text, images, and sounds. Deep learning is a sophisticated machine learning algorithm that achieves far more performance in speech and image recognition than previous technologies.

Deep learning based on convolutional neural networks (CNNs) (the structure is shown in Figure 1) has attracted much attention due to its high performance in image recognition. This technology can apply the image itself to the learning process, without the need for feature extraction before the learning process, and important features can be learned automatically. Therefore, there is great promise for applying this technology to the prediction of clinical radiological images.

Figure 1 CNN structure diagram

(2) Technical background

AI should be used in a wide range of areas of medicine, such as robotics, medical diagnosis, medical statistics, and human biology, up to today's "omics". It has two main branches: virtual and physical. The virtual branch includes informatics methods ranging from deep learning information management to health management system control, including electronic health records, and physician guidance in treatment decisions.

The biggest advances in medicine in the 21st century will be the maturation of precision medicine and the penetration of artificial intelligence in all areas of medicine, based on breakthroughs in molecular biology. In the past five years, the application of "artificial intelligence+" to medical research has become a hot spot in modern scientific and technological research.

Since 2012, our country has paid more attention to and focused on promoting the balanced development of regional medical resources, and telemedicine engineering has become a hot spot. On November 28, 2018, the General Office of the Ministry of Industry and Information Technology issued a notice on the "Work Plan for the Unveiling of Key Tasks for the Innovation of the New Generation of Artificial Intelligence Industry", in which the work goal is "medical image assisted diagnosis system". It's within the eight products, and our project belongs to that product.

(3) Technical highlights

1. Convolutional neural network greatly reduces the amount of computer computing

Convolutional neural network is an important algorithm in deep learning, which is mainly composed of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. Image features can be extracted automatically, eliminating the need to manually define features, saving human resources. Convolutional neural networks greatly reduce the amount of parameters used through weight sharing, so that the computing speed has been greatly improved. In a fully connected neural network, each neuron between two adjacent layers is edge-connected. When the feature dimension of the input layer becomes very high, the parameters that need to be trained for the fully connected network will increase a lot, and the computation speed will become very slow. In the convolutional neural network, the Dropout function (random inactivation) is used, so that the neurons of the convolutional layer are only connected to some of the neuronal nodes of the previous layer, that is, the connections between its neurons are not fully connected, and the weights w and offset b of the connections between some neurons in the same layer are shared. This drastically reduces the number of parameters that need to be trained.

2. Alexnet is used to improve accuracy

We use python language to write relevant code, and build a convolutional neural network under the tensorflow framework developed by Google, which is mainly composed of input layer, convolutional layer, pooling layer, fully connected layer, and Softmax layer. Three convolutional neural networks are to be constructed to simulate the recognition of tooth slices: the first is the VGG-19 network, which has an input size of 224×224 pixels. It consists of 16 convolutional layers and 3 fully connected layers, which are used for feature extraction, and each convolutional layer has a ReLU activation function and a Dropout algorithm. The ReLU activation function is used to increase the nonlinearity of the network and speed up the training, and the Dropout algorithm is set to 0.5 to prevent the model from overfitting. The size of the convolution kernel is 3×3 pixels, the filling is SAME, and the pooling layer adopts the maximum pooling method to remove redundant information and extract important features, with a size of 2×2 pixels. The fully connected layer consists of 1024, 102, and 512 nodes. Stochastic gradient descent SGD was used for training, the batch size was set to 128, the learning rate was 0.0001, and 100 times were experiencedepoch to derive the trained model; The second type of network is the Resnet-34 network, which has an input size of 224×224 pixels. There are 33 convolutional layers and 1 fully connected layer, with the first layer having a convolutional kernel size of 7×7 pixels and the rest of the convolutional kernels being 3×3 in size pixels, the 1st, 8th, 16th, and 28th layers have a convolution step of 2, and the rest of the convolutional layers have a step of 1 , using the ReLU activation function, there will be a residual module every two layers.The network has only one pooling layer, and is trained with stochastic gradient descent SGD, with a batch size set to 128 and a learning rate of 0.0001100 epochs to get the training model; The third type of network is based on the improved network of Alexnet, which has an input size of 400×400 pixels. It consists of 2 convolutional layers and 3 fully connected layers. The convolutional kernel size of layer 1 is 11×11 pixels, the step size is 3, and the size of the pooled layer kernel is 5× 5 pixels in steps of 3; The convolutional kernel size of layer 2 is 5×5 pixels, the step size is 3, and the pooling kernel size is 3×3 pixels in steps of 2. The fully connected layer uses L2 regularization and the Dropout algorithm to prevent model overfitting. The dental pieces were divided into training set, test set, and validation set. The ability of convolutional neural network to automatically extract image features is used to train the training set and save the model. The test set is used to monitor the detection accuracy of the model in real time, and the model is optimized by changing the network-related parameters to make the model optimal. Finally, the validation set data was input into the model to obtain a prediction score to test the accuracy of the model.

3. Use SVM instead of softmax classifier

Use softmax as the classifier for the network in C NN. The classifier uses full connections as a bridge to map the extracted features of CNN to the category output. The softmax classifier can be effectively combined with the backward propagation of the network during network training, which is convenient for the weight update in the CNN model, but when the classification problem is nonlinear, softmax cannot classify well. SVM uses its own kernel function to map the nonlinear classification interface in the original feature space to the high-dimensional feature transformation space, and then generates the linear classification interface, and finally obtains a good classification effect. In order to solve the problem that softmax is not able to handle nonlinear w well, SVM performs well for nonlinear classification problems. This paper uses SVM instead of softmax as the model classifier for CNN, and for CNNThe extracted tooth target features are classified.

(4) Design process

1. Project overview

Given the long time it takes to examine periodontal disease, the high labor consumption, and the possibility of inaccurate diagnosis, the combination with artificial intelligence is obvious. There are various forms of artificial intelligence, such as machine learning, deep learning, etc., and we are developing a computer-aided inspection system based on deep convolutional neural network (CNN) algorithms, and deep convolutional neural networks (CNNs) using python programming language The algorithm analyzes the apical slice of the examinee to determine whether the examinee has periodontal disease, and the doctor diagnoses whether the examinee has periodontal disease based on the results and clinical manifestations. The system can read multiple patients' apical slices at a time or multiple apical slices of one patient quickly and quickly. At the same time, the doctor no longer needs to manually observe the apical slice of each patient, and the system only needs a few seconds to produce the result, which greatly reduces the doctor's labor. First of all, the director of the Department of Stomatology of the Second Affiliated Hospital and his third graduate students communicated with the Information Department of the Second Affiliated Hospital, signed a confidentiality agreement, and imported a large amount of periodontal disease data into the computerThe three students were jointly developed under the guidance of the master's supervisor of the Institute of Translational Medicine, Nanchang University.

2. Product development process

(1) Market research

Through research, about 100 dentists and more than 1,000 periodontitis patients were interviewed, covering Asian and European user groups, to gain an in-depth understanding of the real needs and existing pain points in periodontal management in different regions
(2) Data Collection

1. In the Department of Stomatology of the Second Affiliated Hospital of Nanchang University, the whole mouth curved body lamellar and apical slices of periodontitis patients and periodontal healthy people from January 2016 to January 2020; 2. Full-mouth curved body lamellar of periodontitis patients and periodontal healthy people from the Department of Stomatology, Affiliated Hospital of Jiangxi University of Traditional Chinese Medicine from January 2019 to April 2020. In addition, the collection of dental imaging data from many countries in Asia and Europe has been expanded, including:

16,764 oral panoramas

52,680 apical radiographs

13,248 intraoral photographs

All image data were used for subsequent machine learning modeling, in addition to this, we screened all image images to exclude: 1. Under 14 years old; 2. Image pictures of teeth with severe noise, blurry or severe distortion; Three periodontists with working age of more than 5 years from the Department of Stomatology of the Second Affiliated Hospital of Nanchang University classified periodontitis into all the full-mouth curved body lamellar and full-mouth apical sections by combining periodontal clinical examination, and the films with incomplete diagnosis of the three examiners were excluded. The classification formulated by the American Society of Periodontology in 1999 was used as the diagnostic criteria: (1) mild: periodontal pocket depth ≤ 4 mm, attachment loss 1-2 mm, XThe line shows that alveolar bone resorption does not exceed 1/3 of the root length; (2) Moderate: the depth of the periodontal pocket is ≤6mm, the attachment loss is 3-4mm, and the X-ray shows that the alveolar bone resorption exceeds 1/3 of the root length, but not more than the root length 1/2.There may be mild loosening of the teeth; (3) Severe: periodontal pocket >6mm, attachment loss ≥5mm, X-ray shows alveolar bone resorption more than 1/2 of the root length or even2/3。

(3) Image preprocessing and model construction

The convolutional neural network is built under the tensorflow framework developed by Google. We start by standardizing and normalizing all images. The PAR of the Second Affiliated Hospital of Nanchang University was cropped to remove excess background information and preserve the whole tooth. The original image from the Department of Stomatology of the Affiliated Hospital of Jiangxi University of Traditional Chinese Medicine was adjusted to 400×400 pixels and cropped. Under the guidance of a medical professional, regions of interest (ROIs) of the above data were labeled. As a deep learning algorithm, CNN can automatically extract image features and is suitable for image classification. The CNN structure consists of seven layers with a total of three layers: two convolutional layers, two pooled layers, and three fully connected layers. The fully connected layer uses L2 regularization and discard algorithms to prevent overfitting. The learning rate used in the training set was 0.001. The network input size is 400×400 images and the classification score is calculated on a scale of 0 to 1. Each patient has an average of 10 PERs, which are fed into the Support Vector Machine (SVM) algorithm as 10 features and obtained a final probability score with a range of 0 after the transformationto 1. Based on the Convolutional Neural Network (CNN) model, deep learning training was carried out, and three types of images (panorama, apical film, and intraoral photo) were modeled and analyzed separately to improve the model's ability to recognize different periodontal states.

Figure 2 Flowchart of model construction

(3) Analysis of clinical features

We used clinical data collected from the Second Affiliated Hospital of Nanchang University and the Affiliated Hospital of Jiangxi University of Traditional Chinese Medicine for analysis. We used ggplot2 to plot heat maps to show the clustering of patients' clinical characteristics, and assessed the correlation between features and periodontitis by Spearman grade correlation analysis and t-test/analysis of variance.

Advantages of the method: Combined with the annotation of professional physicians to ensure the accuracy of the target region, the automatic feature extraction of CNN and the SVM classifier work together to improve the diagnostic performance.

(4) Statistical analysis

The training and validation datasets are used to estimate and create the optimal deep CNN algorithm weighting factors. All the deep CNNs in this study are built under the tensorflow framework in the python language. The test dataset was evaluated for diagnostic accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), receiver operating characteristic curve (ROC), and area under the ROC curve (). AUC）。 When comparing clinical features and periodontitis incidence, statistical analysis was performed using the t-test or ANOVA; Spearman grade analysis was performed to identify the classification of periodontitis patients. The P-value <0.05 was considered statistically significant, and a 95% confidence interval (CI) was calculated.

(5) Development and deployment of intelligent service cloud platform

a. System architecture: The front-end interface is developed using the ThinkPHP framework, and the back-end services are deployed on Alibaba Cloud high-performance servers to ensure computing efficiency and data security.

b. Core function: Integrate the trained CNN model, support users to upload oral images (such as apical films, panoramic films) and obtain AI analysis results (such as periodontitis risk scores) in real time. Provide clinical data visualization module (such as heat map display) and statistical report generation function (PPV, NPV and other indicators).

c. Technical support: Alibaba Cloud's elastic computing resources ensure stable services in high-concurrency scenarios, and support rapid model inference and large-scale data storage.

ThinkPHP ensures user-friendliness, and Alibaba Cloud infrastructure meets the high requirements for deep learning in terms of computing power and response speed.

3. Model diagnosis results

(1) Overview and baseline characteristics of the participant

During the study, we selected 11,626 patients and 10,839 healthy patients. After exclusion according to the inclusion criteria, a total of 2843 patients were screened out from the Department of Stomatology of the Second Affiliated Hospital of Nanchang University, and a total of 13834 oral X-ray images were analyzed (Fig. 3), including 13308 (46.01%) healthy patients and 1535 (53.99%) periodontitis patients. In the Department of Stomatology, the Affiliated Hospital of Jiangxi University of Traditional Chinese Medicine screened 238 patients with periodontitis and 113 healthy people, and obtained 351 full-mouth curved body lamina films. So far, a total of 14,185 images have been obtained. We designed a screening model and analyzed it in two steps (Figure 4). The first step of screening was to preliminarily identify potential patients through the PAR-CNN model, using 1924 full-mouth curved laminar slices from the Second Affiliated Hospital. The second step was to establish the PER-CNN model, and at the same time use 11,910 apical slices from 1191 patients for the application of the SVM model. We apply these two steps to accurately diagnose periodontitis.

The training set of the PAR-CNN model contained 1276 patients, the validation set contained 376 patients, and the test set contained 272 patients. Among them, we separately assessed the characteristics of patients in each cluster, such as gender, age, hypertension, diabetes, etc. (Table 1). Similarly, the PER-CNN-SVM combined model training set contains 919 patients, the test set contains 272 patients, and it Baseline characteristic information (Table 2). The results showed that periodontitis in each group was associated with systemic conditions such as age, smoking, hypertension, diabetes, and hereditary periodontal disease (P <0.05)。

Fig.3 Patient flow chart: Flow chart of image image collection, full-mouth curved body lamina (PAR) and apical slice ( PER)

Figure 4: Building a deep learning model to diagnose periodontitis

Through a two-step screening: the first screening was evaluated using the PAR-CNN model to obtain a PAR score; For the second time, the PER-CNN-SVM combination model was used to obtain the PER score to evaluate the results.

Table 1 Basic information of the population included in the PAR-CNN model

Table 2 Basic information of the population included in the PER-CNN-SVM model

(2) Results of the multicenter PAR-CNN model

We used 2,843 patients and 13,834 imaging images from the Second Affiliated Hospital of Nanchang University for analysis. We trained the model using 1276 (66%) full-mouth curved laminar slices recorded from 1276 patients. In the training set, there were 667 periodontitis patients and 609 normal people. A total of 376 patients (20%) were included in the validation set, including 173 patients with periodontitis and 203 patients with normal people. The remaining 272 patients (14%) were included in the test set. The area under the detection curve (AUC) for periodontitis was 0.843 (95% CI: 80.3%-87.8%, Figure 5A) when the model was tested on the PAR of each patient.

A total of 351 full-mouth curved laminar sheets from the Affiliated Hospital of Jiangxi University of Traditional Chinese Medicine were used as the second test set to verify the effectiveness of the model in multi-center detection. The training and validation sets are the same as before. Among them, there were 238 cases of periodontitis and 113 cases of normal people. The area under the curve (AUC) of the ROC curve was 0.793 (95% CI: 74.7%-83.4%, Figure 3A).

Of the total samples, 272 had both full-mouth curved lamellar and apical slices, of which 116 were normal and 156 were periodontitis. Through CNN evaluation, we used 156 PAR images of periodontitis patients as a test set to obtain AI for each personScore. We categorized patients based on this AI score and found that patients could be divided into three groups: 116 normal patients, 94 mild to moderate patients, and severe patients 62 cases (P< 0.001, Spearman=0.802, Figure 5B). AI-based classification makes it possible not only to diagnose periodontitis, but also to classify the severity of the patient. Automated typing can provide ideas for clinical triage. Patients can be automatically classified into different severities to help guide clinical treatment.

(3) The results of the PER-CNN-SVM combination model

The previous full-mouth surface laminar model was used for preliminary screening, and its accuracy needs to be improved. Therefore, we optimized the diagnostic model. We applied the patient's apical slice to the SVM model. Each patient had 10 apical slices (Figure 5C). We perform a CNN analysis on each apical slice to obtain an AI score, which is then put into the SVM model to obtain an SVM score for evaluation.

We included a total of 11,910 apical slices from 1191 patients. We set up 7162 (60%) images as the training set, including 3,789 apical radiographs for periodontitis and 3,373 apical radiographs for normal radiation. The validation set consisted of 2028 (17%) films, including 1073 apical slices for periodontitis and 955 normal apical slices. The test set was 2,720 (23%), of which 1,416 were apical films for periodontitis and 1,304 were normal。 Through the AI score of each apical slice, we found that the apical slice score was significantly higher in periodontitis patients, while the apical slice score was lower in normal patients (Fig. 5D, P < 0.001)。 In addition, we use these SVM scores to evaluate the performance of the CNN model. There were 695 patients with periodontitis and 496 normal patients. There were 919 cases (77%) in the training group, including 539 cases of periodontitis and 380 cases of normal people. The remaining 272 (23%) patients were enrolled in the testing group, which consisted of 156 patients and 116 normal people. The area under the curve (AUC) of this model was 0.977 (95% CI: 92.3%‐99.7%, Figure 3E). In general, with these two steps, we can make an accurate and automated diagnosis of periodontitis patients.

(A) The red line represents the ROC curve of the full-mouth curved body lamellar from the Second Affiliated Hospital of Nanchang University, and its AUC value is 0.843 ( 95% CI: 80.3%-87.8%）。 The blue line is the ROC curve of the full-mouth curved laminar from the Affiliated Hospital of Jiangxi University of Traditional Chinese Medicine, and its AUC value is 0.793 (95% CI: 74.7%-83.4%). (B) After PAR-CNN evaluation, patients were clustered according to their AI scores. (C) Ten PER locations in the oral cavity. (D) Clustering of each AI score. The AI score of the normal population was significantly lower than that of periodontitis patients ((P < 0.001). (E) AREA UNDER THE CURVE OF PER-CNN-SVM combined with ROC was 0.977 (95% CI: 92.3%‐). 99.7%）

(4) The combination of full-mouth curved body layer slice and apical piece

Overall, we used two screening steps to diagnose periodontitis. First, we set up a CNN model of the full-mouth surface layer sheet, and obtained the surface volume layer score. If the surface lamina score is less than 0.2, the patient is considered normal; Otherwise, the patient may develop periodontitis. However, initial screening has the potential to produce false-positive results. To further establish the apical slice-based CNN-SVM model, if the PER-CNN-SVM score is greater than or equal to 0.4, the diagnosis is correct (Figure 4A).。 The combination of CNN and SVM obtained a high diagnostic classification accuracy. The comprehensive modeling performance of the two models was 90.4% sensitivity, 93.1% specificity and 91.5% accuracy.

Fig.6 Combined diagnostic steps for PARS and PERs

(A) We used the PAR-CNN model to screen normal people. However, there is a misdiagnosis. Then, the CNN-SVM combination model was used to evaluate the PERs of the patients to achieve accurate diagnosis. (B) Model confusion matrix diagram. In the ordinate, 1 represents a patient and 0 represents a normal person. In the abscissa, 1 is the patient and 0 is the normal person predicted by the model. Therefore, the false-positive rate is 0.069 and the false-negative rate is 0.096.

and (5) the clinical characteristics of the cluster of periodontitis patients

To further investigate the characteristics of periodontitis, we collected cases of periodontitis. Based on the patient's ten apical slices, we found that the tooth located at E1 had the highest prevalence (Figure 7A). In addition, to explore the relationship between periodontitis and other clinical features, we brought together patients with periodontitis of varying severity. The clinical characteristics of 156 patients with periodontitis in the Second Affiliated Hospital were used for cluster analysis. Based on the clustering results, we found that periodontitis was significantly associated with age and smoking (Figure 7B).

Fig.7 Clinical features of periodontitis

(A) We pooled PERs and found that the two sites with the highest incidence of periodontitis were C1 and E1. (B) Heatmap of clinical features of patients with periodontitis. Periodontitis was found to be significantly associated with age and smoking.

4. R&D plan

At this stage, we are using deep learning technology for rapid diagnosis of periodontitis as the market entry point, with the continuous improvement of data and the maturity of technology, we will stage and grade periodontitis according to the full-mouth curved body lamellar slice + apical slice, and in the future we will use it according to the oral cavityCurved body lamina + apical piece locates the specific location of periodontal damaged teeth.

7. Product introduction

(1) Product positioning

Based on the CNN-SVM dual-fusion big data algorithm, this project develops an auxiliary decision-making system for periodontal disease, and combined with the cloud platform built by the ThinkPHP framework, a periodontal health management platform integrating early screening, diagnosis and prognosis evaluation is developed. The first generation of AI TEETH is in the form of a web interface, and the next few generations of products will produce embedded devices.

(2) The main product concept

The company's project takes the preliminary diagnosis of periodontal disease as the starting point to develop technology and open up the market, after the product technology level is mature, it will develop the staging and grading function of periodontal disease, and in the future, we will locate the specific location of periodontal damaged teeth according to the oral curved body laminar sheet + apical piece.

(3) The target group of the product

The core users of the first generation of this product are hospitals at all levels, aiming to make up for the level of water technology ability of doctors in primary and secondary hospitals, realize the sinking of high-quality medical resources to the grassroots level, and for tertiary hospitals, mainly to alleviate the problems of large demand for treatment and tight time for doctor-patient communication.

(4) Product highlights

(1) Product advantages: high early screening rate (85.3%), accurate diagnosis (97.7%), rapid diagnosis (average 2seconds), good real-time (real-time monitoring of disease changes through follow-up records).

(2) Derivative function: the platform is simple to operate and easy to promote

Develop a personalized health management plan

Realize the standardization and homogeneity of regional medical care

Assist the hospital to establish a periodontal health management platform

It has soft copyright, and has preliminarily developed system 1.0.

8. Strategic analysis

(1) Strategic vision

With AI TEETH products as the starting point and relying on the core technology of artificial intelligence to identify periodontal disease, the company has established a strategic path driven by technological innovation, starting from regional deepening and expanding the global market. In the initial stage, the project will be based in Nanchang City, Jiangxi Province, and will be fully promoted in mainland cities and county-level cities in Jiangxi Province in the first half of 2025 to promote the improvement of local medical informatization and intelligence. In the second half of 2025, with Jiangxi as the center, it will gradually expand to the surrounding second- and third-tier cities, open up regional collaborative application scenarios, and form a stable market foundation. With the gradual maturity of domestic technology verification and business model, the company will further promote its products to remote areas in China from 2026 onwards, improve the oral health management capabilities of grassroots areas, and promote the popularization and application of AI TEETH.

While the national layout is steadily advancing, the team simultaneously launched an internationalization strategy, focusing on the core markets of the Asia-Pacific region, and striving to expand stable income in technology output and local cooperation. The medium and long-term goal is to enter mature medical markets such as Europe, America and Russia, and establish brand influence and market share on a global scale with the help of deep learning technology advantages and doctor annotation data accumulation. Based on SWOT analysis, the company gives full play to its internal advantages such as independent research and development, strong scientific research capabilities, and abundant clinical resources, and responds to external challenges such as rapid changes in digital medical technology and stricter data privacy supervision, and strives to become a pioneer and industry leader in the field of artificial intelligence periodontal health management within three years.

(2) Analysis of the strategic environment

At present, in the medical market at home and abroad, there is no artificial intelligence system for the joint identification of periodontal disease by full-mouth curved body lamina and apical piece, so the track in which this project is located is still blank and has typical blue ocean market characteristics. With the increasing awareness of oral health among residents around the world, "early screening and early intervention" has become a medical trend, coupled with the policy support and investment in digital medicine and intelligent diagnosis technology in various countries, the intelligent oral diagnosis solution based on AI image recognition has broad application prospects in the future. Starting from China, the AI TEETH project will effectively fill the gap in efficiency and accuracy of existing inspection systems. The dual-image fusion scheme of panorama and apical slice used in the project, combined with the high-precision image recognition model constructed by Convolutional Neural Network (CNN), can realize low-cost and high-efficiency large-scale auxiliary diagnosis, which is especially adaptable to developing areas. Compared with the traditional method of relying on manual reading, our AI system can significantly alleviate the problem of doctors' resource constraints, improve the efficiency of diagnosis and treatment, and promote the popularization of intelligent oral diagnosis. Internationally, although some developed countries have made early explorations in the field of AI analysis of dental images, they mostly focus on the direction of dental caries, endodontics or orthodontics, and have not yet formed clinical application products covering the systematic identification of periodontal diseases. With its leading algorithm capabilities and localized large-scale real data training advantages, the project is not only expected to quickly occupy the domestic market, but also has the potential to export technology and services to the developing markets in Asia and high-end markets in Europe and the United States. From the general direction of analysis, as people pay more and more attention to physical examination, the concept of "examination and disease prevention" continues to be deeply rooted in the hearts of the people and the number of medical diagnosis and treatment continues to grow, the problem of mismatch between supply and demand in the medical industry will become more obviousTherefore, high-tech medical diagnostic equipment combined with artificial intelligence will become the direction of the future. Our company's existing technology is based on oral curved body laminar slice + apical slice to detect periodontal disease. The company's products are high-tech measurement systems, and there is no manufacturing cost in the product itself, which greatly saves costs compared to ordinary companies. At the same time, we will concentrate the company's resources on product development, update and market promotion, concentrate advantages, occupy the market and maintain a leading position.

1. Bargaining power of suppliers

Our company's products are self-developed high-tech medical systems, and there is no oppression from upstream suppliers.

2. Customer bargaining power

The main sales objects of our company's products entering the market are hospitals. Considering the convenient, high-speed and low-cost characteristics of our company's existing products, and the lack of a periodontal health management platform integrating early screening, diagnosis and prognosis assessment in the market, our products can effectively enhance the medical capacity and efficiency of hospitals with limited medical conditions; At the same time, our products can not only enhance efficiency, but also reduce unnecessary medical examinations, thereby greatly reducing the financial pressure of patients. Nowadays, in view of people's attention to oral diseases and the high prevalence of periodontal diseases, such products have the advantages of low price, fast and high efficiency, and the price is relatively low.

3. The threat of potential entrants

The medical testing system industry is an industry that has not yet formed a system. At the same time, the industry's entry barriers and technical barriers are high, our company's artificial intelligence full-mouth curved body laminar + apical detection system technology is at the leading level in the country, our company's technology is highly confidential, which is not available to other enterprises for the time being. However, based on future considerations, other companies can also use their own resources to develop related products, and the risks still exist. In view of this risk, our company can take advantage of the existing advantages, constantly update and iterate the system and promote the research and development of new products, and gradually build the brand to form brand benefits to discourage potential entrants from entering the attempt.

4. The threat of substitutes

Existing market-free periodontal health management platform. Our periodontal health management platform for early screening, diagnosis and prognosis assessment is less threatening.

5. The degree of competition in the industry

At present, our company's products belong to a new type of targeted products proposed for the mismatch between social supply and demand, and the degree of competition in the industry is low.

(3) Strategic plan selection (business development strategy).

1. The overall strategy of the enterprise: market development strategy (market expansion strategy).

Overall strategy description: According to the company's current resources, combined with the development trend of the market and technology, start from small, with a level and a target development. Prioritize reaching out to a relatively stable customer base. At the same time, we pay attention to technological development and progress, expand the market and develop customers, and gradually do large-scale and occupy the market. After establishing a relatively stable customer base, it has gradually realized the social benefits of promoting employment and giving back to the society, and gradually became one of the leaders in the industry.

According to the company's current resource capacity, combined with the development trend of market and technology, ITEETH relies on the Second Clinical Medical College of Nanchang University, the Second Affiliated Hospital of Nanchang University and the Translational Medical College of Nanchang University, with its strong faculty and strong scientific research strength, Nanchang City, Jiangxi Province as the entry point to carry out sales, expand market share, and passProvide "high-value, high-quality" products to the society to achieve the company's own development, increase welfare for the society, create value for the medical industry, seek career development opportunities for employees, and bring high return on investment to shareholders.

2. The basic strategy of the enterprise: the strategy of characteristic advantages and the strategy of target agglomeration

The company has established a development path with characteristic advantage strategy and target agglomeration strategy as the core. After completing the clinical trial and verifying the superiority of the product, the company will take the AI TEETH intelligent periodontal diagnosis system as the starting point, give priority to Jiangxi Province, and promote the application of stomatology and private dental institutions in general hospitals and private dental institutions in Shangrao, Jiujiang, Fuzhou, Yichun and other cities around Nanchang. In the first half of 2025, it will be fully promoted in Jiangxi Province, covering municipal and county-level medical institutions, so as to improve the ability and digitalization of primary diagnosis and treatment; In the second half of 2025, the company will expand to surrounding provinces with Jiangxi as the center, focus on the market layout of second- and third-tier cities, build a regional radiation effect, and gradually achieve in-depth coverage in Central China. From 2026, the company will gradually go deep into remote and underdeveloped areas to realize AI TEETHand improve accessibility and inclusiveness through cloud platform remote services. After consolidating the foundation of the domestic market, the company will further accelerate the process of technology optimization and standard internationalization, gradually expand to the markets of major countries in the Asia-Pacific region, enter the European and American markets on the basis of solid brands and channels, and finally form an international strategic layout of multi-center and multi-regional coordination in the world, and strive to become a leading enterprise in the field of intelligent periodontal diagnosis in the world in the next three to five years.

3. Planning for all stages of enterprise growth

Introduction: Establish a brand in major cities in China

Core objectives: To establish brand influence, validate business models, and accumulate clinical data in China's core healthcare market.

Benchmark Hospital Cooperation: Deepen the cooperation with the Second Affiliated Hospital of Nanchang University to create demonstration cases, and expand the cooperation intention between King's College London Dental Institute and Barts and The London School of Medicine and Dentistry.

Market education: Cooperate with KOLs (key opinion leaders) to promote the value of AI periodontal diagnosis. Provide a trial plan for primary hospitals to reduce the use of doors

(3) Policy and certification: Complete the approval of the NMPA (China Food and Drug Administration) Class III certificate to ensure compliance. Apply for medical insurance fee items to enhance the willingness of hospitals to procure.

(4) Brand building: enhance professional influence through industry exhibitions (such as CMEF) and medical journal papers.

Growth stage: Expansion in the Asia-Pacific region to stabilize turnover.

Core goal: to expand into the Asia-Pacific market based in China and establish a stable source of income.

(1) Breakthroughs in key markets

Southeast Asia: Focusing on high-end dental clinics in Thailand and Malaysia (dental tourism demand).

Japan and South Korea: Partnering with local agents to adapt to their high standards of healthcare AI regulatory environment.

(2) Localization adaptation: support multiple languages (English, Thai, Japanese, etc.), in line with the local diagnosis and treatment process. Optimize algorithms for Asia-Pacific populations (e.g., periodontal disease characteristics in Southeast Asian populations).

(3) Business model innovation: cooperation with dental equipment manufacturers: such as Japan's Morita, to achieve "AI software + hardware" bundled sales.

(4) International certification: CE certification (EU) and TFDA certification (Taiwan) have been obtained, paving the way for entering a more stringent market.

3. Mature stage: to the global market - Europe, the United States, Russia, etc

Core goal: to enter the mature markets of Europe and the United States, and become a global leader in periodontal AI diagnosis

High-end market tackling:

Europe: Passed CE certification to enter the procurement system of public hospitals in Germany and France.

United States: Completed FDA certification, focusing on breaking through DSO (Dental Chain Group).

Russia: Partnering with local medical groups to focus on digital upgrades

Construction of technical barriers:

Continue to iterate the algorithm to maintain a leading edge over competing products such as Pearl and DentalX AI.

Explore extended applications such as AI+ robots (e.g., automated periodontal treatment planning).

Data compliance and security: Establish a data management system that complies with GDPR (EU) and HIPAA (US). Deploy on-premise servers overseas to meet data sovereignty requirements.

(8) Risks

With the passage of time, the number of cooperative hospitals increases, and the performance of this product needs to be upgraded, so the technology needs to be constantly updated. At the same time, this product is not the only gold standard diagnosis, it is for reference only, and the final diagnosis needs to be manually performed by a doctor.

(4) SWOT comprehensive analysis

1. Advantages

(1) Team advantage: This product is programmed by computer professionals, and has dental experts as a guide, and the technical level does not lag behind other competitors.

(2) Product advantages: As a periodontal health management platform, this product is convenient and fast for hospitals.

(3) Market advantage: the product market is large, and the market development potential is large. In the initial stage of investment, this product is mainly for medium-sized hospitals, and is committed to occupying the medium-sized hospital market.

(4) Resource advantage: The data comes directly from the hospital, and the Second Affiliated Hospital of Nanchang University and the School of Translational Medicine have reached cooperation intentions, and have rich medical resources.

2. Disadvantages

(1) Lack of market experience: The company's team is a first-time entrepreneur with a relative lack of experience, and the understanding of the characteristics and development of the medical market needs to be strengthened. In the actual market operation process, due to the lack of experience, we will still face many difficulties and encounter many problems.

(2) Weak public relations in the early stage: Due to the lack of accumulation of public relations in the early stage of entrepreneurship, the company will encounter certain obstacles in the process of market development, such as difficulties in financing and lack of product sales channels.

3. Opportunity

(1) National policy opportunities: The State Council put forward the "Opinions on Several Policies and Measures to Vigorously Promote Mass Entrepreneurship and Innovation", and introduced corresponding policies to support the development of innovative enterprises, and the State Council passed the "Internet +" Action Guiding Opinions to deeply transform traditional industries and promote cross-border industrial upgrading. In order to help the economic transformation, Zhiyali Co., Ltd. closely meets the requirements of Internet + medical care and meets the requirements of current development. In addition, the development of artificial intelligence has also brought great opportunities to the periodontal health management platform, providing technical and policy support for the development of the platform.

(2) Market opportunities: At present, the global medical and health industry is in a stage of rapid development, especially in the field of oral medicine, and the market scale continues to expand, bringing unprecedented growth opportunities for innovative medical products. With the continuous improvement of our country's residents' health awareness, the public's attention to oral health care has been significantly enhanced, the diagnosis rate of periodontal disease and related diseases has increased steadily, and the market demand has continued to grow. At the same time, according to the epidemiological data of many countries around the world, periodontal disease has become one of the common chronic diseases that affect the quality of life of the population, especially in developing countries and aging countries, and more efficient and low-cost early screening and management methods are urgently needed. The rising incidence of periodontal disease in our country not only brings a broad domestic market space for the company, but also indicates the feasibility and necessity of expanding to the international market. In this context, the rapid development of artificial intelligence and Internet technology has provided a strong technical support and basic environment for the construction and global deployment of periodontal intelligent diagnosis system, laying a solid foundation for the company to go international and participate in the construction of global digital medical ecology.

4. Threats

(1) The threat of substitutes: for the company's products, the main substitutes on the market are the existing standard of routine clinical examination and tooth mouth curved body lamellar + apical piece, because the results of our company's products are not the gold standard, with insufficient authority, which will pose a certain threat to our company's products.

(2) The threat of competitors: As intelligent technology products have received more and more attention from the state and society, with the introduction of relevant laws and regulations, this field will have greater development advantages and space. In particular, the entry of some foreign and domestic brands with mature technology and rich experience will cause certain pressure and threats to the company's periodontal health management platform.