Dental film deep caries and pulpitis differential diagnosis model training method and diagnosis system
By training a deep learning model based on a convolutional neural network and using Grad-CAM technology to generate heat maps, the problem of low accuracy in differentiating between deep caries and pulpitis was solved, achieving efficient and accurate differential diagnosis of dental radiographs, which is suitable for auxiliary diagnosis in primary healthcare institutions and dental emergency departments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-29
- Publication Date
- 2026-03-10
AI Technical Summary
Existing diagnostic methods for deep caries and pulpitis have low accuracy and cannot directly present diagnostic evidence, leading to a high rate of misdiagnosis, especially in primary hospitals or emergency departments where there is a lack of highly accurate and efficient auxiliary diagnostic tools.
A deep learning model based on convolutional neural networks is used, combined with attention mechanism, multi-task learning mechanism and transfer learning mechanism, to train a dental X-ray image dataset. Grad-CAM technology is used to generate heat maps to improve the model's visualization and interpretation capabilities. Data preprocessing and 5-fold cross-validation are used to improve model performance.
It enables efficient and accurate differential diagnosis between deep caries and pulpitis, improves the transparency and credibility of diagnosis, reduces the misdiagnosis rate, and is suitable for auxiliary diagnosis in primary healthcare institutions and dental emergency departments.
Smart Images
Figure CN121640123A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of artificial intelligence oral diagnosis, in particular to a dental film deep caries and pulpitis differential diagnosis model training method and a diagnosis system. BACKGROUND
[0002] Deep caries and pulpitis are very common and challenging diseases in oral clinical practice, especially in oral emergency. Deep caries and pulpitis are usually caused by bacterial infection, commonly seen in untreated caries (tooth decay). Deep caries refers to the decayed tooth cavity that has penetrated into the dentin or even approached the pulp tissue, while pulpitis refers to inflammation of the pulp tissue, often caused by bacterial invasion, causing severe pain and discomfort.
[0003] The differential diagnosis of deep caries and pulpitis has always been a difficult point in clinical work. Both of them show deep and large caries, probing sensitivity, and positive signs when cold testing, etc. Relying on subjective symptoms and traditional imaging examination, even experienced doctors have a certain misdiagnosis rate. Deep caries can be treated with pulp preservation (such as indirect pulp capping followed by direct filling), while pulpitis requires root canal treatment, and misdiagnosis will directly lead to incorrect treatment plan. If deep caries is misdiagnosed as pulpitis, the healthy pulp will be treated with root canal treatment after pulp opening, resulting in loss of tooth vitality, decreased anti-fracture ability, and lifelong crown repair protection, causing waste of medical resources and increased costs; if pulpitis is misdiagnosed as deep caries, the wrong pulp preservation treatment will lead to treatment failure, causing acute attack of chronic pulpitis or acute periapical periodontitis, and exacerbating the patient's pain, prolonging the course of treatment. In primary hospitals or emergency departments, due to the lack of experienced experts, the ability to interpret images is often limited, therefore, there is an urgent need for high-accuracy and high-efficiency auxiliary diagnosis tools to help medical personnel improve their diagnostic level and ensure timely and accurate treatment decisions.
[0004] In recent years, artificial intelligence technology has made breakthroughs in the field of medical imaging, providing a solid technical foundation for the differential diagnosis of deep caries and pulpitis. Models based on convolutional neural networks (CNN) have shown excellent performance in lesion detection and classification of X-ray films (such as chest films or bone films), CT, and MRI. Their performance has even surpassed that of experienced experts, demonstrating the feasibility of AI learning complex features from medical images and making accurate judgments. Currently, a large number of studies have confirmed that AI can be used for caries detection in dental films, identification of apical lesions, and periodontal bone height measurement. For example, Yoon Kyubaek et al. developed an AI system for caries detection, with an average mAP score of 0.769 and individual scores ranging from 0.695 to 0.893. Multiple AI models can automatically identify full-mouth tooth positions, demonstrating the feasibility of dental image analysis. In addition, open-source frameworks and models such as PyTorch, MONAI, and TeethNet provide a strong technical foundation, greatly reducing the development threshold and ensuring the technical advancement and reproducibility of research. Recent studies have shown that image-based methods have made significant progress in differentiating deep caries and pulpitis, but this method relies on complex feature engineering, resulting in low accuracy in differentiating deep caries and pulpitis and the inability to directly display diagnostic evidence. SUMMARY
[0005] To this end, the present application provides a dental film deep caries and pulpitis differential diagnosis model training method and diagnosis system to solve the problem of low accuracy in differentiating deep caries and pulpitis and the inability to directly display diagnostic evidence in the prior art.
[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions:
[0007] In a first aspect, a dental film deep caries and pulpitis differential diagnosis model training method is provided, comprising:
[0008] Step 1: Obtain a dental film image dataset verified by pathology or clinical treatment results;
[0009] Step 2: Perform data desensitization processing on the dental film image dataset, and perform image-level diagnosis label annotation and pixel-level lesion annotation to obtain an annotated dataset;
[0010] Step 3: Perform data preprocessing on the annotated dataset and divide it into a training set and a test set;
[0011] Step 4: Input the training set into a deep learning model for training to obtain an initial differential diagnosis model; the deep learning model uses attention mechanisms, multi-task learning mechanisms, and transfer learning mechanisms during training;
[0012] Step 5: Input the test set into the initial differential diagnosis model to obtain the model output results;
[0013] Step 6: Calculate the model performance evaluation index based on the model output results, generate a heat map using Grad-CAM technology, verify the initial differential diagnosis model based on the heat map, and obtain the final differential diagnosis model for deep caries and pulpitis on dental radiographs.
[0014] Preferably, in step 1, the sample size of the dental radiograph dataset is calculated using a diagnostic test sample size formula, which is:
[0015]
[0016] Where n represents the number of samples required for each group, Z 1-α / 2 The Z-value represents the significance level. 1-β The Z-value represents the test power, p1 represents the expected sensitivity or specificity, and p2 represents the minimum acceptable performance. This represents the mean of p1 and p2.
[0017] Preferably, in step 3, the data preprocessing includes scaling, size normalization, grayscale normalization, and image enhancement.
[0018] Preferably, in step 4, the deep learning model uses a convolutional neural network based on the PyTorch framework.
[0019] Preferably, the convolutional neural network uses a ResNet18 pre-trained model.
[0020] Preferably, in step 4, the deep learning model uses the FocalLoss loss function during training.
[0021] Preferably, in step 4, the attention mechanism adopts the SE Block attention mechanism or the CBAM attention mechanism.
[0022] Preferably, in step 6, the model performance evaluation metrics include AUC, sensitivity, and specificity.
[0023] Preferably, in step 6, 5-fold cross-validation is used when validating the initial differential diagnosis model.
[0024] Secondly, a system for differentiating between deep dental caries and pulpitis on dental X-rays includes:
[0025] The data acquisition module is used to acquire patients' dental X-ray images;
[0026] The data annotation module is used to perform data desensitization processing on the dental radiograph data, and to perform image-level diagnostic labeling and pixel-level lesion annotation to obtain labeled data;
[0027] The data preprocessing module is used to preprocess the labeled data;
[0028] The differential diagnosis module is used to input the preprocessed labeled data into the differential diagnosis model of deep caries and pulpitis on dental radiographs for diagnosis and to obtain the diagnosis result; the differential diagnosis model of deep caries and pulpitis on dental radiographs is trained by the differential diagnosis model training method of deep caries and pulpitis on dental radiographs.
[0029] The heatmap generation module is used to generate diagnostic heatmaps based on diagnostic results using Grad-CAM technology.
[0030] Compared with the prior art, this application has at least the following beneficial effects:
[0031] This application provides a method for training a differential diagnostic model for deep dental caries and pulpitis on dental radiographs, comprising: acquiring a dataset of dental radiograph images verified by pathological or clinical treatment results; desensitizing the dental radiograph image dataset and annotating it with image-level diagnostic labels and pixel-level lesion labels to obtain an annotated dataset; preprocessing the annotated dataset and dividing it into a training set and a test set; inputting the training set into a deep learning model for training to obtain an initial differential diagnostic model; inputting the test set into the initial differential diagnostic model to obtain the model output results; calculating the model performance evaluation index based on the model output results, and using Grad-CAM technology to generate a heatmap to verify the initial differential diagnostic model, thereby obtaining the final differential diagnostic model for deep dental caries and pulpitis on dental radiographs. The differential diagnostic model for deep dental caries and pulpitis on dental radiographs trained using the method provided in this application can efficiently and accurately distinguish between deep caries and pulpitis, and has visual interpretation capabilities, improving the transparency and credibility of clinical diagnosis. Attached Figure Description
[0032] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).
[0033] Figure 1 This is a flowchart of a method for training a diagnostic model for differentiating deep dental caries and pulpitis from dental radiographs, provided in Embodiment 1 of this application.
[0034] Figure 2 This is a structural block diagram of a training method for a differential diagnostic model of deep dental caries and pulpitis provided in Embodiment 1 of this application;
[0035] Figure 3 A flowchart illustrating the inclusion and exclusion criteria provided in Embodiment 1 of this application;
[0036] Figure 4 This is a schematic diagram of the confusion matrix provided in Embodiment 1 of this application;
[0037] Figure 5 This is a code diagram provided for Embodiment 1 of this application;
[0038] Figure 6 The training and verification curves provided in Embodiment 1 of this application. Detailed Implementation
[0039] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0040] In the description of this application: unless otherwise stated, "a plurality of" means two or more. The terms "first," "second," "third," etc., in this application are intended to distinguish the objects referred to and do not have any special meaning in terms of technical connotation (e.g., they should not be construed as an emphasis on importance or order). Expressions such as "comprising," "including," and "having" also mean "not limited to" (certain units, components, materials, steps, etc.).
[0041] The terms used in this application, such as "upper," "lower," "left," "right," and "middle," are generally used to indicate the general relative positional relationship for the purpose of intuitive understanding by referring to the accompanying drawings, and are not absolute limitations on the positional relationship in the actual product.
[0042] Example 1
[0043] Please see Figure 1 and Figure 2 This embodiment provides a method for training a diagnostic model to differentiate between deep dental caries and pulpitis, including:
[0044] S1: Obtain a dataset of dental radiographs validated by pathological or clinical treatment results;
[0045] Specifically, the dental radiograph dataset in this embodiment consists of the imaging and clinical data of patients who visited the Emergency Comprehensive Treatment Center of Beijing Stomatological Hospital, had electronic periapical radiographs taken, were initially diagnosed with "deep caries" or "pulpitis" by clinical practice, and whose final diagnosis was verified by postoperative treatment.
[0046] This embodiment of the study utilizes Beijing Stomatological Hospital to obtain a large amount of high-quality dental radiographs and case data verified by pathological or clinical treatment results. This "gold standard" data is a core prerequisite for training a highly reliable AI model. The hospital's research team possesses extensive clinical experience in stomatology and can accurately define the subtle radiographic differences between deep caries and pulpitis (e.g., the relationship between caries depth and pulp cavity, whether the periapical ligament is widened, and the presence of localized radiolucentism), ensuring the scientific rigor of the AI model's learning objectives.
[0047] In this embodiment, the total sample size of the dental radiograph dataset is 1000 periapical radiographs of affected teeth, with a target sample size ratio of 1:1 for the deep caries group and the pulpitis group (i.e., 500 cases in each group). Because sample size estimation needs to ensure sufficient statistical power, the sample size in this embodiment is calculated using the diagnostic test sample size formula, which is:
[0048]
[0049] Where n represents the number of samples required for each group, Z 1-α / 2 The Z-value represents the significance level (usually α = 0.05, then Z...). 0.975 =1.96), Z 1-β The Z-value represents the test power (usually β = 0.1 or 0.2, then Z...). 0.9 =1.28, Z 0.8 =0.84), p1 represents the expected sensitivity or specificity (the model is expected to have a sensitivity of 0.90), and p2 represents the minimum acceptable performance (usually set to 0.80). This represents the mean of p1 and p2.
[0050] Suppose this embodiment aims to verify whether the model's sensitivity reaches 90% (p1 = 0.9), and sets the minimum acceptable value to 80% (p2 = 0.8). Let α = 0.05 and β = 0.2 (i.e., efficacy of 80%), then:
[0051]
[0052] In summary, when α = 0.05, β = 0.2, and the expected sensitivity is 0.90, the minimum acceptable value is 0.80, which calculates to a minimum sample size of 199 cases per group. To simultaneously ensure the power of the specificity test and the robustness of model training, this embodiment increases the sample size to 500 cases per group. Therefore, to train a high-performance deep learning model and avoid overfitting, it is empirically recommended that at least 500-1000 samples be required per category, and the target sample size in this embodiment meets this widely accepted standard.
[0053] Once the sample size is determined, appropriate samples need to be selected based on the inclusion and exclusion criteria. Please refer to [link / reference needed]. Figure 3 The sample inclusion criteria in this embodiment are as follows:
[0054] (1) Select samples that were initially diagnosed as "deep caries" or "pulpitis" after clinical examination and periapical radiographs;
[0055] (2) Samples that have undergone corresponding dental and endodontic treatment (filling / indirect pulp capping for deep caries, root canal treatment for pulpitis);
[0056] (3) Standard projection angle for periapical radiographs, with clear images (samples of coronal caries, pulp chamber, root canal, and periapical tissue).
[0057] The specific sample exclusion criteria are as follows:
[0058] (1) Samples with severe artifacts, distortion, breakage, or under / overexposure in the root tip film, which affect the interpretation;
[0059] (2) Samples with other concurrent diseases in the affected tooth, such as: acute / chronic periapical periodontitis, periodontitis, root longitudinal fracture, internal resorption, external resorption, history of root canal treatment, etc.
[0060] (3) Samples of teeth with special dental anatomy (such as malformed central cusp, tooth in tooth) or covered by restorations (such as full crown, post and core) that cannot be accurately evaluated;
[0061] (4) Samples of patients who sought medical attention a second time due to postoperative pain, and whose clinical and imaging examinations overturned the original diagnosis.
[0062] (5) The patient’s clinical medical records are incomplete, making it impossible to confirm the sample for the final diagnosis.
[0063] S2: Desensitize the dental X-ray image dataset, and perform image-level diagnostic labeling and pixel-level lesion labeling to obtain a labeled dataset;
[0064] Specifically, in this embodiment, 500 clinically validated dental X-rays of deep caries and pulpitis were collected, and then expert annotation and consistency checks were performed. During annotation, a high-quality annotated dataset was constructed by combining a multi-level annotation mechanism (image-level diagnostic labels and pixel-level lesion annotations), thereby improving the model's learning accuracy.
[0065] S3: Preprocess the labeled dataset and divide it into training and test sets;
[0066] Specifically, this embodiment performs standardized preprocessing on the dental X-ray image data, including scaling, size normalization, grayscale standardization, and image enhancement. After data preprocessing, the dataset is divided into a training set and a test set, with the training set accounting for 80% and the test set accounting for 20%.
[0067] S4: Input the training set into the deep learning model for training to obtain the initial identification and diagnosis model; the deep learning model adopts attention mechanism, multi-task learning mechanism and transfer learning mechanism during the training process;
[0068] Specifically, in this embodiment, the deep learning model uses a convolutional neural network (such as ResNet, DenseNet, etc.) within the PyTorch framework. Preferably, the convolutional neural network uses a ResNet18 pre-trained model. In this embodiment, ONNXRuntime is used for lightweight model deployment, and a Python+PyQt desktop interface is developed. Model pruning, quantization, and other techniques are employed to ensure real-time operation on ordinary hardware.
[0069] When training a deep learning model, the FocalLoss loss function is introduced to address the class imbalance problem, and attention mechanisms, multi-task learning, and transfer learning are employed to enhance the model's focus on key regions, thereby improving the model's discrimination ability. Specifically, the attention mechanism used is either the SE Block attention mechanism or the CBAM attention mechanism.
[0070] S5: Input the test set into the initial differential diagnosis model and obtain the model output results;
[0071] S6: Calculate the model performance evaluation index based on the model output results, generate a heat map using Grad-CAM technology, verify the initial differential diagnosis model based on the heat map, and obtain the final differential diagnosis model of deep caries and pulpitis on dental X-ray.
[0072] Specifically, this embodiment employs 5-fold cross-validation and calculates indices such as AUC, sensitivity, and specificity. Grad-CAM technology is used to generate a heatmap, which is then used to validate the initial differential diagnostic model, ultimately yielding the final differential diagnostic model for deep caries and pulpitis in dental radiographs.
[0073] This embodiment integrates Grad-CAM technology, which generates heatmaps to highlight areas of interest in the model, giving it visual interpretability and enhancing the model's interpretability and physician trust.
[0074] Please see Figure 4 and Figure 5 , Figure 4 The confusion matrix in the test set shows the classification results of the initial differential diagnosis model. It intuitively reflects the model's accurate classification and misclassification of the two categories of "deep caries" and "pulpitis", proving that the model has good discrimination ability.
[0075] Please seeFigure 6 , Figure 6 The training and validation curves include the loss curve and the accuracy curve. This figure shows the convergence of the model during training. The validation loss decreases steadily with each training epoch without overfitting, and the validation accuracy continues to rise and tends to stabilize, demonstrating the effectiveness of the training process and the rationality of the model structure.
[0076] The method for training a model to differentiate between deep dental caries and pulpitis provided in this embodiment has the following advantages:
[0077] (1) High accuracy: The model AUC>0.80, and the sensitivity and specificity are both higher than 85%, which is better than the level of intermediate physicians;
[0078] (2) High interpretability: The diagnostic basis is visually presented through heat maps, which improves clinical acceptance;
[0079] (3) Strong generalization ability: It supports images acquired by multiple centers and multiple devices and has good external verification performance;
[0080] (4) Easy to deploy: Lightweight design, can run on ordinary computers or in the cloud, suitable for use in primary healthcare institutions.
[0081] Therefore, it can be seen that the differential diagnostic model for deep caries and pulpitis trained using the training method provided in this embodiment can efficiently and accurately distinguish between deep caries and pulpitis, and has the ability to provide visual interpretation, thereby improving the transparency and credibility of clinical diagnosis. It is also applicable to various scenarios such as initial screening in the oral emergency department, auxiliary diagnosis in primary healthcare institutions, oral medicine teaching and training, and large-scale oral public health screening.
[0082] Furthermore, the method for training a diagnostic model for differentiating deep caries and pulpitis from dental radiographs provided in this embodiment has certain scientific and social value:
[0083] Scientific value:
[0084] (1) AI research exploring subtle differential diagnosis in oral medicine:
[0085] Current dental AI research largely focuses on detection (such as the presence of caries or periapical periodontitis), while this example focuses on the more complex area of differentiation, namely distinguishing between deep caries and pulpitis, which have similar radiographic appearances. This represents a further enhancement of the cognitive capabilities required of AI and has significant academic research value.
[0086] The research findings will form an AI solution for medical images with high complexity and subtle differences, and its methodology can be transferred to other similar medical differential diagnosis scenarios (such as early cancer identification and differentiation of tumors of different nature).
[0087] (2) Construct a high-quality, professionally annotated oral imaging dataset:
[0088] The "deep caries-pulpitis" image dataset constructed in this embodiment and annotated by experts will itself become a valuable research asset. This dataset can be open-sourced or used for further in-depth research and development (such as predicting pulpitis outcomes and assessing the success rate of vital pulp treatment).
[0089] This embodiment will apply visualization technologies such as Grad-CAM to generate a "heatmap" to show the imaging regions that the AI focuses on when making decisions. This not only enhances the credibility of the model, but also inspires clinicians to discover previously overlooked subtle imaging features, deepen their understanding of the disease itself, and achieve "human-machine interaction."
[0090] Social value:
[0091] (1) Improve the quality and equity of healthcare:
[0092] Providing a "tireless expert assistant" to grassroots hospitals and dental clinics that lack senior specialists can significantly improve their diagnostic capabilities and narrow the gap in treatment levels between different levels of medical institutions. This is an effective technical means to implement the "hierarchical medical system" policy.
[0093] AI-assisted diagnosis helps reduce subjective interpretation differences among different doctors, promotes the standardization and homogenization of oral disease diagnosis, and enables patients to enjoy stable and reliable diagnostic services no matter where they are.
[0094] (2) Optimize resource allocation and reduce costs and increase efficiency:
[0095] For patients: The direct benefit is avoiding the physical and psychological harm and additional financial burden caused by misdiagnosis. An unnecessary root canal treatment costs far more than a filling and can cause irreversible damage to the lifespan of the tooth. AI assistance can effectively protect patients' interests.
[0096] For doctors: It greatly reduces the pressure of reading images in emergency and routine outpatient clinics, improves work efficiency, and allows them to devote more energy to complex treatment procedures and doctor-patient communication, thus reducing professional burnout.
[0097] For medical examination institutions and public health systems: In large-scale oral health screenings, AI software can complete the initial screening, efficiently and quickly identifying suspected cases from a massive pool of people undergoing examinations, which are then reviewed by experts. This optimizes the allocation of human resources, allowing limited expert resources to be focused where they are most needed, resulting in significant public health benefits.
[0098] (3) New medical education model:
[0099] Become an "AI Mentor": Provide dental students and interns with a training platform for image interpretation that offers real-time feedback. They can practice on the system and receive instant AI interpretation results and confidence level prompts, accelerating their learning curve.
[0100] Provides typical teaching cases: The software's built-in typical case library and misdiagnosis case analysis can serve as valuable digital teaching resources for case-based teaching and assessment.
[0101] Example 2
[0102] This embodiment provides a system for differentiating between deep dental caries and pulpitis on dental X-rays, including:
[0103] The data acquisition module is used to acquire patients' dental X-ray images;
[0104] The data annotation module is used to desensitize dental X-ray image data and perform image-level diagnostic labeling and pixel-level lesion annotation to obtain labeled data.
[0105] The data preprocessing module is used to preprocess the labeled data;
[0106] The differential diagnosis module is used to input the preprocessed labeled data into the differential diagnosis model of deep caries and pulpitis on dental radiographs for diagnosis and to obtain the diagnosis results; wherein, the differential diagnosis model of deep caries and pulpitis on dental radiographs is trained by a differential diagnosis model training method for deep caries and pulpitis on dental radiographs provided in Example 1;
[0107] The heatmap generation module is used to generate diagnostic heatmaps based on diagnostic results using Grad-CAM technology.
[0108] The dental X-ray system for differentiating between deep caries and pulpitis provided in this embodiment also supports users to import DICOM and common image formats into the user interface and analyze and output diagnostic reports with one click, thereby assisting doctors in making decisions.
[0109] The specific implementation details of the training method for a differential diagnostic model of deep dental caries and pulpitis can be found in the limitations in Example 1 above, and will not be repeated here.
[0110] The technical features of the above embodiments can be combined in any way (as long as there is no contradiction in the combination of these technical features). For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; these embodiments not explicitly written should also be considered to be within the scope of this specification.
Claims
1. A training method for a diagnostic model to differentiate between deep dental caries and pulpitis, characterized in that, The method comprises the following steps: Step 1: obtaining a dental film image data set verified by pathology or clinical treatment results; Step 2: data desensitization processing is performed on the dental film image data set, and image-level diagnosis label annotation and pixel-level lesion annotation are performed to obtain an annotated data set; Step 3: data preprocessing is performed on the annotated data set, and the data set is divided into a training set and a test set; Step 4: the training set is input into a deep learning model for training to obtain an initial differential diagnosis model; the deep learning model adopts an attention mechanism, a multi-task learning mechanism and a transfer learning mechanism during training; Step 5: the test set is input into the initial differential diagnosis model to obtain a model output result; Step 6: model performance evaluation indicators are calculated according to the model output result, and a heat map is generated using Grad-CAM technology, the initial differential diagnosis model is verified according to the heat map, and a final dental film deep caries and pulpitis differential diagnosis model is obtained.
2. The tooth slice deep caries and pulpitis differential diagnosis model training method according to claim 1, characterized in that, In step 1, the sample size of the dental film image data set is calculated by a diagnostic test sample size formula, and the diagnostic test sample size formula is: where n represents the required sample size per group, Z 1-α / 2 Z value representing the level of significance, Z 1-β Z value representing the test power, pi represents the expected sensitivity or specificity, p2 represents the acceptable minimum performance, represents the mean of pi and p2.
3. The tooth slice deep caries and pulpitis differential diagnosis model training method according to claim 1, characterized in that, In step 3, the data preprocessing includes scaling, size normalization, gray scale standardization and image enhancement.
4. The tooth slice deep caries and pulpitis differential diagnosis model training method according to claim 1, characterized in that, In step 4, the deep learning model adopts a convolutional neural network of a PyTorch framework.
5. The tooth slice deep caries and pulpitis differential diagnosis model training method according to claim 4, characterized in that, The convolutional neural network adopts a ResNet18 pre-training model. 6.The method of claim 1, wherein the method further comprises: obtaining a plurality of training data sets, each of which includes a plurality of images of a tooth slice and a corresponding label indicating whether the tooth slice is deep caries or pulpitis; and training the model using the plurality of training data sets. In step 4, the deep learning model adopts a Focal Loss loss function during training.
7. The tooth slice deep caries and pulpitis differential diagnosis model training method according to claim 1, characterized in that, In step 4, the attention mechanism adopts an SE Block attention mechanism or a CBAM attention mechanism. 8.The method of claim 1, wherein the method further comprises: determining a feature value of each of the plurality of features; and determining a weight of each of the plurality of features based on the feature value of each of the plurality of features. In step 6, the model performance evaluation indicators include AUC, sensitivity and specificity. 9.The method of claim 1, wherein the method further comprises: training the model using the training data. In step 6, 5-fold cross-validation is adopted to verify the initial differential diagnosis model.
10. A system for differentiating between deep caries and pulpitis in a dental film, characterized by, The method comprises the following steps: A data acquisition module is configured to acquire patient dental film image data; A data annotation module is configured to perform data desensitization processing on the dental film image data, and perform image-level diagnosis label annotation and pixel-level lesion annotation to obtain annotated data; A data preprocessing module is configured to perform data preprocessing on the annotated data; A differential diagnosis module is configured to input the preprocessed annotated data into a dental film deep caries and pulpitis differential diagnosis model for diagnosis, and obtain a diagnosis result; the dental film deep caries and pulpitis differential diagnosis model is trained by the dental film deep caries and pulpitis differential diagnosis model training method of any one of claims 1-9; A heat map generation module is configured to generate a diagnosis heat map according to the diagnosis result using Grad-CAM technology.