A multimodal periodontal image recognition method and system based on inductive learning

Abstract

The present invention discloses a multimodal periodontal image recognition method and system based on inductive learning. First, labeled periodontal disease pictures are obtained, and a pre-trained model is updated in a supervised manner; then, the knowledge of the severity of periodontal disease is formalized to form a knowledge base; then, pseudo-labels are generated for unlabeled periodontal disease picture data based on the pre-trained model, and the pseudo-labels are compared with the knowledge base, and the pseudo-labels are corrected and updated with the results of logical reasoning, and the model is retrained using the corrected data; finally, the trained model is used to generate the final label corresponding to the picture, and the final recognition result is obtained using a multimodal fusion method. The present invention uses the framework of inductive learning, and in a scenario with only a small amount of labeled data, by formalizing and utilizing expert knowledge, while accelerating the training process, the model can achieve a higher recognition accuracy.

Description

Technical Field

[0001] The present invention relates to a multimodal periodontal image recognition method and system based on inductive learning, belonging to the technical field of image data processing. Background Art

[0002] In current scientific research and clinical applications, periodontal disease classification based on image recognition is mainly based on a "two-stage" process: the first step is to use target detection models such as Fast-RCNN to first locate the areas where lesions may exist in the patient's periodontal image; the second step is to further classify the detected potential diseased areas based on image classification models such as ResNet, and integrate all the prediction results on a patient's periodontal image to obtain the final judgment result. However, this method has obvious shortcomings in real-world applications. In clinical applications, medical imaging data must be obtained from actual patients, so it is often difficult to collect them in batches; secondly, the categories of patient data actually collected are often related to the severity and incidence of the disease, so there is a serious imbalance between data categories; at the same time, the annotation of medical imaging data requires the doctor's professional knowledge and fine annotation, and the manpower and material resources required are also higher than those of ordinary image data. Therefore, medical imaging data often has many problems such as small quantity, imbalance, and few annotations. In the "two-stage" method of periodontal disease classification, the accuracy of the potential lesion area detected in the first stage greatly affects the accuracy of the judgment results in the second stage, and target detection often relies on a large amount of fine-grained data annotation. Such a large number of fine-grained annotated samples are difficult to obtain in practical applications, and their transferability is poor; at the same time, the classification model in the second stage also often relies on a large amount of annotated data to improve accuracy, and the imbalance in the incidence of periodontal diseases leads to the problem that some categories of data are inherently difficult to collect, which makes it difficult to improve the accuracy of the classification model. In summary, although the "two-stage" periodontal disease classification model has achieved certain results in scientific research, the problems of small number, imbalance, and few annotations of medical imaging data in practice make it difficult to apply it in actual clinical applications. In view of the above defects, the present invention proposes a multimodal periodontal image recognition method and system based on inductive learning. The method integrates the role of human expert knowledge, so that in the clinical application environment with small number, imbalance, and few annotations of training data, the deep model can still achieve good classification and discrimination results. Summary of the invention

[0003] Purpose of the invention: Medical imaging data has many problems such as small quantity, imbalance, and few annotations. It is difficult for existing deep learning algorithms to train models with good generalization performance on them. To address this problem, the present invention adopts the framework of abductive learning, and accelerates the training speed of the model in semi-supervised scenarios by introducing expert knowledge and logical reasoning in the training process of the neural network, thereby improving the final prediction accuracy of the model. Abductive learning is a learning framework that aims to combine logical reasoning with neural perception, and is mainly manifested in the mutual improvement of the logical knowledge base and the neural network: in the abductive stage, the knowledge base will infer the sample pseudo-labels output by the neural network model, thereby achieving correction of the perception results, and use the corrected results to further train the model to improve the prediction accuracy of the model; at the same time, the results of the neural network model can also be used as a basis to improve the quality of the knowledge base, including eliminating noise in the knowledge base and resolving conflicts in the knowledge base. In actual medical scenarios, the abductive learning framework can, with the help of expert knowledge, make better use of information from a small amount of labeled data and a large amount of unlabeled data, improve the quality of pseudo-labeling in the semi-supervised learning process, and improve the convergence speed of the algorithm, so that the model can still achieve training efficiency and accuracy comparable to supervised learning in semi-supervised scenarios. In typical application scenarios, the algorithm system based on abductive learning can help neural network models get rid of their dependence on large amounts of labeled data while also helping machine learning tasks collect more high-quality data and corresponding high-confidence pseudo-labeling, providing a guarantee for the development of other types of supervised learning research in the future.

[0004] Due to task limitations, this method assumes that the knowledge base is accurate and complete, that is, in the process of inductive learning, it is not necessary to modify the logical knowledge base according to the perception results, and the output results of the default logical knowledge base are always accurate. In the case of limited periodontal image data, the method of the present invention can help the deep learning model to quickly learn the current task using the information of unsupervised data, while ensuring the stability of the accuracy, greatly reducing the workload of doctors in the data annotation stage.

[0005] Technical solution: A multimodal periodontal image recognition method based on inductive learning, which includes four parts: model initialization and supervised training, construction of expert knowledge base, pseudo-label generation and model update based on inductive learning framework, and generation of periodontal image recognition results.

[0006] In the initialization and supervised training of the model: collect a small number of labeled multimodal periodontal images of patients; prepare a pre-trained model; use the labeled periodontal image data to perform multiple rounds of gradient descent and back propagation to update the parameters of the pre-trained model to form an initial classification model M_0;

[0007] In the construction of the expert knowledge base: the basis for the identification of periodontal disease classification in periodontal images is collected and abstracted in the form of first-order predicate logic to form an expert knowledge base, denoted as KB;

[0008] In the pseudo-label generation and model update based on the abductive learning framework: first, a large amount of unlabeled multimodal periodontal image data is collected, denoted as the data set S_0; then, the initial classification model M_0 is used to generate pseudo-labels for the unlabeled data in S_0, and the pseudo-labeled data set is denoted as S'_0; the pseudo-labels corresponding to S'_0 are sent to the knowledge base KB for reasoning. For the periodontal image sample s, if the pseudo-label of s is consistent with the logic of the knowledge base, no modification is made to its pseudo-label; if the pseudo-label of the periodontal image sample s is inconsistent with the logic of the expert knowledge base, all unsatisfied logical rules are found, and the pseudo-labels are greedily searched. Search, with the lowest modification cost and the most obvious reduction in inconsistency after modification as the criterion, traverse and modify the pseudo-labels that do not conform to the logical rules one by one, until the inconsistency between the pseudo-label combination and the knowledge base meets the preset requirements; the dataset after the pseudo-label correction is recorded as S"_0, and it is combined with the initial annotated periodontal image data to synthesize the dataset S_1, and S_1 is used to perform a round of gradient descent and back propagation to update the parameters of the initial classification model M_0, and the final model is recorded as M_1; the model M_1 is used to generate pseudo-labels for the unlabeled periodontal image data in S_0, and the above operation is repeated k rounds until the model converges, and the final classification model is recorded as M;

[0009] In the generation of periodontal image recognition results: collect the periodontal image test set T; use the classification model M to generate recognition results for the test set T, and record the final output logit of each sample; collect the recognition results of multimodal images from the same patient, and use the multimodal fusion integration method to generate the patient's final periodontal image recognition results.

[0010] The specific implementation process of model initialization and supervised training is as follows:

[0011] Step 100 , collecting a small number of periodontal image sequences taken by digital cameras, extracting multimodal periodontal images, and annotating them to form an initial supervised data set;

[0012] Step 101, prepare a pre-trained model on ImageNet;

[0013] Step 102, using the cross entropy loss function and the labeled data in the initial supervised data set, perform multiple rounds of gradient descent and back propagation to update the parameters of the pre-trained model until the model is overfitted on the training set and the accuracy on the validation set no longer increases significantly. At this time, the training is considered to be completed and the current model is recorded as the initial classification model M_0.

[0014] The construction process of the expert knowledge base is as follows:

[0015] Step 200, collecting the professional basis and judgment process for judging the severity of periodontal disease as a record;

[0016] Step 201, extracting logical elements used for professional judgment from the records, and recording them as constants, variables, functions or predicates according to their attributes;

[0017] Step 202, abstracting the constants, variables, functions, predicates, etc. extracted in step 201 into symbols used in first-order logic;

[0018] Step 203, using the symbols abstracted in step 202, the reasoning process of judging the severity of periodontal disease is formulated using first-order logic to form an expert knowledge base, which is recorded as KB.

[0019] The pseudo-label generation and model update based on the inductive learning framework are as follows:

[0020] Step 300, collecting a multimodal periodontal image dataset of a patient to form a dataset S_0, which does not need to be labeled;

[0021] Step 301, using the initial classification model M_0 obtained in step 102 to generate pseudo labels for the unlabeled data in S_0, and the pseudo-labeled data set is recorded as S'_0;

[0022] Step 302: Send the pseudo-label corresponding to S'_0 to the knowledge base KB for inference. For sample s in data set S'_0: if the pseudo-label of s is consistent with the logic of the knowledge base, do not modify its pseudo-label and jump to step 304; if the pseudo-label of s is inconsistent with the logic of the expert knowledge base, jump to step 303;

[0023] Step 303: If the pseudo-label is inconsistent with the knowledge base, find out all the logical rules that the pseudo-label does not satisfy, and start greedy search for the pseudo-label. Take the lowest modification cost and the most obvious reduction of inconsistency after modification as the criterion, traverse and modify the pseudo-labels that do not meet the logical rules one by one, until the inconsistency between the pseudo-label combination and the knowledge base meets the preset requirements, and return the corrected pseudo-label obtained by inversion.

[0024] Step 304, the dataset after pseudo-labeling correction is recorded as S"_0, and it is combined with the initial supervised data to form a dataset S_1;

[0025] Step 305, using S_1 to perform a round of gradient descent and back propagation to update parameters of the initial classification model M_0, and the obtained model is recorded as M_1;

[0026] Step 306, if the training process meets the stop requirement at this time, jump to step 307, otherwise jump to step 301;

[0027] Step 307, output the classifier M finally obtained by the inverse training process.

[0028] In the pseudo-label generation and model updating process based on the inductive learning framework, the method for judging the stop of the training process is to reach a preset number of training rounds n=50, or the proportion of pseudo-labels that need to be corrected in the inductive process is less than a fixed value, and the model reaches overfitting.

[0029] The generation of periodontal image recognition results is as follows:

[0030] Step 400, collecting an unlabeled multimodal periodontal image test set T;

[0031] Step 401: for each sample t in the test set T, use the classification model M to generate its classification recognition result, and record the final output logit of each sample;

[0032] Step 402: group the multimodal periodontal images from the same patient into a group, denoted as Z k _i, where i represents the patient serial number corresponding to the group of periodontal images, and k represents the number of multimodal periodontal images in the group;

[0033] Step 403: for each set of multimodal periodontal image data Z k _i, calculate the average of k outputs where logit i.k Represents the original logit value corresponding to the kth sample in the i-th group, and logit i The category with the highest confidence is taken as the periodontal disease classification and recognition result of the patient with serial number i.

[0034] Among them, the periodontal image samples can be further preprocessed in the training and recognition stages, including scaling, adding noise, random perturbation, random cropping, normalization and other operations. The purpose of this is to increase the diversity of the sample space, while improving the generalization ability of the model, it can also improve the prediction performance of the model through integration.

[0035] The gradient descent refers to the stochastic gradient descent method, which iteratively updates the model parameters by calculating the gradient of the loss function on a small batch of data to help the model converge to a (local) optimal solution.

[0036] A multimodal periodontal image recognition system based on inductive learning includes: a model initialization and supervised training module, an expert knowledge base construction module, a pseudo-label generation and model update module based on an inductive learning framework, and a periodontal image recognition result generation module.

[0037] Model initialization and supervised training module: collect a small number of labeled multimodal periodontal images of patients; prepare a pre-trained model; use the labeled periodontal image data to perform multiple rounds of gradient descent and back propagation to update the parameters of the pre-trained model to form an initial classification model M_0;

[0038] The construction module of the expert knowledge base: collect the basis for the identification and classification of periodontal diseases in periodontal images, express them abstractly in the form of first-order predicate logic, and form an expert knowledge base, denoted as KB;

[0039] Pseudo-label generation and model updating module based on the abductive learning framework: First, a large amount of unlabeled multimodal periodontal image data is collected, denoted as the data set S_0; then, the initial classification model M_0 is used to generate pseudo-labels for the unlabeled data in S_0, and the pseudo-labeled data set is denoted as S'_0; the pseudo-labels corresponding to S'_0 are sent to the knowledge base KB for reasoning. For periodontal image sample s, if the pseudo-label of s is consistent with the logic of the knowledge base, no modification is made to its pseudo-label; if the pseudo-label of periodontal image sample s is inconsistent with the logic of the expert knowledge base, all unsatisfied logical rules are found and the pseudo-labels are updated. Greedy search, with the lowest modification cost and the most obvious reduction in inconsistency after modification as the criterion, traverse and modify the pseudo-labels that do not conform to the logical rules one by one until the inconsistency between the pseudo-label combination and the knowledge base meets the preset requirements; the dataset after the pseudo-label correction is recorded as S"_0, and it is combined with the initial supervised data to synthesize the dataset S_1, and S_1 is used to perform a round of gradient descent and back propagation to update the parameters of the initial classification model M_0, and the final model is recorded as M_1; the model M_1 is used to generate pseudo-labels for the unlabeled periodontal image data in S_0, and the above operation is repeated k rounds until the model converges, and the final classification model is recorded as M;

[0040] The module for generating periodontal image recognition results: collect the periodontal image test set T; use the classification model M to generate recognition results for the test set T, and record the final logit output of each sample; collect the recognition results of multimodal images from the same patient, and use the multimodal fusion integration method to generate the patient's final periodontal image recognition results.

[0041] The implementation process of the multimodal periodontal image recognition system based on inductive learning is the same as the above method and will not be repeated here.

[0042] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the multimodal periodontal image recognition method based on inductive learning as described above is implemented.

[0043] A computer-readable storage medium stores a computer program for executing the multimodal periodontal image recognition method based on inverse learning as described above.

[0044] The initialization model required in the inductive learning stage can be trained with only a small amount of labeled data to meet the requirements.

[0045] In the inductive learning stage, using expert knowledge to correct the pseudo-labels of unlabeled data can improve training efficiency and model accuracy and make better use of the information of unlabeled data.

[0046] During the evaluation phase, a multimodal fusion method is used on the model output to integrate the prediction results of the original patients’ multimodal data to improve the accuracy of the prediction.

[0047] Beneficial effects: Compared with the prior art, the multimodal periodontal image recognition method and system based on inductive learning provided by the present invention proposes a feasible solution to the problems existing in the field of intelligent assisted medicine, such as the small amount of image data, few annotations, and the unexplainability of the model. It has been successfully applied in the auxiliary diagnosis of periodontal image diseases, while improving the model prediction accuracy and model training efficiency, greatly reducing the workload of doctors, and providing the possibility for the collection of large amounts of data. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 Flow chart of initialization and supervised training of the model in an embodiment of the present invention;

[0049] Figure 2 This is a flow chart of constructing an expert knowledge base in an embodiment of the present invention.

[0050] Figure 3 Flowchart of pseudo-label generation and model updating based on the inductive learning framework in an embodiment of the present invention.

[0051] Figure 4 This is a flow chart of generating periodontal disease prediction results in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The present invention is further explained below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0053] A multimodal periodontal image recognition method based on inductive learning, which includes four major processes: model initialization and supervised training, construction of expert knowledge base, pseudo-label generation and model update based on inductive learning framework, and generation of periodontal image recognition results.

[0054] The process of model initialization and supervised training is as follows Figure 1 As shown. First, collect a small number of patient periodontal image sequences taken by digital cameras, extract the patient's multimodal periodontal images, and invite hospital experts to annotate them to form an initial supervised data set (step 10); at the same time, prepare a pre-trained model on ImageNet (step 11); then, use the cross entropy loss function and the annotated data in the initial supervised data set to perform multiple rounds of gradient descent and back propagation to update the parameters of the pre-trained model (step 12); if the model is overfitted on the training set and the accuracy on the validation set no longer increases significantly, the training is considered to be completed and jump to step 14, otherwise jump to step 12 (step 13); finally, output the initial classification model M_0 (step 14);

[0055] The process of building the expert knowledge base is as follows: Figure 2 As shown. First, ask the hospital experts about their professional basis and judgment process when judging the severity of periodontal disease, and make detailed records (step 20); second, extract the actual logical elements used by the doctor to make professional judgments from the records, and record them as constants, variables, functions or predicates according to their attributes (step 21); then, abstract the extracted constants, variables, functions and predicates into symbols used in first-order logic (step 22); finally, use the abstract symbols to formulate the reasoning process of the doctor's judgment of the severity of periodontal disease using first-order logic to form an expert knowledge base (step 23).

[0056] The process of pseudo-label generation and model updating based on the inductive learning framework is as follows: Figure 3As shown. First, collect the multimodal periodontal image dataset of patients, which does not need to be labeled (step 30); then, use the initial classification model to generate pseudo-labels for the unlabeled data in the dataset (step 31); then, send the generated pseudo-labels to the knowledge base for reasoning (step 32); then make a judgment, if the pseudo-label of a sample is consistent with the logic of the knowledge base, do not modify its pseudo-label, and jump to step 35, otherwise jump to step 34 (step 33); if the pseudo-label is inconsistent with the knowledge base, find all the logical rules that the pseudo-label does not satisfy, and perform greedy search on the pseudo-label, and choose the one with the lowest modification cost and the most obvious reduction in inconsistency after modification. Then, traverse and modify the pseudo-labels that do not conform to the logical rules one by one until the inconsistency between the pseudo-label combination and the knowledge base meets the preset requirements, and return the corrected pseudo-labels obtained by inversion; (steps 34-1, 34-2); then, synthesize the data set after the pseudo-label correction with the initial supervised data into a new labeled data set (step 35); then, use the labeled data set to perform a round of gradient descent and back propagation to update the parameters of the classification model to obtain an updated classification model (step 36); if the training process meets the stopping requirements at this time, jump to step 38, otherwise jump to step 31 (step 37); finally, output the classification model finally obtained by the inverse training process (step 38).

[0057] The generation of periodontal image recognition results, the process of prediction and diagnosis of periodontal diseases Figure 4 As shown. First, the unlabeled data set of patients in the actual scenario is re-collected by using the pseudo-label generation and model update step 30 to form a test set (step 40); then, for each sample in the test set, the classification model is used to generate its recognition result, and the logit of each sample is finally output (step 41); then, the multimodal images from the same patient are divided into a group (step 42); finally, for each group of multimodal image data, the average logit of its output is calculated, and the category with the highest confidence in the logit is used as the patient's periodontal disease classification and recognition result (step 43).

[0058] A multimodal periodontal image recognition system based on inductive learning includes: a model initialization and supervised training module, an expert knowledge base construction module, a pseudo-label generation and model update module based on an inductive learning framework, and a periodontal image recognition result generation module.

[0059] Model initialization and supervised training module: collect a small number of multimodal periodontal images of patients annotated by medical experts; prepare a pre-trained model; use the annotated data to perform multiple rounds of gradient descent and back propagation to update the parameters of the pre-trained model to form an initial classification model M_0;

[0060] The construction module of the expert knowledge base: collect the basis for medical experts to judge periodontal diseases, abstract them, extract the constants, variables, functions and predicates involved, and symbolize them; use the abstract symbols to formalize the reasoning logic used by experts in judgment, and form an expert knowledge base, recorded as KB;

[0061] Pseudo-label generation and model updating module based on the abductive learning framework: First, a large amount of unlabeled multimodal periodontal image data is collected, denoted as the data set S_0; then, the initial classification model M_0 is used to generate pseudo-labels for the unlabeled data in S_0, and the pseudo-labeled data set is denoted as S'_0; the pseudo-labels corresponding to S'_0 are sent to the knowledge base KB for reasoning. For sample s, if the pseudo-label of s is consistent with the logic of the knowledge base, no modification is made to its pseudo-label; if the pseudo-label of s is inconsistent with the logic of the expert knowledge base, all unsatisfied logical rules are found, and the pseudo-labels are greedily updated. Search, with the lowest modification cost and the most obvious reduction in inconsistency after modification as the criterion, traverse and modify the pseudo-labels that do not conform to the logical rules one by one, until the inconsistency between the pseudo-label combination and the knowledge base meets the preset requirements; the dataset after the pseudo-label correction is recorded as S"_0, and it is combined with the initial supervised data to synthesize the dataset S_1, and S_1 is used to perform a round of gradient descent and back propagation to update the parameters of the classification model M_0, and the final model is recorded as M_1; the model M_1 is used to generate pseudo-labels for the unlabeled data in S_0, and the above operation is repeated k rounds until the model converges, and the final classification model is recorded as M;

[0062] The module for generating periodontal image recognition results: collect the periodontal image test set T; use the classification model M to generate prediction results for the test set T, and record the final output logit of each sample; collect the recognition results of multimodal images from the same patient, and use the multimodal fusion integration method to generate the patient's final classification recognition results.

[0063] Obviously, those skilled in the art should understand that the various steps of the multimodal periodontal image recognition method based on inductive learning or the various modules of the multimodal periodontal image recognition system based on inductive learning of the above-mentioned embodiments of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, optionally, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order from that herein, or they can be made into individual integrated circuit modules respectively, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. In this way, the embodiments of the present invention are not limited to any specific hardware and software combination.

Claims

1. A multimodal periodontal image recognition method based on inductive learning, characterized in that: It includes model initialization and supervised training, construction of expert knowledge base, pseudo-label generation and model update based on the inductive learning framework, and generation of periodontal image recognition results; During the initialization and supervised training of the model: a small number of annotated multimodal periodontal images of patients were collected; Prepare a pre-trained model; use the labeled periodontal image data to perform multiple rounds of gradient descent and back propagation to update parameters of the pre-trained model to form an initial classification model M_0; In the construction of the expert knowledge base: the basis for the identification and classification of periodontal diseases in periodontal images is collected and abstracted in the form of first-order predicate logic to form an expert knowledge base, denoted as KB; In the pseudo-label generation and model update based on the abductive learning framework: first, a large amount of unlabeled multimodal periodontal image data is collected, denoted as the data set S_0; then, the initial classification model M_0 is used to generate pseudo-labels for the unlabeled data in S_0, and the pseudo-labeled data set is denoted as S'_0; the pseudo-labels corresponding to S'_0 are sent to the knowledge base KB for reasoning. For the periodontal image sample s, if the pseudo-label of s is consistent with the logic of the knowledge base, no modification is made to its pseudo-label; if the pseudo-label of the periodontal image sample s is inconsistent with the logic of the expert knowledge base, all unsatisfied logical rules are found, and the pseudo-labels are greedily searched. Search, with the lowest modification cost and the most obvious reduction in inconsistency after modification as the criterion, traverse and modify the pseudo-labels that do not conform to the logical rules one by one, until the inconsistency between the pseudo-label combination and the knowledge base meets the preset requirements; the dataset after the pseudo-label correction is recorded as S''_0, and it is combined with the initial annotated periodontal image data to synthesize the dataset S_1, and S_1 is used to perform a round of gradient descent and back propagation to update the parameters of the initial classification model M_0, and the final model is recorded as M_1; the model M_1 is used to generate pseudo-labels for the unlabeled periodontal image data in S_0, and the above operation is repeated k rounds until the model converges, and the final classification model is recorded as M; In the generation of periodontal image recognition results: collect the periodontal image test set T; use the classification model M to generate recognition results for the test set T, and record the final output logit of each sample; The recognition results of multimodal images from the same patient are collected, and the final periodontal image recognition results of the patient are generated using the multimodal fusion integration method.

2. The multimodal periodontal image recognition method based on inductive learning according to claim 1 is characterized in that: The specific implementation process of model initialization and supervised training is as follows: Step 100 , collecting a small number of periodontal image sequences taken by digital cameras, extracting multimodal periodontal images, and annotating them to form an initial supervised data set; Step 101, prepare a pre-trained model on ImageNet; Step 102, using the cross entropy loss function and the labeled data in the initial supervised data set, perform multiple rounds of gradient descent and back propagation to update the parameters of the pre-trained model until the model is overfitted on the training set and the accuracy on the validation set no longer increases significantly. At this time, the training is considered to be completed and the current model is recorded as the initial classification model M_0.

3. The multimodal periodontal image recognition method based on inductive learning according to claim 1, characterized in that: The specific process of constructing the expert knowledge base is as follows: Step 200, collecting the professional basis and judgment process for judging the severity of periodontal disease as a record; Step 201, extracting logical elements used for professional judgment from the record, and recording them as constants, variables, functions or predicates according to their attributes; Step 202, abstracting the constants, variables, functions and predicates extracted in step 201 into symbols used in first-order logic; Step 203, using the symbols abstracted in step 202, the reasoning process of judging the severity of periodontal disease is formulated using first-order logic to form an expert knowledge base, denoted as KB.

4. The multimodal periodontal image recognition method based on inductive learning according to claim 1 or 2, characterized in that: The pseudo-label generation and model update based on the inductive learning framework are as follows: Step 300, collecting a multimodal periodontal image dataset of a patient to form a dataset S_0, which does not need to be labeled; Step 301, using the initial classification model M_0 to generate pseudo labels for the unlabeled data in S_0, and the pseudo-labeled data set is recorded as S'_0; Step 302, send the pseudo-label corresponding to S'_0 to the knowledge base KB for reasoning; for sample s in the data set S'_0: if the pseudo-label of s is consistent with the logic of the knowledge base, do not modify its pseudo-label and jump to step 304; if the pseudo-label of s is inconsistent with the logic of the expert knowledge base, jump to step 303; Step 303: If the pseudo-label is inconsistent with the knowledge base, find out all the logical rules that the pseudo-label does not satisfy, and perform greedy search on the pseudo-label. Take the lowest modification cost and the most obvious reduction of inconsistency after modification as the criterion, traverse and modify the pseudo-labels that do not conform to the logical rules one by one, until the inconsistency between the pseudo-label combination and the knowledge base meets the preset requirements, and return the corrected pseudo-label obtained by inversion. Step 304, the dataset after pseudo-labeling correction is recorded as S''_0, and it is combined with the initial supervised data to form a dataset S_1; Step 305, using S_1 to perform a round of gradient descent and back propagation to update parameters of the initial classification model M_0, and the obtained model is recorded as M_1; Step 306, if the training process meets the stop requirement at this time, jump to step 307, otherwise jump to step 301; Step 307, output the classifier M finally obtained by the inverse training process.

5. The multimodal periodontal image recognition method based on inductive learning according to claim 4 is characterized in that: The generation of periodontal image recognition results is as follows: Step 400, collecting an unlabeled multimodal periodontal image test set T; Step 401: for each sample t in the test set T, use the classification model M to generate its classification recognition result, and record the final output logit of each sample; Step 402: group the multimodal periodontal images from the same patient into a group, denoted as Z k _i, where i represents the patient serial number corresponding to the group of periodontal images, and k represents the number of multimodal periodontal images in the group; Step 403: for each set of multimodal periodontal image data Z k _i, calculate the average of k outputs ,in Represents the original logit value corresponding to the kth sample in the i-th group, and The category with the highest confidence is taken as the periodontal disease classification and recognition result of the patient with serial number i.

6. The multimodal periodontal image recognition method based on inductive learning according to claim 4, characterized in that: In the pseudo-label generation and model updating process based on the inverse learning framework, the method for judging the stop of the training process is to reach a preset number of training rounds, or the proportion of pseudo-labels that need to be corrected in the inverse process is less than a fixed value, and the model reaches overfitting.

7. A multimodal periodontal image recognition system based on inductive learning, characterized in that: It includes: model initialization and supervised training module, expert knowledge base construction module, pseudo-label generation and model update module based on inductive learning framework, and periodontal image recognition result generation module; Model initialization and supervised training module: collect a small number of annotated multimodal periodontal images of patients; Prepare a pre-trained model; use the labeled periodontal image data to perform multiple rounds of gradient descent and back propagation to update parameters of the pre-trained model to form an initial classification model M_0; The construction module of the expert knowledge base: collect the basis for the identification and classification of periodontal diseases in periodontal images, express them abstractly in the form of first-order predicate logic, and form an expert knowledge base, denoted as KB; Pseudo-label generation and model updating module based on the abductive learning framework: First, a large amount of unlabeled multimodal periodontal image data is collected, denoted as data set S_0; then, the initial classification model M_0 is used to generate pseudo-labels for the unlabeled data in S_0, and the pseudo-labeled data set is denoted as S'_0; the pseudo-labels corresponding to S'_0 are sent to the knowledge base KB for reasoning. For periodontal image sample s, if the pseudo-label of s is consistent with the logic of the knowledge base, no modification is made to its pseudo-label; if the pseudo-label of periodontal image sample s is inconsistent with the logic of the expert knowledge base, all unsatisfied logical rules are found, and a greedy search is performed on the pseudo-labels. The pseudo-labels that do not conform to the logical rules are traversed and modified one by one based on the criteria of the lowest modification cost and the most obvious reduction in inconsistency after modification, until the inconsistency between the pseudo-label combination and the knowledge base meets the preset requirements; The dataset after pseudo-label correction is recorded as S''_0, and it is combined with the initial supervised data to synthesize the dataset S_1. S_1 is used to perform a round of gradient descent and back propagation to update the parameters of the initial classification model M_0, and the final model is recorded as M_1. The model M_1 is used to generate pseudo-labels for the unlabeled periodontal image data in S_0, and the above operation is repeated k times until the model converges. The final classification model is recorded as M. The module for generating periodontal image recognition results: collects the periodontal image test set T; uses the classification model M to generate recognition results for the test set T, and records the final output logit of each sample; The recognition results of multimodal images from the same patient are collected, and the final periodontal image recognition results of the patient are generated using the multimodal fusion integration method.

8. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the multimodal periodontal image recognition method based on inductive learning as described in any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program for executing the multimodal periodontal image recognition method based on inverse learning as described in any one of claims 1 to 6.

Images