Intelligent oral disease triage method and system based on oral image and terminal

Abstract

The invention discloses an intelligent oral disease triage method and system based on oral images and a terminal, and the method comprises the following steps: S1, building a two-dimensional convolutional neural network model based on a dynamic skeleton, which comprises a plurality of deep learning skeletons and a dynamic feature extractor for extracting corresponding features according to different deep learning skeletons, the deep learning skeleton is a neural network which takes a patient oral medical image as an input and takes a disease feature corresponding to the patient oral medical image as an output; s2, acquiring oral medical images of a patient; and S3, inputting the oral medical image picture of the patient into the two-dimensional convolutional neural network model based on the dynamic skeleton, and outputting a recommended department matched with the corresponding disease feature. The two-dimensional convolutional neural network model based on the dynamic skeleton considers the difference of different deep learning picture processing frameworks for pictures with different characteristics, and can help to more accurately carry out oral disease triage.

Description

Technical Field

[0001] The present invention relates to the field of hospital guidance and the field of deep learning image processing algorithm design in computer science, and specifically to an intelligent oral disease triage method, system and terminal based on oral imaging pictures. Background Art

[0002] According to the Global Burden of Disease data, the global incidence of oral diseases increased by 40% from 1990 to 2015, and the disability-adjusted life years caused by these diseases increased by 64% to 16.9 million years. In China, the "2019 Health Development Statistics Report" shows that the number of patient visits reached 8.72 billion, an increase of 410 million over the previous year. This shows that oral diseases have an increasingly serious impact on personal health and daily life functions. With the increasing incidence of oral diseases and the shortage of medical resources, the traditional triage method of oral diseases has been unable to effectively meet the needs of patients and make rational use of hospital resources.

[0003] Traditional triage methods often rely on doctors’ experience and manual operation, which may lead to low triage efficiency, waste of medical resources and increased patient dissatisfaction. Optical image analysis and text classification system based on patient self-report.

[0004] Traditional X-ray image analysis mainly relies on the doctor's experience and skills, which may lead to subjective assessment and inaccuracy. Text-based automatic triage systems may be limited by inaccurate patient descriptions. In addition, the scarcity and uneven distribution of oral medical resources may prevent patients from receiving the required medical intervention in a timely manner.

[0005] At present, intelligent triage systems in the medical field mainly rely on machine learning and image processing technology to assist diagnosis and triage by analyzing medical images. Although some research and applications have made progress in this field, there are still some challenges. For example, existing methods may encounter problems such as insufficient image feature extraction and insufficient classification accuracy when dealing with multi-label classification tasks, thus affecting the accuracy and efficiency of triage.

[0006] For example, Chinese patent CN202410193253.9 discloses a method and system for oral medical guidance based on plug-in instruction fine-tuning of a large model. The above patent solution constructs different plug-in instruction data sets to enable the large model to simulate the doctor's guidance method after multiple rounds of inquiries to the patient. The above patent is easily misleading due to the patient's subjective description. Summary of the invention

[0007] In view of the deficiencies in the prior art, the object of the present invention is to provide an intelligent oral disease triage method, system and terminal based on oral imaging pictures.

[0008] To achieve the above object, the present invention provides the following technical solutions: An intelligent oral disease triage method based on oral imaging pictures comprises the following steps: S1, establishing a two-dimensional convolutional neural network model based on a dynamic skeleton, which includes a plurality of deep learning skeletons and a dynamic feature extractor for extracting corresponding features according to different deep learning skeletons, wherein the deep learning skeleton takes a patient's oral medical image as input and takes the disease features corresponding to the patient's oral medical image as output of a neural network; S2, obtaining oral medical imaging images of patients; S3, inputs the patient's oral medical imaging images into the two-dimensional convolutional neural network model based on the dynamic skeleton, and outputs the recommended department that matches the corresponding disease characteristics.

[0009] The deep learning framework includes an independent Deep learning frameworks, Deep learning frameworks, Deep Learning Frameworks and Deep Learning Frameworks .

[0010] The dynamic skeleton-based two-dimensional convolutional neural network model dynamically switches among different deep learning skeletons, and it dynamically switches based on whether the training accuracy of the deep learning skeleton in the current stage reaches a preset value.

[0011] In S2, the acquired oral medical imaging images of the patient are preprocessed, and the steps are as follows: 1) Read as a grayscale image and use adaptive histogram equalization to process the image; 2) Flip the image horizontally, with the center of the image as the rotation center Rotate the augmented data by a random angle within ; 3) Standardize the image data, unify the image size to 224×224, and scale the pixel values ​​to the range of [0,1]; 4) Convert image data and label data into tensors.

[0012] The steps for constructing the two-dimensional convolutional neural network model based on the dynamic skeleton are as follows: 1) Image serialization, given an image ,in is the height of the image, is the width of the image, is the number of channels of the image. Convert the image data to a PyTorch tensor and adjust its dimension order to , each image is processed as a data sample containing a normalized and transformed tensor; 2) Deep learning framework encapsulation and calling. The deep learning framework is pre-trained and encapsulated in four In the file, the model with the highest accuracy is selected for image data with different characteristics, as shown below, where the accuracy : ; 3) Improvement of loss function. Binary cross entropy BCE is used as the basic loss function, and an unbalanced weight strategy is introduced, where P is the positive class probability predicted by the model, y is the true label, and its value is 1. is the weight of the positive class, is the weight of the negative class:

[0013] .

[0014] The recommended departments include one or more of oral implantology, oral rehabilitation, oral periodontics, oral orthodontics, oral and maxillofacial alveolar surgery, and oral endodontics.

[0015] The patient's oral medical imaging pictures are oral Image picture.

[0016] A system based on the above-mentioned intelligent oral disease triage method based on oral imaging pictures comprises: An acquisition module is used to acquire oral medical imaging images of patients; A processing module, comprising a plurality of deep learning skeletons and a dynamic feature extractor for extracting corresponding features according to different deep learning skeletons, wherein the deep learning skeleton takes a patient's oral medical image as input and a disease feature corresponding to the patient's oral medical image as output; The determination module determines the recommended department that matches the corresponding disease characteristics based on the disease characteristics output by the processing module and the department label.

[0017] A device comprising a processor and a memory; The memory is used to store program codes and transmit the program codes to the processor; The processor is used to execute the above-mentioned intelligent oral disease triage method based on oral imaging pictures according to the instructions in the program code.

[0018] A computer-readable storage medium includes instructions, which, when executed on a computer, enable the computer to execute the above-mentioned intelligent oral disease triage method based on oral imaging pictures.

[0019] Beneficial effects of the present invention: 1) Based on more objective oral medical images, the present invention can avoid the subjective influence of patients' oral triage, and has a higher accuracy rate than other existing intelligent triage systems based on text or medical knowledge graphs.

[0020] 2) Based on deep learning algorithms, the present invention can reduce the process of manual triage, provide higher accuracy for oral triage, and reduce the waste of medical resources caused by wrong appointments.

[0021] 3) The deep learning algorithm consists of a two-dimensional convolutional neural network with a dynamic skeleton, which takes into account the differences in different deep learning image processing frameworks for images with different characteristics, and can help to triage oral diseases more accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a two-dimensional convolutional neural network model based on dynamic skeleton.

[0023] Figure 2 A schematic diagram of the training process of the dynamic skeleton-based two-dimensional convolutional neural network model of the present invention. DETAILED DESCRIPTION

[0024] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0025] It should be noted that all directional indications in the embodiments of the present invention (such as up, down, left, right, front, back, etc.) are only used to explain the relative position relationship, movement status, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0026] This paper establishes a two-dimensional convolutional neural network model based on a dynamic skeleton. The core feature of this model is that it introduces a dynamic skeleton replacement mechanism. This mechanism allows us to flexibly switch and adjust the skeleton structure of the model at different training stages or for specific data set characteristics. In order to enhance the representation ability and adaptability of the model, four deep learning frameworks that have performed well in the field of image recognition are used as the skeleton: , , as well as .

[0027] in Feature reuse and gradient flow are enhanced through dense connections, and each layer receives the output of all previous layers to form a dense structure. It is a deep residual network that can solve the gradient vanishing problem through residual learning. It is an efficient network architecture based on compound scaling, which improves performance by uniformly scaling depth, width and resolution, and uses automatic data enhancement to improve generalization capabilities. It is a lightweight network architecture that uses inverted residual and linear bottleneck layers to reduce calculations and speed up training. Light image speed, reduce memory usage.

[0028] A new deep learning framework can also be added as the skeleton, and each deep learning framework is independent of each other. Each deep learning framework targets different features, so a dynamic feature extractor is also proposed. According to the different deep learning frameworks called, the dynamic feature extractor is used to dynamically extract the corresponding features.

[0029] The dynamic feature extractor can flexibly switch and adjust the model structure during the training process. This extractor loads four pre-trained models (DenseNet121, ResNet50, EfficientNet-B7, MobileNet-V2) and manages these models through a custom DynamicFeatureExtractor class. During the training process, the extractor trains and evaluates each model, and dynamically switches to another model based on the performance of the model at different stages. Each time the switch occurs, the system retains the current optimal model and continues to track the accuracy of each model, and finally selects and retains the best performing model for subsequent feature extraction tasks. This framework provides great flexibility for adjusting the model structure and can optimize the performance of the model at different training stages.

[0030] Training and dynamic switching of dynamic feature extractors 1. Detailed description of the training process Suppose we have a dataset containing thousands of images, and the task is to classify these images. We select the DenseNet121 model from four pre-trained models (DenseNet121, ResNet50, EfficientNet-B7, MobileNet-V2) for initial training.

[0031] Phase 1: At the beginning, we select DenseNet121 as the initial model and load it into the GPU. We use the dataset to train the model and gradually update the parameters of the model so that it can better adapt to the classification task. In this phase, the system improves the model through forward propagation, loss calculation, back propagation, and parameter optimization.

[0032] Evaluation: After each epoch, the system will use the validation set to evaluate the current model and calculate the classification accuracy of the model on the validation set.

[0033] Dynamic switching: Assume that in the first stage, we trained for 5 epochs and obtained the preliminary accuracy of DenseNet121. Next, the system will determine whether to switch to the next model, such as ResNet50, based on this accuracy. The basis for switching can be whether the performance of the current model meets expectations, or switching to a new model at each stage according to the preset strategy.

[0034] 2. Dynamic switching between different stages Dynamic switching refers to switching the model used at different stages of training. For example, DenseNet121 may be used in the initial stage of training, ResNet50 in the middle stage, and EfficientNet-B7 in the final stage. Each stage can be several epochs, and can also be dynamically adjusted based on the performance of the model. For example: Stage 1: DenseNet121 has an accuracy of 80% after 5 epochs.

[0035] Phase 2: Switch to ResNet50 and continue training for 5 epochs, with the accuracy increased to 82%.

[0036] Stage 3: Switch to EfficientNet-B7 and train for another 5 epochs, with the accuracy further improved to 85%.

[0037] That is, the two-dimensional convolutional neural network model based on the dynamic skeleton established in the present invention can autonomously switch between different deep learning frameworks to achieve the highest accuracy. The dynamic feature extractor has different structures for each pre-trained model according to the different deep learning frameworks, so the features they extract are also different. For example, DenseNet121 may be better at capturing detailed textures in an image, while ResNet50 is better at recognizing the overall shape and pattern of an image.

[0038] When the dynamic feature extractor switches to a different model, it uses the feature extraction capabilities of the current model to process the input image. This means that the features extracted at each stage are determined by the architecture and characteristics of the currently selected model. Therefore, feature extractors of different models can capture different image features at different training stages, thereby optimizing the final classification performance.

[0039] During the model training process, considering that the uneven data quantity and quality of the local data set we put into training will affect the accuracy of the guidance classification results, the model can automatically call four pre-trained models DenseNet121, ResNet50, EfficientNet-B7, and MobileNet-V2 to train the local data set independently. In addition, we have also carefully designed a dynamic feature extractor in the guidance system. This dynamic feature extractor can select four models for feature extraction of oral imaging pictures according to the accuracy. This mechanism allows us to flexibly switch and adjust the structure of the model at different training stages. Finally, the accuracy of the guidance classification results of different models is compared on the test set and the validation set, and the model with the best feature combination is selected as the system choice. This design allows our system to adapt to the differences in different local data sets caused by shooting equipment, shooting techniques, image quality, number of pictures, etc.

[0040] During the patient's use, since the model has been trained, when the patient uploads his oral image through the app, our system will output the internal guidance and classification results for each of them. At the same time, the system will compare the internal outputs of the four models, and output the guidance and classification results of the optimal model based on the comprehensive consideration of the recognition accuracy of the model in the local data set and the similarity of the internal outputs of the four models. This design not only enables us to have a high accuracy rate, but also takes into account the advantages of each model.

[0041] Judgment and comparison of accuracy Each model is trained independently by calculating its accuracy on the validation set. Accuracy is usually determined by comparing the classification accuracy of the model on the validation set. At the end of each stage, the accuracy of the current model is compared with the accuracy of all previous models.

[0042] Individual accuracy judgment: At each stage, the system records the accuracy of the current model and compares it with the performance of the previous model. If the current model has a higher accuracy, it is saved as the current optimal model.

[0043] Final model selection: After multiple stages, the best performing model (i.e., the model with the highest accuracy) is retained and used for the final feature extraction and classification tasks.

[0044] This process ensures that throughout the training process, the system is able to dynamically select and adjust the most suitable model, thereby improving the overall accuracy of the classification task.

[0045] The present invention also discloses an intelligent oral disease triage method based on oral imaging pictures, which comprises the following steps: S1, establishing a two-dimensional convolutional neural network model based on a dynamic skeleton, which includes a plurality of deep learning skeletons and a dynamic feature extractor for extracting corresponding features according to different deep learning skeletons, wherein the deep learning skeleton takes a patient's oral medical image as input and takes the disease features corresponding to the patient's oral medical image as output of a neural network; The steps for constructing the two-dimensional convolutional neural network model based on the dynamic skeleton are as follows: 1) Image serialization, given an image ,in is the height of the image, is the width of the image, is the number of channels of the image. Convert the image data to a PyTorch tensor and adjust its dimension order to , each image is processed as a data sample containing a normalized and transformed tensor; 2) Deep learning framework encapsulation and calling. The deep learning framework is pre-trained and encapsulated in four In the file, the model with the highest accuracy is selected for image data with different characteristics, as shown below, where the accuracy : ; 3) Improvement of loss function. Binary cross entropy BCE is used as the basic loss function, and an unbalanced weight strategy is introduced, where P is the positive class probability predicted by the model, y is the true label, and its value is 1. is the weight of the positive class, is the weight of the negative class:

[0046] .

[0047] S2, obtaining a medical image of the patient's oral cavity, wherein the medical image of the patient's oral cavity is Image pictures; and the patient's oral medical image pictures need to be preprocessed, the specific preprocessing steps are as follows: 1. Read as grayscale image and process the image using adaptive histogram equalization (CLAHE).

[0048] 2. Flip the image horizontally, with the image center as the rotation center. The augmented data is rotated by a random angle within .

[0049] 3. Standardize the image data and unify the image size into , and scale the pixel values ​​to within the range.

[0050] 4. Convert image data and label data into tensors.

[0051] S3, inputs the patient's oral medical imaging images into the two-dimensional convolutional neural network model based on the dynamic skeleton, and outputs the recommended department that matches the corresponding disease characteristics.

[0052] The output results include the following secondary oral departments: oral implantology, oral rehabilitation, oral periodontology, oral orthodontics, oral and maxillofacial alveolar surgery, and oral endodontics. This technology innovatively uses objective oral medical imaging images as the basis for triage, avoiding the misleading oral description of patients and the cumbersome manual triage process.

[0053] The specific process of the deep learning framework used in this application is as follows: DenseNet121: Image input->Conv2d (7x7 convolution)->BatchNorm2d->ReLU->MaxPool2d (3x3 pooling)->DenseBlock1 (6 DenseLayer)->Transition1->DenseBlock2 (12 DenseLayer)->Transition2->DenseBlock3 (24 DenseLayer)->Transition3->DenseBlock4 (16 DenseLayer)->BatchNorm2d->classifier->classification label output Among them, the structure of each DenseLayer is: BatchNorm2d ->ReLU ->Conv2d (1x1 convolution) -> BatchNorm2d ->ReLU ->Conv2d (3x3 convolution), and the structure of each Transition layer is: BatchNorm2d ->ReLU ->Conv2d (1x1 convolution) ->AvgPool2d (2x2 pooling) illustrate: Conv2d: 2D convolutional layer; BatchNorm2d: 2D batch normalization layer; ReLU: ReLU activation function; MaxPool2d: 2D max pooling layer; DenseBlock: dense block; DenseLayer: dense layer; DenseLayer: transition layer; Classifier: Classifier.

[0054] ResNet50: Image input->Conv2d (7x7 convolution)->BatchNorm2d->ReLU->MaxPool2d (3x3 pooling)->layer1 (3 Bottleneck)->layer2 (4 Bottleneck)->layer3 (6 Bottleneck)->layer4 (3 Bottleneck)->AvgPool2d->linear fully connected layer->classification label output The structure of each Bottleneck is Conv2d (1x1 convolution) -> BatchNorm2d -> Conv2d (3x3 convolution) -> BatchNorm2d -> Conv2d (1x1 convolution) -> BatchNorm2d -> ReLU. If the number of input channels of the Bottleneck module is inconsistent with the number of output channels, downsample will be added to adjust the number of input channels to make them consistent through 1x1 convolution.

[0055] illustrate: Conv2d: 2D convolutional layer; BatchNorm2d: 2D batch normalization layer; ReLU: ReLU activation function; MaxPool2d: 2D max pooling layer; Bottleneck: bottleneck layer; Downsample: downsampling; AvgPool2d: 2D global average pooling layer; Linear fully connected layer: Linear fully connected layer.

[0056] EfficientNet-B7: Image input->Conv2d (3x3 convolution)->BatchNorm2d->SiLU->Sequential1 (4 MBConv blocks)->Sequential2 (7 MBConv blocks)->Sequential3 (7 MBConv blocks)->Sequential4 (10 MBConv blocks)->Sequential5 (10 MBConv blocks)->Sequential6 (13 MBConv blocks)->Sequential7 (4 MBConv blocks)->Conv2d (1x1 convolution)->BatchNorm2d->SiLU->AvgPool2d->classifier->classification label output Each MBConv consists of five parts: the first part: Conv2d (1x1 convolution) -> BatchNorm2d -> SiLU; the second part: Conv2d (3x3 convolution) -> BatchNorm2d -> SiLU; the third part: Squeeze-and-Excitation (SE) module; the fourth part: Conv2d (1x1 convolution) -> BatchNorm2d; the fifth part: StochasticDepth random depth. illustrate: Conv2d: 2D convolutional layer; BatchNorm2d: 2D batch normalization layer; SiLU: SiLU activation function; MBConv: Mobile Flip Bottleneck Convolution Block; AvgPool2d: 2D global average pooling layer; ReLU6: ReLU6 activation function.

[0057] MobileNet-V2: Image input->Conv2d (3x3 convolution)->BatchNorm2d ->ReLU6 ->I nvertedResidual1 ->InvertedResidual2->InvertedResidual3 ->InvertedResidual4->InvertedResidual5->InvertedResidual6->InvertedResidual7->InvertedResidual8->InvertedResidual9->InvertedResidual10->InvertedResidual11->InvertedResi dual12->InvertedResidual13->InvertedResidual14->InvertedResidual15->InvertedResidual16->InvertedResidual17->Conv2d (1x1 convolution)->BatchNorm2d->ReLU6->classifier(linear)->Classification label output Among them, the InvertedResidual blocks are composed of the inverted residual structure of Conv2d (1x1 convolution) -> BatchNorm2d -> ReLU6 -> Conv2d (3x3 convolution) -> BatchNorm2d -> ReLU6 -> Conv2d (1x1 convolution) -> BatchNorm2d.

[0058] illustrate: InvertedResidual: transposed residual block; Conv2d: 2D convolutional layer; BatchNorm2d: 2D batch normalization layer; classifier(linear): Linear classifier.

[0059] like Figure 2As shown in the figure, the two-dimensional convolutional neural network model based on dynamic skeleton proposed in the present application includes four independent deep learning skeletons and a dynamic feature extractor for extracting corresponding features according to different deep learning skeletons. In actual training, DenseNet121 is selected as the initial model from the pre-trained models of four independent deep learning skeletons, namely DenseNet121, ResNet50, EfficientNet-B7, and MobileNet121. At the same time, the dynamic feature extractor extracts image features according to the features required by DenseNet121, and then loads the selected DenseNet121 into the GPU, trains the model using the data set, and performs forward propagation. The model parameters are gradually updated through propagation, loss calculation, back propagation and parameter optimization to make it more suitable for classification tasks. Then, after each epoch, the validation set is used to evaluate the current model, and the classification accuracy of the model on the validation set is calculated. After several epochs, the preliminary accuracy of the current model is evaluated. If the current model performance meets expectations, it will continue to be used. If it does not meet expectations or according to the preset strategy, consider switching to the next model and repeating the training to complete the model training. Then, based on the oral imaging pictures uploaded by the patient, the recognition accuracy of the four independent models and the dynamic skeleton model on the local data set and the similarity of the internal output can be evaluated respectively, and then the guidance classification result output of the optimal model can be selected. Specific embodiments ① Oral medical image preprocessing The original dcm Data is converted to png The data is read as a grayscale image. Adaptive histogram equalization (CLAHE) is used to adjust the contrast of the image, which can maintain the details of the local image while avoiding the over-enhancement problem that may be caused by global equalization. After the image is uniformly resized, the image is processed using the white balance algorithm. In order to expand the training set, the image is horizontally flipped and rotated with the image center as the center. The augmented data is rotated by a random angle within .

[0061] ② Construction of a two-dimensional convolutional neural network model for dynamic skeletons , , as well as Four deep learning frameworks that have performed well in the field of image recognition are used as flexible and replaceable dynamic skeletons after pre-training. This mechanism allows us to adjust the model skeleton at different training stages or for specific dataset characteristics.

[0062] In order to train and verify the accuracy of the model's triage, senior doctors manually annotated the recommended clinic for each oral image. The data set was randomly divided into three parts, with a training set accounting for 64%, a test set accounting for 20%, and a validation set accounting for 16%. The training set was used for model training and parameter optimization; the validation set was used to adjust hyperparameters and select the best model to prevent overfitting; the test set was used for the final evaluation of the model to test its generalization ability on unseen data.

[0063] ③ Mobile app construction As a supplement to the above model, we designed a mobile application for the Android platform to ensure that users can access the system no matter where they are. It can output several dental clinics as follows: Oral Implantology, Oral Prosthodontics, Oral Periodontics, Oral Orthodontics, Oral and Maxillofacial Alveolar Surgery, and Oral Endodontics.

[0064] That is, when patients register for the first time, they take oral X-ray images according to the advice of the general dentist. Then, patients only need to upload their oral X-ray images in the app, and this intelligent oral triage system can recommend several secondary oral clinics based on the potential lesions in the images. Based on this, patients can go to the dental clinic with a higher degree of fit.

[0065] The present invention also provides a system based on the above-mentioned intelligent oral disease triage method based on oral imaging pictures, which comprises: An acquisition module is used to acquire the patient's oral medical imaging pictures. The acquisition module can be an APP downloaded on the mobile terminal. By taking oral X-ray images and inputting them into the APP, the image acquisition can be completed; A processing module, comprising a plurality of deep learning skeletons and a dynamic feature extractor for extracting corresponding features according to different deep learning skeletons, wherein the deep learning skeleton takes a patient's oral medical image as input and a disease feature corresponding to the patient's oral medical image as output; The determination module determines the recommended department that matches the corresponding disease characteristics based on the disease characteristics output by the processing module and the department label. The determination module can also be an APP set on the mobile terminal. By inputting the oral X-ray image, it is directly processed by the processing module and outputs the corresponding recommended department through the determination module.

[0066] A device, which may be a server, includes a processor and a memory; The memory is used to store program codes and transmit the program codes to the processor; The processor is used to execute the above-mentioned intelligent oral disease triage method based on oral imaging pictures according to the instructions in the program code.

[0067] The server may vary greatly due to different configurations or performances, and may include one or more central processing units (e.g., one or more processors) and memories, and one or more storage media (e.g., one or more mass storage devices) for storing applications or data. The memories and storage media may be short-term storage or persistent storage.

[0068] A computer-readable storage medium includes instructions, which, when executed on a computer, enable the computer to execute the above-mentioned intelligent oral disease triage method based on oral imaging pictures.

[0069] The memory may include a random access memory or a read-only memory. Optionally, the memory includes a non-transitory computer-readable medium. The memory may be used to store instructions, programs, codes, code sets or instruction sets. The memory may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area may store data involved in the above-mentioned various methods, etc. The memory may also be optionally at least one storage device located away from the aforementioned processor. As a computer storage medium, the memory may include an operating system, a network communication module, a user interface module, and an application for an intelligent oral disease triage method based on oral imaging pictures.

[0070] The embodiments should not be regarded as limiting the present invention, but any improvements made based on the spirit of the present invention should be within the protection scope of the present invention.

Claims

1. An intelligent oral disease triage method based on oral imaging pictures, characterized in that: It includes the following steps: S1, establishing a two-dimensional convolutional neural network model based on a dynamic skeleton, which includes a plurality of deep learning skeletons and a dynamic feature extractor for extracting corresponding features according to different deep learning skeletons, wherein the deep learning skeleton takes a patient's oral medical image as input and takes the disease features corresponding to the patient's oral medical image as output of a neural network; S2, obtaining oral medical imaging images of patients; S3, inputs the patient's oral medical imaging images into the two-dimensional convolutional neural network model based on the dynamic skeleton, and outputs the recommended department that matches the corresponding disease characteristics.

2. The intelligent oral disease triage method based on oral imaging pictures according to claim 1, characterized in that: The deep learning framework includes an independent Deep learning frameworks, Deep learning frameworks, Deep Learning Frameworks and Deep Learning Frameworks .

3. The intelligent oral disease triage method based on oral imaging pictures according to claim 2, characterized in that: The dynamic skeleton-based two-dimensional convolutional neural network model dynamically switches among different deep learning skeletons, and it dynamically switches based on whether the training accuracy of the deep learning skeleton in the current stage reaches a preset value.

4. The intelligent oral disease triage method based on oral imaging pictures according to claim 1, 2 or 3, characterized in that: In S2, the acquired oral medical imaging images of the patient are preprocessed, and the steps are as follows: 1) Read as a grayscale image and use adaptive histogram equalization to process the image; 2) Flip the image horizontally, with the center of the image as the rotation center Rotate the augmented data by a random angle within ; 3) Standardize the image data, unify the image size to 224×224, and scale the pixel values ​​to the range of [0,1]; 4) Convert image data and label data into tensors.

5. The intelligent oral disease triage method based on oral imaging pictures according to claim 1, 2 or 3, characterized in that: The steps for constructing the two-dimensional convolutional neural network model based on the dynamic skeleton are as follows: 1) Image serialization, given an image ,in is the height of the image, is the width of the image, is the number of channels of the image. Convert the image data to a PyTorch tensor and adjust its dimension order to , each image is processed as a data sample containing a normalized and transformed tensor; 2) Deep learning framework encapsulation and calling. The deep learning framework is pre-trained and encapsulated in four In the file, the model with the highest accuracy is selected for image data with different characteristics, as shown below, where the accuracy : ； 3) Improvement of loss function. Binary cross entropy BCE is used as the basic loss function, and an unbalanced weight strategy is introduced, where P is the positive class probability predicted by the model, y is the true label, and its value is 1. is the weight of the positive class, is the weight of the negative class: 。 6. The intelligent oral disease triage method based on oral imaging pictures according to claim 1, characterized in that: The recommended departments include one or more of oral implantology, oral rehabilitation, oral periodontics, oral orthodontics, oral and maxillofacial alveolar surgery, and oral endodontics.

7. The intelligent oral disease triage method based on oral imaging pictures according to claim 1, 2 or 3, characterized in that: The patient's oral medical imaging pictures are oral Image picture.

8. A system based on the intelligent oral disease triage method based on oral imaging pictures according to any one of claims 1 to 7, characterized in that: It includes: An acquisition module is used to acquire oral medical imaging images of patients; A processing module, comprising a plurality of deep learning skeletons and a dynamic feature extractor for extracting corresponding features according to different deep learning skeletons, wherein the deep learning skeleton takes a patient's oral medical image as input and a disease feature corresponding to the patient's oral medical image as output; The determination module determines the recommended department that matches the corresponding disease characteristics based on the disease characteristics output by the processing module and the department label.

9. A terminal, characterized in that: The device includes a processor and a memory; The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the intelligent oral disease triage method based on oral imaging pictures as described in any one of claims 1 to 7 according to the instructions in the program code.

10. A computer-readable storage medium, characterized in that: It includes instructions, which, when executed on a computer, enable the computer to execute the intelligent oral disease triage method based on oral imaging pictures as described in any one of claims 1 to 7.

Images