An oral image multi-tissue full-automatic segmentation method based on artificial intelligence

Abstract

The application discloses an oral cavity image multi-tissue full-automatic segmentation method based on artificial intelligence, relates to the technical field of oral cavity image multi-tissue full-automatic segmentation, and comprises the following steps: collecting and enriching annotated oral cavity image data; designing a data enhancement method to improve image quality; constructing a convolutional neural network model, introducing an attention mechanism to improve accuracy; designing a combined loss function, dynamically adjusting weights to balance multi-class segmentation; predicting and generating results for new samples, and performing post-processing optimization; evaluating different tissue segmentation effects to further improve the integrated model. The application greatly reduces the artificial annotation workload and improves the data set construction efficiency; realizes a dynamic weight adjustment mechanism of the loss function, can be suitable for multi-class medical image segmentation tasks, and improves the model generalization; and an automatic model integration and optimization method is proposed, which can continuously improve the model performance and output more accurate results.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oral image multi-tissue full-automatic segmentation, in particular to an oral image multi-tissue full-automatic segmentation method based on artificial intelligence. BACKGROUND

[0002] At present, the application of deep learning technology in oral CT / MRI image segmentation has made great progress, and can realize the automatic segmentation of teeth, maxilla, soft tissue and other regions. Convolutional neural network UNet, Attention UNet and other structures are successfully applied to oral segmentation tasks, the segmentation accuracy is continuously improved, and the open source oral image dataset has a certain scale.

[0003] However, compared with other medical images, it is still small, although commercial companies have developed oral image automatic segmentation products for auxiliary diagnosis and analysis, but it still faces challenges in wide clinical application, and the segmentation quality of different oral anatomical structures is uneven, fine structures are easily blocked and cannot be segmented, there are problems such as over-segmentation and missed segmentation, and post-processing still has room for improvement. SUMMARY

[0004] In view of the problems existing in the above automatic segmentation and identification, the present application is proposed.

[0005] Therefore, the problem to be solved by the present application is how to provide a method for realizing design of class difficulty perception loss function, dynamic balance of different class samples, and realization of fine sample weight control.

[0006] To solve the above technical problems, the present application provides the following technical scheme:

[0007] In a first aspect, the present application provides an oral image multi-tissue full-automatic segmentation method based on artificial intelligence, which comprises: collecting and labeling oral image data; designing a data enhancement method to enhance the oral image data; constructing a convolutional neural network model; designing a combined loss function and dynamically adjusting the weight; inputting the oral image data into the combined loss function for prediction, evaluating the segmentation effect of different tissues according to the prediction result, and optimizing the neural network model.

[0008] As a preferred scheme of the oral image multi-tissue full-automatic segmentation method based on artificial intelligence, wherein: the labeling process comprises formulating detailed labeling instructions, giving clear definitions and labeling requirements for various oral tissues and structures; designing a reasonable labeling quality inspection mechanism; using repeated labeling with a set proportion to evaluate the consistency between labelers; statistical time data of different labelers to optimize the labeling process; the inspection mechanism comprises: primary inspection: using a full convolution network model to automatically analyze the labeled image, judging whether the labeling is complete and whether it follows the labeling specification, and outputting the integrity test result; the same image is labeled by two labelers, and the computer vision algorithm automatically compares the two labels to find inconsistent areas; if the matching degree is higher than the first threshold, the system automatically confirms the consistency and passes the inspection; secondary inspection: if the matching degree is between the first threshold and the second threshold, the labeler discussion mechanism is automatically triggered, and if the matching degree is lower than the second threshold, it is automatically submitted to algorithm review; for low matching images, use contrast learning algorithm to automatically judge the error reason and output the review result; overall inspection: record the test results of each labeler, once a problem is found, automatically trigger the full job inspection, and output the overall review report, wherein the second threshold is less than the first threshold.

[0009] As a preferred scheme of the oral image multi-tissue full-automatic segmentation method based on artificial intelligence, wherein: the convolutional neural network model comprises an encoder using ResNet50 and a decoder using up-sampling and transpose convolution layers; an attention module is added to the second-to-last layer of the encoder, and CBAM attention is used to obtain spatial weights; a skip connection is set to connect the output of the middle layer of the encoder to the corresponding decoder layer; an Adam optimizer is used, and the learning rate is set to 1e-4 and the Batch Size is between [16, 32].

[0010] As a preferred scheme of the oral image multi-tissue full-automatic segmentation method based on artificial intelligence, wherein: the design process of the combined loss function is: defining K binary cross-entropy losses Lossk according to the number K of segmentation categories; calculating the number of samples of each category to automatically set the initial weight k reflecting sample balance; record the change curve of each Lossk during training, if a certain item rises too fast, increase its k , the specific process is: statistics of the slope of each loss function Lossk of dentin, enamel, background, alveolar cortical bone, alveolar cancellous bone, mandibular nerve canal and wisdom tooth in the last n periods, the formula is:

[0011] slope k =(Lossk[n]-Lossk[n-1]) / (step[n]-step[n-1])

[0012] Calculate the mean u and standard deviation σ of all category slopes; if the slope of a certain category exceeds u+σ, it is considered to be rising too fast; if the slope of a certain category exceeds the mean by one standard deviation, increase its category weight α k ; k ;

[0013] Recalculate the current sample size every 3 cycles, adjust α k ; when all Lossk curves tend to be stable, save the current α k as the final weight; construct a dynamically adjusted combined loss:

[0014] Loss=Σα k *Lossk

[0015] Continue training, if a Lossk rebounds, then re-increase the corresponding α k intervention.

[0016] As a preferred scheme of the artificial intelligence-based oral image multi-tissue full-automatic segmentation method, the weight training process comprises: default weight α k =[α1, α2,..., α K ]; in each batch of training, calculate the average prediction probability p i of each category sample in the current batch, i=1, 2,..., K; introduce a temperature coefficient T to calculate the dynamic weight: α i =1 / (1+exp(p i / T)), β i =1-α i ; calculate the adjusted α i ' and β i ' of the current batch:

[0017] α i '=α i *α i

[0018] β i '=β i *(1-α i )

[0019] Then the category sample weight of this batch is: α'=[α1', α2',..., α K '], β'=[β1', β2',..., β K ']; combine α', β' and each Lossk to do backward and parameter update; repeat the process to continuously adjust the category weights.

[0020] As a preferred scheme of the oral image multi-tissue full-automatic segmentation method based on artificial intelligence, when the segmentation effect of a specific category is low, the corresponding p i is small, and a is large; when the specific category is easy to predict, p i is large, and a is small; to evaluate the difficulty of the category, a threshold value θ is set, when p i < θ, the category is marked as a difficult sample; if the number of difficult sample categories exceeds a certain limit, the T value is reduced, and the dynamic weight a is increased, so as to focus on optimizing the difficult sample; if the number of easy sample categories exceeds the limit, the T value is increased, and the dynamic weight a is reduced.

[0021] As a preferred scheme of the oral image multi-tissue full-automatic segmentation method based on artificial intelligence, the threshold value θ is set as follows: the distribution histogram of the prediction probability p i of each category sample is counted on the training set; the mean and standard deviation of the probability distribution of each category are calculated; the threshold value θ is initialized as the mean of the probability distribution minus 0.5 standard deviations; different threshold values θ are tested on the verification set, and the proportion of each category being incorrectly marked as a difficult sample is recorded; the θ value that makes the incorrect marking rate lowest is selected as the threshold value.

[0022] In a second aspect, to further solve the problems existing in the prior art automatic segmentation and recognition, the present application provides an oral image multi-tissue full-automatic segmentation system based on artificial intelligence, which comprises a data set construction module, a data enhancement module, a model training module, a model prediction module and a model evaluation module.

[0023] In a third aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, any step of the oral image multi-tissue full-automatic segmentation method based on artificial intelligence according to the first aspect of the present application is implemented.

[0024] In a fourth aspect, the present application provides a computer readable storage medium, which stores a computer program, and when the computer program is executed by a processor, any step of the oral image multi-tissue full-automatic segmentation method based on artificial intelligence according to the first aspect of the present application is implemented.

[0025] The application has the advantages that the application greatly reduces the artificial labeling workload and improves the data set construction efficiency; a dynamic weight adjustment mechanism of a loss function is realized, which is applicable to a segmentation task of multi-class medical images and improves the model generalization; and an automatic model integration and optimization method is provided, which can continuously improve the model performance and output more accurate results. BRIEF DESCRIPTION OF DRAWINGS

[0026] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor. Among them:

[0027] Fig. 1 Flowchart of the oral image multi-tissue full-automatic segmentation method based on artificial intelligence in embodiment 1.

[0028] Fig. 2 Specific class comparison process chart of the oral image multi-tissue full-automatic segmentation method based on artificial intelligence in embodiment 1. DETAILED DESCRIPTION

[0029] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings of the specification.

[0030] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the connotation of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0031] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment that excludes other embodiments.

[0032] Embodiment 1

[0033] Reference Figs. 1-2 For the first embodiment of the present application, the embodiment provides an oral image multi-tissue full-automatic segmentation method based on artificial intelligence, and the specific steps are as follows:

[0034] S1: Collect and label oral image data.

[0035] Preferably, the labeling process includes formulating detailed labeling instructions, giving clear definitions and labeling requirements for various oral tissues and structures (including dentin, enamel, background, cortical bone, cancellous bone, mandibular nerve canal, wisdom teeth, and multiple categories); designing a reasonable labeling quality inspection mechanism; using a set proportion of repeated labeling to evaluate the consistency between labelers; and statistically analyzing the time data of different labelers to optimize the labeling process.

[0036] Further, the inspection mechanism includes: primary inspection: using a fully convolutional network model to automatically analyze the labeled image, judging whether the labeling is complete and whether it follows the labeling specification, and outputting the integrity test result; the same image is labeled by two labelers, and a computer vision algorithm automatically compares the two labels to find inconsistent areas; calculate the matching degree of the two labels, if the matching degree is higher than the first threshold, the system automatically confirms the consistency and passes the inspection; secondary inspection: if the matching degree is between the first threshold and the second threshold, the labeler discussion mechanism is automatically triggered, and if the matching degree is lower than the second threshold, it is automatically submitted for algorithm review; for low matching images, use a contrast learning algorithm to automatically judge the error reason and output the review result; overall inspection: record the test results of each labeler, once a problem is found, automatically trigger the full job inspection, output the overall review report, collect all the interaction of the labelers, analyze the common labeling difficulties and errors, and conduct key training; continuously analyze the actual effect, optimize the inspection process, and realize automation. The second threshold is less than the first threshold, in this embodiment, the first threshold is 90%, and the second threshold is 80%.

[0037] Further, the specific steps of using a set proportion of repeated labeling to evaluate the consistency between labelers include: checking a part of images to evaluate whether the labeling result meets the labeling specification and whether there is obvious missing or wrong labeling; comparing the labeling results of different labelers on similar images to determine whether the labeling has consistent standards and whether there is serious subjective bias; statistically analyzing the labeling time of each labeler to determine whether the time is reasonable and whether there is a difference in business level; using an automatic pixel comparison algorithm to calculate the overlapping area of different labels of the same image, and evaluating the consistency through quantitative indicators; constructing a labeling quality model to automatically rate the labelers based on historical labeling data, and improving the problem solvers.

[0038] Further, collect the average annotation duration data of each annotator on different images, and count the time distribution of all staff; estimate the reasonable expected annotation duration range of different types of images according to the difficulty of the images; compare the average duration of each annotator with the expected duration range, and the time obviously deviating from the expected range may be unreasonable; check the annotation quality for the cases of obviously too long or too short time, and feedback if problems are found; count the time difference of different annotators processing the same image, and analyze whether there is individual difference in time utilization efficiency; construct a learning curve model, fit the time learning curve of the annotator, and judge whether the time consumption decline trend is normal; organize annotation test, evaluate the business level of the annotator, and perform correlation analysis with the time consumption data.

[0039] S2: Design a data enhancement method to enhance the oral image data.

[0040] Preferably, the training set image is subjected to data enhancement, such as rotation, scaling, flipping or noise addition, etc., to expand the data set size and sample diversity, and to improve the generalization ability of the model.

[0041] S3: Construct a convolutional neural network model.

[0042] Preferably, the encoder uses ResNet50, the decoder uses up-sampling and transpose convolution layer; an attention module is added to the second-to-last layer of the encoder, and CBAM attention is used to obtain spatial weight; a skip connection is set to connect the middle layer output of the encoder to the corresponding decoder layer; an Adam optimizer is used, and the learning rate is set to 1e-4 and the Batch Size is between [16, 32].

[0043] S4: Design a combined loss function and dynamically adjust the weight.

[0044] Preferably, the specific steps are: define K binary cross-entropy losses Lossk according to the number of segmentation categories K; calculate the number of samples of each category, and automatically set the initial weight α k reflecting sample balance; record the change curve of each Lossk during training, and if a certain item rises too fast, increase its α k , the specific process is: calculate the slope of each loss function Lossk of dentin, enamel, background, cortical bone of alveolar bone, cancellous bone of alveolar bone, mandibular nerve canal and wisdom tooth in the last n periods, the formula is:

[0045] slope k =(Lossk[n]-Lossk[n-1]) / (step[n]-step[n-1])

[0046] Calculate the mean u and standard deviation σ of the slopes of all categories; if the slope of a certain category is greater than u+σ, increase its αk If the slope of one category is more than one standard deviation of the average, increase the weight of this category k ; every 3 epochs, recalculate the current sample size, adjust a k ; when all Lossk curves tend to be stable, save the current a k as the maximum weight; construct a dynamic adjustment of the combined loss: Loss = åa k *Lossk; continue training, if a certain Lossk rebounds, then re-increase the corresponding a k intervention; this dynamic combination loss mechanism will adaptively adjust the weight of each category a k ; ultimately obtain the segmentation model with balanced improvement of each category segmentation accuracy on the validation set; save a k as the importance of each category, applied to new tasks.

[0047] Further, the default weight a k = [a1, a2,..., a K ]; in each training, calculate the average prediction probability p i of each category sample in the current batch, i = 1, 2,..., K; introduce the temperature coefficient T, calculate the dynamic weight: a i = 1 / (1+exp(p i / T)), b i = 1-a i ; calculate the adjusted a i ' of the current batch: a i ' = a i * a i ; similarly, get the adjusted b i ' : b i ' = b i *(1-a i ); then the category sample weight of this batch is: a' = [a1', a2',..., a K '], b' = [b1', b2',..., b K ']; according to a', b' and each Lossk, do weighted combination, backward and parameter update; repeat the above process, continue to dynamically adjust the weight of each category.

[0048] When the segmentation effect of a specific category is low, the corresponding p i is small, and a is large at this time; when a specific category is easy to predict, p i is large, and a is small; to evaluate the difficulty of the category, set a threshold value q, when p iIf the number of difficult sample categories exceeds a certain limit, the value of T is reduced, and thus the value of a is increased, focusing on optimizing difficult samples; if the number of easy sample categories exceeds the limit, the value of T is increased, and the value of a is reduced, avoiding over-optimizing easy samples.

[0049] Further, the setting process of the threshold value θ is as follows: on the training set, the distribution histogram of the prediction probability p of each category sample is counted; the mean and standard deviation of the probability distribution of each category are calculated; θ is initialized as the mean of the probability distribution minus 0.5 standard deviations; different values of θ are tested on the validation set, and the proportion of each category being incorrectly marked as a difficult sample is recorded; the value of θ that makes the incorrect marking rate the lowest is selected as the threshold value; if there are too many easy samples in a category, θ can be appropriately adjusted lower; if there are too many difficult samples, θ can be increased. i t Further, a dynamic threshold value θ can also be set, and θ is gradually increased with the training round requirement; the setting process of θ and the validation results are recorded as a reference for future tasks; in application, if it is found that the determination of difficult and easy samples is inaccurate, θ is adjusted. t

[0050] Further, a dynamic threshold value θ can also be set, and θ is gradually increased with the training round requirement; the setting process of θ and the validation results are recorded as a reference for future tasks; in application, if it is found that the determination of difficult and easy samples is inaccurate, θ is adjusted.

[0051] S5: inputting the oral image data into the combined loss function for prediction, evaluating the segmentation effects of different tissues according to the prediction results, and optimizing the neural network model.

[0052] Specifically, the segmentation effects of each tissue category are evaluated on the test set, error analysis is performed on the tissue categories with poor segmentation effects, problems such as fuzzy boundaries and insufficient samples are determined, the sample loss function weight of the tissue category is adjusted, or a special attention module is designed, and the segmentation performance of the category is focused on improving.

[0053] The embodiment also provides an oral image multi-tissue full-automatic segmentation system based on artificial intelligence, which comprises a data set construction module, a data enhancement module, a model training module, a model prediction module and a model evaluation module.

[0054] The embodiment also provides a computer device suitable for the oral image multi-tissue full-automatic segmentation method based on artificial intelligence, which comprises a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the oral image multi-tissue full-automatic segmentation method based on artificial intelligence proposed in the above embodiment. ​​

[0055] The computer device can be a terminal, which includes a processor, a memory, a communication interface, a display screen and an input device connected by a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner. The wireless manner can be achieved by WIFI, operator network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0056] The embodiment also provides a storage medium having a computer program stored thereon, the program being executed by a processor to implement the method for automatically segmenting a plurality of tissues in an oral image based on artificial intelligence as proposed in the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic storage, a flash memory, a magnetic disk or an optical disk.

[0057] To sum up, the present application greatly reduces the workload of manual annotation and improves the efficiency of data set construction. The dynamic weight adjustment mechanism of the loss function can be applied to the segmentation task of multi-class medical images and improve the model generalization. In addition, an automatic model integration and optimization method is proposed, which can continuously improve the model performance and output more accurate results.

[0058] Embodiment 2

[0059] Referring to Tables 1 and 2, for the second embodiment of the present application, on the basis of the first embodiment, in order to verify its beneficial effects, specific application examples and comparative data of the present application are provided.

[0060] First, 3000 MRI images were collected from the public oral image dataset, the images contain patients of different ages and genders, and are annotated in detail according to the anatomical structure of teeth, gums, tongue, oral maxilla and oral mandible 5 categories, wherein the annotation format follows the DICOM standard.

[0061] A variety of data augmentation methods are used to expand the dataset to 5000, as shown in the following table:

[0062] Table 1 data augmentation example table

[0063]

[0064] An Attention UNet network is constructed, which contains a compression path and an attention mechanism; a combined loss with dynamic weight adjustment is designed, a loss function is established, the weights are dynamically adjusted to balance the multi-class segmentation, the data is input, and the indicators of the model on the test set are evaluated, as shown in Table 2:

[0065] Table 2 model evaluation results

[0066]

[0067] The present application uses more comprehensive indicators, and the prior art only uses the Dice coefficient, so the present application can achieve more accurate evaluation; the automatic model integration and optimization method of the present application can continuously improve the model performance and output more accurate results.

[0068] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. An artificial intelligence-based oral image multi-tissue full-automatic segmentation method, characterized in that: Comprise: Collect and label oral image data; Design a data augmentation method to enhance the oral image data; Construct a convolutional neural network model; Design a combined loss function to dynamically adjust the weight; Input the oral image data into the combined loss function for prediction, evaluate the segmentation effect of different tissues according to the prediction result, and optimize the neural network model; The labeling process includes formulating detailed labeling instructions, giving clear definitions and labeling requirements for various oral tissues and structures; Design a reasonable labeling quality inspection mechanism; Use a set proportion of repeated labeling to evaluate the consistency between labelers; Statistical data of different labelers' time, optimize the labeling process; The inspection mechanism includes: primary inspection: using a full convolutional network model to automatically analyze the labeled image, judging whether the labeling is complete and whether it follows the labeling specification, and outputting the integrity test result; The same image is labeled by two labelers, and the computer vision algorithm automatically compares the two labels to find inconsistent areas; If the matching degree is higher than the first threshold, the system automatically confirms that it is consistent and passes the inspection; Secondary inspection: if the matching degree is between the first threshold and the second threshold, the labeler discussion mechanism is automatically triggered; If the matching degree is lower than the second threshold, it is automatically submitted to the algorithm for review; For low matching images, use contrast learning algorithm to automatically judge the error reason and output the review result; Overall inspection: record the test results of each labeler, if a problem is found, automatically trigger the full job inspection, output the overall review report, and the second threshold is less than the first threshold; The design process of the combination loss function is as follows: defining K binary cross-entropy losses Lossk according to the number K of segmentation categories; calculating the number of samples of each category to automatically set the initial weight α k Reflecting sample balance; record the change curve of each Lossk during the training process, and if a certain item rises too fast, increase α k The specific process is as follows: the slopes of each loss function Lossk of dentin, enamel, background, alveolar cortical bone, alveolar cancellous bone, mandibular nerve canal and wisdom tooth in the last n periods are counted, and the formula is: slope k = (Lossk[n] - Lossk[n-1]) / (step[n] - step[n-1]) Calculate the mean u and standard deviation σ of the slopes for all categories; if the slope of a certain category is... k If the slope of a category exceeds u+σ, then the category is considered to be rising too rapidly; if the slope of a category exceeds one standard deviation from the mean, then the weight α of that category is increased. k The current sample size is recalculated every 3 periods, and α is adjusted accordingly. k When all Lossk curves tend to plateau, save the current α. k As the final weights; construct a dynamically adjusted combined loss: Loss = Σα k *Lossk Continue training, if a Lossk rebounds, then re-increase the corresponding α k Intervene; The training process for the weights includes: default weight α k =[α1, α2, ..., α K In each training batch, calculate the average predicted probability p of each class of samples in the current batch. i Let i = 1, 2, ..., K; introduce a temperature coefficient T, and calculate the dynamic weight: α i = 1 / (1+ exp(p i / T)), β i = 1-α i ; Calculate the adjusted α for the current batch i 'and β i ': α i ' =α i *α i β i ' = β i * (1-α i ) Then the sample weights of this batch of categories are: α' = [α1', α2',..., α K '], β' = [β1', β2',..., β K '];α', β' are combined with each Lossk by weighting, and backward and parameter updating are performed; the process is repeated to continuously dynamically adjust the weights of each category.

2. The artificial intelligence-based oral image multi-tissue full-automatic segmentation method of claim 1, wherein: The convolutional neural network model includes ResNet50 as the encoder and up-sampling and transpose convolutional layer as the decoder; Add an attention module to the second-to-last layer of the encoder to obtain spatial weights using CBAM attention; Set the skip connection to connect the middle layer output of the encoder to the corresponding decoder layer; Use Adam optimizer with a learning rate of 1e-4 and a Batch Size of [16, 32].

3. The artificial intelligence-based oral image multi-tissue full-automatic segmentation method of claim 2, wherein: To evaluate the difficulty of the category, a threshold θ is set, when p i <θ, it is marked as a difficult sample category; If the number of difficult sample categories exceeds the limit, reduce the T value, then increase the dynamic weight α, and focus on optimizing difficult samples; If the number of easy sample categories exceeds the limit, increase the T value and reduce the dynamic weight α. 4.The method of claim 3, wherein the method comprises: The threshold θ setting process is: On the training set, the distribution histogram of the prediction probability p of each class sample is counted. i On the training set, the distribution histogram of the prediction probability p of each class sample is counted. Calculate the mean and standard deviation of each class probability distribution; Initialize the threshold θ to the mean of the probability distribution minus 0.5 standard deviations; Test different threshold θ values on the validation set and record the proportion of each class incorrectly labeled as difficult samples; Select the θ value with the lowest error labeling rate as the threshold.

5. An artificial intelligence-based full-automatic segmentation system for multiple tissues in oral images, based on any one of the artificial intelligence-based full-automatic segmentation methods for multiple tissues in oral images according to claims 1-4. Comprise: A data set construction module for collecting and organizing multi-channel oral medical images, and designing automatic and manual inspection and labeling quality processes; A data augmentation module for image enhancement and using GAN to synthesize a data set; A model training module for designing and building a CNN network structure, constructing a training loss function, and designing a weight adjustment mechanism; A model prediction module for pre-processing input images, loading models, and outputting prediction results; a model evaluation module for designing evaluation metrics to measure segmentation performance.

Images