Method and system for identifying odontogenic jaw cyst and tumor based on deep learning

Abstract

The invention relates to the technical field of medical image processing, in particular to a method and system for identifying odontogenic jaw cyst and tumor based on deep learning, and the method comprises the steps: obtaining original jaw CBCT image data, carrying out the preprocessing, and generating a combined annotation data set containing lesion types, tooth root absorption form annotations and CT value annotations; constructing a multi-branch neural network model, and training the multi-branch neural network model by using the joint labeling data set to obtain an odontogenic jaw lesion recognition model; and obtaining jaw CBCT image data of a target patient, inputting the jaw CBCT image data into the odontogenic jaw lesion recognition model, and outputting lesion type classification results, tooth root absorption forms and bone destruction volumes. The diagnosis capability of minimal lesions can be improved, missed diagnosis is reduced, whole-process auxiliary diagnosis from qualitative diagnosis to quantitative analysis is realized, and the diagnosis speed is increased.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing technology, and in particular to a method and system for identifying odontogenic jaw cysts and tumors based on deep learning. Background Art

[0002] Odontogenic jaw tumors (such as ameloblastoma) and odontogenic jaw cysts (such as odontogenic keratocyst, dentigerous cyst, radicular cyst) are common diseases of the oral and maxillofacial region. Early diagnosis and precise treatment are crucial to the patient's prognosis.

[0003] However, traditional diagnostic methods mainly rely on doctors' clinical experience and subjective judgment of imaging examinations (such as X-rays, CT, etc.), and have the following limitations: Highly subjective: Different doctors may interpret the same image differently, resulting in inconsistent diagnostic results. In addition, when reading CBCT images, the grayscale value is device-dependent, which may cause an impact when doctors adjust the grayscale value to observe the lesion density.

[0004] Inefficiency: Doctors need to spend a lot of time reading films, which makes it difficult to meet the growing medical needs.

[0005] Difficulty in early diagnosis: Some odontogenic jaw cysts and tumors have no obvious symptoms in the early stage, and the imaging manifestations are atypical, which makes them easy to be missed or misdiagnosed. In particular, some inexperienced junior dentists tend to ignore or even find it difficult to identify odontogenic jaw cysts and tumors, resulting in missed diagnosis, misdiagnosis, and incorrect treatment, and patients fail to receive timely and appropriate treatment in the early stage.

[0006] Although there are methods for diagnosing lesions with the assistance of models, there is currently no auxiliary diagnosis solution for odontogenic lesions, and the existing models still have the following limitations: insufficient classification performance, difficulty in distinguishing lesions with blurred boundaries (such as ameloblastoma and keratocyst), poor adaptability to complex anatomical structures (such as loculated tumors and oval cysts); lack of multimodal feature fusion, the existing models only focus on the overall characteristics of the image, and do not jointly analyze detailed information such as root absorption patterns and dynamic CT values, and still require manual further detailed diagnosis, wasting patients' consultation time.

[0007] Therefore, there is an urgent need for methods and systems based on deep learning to identify odontogenic jaw cysts and tumors, which can improve the diagnostic capabilities of minor lesions, reduce missed diagnoses, misdiagnoses, and incorrect treatments, achieve full-process auxiliary diagnosis from qualitative diagnosis to quantitative analysis, and improve the speed of diagnosis. Summary of the invention

[0008] In order to solve the above technical problems, the present invention provides a method and system for identifying odontogenic jaw cysts and tumors based on deep learning, which can improve the diagnostic ability of minor lesions, reduce missed diagnosis, misdiagnosis, and mistreatment, realize full-process auxiliary diagnosis from qualitative diagnosis to quantitative analysis, and improve the diagnosis speed.

[0009] The present invention provides a method for identifying odontogenic jaw cysts and tumors based on deep learning, comprising the following steps: S1. Obtain the original jaw CBCT image data, perform preprocessing, and generate a joint annotation dataset including lesion type, root resorption morphology annotation, and CT value annotation; S2. construct a multi-branch neural network model, and use the joint annotation data set to train the multi-branch neural network model to obtain an odontogenic jaw lesion recognition model; wherein the multi-branch neural network model includes: a main branch and two auxiliary branches; The first auxiliary branch is a tooth root resorption interpretation module, which is used to output the tooth root resorption morphology according to the input jaw CBCT image data; The second auxiliary branch is a dynamic CT value analysis module, which is used to output the bone destruction volume according to the input jaw CBCT image data; The main branch is the odontogenic jaw lesion recognition module, which is used to output lesion type prediction results based on the input jaw CBCT image data, root resorption morphology and bone destruction volume; S3. Obtain the target patient's jaw CBCT image data, input the odontogenic jaw lesion recognition model, and output the lesion type classification results, root resorption morphology, and bone destruction volume.

[0010] Furthermore, in S1, the original jaw CBCT image data is obtained and preprocessed to generate a joint annotation data set including lesion type, root resorption morphology annotation and CT value annotation, including: S11, performing anatomically guided multi-region cropping on the original jaw CBCT image data to generate a number of image blocks covering the complete tooth root and the bone structure within a preset range of the tooth root; S12, annotating the lesion type, the category label and the CT value of the root absorption morphology based on the image block; wherein the lesion type includes cyst and tumor; the root absorption morphology includes no absorption, oblique absorption, serrated absorption and truncated absorption; S13. Construct a joint annotation dataset based on the annotated image blocks.

[0011] Furthermore, in S2, the loss function of the multi-branch neural network model adopts a joint loss function, and the calculation formula is as follows: L = αL main +βL aux1 +γL aux2 ; Among them, L represents the joint loss function, α, β, and γ are the influence weights of each branch, and L main represents the loss function of the main branch, L aux1 represents the loss function of the first auxiliary branch, L aux2 represents the loss function of the second auxiliary branch.

[0012] Furthermore, the odontogenic jaw lesion recognition module adopts the ResUnet neural network, and embeds a dynamic attention gating module before each upsampling operation of the decoder in the ResUnet neural network, which is used to receive the features output by the encoder and the features output by the two auxiliary branches, and generate spatial attention weights, and obtain the multi-branch fused features according to the spatial attention weights and the features output by the encoder.

[0013] Furthermore, the root resorption judgment module adopts CNN neural network.

[0014] Furthermore, the dynamic CT value analysis module uses wavelet transform to calculate the bone destruction volume.

[0015] Furthermore, the bone destruction volume is calculated using wavelet transform including: Sa, obtain the CT value of CBCT image data; Sb, using wavelet transform to decompose the CT value of CBCT image data to obtain wavelet coefficients of different scales; Sc, extract gradient features according to wavelet coefficients and perform normalization; Sd, determine the voxel points belonging to the bone destruction area according to the normalized gradient characteristics and the preset threshold; Se. Determine the total volume of bone destruction based on all voxel points belonging to the bone destruction area.

[0016] The present invention also provides a system for identifying odontogenic jaw cysts and tumors based on deep learning, which is used to implement the above-mentioned method for identifying odontogenic jaw cysts and tumors based on deep learning. The system includes the following modules: The data acquisition module is used to acquire the original jaw CBCT image data and perform preprocessing to generate a joint annotation data set including lesion type, root resorption morphology annotation and CT value gradient annotation; A model building module, connected to the data acquisition module, is used to build a multi-branch neural network model, and train the multi-branch neural network model using a joint annotation data set to obtain an odontogenic jaw lesion recognition model; The output module is connected to the model building module and is used to obtain the jaw CBCT image data of the target patient, input the odontogenic jaw lesion recognition model, and output the lesion type classification results, root resorption morphology and bone destruction volume.

[0017] The embodiments of the present invention have the following technical effects: This study used the original jaw CBCT The image data is cropped in multiple regions guided by the anatomical structure to generate image blocks covering the complete tooth root and the bone structure within the preset range of the tooth root, so that the subsequent annotation and analysis are more targeted and the key information related to the odontogenic jaw lesions can be effectively extracted. By constructing a multi-branch neural network model and using a joint annotation data set containing lesion type, root absorption morphology annotation and CT value annotation for training, the model can make judgments based on comprehensive information from multiple aspects, and can more comprehensively and accurately identify the lesion type, root absorption morphology and bone destruction volume, thereby improving the diagnostic accuracy of odontogenic jaw cysts and tumors, and can improve the diagnostic ability of minor lesions, reduce missed diagnoses, and achieve full-process auxiliary diagnosis from qualitative diagnosis to quantitative analysis, thereby improving the speed of diagnosis. The root absorption interpretation module and the dynamic CT value analysis module focus on the analysis of root absorption morphology and bone destruction volume, respectively, and provide supplementary information for the main branch odontogenic jaw lesion recognition module. The odontogenic jaw lesion recognition module embeds a dynamic attention gating module before each upsampling operation of the decoder, which can fuse the features output by the encoder and the features output by the two auxiliary branches, and highlight the features that play an important role in lesion recognition by generating spatial attention weights, thereby further improving the accuracy of lesion recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 is a flow chart of a method for identifying odontogenic jaw cysts and tumors based on deep learning provided by an embodiment of the present invention; Figure 2 is a schematic diagram of a panoramic view of a normal mandibular and tooth anatomical structure and a cross-sectional view of a dental arch provided by an embodiment of the present invention; Figure 3 It is a schematic diagram of a tooth root serrated resorption odontogenic jaw tumor and a panoramic view of the tooth anatomical structure and a dental arch cross-sectional view provided by an embodiment of the present invention; Figure 4 It is a schematic diagram of a tooth root truncation-like resorption odontogenic jaw tumor and a panoramic view of the tooth anatomical structure and a dental arch cross-sectional view provided by an embodiment of the present invention; Figure 5It is a schematic diagram of a tooth root non-absorption odontogenic jaw cyst and a panoramic view of the tooth anatomical structure and a cross-sectional view of the dental arch provided by an embodiment of the present invention; Figure 6 It is a schematic diagram of an odontogenic jaw cyst with bevel-shaped absorption of tooth root and a panoramic view of the tooth anatomical structure and a cross-sectional view of the dental arch provided by an embodiment of the present invention; Figure 7 It is a schematic diagram of the structure of a system for identifying odontogenic jaw cysts and tumors based on deep learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be described clearly and completely below. 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 belong to the scope of protection of the present invention.

[0021] The embodiment of the present invention provides a method for identifying odontogenic jaw cysts and tumors based on deep learning. Figure 1 is a flowchart of a method for identifying odontogenic jaw cysts and tumors based on deep learning provided by an embodiment of the present invention, Figure 2 is a schematic diagram of a normal mandibular and tooth anatomical structure panoramic view and dental arch cross-sectional view provided by an embodiment of the present invention, see Figure 1 and Figure 2 , including: S1. Obtain the original jaw CBCT (Cone Beam Computed Tomography) image data and perform preprocessing to generate a joint annotation dataset containing lesion type, root resorption morphology annotation, and CT value (Computed Tomography Value) annotation.

[0022] In some embodiments, S1 may include the following sub-steps: S11. Perform anatomically guided multi-region cropping on the original mandibular CBCT image data to generate a number of image blocks covering the complete tooth root and the bone structure within a preset range of the tooth root.

[0023] The original mandibular CBCT image is cropped according to different anatomical regions (such as maxillary and mandibular) and the specific position of each tooth to ensure that each cropped image block contains a complete tooth root and its surrounding bone structure; illustratively, the preset range can be from the top of the alveolar ridge to the lower edge of the mandible, and the lateral width is the bone structure within the distance from the root surface to the outer bone plate of the mandible.

[0024] S12. Label the lesion type, root resorption morphology, and CT value based on the image block.

[0025] Figure 3 It is a schematic diagram of a tooth root serrated resorption odontogenic jaw tumor and a panoramic view of the tooth anatomical structure and a dental arch cross-sectional view provided by an embodiment of the present invention, Figure 4 It is a schematic diagram of a tooth root truncation-like resorption odontogenic jaw tumor and a panoramic view of the tooth anatomical structure and a dental arch cross-sectional view provided by an embodiment of the present invention, Figure 5 It is a schematic diagram of a tooth root non-absorption odontogenic jaw cyst and a panoramic view of the tooth anatomical structure and a cross-sectional view of the dental arch provided by an embodiment of the present invention, Figure 6 is a schematic diagram of a tooth root oblique absorption odontogenic jaw cyst and a panoramic view of the tooth anatomical structure and a cross-sectional view of the dental arch provided by an embodiment of the present invention, see Figure 3-Figure 6 The types of odontogenic jaw bone lesions include cysts and tumors; the root resorption morphologies include no resorption, oblique resorption, serrated resorption and truncated resorption.

[0026] Among them, the lesion structure of odontogenic jaw tumor (ameloblastoma) is: mostly in the shape of chambers, with relatively high CT values, and the root is absorbed in a serrated or truncated shape; The lesion structure of odontogenic jaw cyst (radicular cyst, tooth-containing cyst, keratotic cyst) is a round or oval transparent shadow with neat edges and relatively low CT value. Root resorption rarely occurs, and the absorption pattern is mostly oblique.

[0027] CT value is an important parameter in imaging data. It can reflect the density information of tissue and plays an important role in judging the nature of lesions.

[0028] In some embodiments, before CT value annotation, the original jaw CBCT image data is first subjected to dynamic CT value normalization, and the grayscale range is adjusted according to different density partitions of the jaw to highlight the contrast between the lesion area and the surrounding tissue. The normalization formula is as follows: ; Among them, I norm (x, y, z) represents the CT value of the voxel point with the coordinates (x, y, z) after standardization, I(x, y, z) represents the original CT value of the voxel point with the coordinates (x, y, z), and I max Indicates the maximum CT value in the original jaw CBCT image data, I min Indicates the minimum CT value in the original jaw CBCT image data.

[0029] S13. Construct a joint annotation dataset based on the annotated image blocks.

[0030] S2. Construct a multi-branch neural network model and use the joint annotation dataset to train the multi-branch neural network model to obtain an odontogenic jaw lesion recognition model.

[0031] In some embodiments, the multi-branch neural network model includes: a main branch and two auxiliary branches; The first auxiliary branch is a tooth root resorption interpretation module, which is used to output the tooth root resorption morphology according to the input jaw CBCT image data; The second auxiliary branch is a dynamic CT value analysis module, which is used to output the bone destruction volume according to the input jaw CBCT image data; The main branch is the odontogenic jaw lesion recognition module, which is used to output lesion type prediction results based on the input jaw CBCT image data, root resorption morphology and bone destruction volume.

[0032] In some embodiments, the odontogenic jaw lesion recognition module adopts a ResUnet neural network, and a dynamic attention gating module is embedded before each upsampling operation of the decoder in the ResUnet neural network, which is used to receive the features output by the encoder and the features output by the two auxiliary branches, and generate spatial attention weights, and obtain multi-branch fused features based on the spatial attention weights and the features output by the encoder.

[0033] Specifically, the dynamic attention gating module calculates the features after multi-branch fusion based on the features output by the encoder and the features output by the two auxiliary branches as follows: ; ; Among them, σ represents the sigmoid activation function, represents the feature concatenation operation, Conv represents the convolution operation, and W att represents the spatial attention weight, F absorb represents the output feature of the root resorption discrimination module, F grad Represents the characteristics of the dynamic CT value analysis module output, F main represents the characteristics of the encoder output, F en Represents the features after multi-branch fusion, represents element-wise multiplication, and + represents a residual connection.

[0034] In some embodiments, the root resorption judgment module adopts a CNN neural network.

[0035] In some embodiments, the dynamic CT value analysis module uses wavelet transform to calculate the bone destruction volume.

[0036] In some embodiments, the loss function of the multi-branch neural network model adopts a joint loss function, and the calculation formula is as follows: L = αL main +βL aux1 +γL aux2 ; Among them, L represents the joint loss function, α, β, and γ are the influence weights of each branch, and L main represents the loss function of the main branch, L aux1 represents the loss function of the first auxiliary branch, L aux2 represents the loss function of the second auxiliary branch.

[0037] Exemplarily, the loss function of the main branch may adopt the Focal loss function, thereby alleviating the problem of imbalanced lesion category samples and balancing the model's attention to the two samples; The loss function of the first auxiliary branch can be a loss function that is a fusion of the Dice loss function and the weighted cross entropy, which optimizes the segmentation area from a low amount and alleviates the category imbalance problem when the absorption area accounts for a small proportion; The loss function of the second auxiliary branch can adopt MSE (mean square error) as the loss function.

[0038] In some embodiments, calculating the bone destruction volume using wavelet transform includes: Sa. Obtain the CT value of CBCT image data.

[0039] In some embodiments, wavelet transform calculation may be performed based on the standardized CT value.

[0040] Sb, use wavelet transform to decompose the CT value of CBCT image data to obtain wavelet coefficients of different scales.

[0041] In some embodiments, the Daubechies wavelet function may be used; the calculation formula is as follows: ; Among them, j represents the scale layer of the wavelet function, W j (x, y, z) represents the wavelet coefficients of the jth level, I norm (m,n,p) represents the normalized CT value of the voxel point with coordinates (m,n,p), (m,n,p) is the coordinate of a reference voxel point, (x,y,z) represents the coordinate of the current voxel point, φ j (xm,yn,zp) represents the wavelet basis function at the j-th scale, which indicates the offset of the current voxel point relative to the reference voxel point in three-dimensional space.

[0042] Sc, extract gradient features based on wavelet coefficients and perform normalization.

[0043] Calculate the gradient of each layer of wavelet coefficients to highlight the local changes in influence. The calculation formula is as follows: ; Among them, G j (x, y, z) represents the gradient of the wavelet coefficients at the jth scale.

[0044] The calculation formula for normalizing the gradient is as follows: ; Among them, G norm,j (x, y, z) represents the normalized gradient of the wavelet coefficients at the jth level, G max,j represents the maximum gradient of the wavelet coefficients at the jth level, G min,j Represents the minimum gradient of the wavelet coefficients at the jth level.

[0045] Sd, determines the voxel points belonging to the bone destruction area based on the normalized gradient features and the preset threshold.

[0046] In some embodiments, the damage threshold T may be set based on expert experience or historical data.

[0047] For each voxel point, determine whether its normalized gradient value is greater than the damage threshold T. If so, the voxel point is determined to belong to the bone destruction area: ; Among them, D(x,y,z) represents the category to which the voxel point belongs. When D(x,y,z)=1, it means that the voxel point belongs to the bone destruction area, and when D(x,y,z)=0, it means that the voxel point belongs to the normal area.

[0048] Se. Determine the total volume of bone destruction based on all voxel points belonging to the bone destruction area.

[0049] Calculate the number of all voxel points with D(x,y,z)=1 to obtain the total volume of bone destruction. The calculation formula is as follows: ; Among them, V destroy Represents the total volume of bone destruction.

[0050] S3. Obtain the target patient's jaw CBCT image data, input the odontogenic jaw lesion recognition model, and output the lesion type classification results, root resorption morphology, and bone destruction volume.

[0051] This study used the original jaw CBCT The image data is cropped in multiple regions guided by the anatomical structure to generate image blocks covering the complete tooth root and the bone structure within the preset range of the tooth root, so that the subsequent annotation and analysis are more targeted and the key information related to the odontogenic jaw lesions can be effectively extracted. By constructing a multi-branch neural network model and using a joint annotation data set containing lesion type, root absorption morphology annotation and CT value annotation for training, the model can make judgments based on comprehensive information from multiple aspects, and can more comprehensively and accurately identify the lesion type, root absorption morphology and bone destruction volume, thereby improving the diagnostic accuracy of odontogenic jaw cysts and tumors, and can improve the diagnostic ability of minor lesions, reduce missed diagnoses, and achieve full-process auxiliary diagnosis from qualitative diagnosis to quantitative analysis, thereby improving the speed of diagnosis. The root absorption interpretation module and the dynamic CT value analysis module focus on the analysis of root absorption morphology and bone destruction volume, respectively, and provide supplementary information for the main branch odontogenic jaw lesion recognition module. The odontogenic jaw lesion recognition module embeds a dynamic attention gating module before each upsampling operation of the decoder, which can fuse the features output by the encoder and the features output by the two auxiliary branches, and highlight the features that play an important role in lesion recognition by generating spatial attention weights, thereby further improving the accuracy of lesion recognition.

[0052] The embodiment of the present invention provides a system for identifying odontogenic jaw cysts and tumors based on deep learning. Figure 7 is a schematic diagram of the structure of a system for identifying odontogenic jaw cysts and tumors based on deep learning provided by an embodiment of the present invention, see Figure 7 The system is used to implement the above-mentioned method for identifying odontogenic jaw cysts and tumors based on deep learning. The system includes the following modules: The data acquisition module is used to acquire the original jaw CBCT image data and perform preprocessing to generate a joint annotation data set including lesion type, root resorption morphology annotation and CT value gradient annotation; A model building module, connected to the data acquisition module, is used to build a multi-branch neural network model, and train the multi-branch neural network model using a joint annotation data set to obtain an odontogenic jaw lesion recognition model; The output module is connected to the model building module and is used to obtain the jaw CBCT image data of the target patient, input the odontogenic jaw lesion recognition model, and output the lesion type classification results, root resorption morphology and bone destruction volume.

[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying odontogenic jaw cysts and tumors based on deep learning, characterized in that: The steps include: S1. Obtain the original jaw CBCT image data, perform preprocessing, and generate a joint annotation dataset including lesion type, root resorption morphology annotation, and CT value annotation; S2, constructing a multi-branch neural network model, and using the joint annotation data set to train the multi-branch neural network model to obtain an odontogenic jaw lesion recognition model; wherein the multi-branch neural network model includes: a main branch and two auxiliary branches; The first auxiliary branch is a tooth root resorption interpretation module, which is used to output the root resorption morphology according to the input jaw CBCT image data; The second auxiliary branch is a dynamic CT value analysis module, which is used to output the bone destruction volume according to the input jaw CBCT image data; The main branch is an odontogenic jaw lesion recognition module, which is used to output lesion type prediction results based on input jaw CBCT image data, root resorption morphology and bone destruction volume; S3. Obtain the jaw CBCT image data of the target patient, input the odontogenic jaw lesion recognition model, and output the lesion type classification result, root resorption morphology, and bone destruction volume.

2. The method for identifying odontogenic jaw cysts and tumors based on deep learning according to claim 1, characterized in that: In S1, the original jaw CBCT image data is obtained and preprocessed to generate a joint annotation data set including lesion type, root resorption morphology annotation and CT value annotation, including: S11, performing anatomically guided multi-region cropping on the original jaw CBCT image data to generate a plurality of image blocks covering the complete tooth root and the bone structure within a preset range of the tooth root; S12, annotating the lesion type, the category label of the root absorption morphology and the CT value based on the image block; wherein the lesion type includes cysts and tumors; the root absorption morphology includes no absorption, oblique absorption, serrated absorption and truncated absorption; S13. Construct a joint annotation dataset based on the annotated image blocks.

3. The method for identifying odontogenic jaw cysts and tumors based on deep learning according to claim 1, characterized in that: In S2, the loss function of the multi-branch neural network model adopts a joint loss function, and the calculation formula is as follows: L=αL main +βL aux1 +γL aux2 ; Among them, L represents the joint loss function, α, β, and γ are the influence weights of each branch, and L main represents the loss function of the main branch, L aux1 represents the loss function of the first auxiliary branch, L aux2 represents the loss function of the second auxiliary branch.

4. The method for identifying odontogenic jaw cysts and tumors based on deep learning according to claim 3, characterized in that: The odontogenic jaw lesion recognition module adopts a ResUnet neural network, and embeds a dynamic attention gating module before each upsampling operation of the decoder in the ResUnet neural network, which is used to receive the features output by the encoder and the features output by two auxiliary branches, and generate spatial attention weights, and obtain multi-branch fused features according to the spatial attention weights and the features output by the encoder.

5. The method for identifying odontogenic jaw cysts and tumors based on deep learning according to claim 3, characterized in that: The tooth root resorption judgment module adopts a CNN neural network.

6. The method for identifying odontogenic jaw cysts and tumors based on deep learning according to claim 3, characterized in that: The dynamic CT value analysis module uses wavelet transform to calculate the bone destruction volume.

7. The method for identifying odontogenic jaw cysts and tumors based on deep learning according to claim 6, characterized in that: The method of calculating the bone destruction volume by wavelet transform comprises: Sa, obtain the CT value of CBCT image data; Sb, using wavelet transform to decompose the CT value of CBCT image data to obtain wavelet coefficients of different scales; Sc, extract gradient features according to wavelet coefficients and perform normalization; Sd, determine the voxel points belonging to the bone destruction area according to the normalized gradient characteristics and the preset threshold; Se. Determine the total volume of bone destruction based on all voxel points belonging to the bone destruction area.

8. A system for identifying odontogenic jaw cysts and tumors based on deep learning, used to implement the method for identifying odontogenic jaw cysts and tumors based on deep learning as described in any one of claims 1 to 7, characterized in that: The system includes the following modules: The data acquisition module is used to acquire the original jaw CBCT image data and perform preprocessing to generate a joint annotation data set including lesion type, root resorption morphology annotation and CT value gradient annotation; A model building module, connected to the data acquisition module, for building a multi-branch neural network model, and using the joint annotation data set to train the multi-branch neural network model to obtain an odontogenic jaw lesion recognition model; The output module is connected to the model building module, and is used to obtain the jaw CBCT image data of the target patient, input the odontogenic jaw lesion recognition model, and output the lesion type classification result, root resorption morphology and bone destruction volume.

Images