A method, device and medium for identifying human tooth cracks based on deep learning

Abstract

The invention discloses a method, device and medium for identifying human tooth cracks based on deep learning, wherein a sample training set is obtained by acquiring slice data of ex vivo teeth; a human tooth crack identification model is constructed, and the sample training set is input into the human tooth crack identification model to obtain a complete human tooth crack identification model; the crack prediction result of the slice data is predicted by the human tooth crack identification model, and a loss function is constructed to optimize the crack prediction result to obtain an optimized human tooth crack identification model; a learning rate is set by an SGD optimizer, and the optimized human tooth crack identification model is trained based on the learning rate; the trained human tooth crack identification model is obtained, and crack judgment is performed on the slice data of ex vivo teeth. The slice data is processed by the human tooth crack identification model, and crack judgment is performed on the slice data of ex vivo teeth, thereby improving the recognition ability and recognition accuracy of tooth cracks under the condition of limited sensitivity to tooth cracks.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a method, device and medium for recognizing human tooth cracks based on deep learning. Background Art

[0002] Early identification and management of tooth cracks have always been a key challenge in clinical practice. Clinical identification relies heavily on the knowledge and experience of clinicians and is prone to inaccuracies due to visual fatigue or misjudgment. In addition, even experienced clinicians can only identify cracks larger than 50μm, indicating that the human eye has limited sensitivity to details. Accurately identifying tiny cracks and determining their extent is crucial for treatment planning and prognosis, but it remains extremely challenging. It has long been expected to develop a well-validated tool to assist clinicians in diagnosing tooth cracks, formulating treatment plans, and assessing prognosis.

[0003] In recent years, the rapid development of artificial intelligence has attracted great attention. Although artificial intelligence has been widely used in the medical field, few studies have explored the use of image processing technology and deep learning to detect tooth cracks. This research gap may be attributed to the limited quality of clinical images, which seriously affects the performance of detection. Therefore, few related studies have turned to multi-source high-precision images, including oral microscope images and high-resolution CBCT. However, the resolution of commonly used clinical CBCT is quite low, and previous studies on tooth crack detection have not yet attempted to be applied to clinical CBCT. Summary of the invention

[0004] The technical problem to be solved by the present invention is that it is difficult to identify tooth cracks in clinical practice and the recognition accuracy is low. The purpose is to provide a method, device and medium for identifying human tooth cracks based on deep learning. By acquiring ex vivo tooth slice data and using ex vivo tooth images as a learning basis, a human tooth crack recognition model is constructed to process the slice data, and crack judgment is performed on the ex vivo tooth slice data, thereby improving the recognition ability and recognition accuracy of tooth cracks under conditions of limited sensitivity to tooth cracks.

[0005] The present invention is achieved through the following technical solutions:

[0006] The first aspect of the present invention provides a method for identifying human tooth cracks based on deep learning, comprising the following specific steps:

[0007] Acquire slice data of ex vivo teeth, and pre-process the slice data to obtain a sample training set;

[0008] Construct a human tooth crack recognition model and input the sample training set into the human tooth crack recognition model:

[0009] Perform convolution and downsampling operations on the slice data to obtain the feature map of the slice;

[0010] Perform position encoding on the feature map, flatten the feature matrix on the encoded feature map, and obtain a one-dimensional feature vector;

[0011] Perform attention calculation on the one-dimensional feature vector to obtain the feature vector;

[0012] Get the slice before the current slice, concatenate and convolve the feature vector with the data of the previous slice, and obtain a feature map after feature fusion;

[0013] The spatial memory queue is obtained by taking the previous slice as the target slice;

[0014] Perform attention calculation, feature extraction and nonlinear transformation on the feature vector, the feature map after feature fusion and the spatial memory queue to obtain the output feature vector of the current target slice;

[0015] Perform maximum pooling convolution and linear transformation on the output feature vector to obtain the crack prediction result of the slice data;

[0016] A loss function is constructed to optimize the crack prediction results, and an optimized human tooth crack recognition model is obtained;

[0017] Use the SGD optimizer to set the learning rate, and train the optimized human tooth crack recognition model based on the learning rate;

[0018] The trained human tooth crack recognition model is obtained, and crack judgment is performed on the slice data of the ex vivo tooth.

[0019] The present invention acquires ex vivo tooth slice data, uses ex vivo tooth images as a learning basis, constructs a human tooth crack recognition model to process the slice data, and performs crack judgment on the ex vivo tooth slice data, thereby improving the recognition ability and recognition accuracy of tooth cracks under conditions with limited sensitivity to tooth cracks.

[0020] Furthermore, the step of obtaining the sample training set includes:

[0021] Obtaining ex vivo tooth Micro-CT and CBCT images, slicing the images, and obtaining slice data;

[0022] Based on the crack location information provided by Micro-CT, the crack distribution information in the corresponding single tooth CBCT data is obtained and converted into labels;

[0023] The CBCT image of each tooth is integrated with the corresponding label data to obtain a sample training set.

[0024] Furthermore, the constructed human tooth crack recognition model includes:

[0025] The feature extractor includes the first four downsampling stages of the ConvNext network, where each stage includes a normalization layer, a convolutional layer, a depthwise convolutional layer, and an activation function layer;

[0026] The neighbor perception module includes multiple hybrid modules with multi-head attention layers and normalization layers;

[0027] Deep spatial guidance module, including convolutional layers, multi-head attention layers, normalization layers and fully connected layers.

[0028] Furthermore, the size of the slice is , the feature map size of the slice is , the size of the one-dimensional feature vector is , where C represents the number of output channels of the first downsampling stage in the feature extractor module, and W and H represent the width and height of the feature map, respectively.

[0029] Furthermore, the attention calculation is performed on the one-dimensional feature vector to obtain the feature vector, and the specific calculation steps include:

[0030] ;

[0031] in, represents the attention calculation of the neighbor perception module, represents the normalized exponential function, The feature map representing the target slice is flattened to obtain a one-dimensional feature vector. The feature map of the previous slice representing the target slice is flattened to obtain a one-dimensional feature vector. The feature map of the next slice representing the target slice is flattened to obtain a one-dimensional feature vector, express The transpose of is a flag used to indicate that the current one-dimensional vector belongs to the target slice. is an identifier used to indicate that the current one-dimensional vector belongs to the previous slice adjacent to the target slice. is an identifier used to indicate that the current one-dimensional vector belongs to the next slice adjacent to the target slice. Represents a one-dimensional feature vector Dimension.

[0032] Furthermore, the attention calculation of the feature vector, the feature map after feature fusion and the spatial memory queue specifically includes:

[0033] Get the feature fusion matrix size and spatial memory queue size;

[0034] The feature vector is processed by convolution and SoftMax function, and the processed feature vector is multiplied with the unprocessed feature vector itself through residual connection;

[0035] The multiplied feature vector is concatenated with the output feature vector of the current target slice in the channel dimension to obtain a feature fusion matrix with a size of ;

[0036] After convolution processing, the feature fusion matrix size is adjusted to ;

[0037] Get the size of the spatial memory queue ;

[0038] Based on the feature fusion matrix size and the spatial memory queue size, attention calculation is performed on the acquired feature vector, the feature map after feature fusion, and the spatial memory queue. The specific calculation steps include:

[0039] ;

[0040] in, represents the multi-head attention computation in the deep spatial guidance module, Represents the feature map after feature fusion, Indicates the output feature vector after attention calculation, express As the spatial memory queue obtained when slicing the target, is the sequence number of the target slice, Represents the feature vector The dimensions of It is an identifier used to indicate the feature vector obtained after attention calculation.

[0041] Furthermore, in the process of updating the spatial memory queue, when the queue reaches the optimal queue length, the farthest slice feature information in the current queue is automatically deleted from the queue head. The specific calculation steps include:

[0042] ;

[0043] in represents the updated spatial memory queue, n represents the optimal queue length, and idx represents the current target slice idx+1 represents the sequence number of the previous slice adjacent to the current target slice. Represents the spatial memory queue To update, Integrate into The new spatial memory queue obtained later, ,Right now , used to determine whether a is greater than b. If a is greater than b, execute , otherwise execute D, in It is used to determine whether idx is greater than n. If idx is greater than n, execute , otherwise execute 0, ∃ represents an existential quantifier, Indicates the target slice with serial number idx-n The output feature vector of Indicates the target slice with serial number idx-n. Indicates the target slice with the serial number idx The output feature vector of is an exponential decay weight.

[0044] Furthermore, the loss function is constructed to optimize the crack prediction result, and the specific steps include:

[0045] The loss function is constructed using CrossEntropy Loss function and SmoothL1 Loss function;

[0046] The predicted probability value of the entire slice sequence is compared with the target value of the entire slice sequence, and the predicted probability value of each tooth slice is adjusted based on the loss function. The specific calculation steps include:

[0047] ;

[0048] Where n represents the number of slices of a single tooth, It indicates the probability value predicted by the model that the i-th slice has a crack, and pred is an identifier used to indicate that the probability value is the probability value predicted by the network. Indicates the target value of the SmoothL1 Loss function for the i-th slice. Target is a flag used to indicate the target value. Represents the category number of the i-th slice, where true is a flag used to indicate the true category.

[0049] A second aspect of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a method for identifying human tooth cracks based on deep learning is implemented.

[0050] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for identifying human tooth cracks based on deep learning.

[0051] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0052] By obtaining ex vivo tooth slice data and using ex vivo tooth images as a learning basis, a human tooth crack recognition model is constructed to process the slice data, and crack judgment is performed on the ex vivo tooth slice data, thereby improving the recognition ability and accuracy of tooth cracks under conditions with limited sensitivity to tooth cracks. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without creative work. In the drawings:

[0054] Figure 1 is an operation flow chart in an embodiment of the present invention;

[0055] Figure 2 is a diagram showing the structure of a detection method in an embodiment of the present invention;

[0056] Figure 3 The figure is a comparison diagram of the output results of the embodiment of the present invention and the existing method. DETAILED DESCRIPTION

[0057] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0058] As a possible implementation, Figure 1 and 2 As shown, this embodiment provides a method for identifying human tooth cracks based on deep learning, which specifically includes:

[0059] S1. Obtain training dataset

[0060] The ex vivo tooth Micro-CT and CBCT images were obtained, the image data were sliced ​​from the horizontal plane dimension, and the sliced ​​data were output in DICOM format. Based on the crack location information provided by Micro-CT, the crack distribution information in the corresponding single tooth CBCT data was obtained and converted into labels. In the labels, slices belonging to different categories were assigned the same number. The category number of slices with tooth cracks was 1, and the category number of slices without tooth cracks was 0. The CBCT image of each tooth and the corresponding label information constituted a sample training set, which included slices with cracks and slices without cracks.

[0061] S2: Human tooth crack recognition model

[0062] The model framework includes:

[0063] The feature extractor includes the first four downsampling stages of the ConvNext network, where each stage includes a normalization layer, a convolutional layer, a depthwise convolutional layer, and an activation function layer;

[0064] The neighbor perception module includes multiple hybrid modules with multi-head attention layers and normalization layers;

[0065] Deep spatial guidance module, including convolutional layers, multi-head attention layers, normalization layers and fully connected layers.

[0066] The construction steps and iterative process of the human tooth crack recognition model are as follows:

[0067] S2-1: The slices of the whole tooth to be trained are input into the human tooth crack recognition model in the order of slices. The human tooth crack recognition model trains each slice as a target slice in turn. The feature extractor trains the target slices in turn. , the slice before the target slice and the slice after the target slice Convolution and downsampling operations are performed, where idx represents the serial number of the current target slice, and then the feature maps of the target slices are obtained respectively. , the feature map of the previous slice of the target slice And the feature map of the next slice of the target slice And serve as the input of the proximity perception module;

[0068] S2-2: The feature map obtained by the feature extractor , and Position encoding is performed separately. Represents the feature map of the target slice currently being trained, where is a flag used to indicate that the current feature map belongs to the target slice. Represents the feature map of the previous slice adjacent to the target slice, is an identifier used to indicate that the current feature map belongs to the previous slice adjacent to the target slice. Represents the feature map of the next slice adjacent to the target slice, is used to indicate that the current feature map belongs to the next slice adjacent to the target slice, and then the feature map , and Flatten the feature matrices respectively to obtain the corresponding one-dimensional feature vectors , and ,in is a flag used to indicate that the current one-dimensional vector belongs to the target slice. is an identifier used to indicate that the current one-dimensional vector belongs to the previous slice adjacent to the target slice. is an identifier used to indicate that the current one-dimensional vector belongs to the next slice adjacent to the target slice. and As the query and key of the multi-head attention module, As the value of the multi-head attention module, the feature vector is output after attention calculation on Query, Key and Value. ,in is an identifier used to indicate the feature vector obtained after attention calculation;

[0069] S2-3: Same as the previous slice Output result when used as target slice Perform concatenation and convolution operations to obtain the feature map after feature fusion ,in To indicate the and The identification of the feature map after feature fusion is As the query of the multi-head attention module in the deep spatial guidance module, the previous slice The spatial memory queue obtained when used as the target slice As the key of the multi-head attention module in the deep spatial guidance module, express As the spatial memory queue obtained when slicing the target, is the sequence number of the target slice, As the value of the multi-head attention module in the deep spatial guidance module, , and After attention calculation, feature extraction and nonlinear transformation are performed through the fully connected layer and normalization layer to obtain the output feature vector of the current target slice. ;

[0070] S2-4: Spatial Memory Queue To update, Integrate into , get a new spatial memory queue ;

[0071] S2-5: Yeah Perform maximum pooling convolution and linear transformation output detection method to slice the target The prediction results are outputted one by one, and finally each slice of the whole tooth is outputted as the prediction result of the target slice. A loss function is constructed to calculate the error between the output results of all slices of the whole tooth and the label. Back propagation is performed according to the calculated results to optimize the human tooth crack recognition model.

[0072] In some possible embodiments, in step S2-1:

[0073] Input target slice , the slice before the target slice and the slice after the target slice The dimensions are , the feature map obtained after the feature extractor module , and The sizes are , , , where C represents the number of output channels of the first downsampling stage in the feature extractor module, and W and H represent the feature maps width and length.

[0074] In some possible embodiments, in step S2-2, , and After flattening, we get one-dimensional feature vectors , and The sizes are , and ,right , and Output feature vector after attention calculation , and Restore size to , the attention calculation formula is expressed as:

[0075] ;

[0076] in Represents a one-dimensional feature vector The dimension of It is an identifier used to indicate that the current one-dimensional vector belongs to the next slice adjacent to the target slice.

[0077] In some possible embodiments, in step S2-3, After a convolution kernel The convolutional layer and SoftMax function are processed, and then connected with Multiply it by itself and then The feature fusion matrix is ​​obtained by splicing in the channel dimension. Size , and then through a convolution kernel The convolutional layer will Resize to ;

[0078] Previous slice The spatial memory queue obtained when used as the target slice The size is , the attention calculation formula in the deep space guidance module is expressed as:

[0079] ;

[0080] in Represents the feature vector The dimensions of It is an identifier used to indicate the feature vector obtained after attention calculation.

[0081] In some possible embodiments, in step S2-4, the spatial memory queue The optimal queue length set in is 20. In the process of updating the spatial memory queue, when the optimal queue length is reached, the farthest slice feature information in the current queue is automatically deleted from the queue head. The formula is as follows:

[0082] ;

[0083] in Represents the updated spatial memory queue, represents the optimal queue length, Indicates the current target slice The serial number, Indicates the sequence number of the previous slice adjacent to the current target slice. Represents the spatial memory queue To update, Integrate into The new spatial memory queue obtained later, ,Right now , used to determine whether a is greater than b. If a is greater than b, execute , otherwise execute D, in It is used to determine whether idx is greater than n. If idx is greater than n, execute , otherwise execute 0, ∃ represents an existential quantifier, Indicates the target slice with serial number idx-n The output feature vector of Indicates the target slice with serial number idx-n. Indicates the target slice with the serial number idx The output feature vector of is the exponential decay weight, set to 0.2.

[0084] In some possible embodiments, in step S2-5, the loss function is composed of a CrossEntropy Loss function and a SmoothL1 Loss function. The accuracy of the model for the overall detection of a single tooth slice is improved by adjusting the predicted probability value of each tooth slice. For slices with cracks, we set the target value in the SmoothL1 Loss function to 1, and set the target value of the SmoothL1 Loss function for slices without cracks to 0. The predicted probability value of the entire slice sequence is compared with the target value of the entire slice sequence. The Loss function calculation formula is as follows:

[0085] ;

[0086] in represents the number of slices of a single tooth, The model predicts the The probability value of a slice having a crack, is an identifier used to indicate that the probability value is the probability value predicted by the network, Indicates the setting The target value of the SmoothL1 Loss function for each slice, is a marker for indicating the target value, Indicates The category number of the slice, where is a label used to indicate the true category.

[0087] S3: Training the optimized human tooth crack recognition model

[0088] Using the sample training set constructed in step S1 and the human tooth crack recognition model in step S2, the SGD optimizer is used to set the learning rate. The learning rate decreases linearly with the number of iterations during the training process, and the maximum number of iterations is 100,000. Finally, the parameter weights of the human tooth crack recognition model are obtained. ;

[0089] S4: Testing using the trained human tooth crack recognition model;

[0090] Use the human tooth crack recognition model to identify cracks in the input tooth image, combined with parameter weights , the whole tooth CBCT image is sliced ​​from the horizontal plane dimension to obtain multiple slices, which are input into the human tooth crack recognition model, and whether each slice contains cracks is judged to obtain the recognition result.

[0091] As a possible implementation, this embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, a method for identifying human tooth cracks based on deep learning is implemented.

[0092] As a possible implementation, this embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for identifying human tooth cracks based on deep learning.

[0093] As a possible implementation, Figure 3 As shown in the figure, the detection method proposed in this application and other methods are tested using test data sets of three tooth positions: molars, premolars and incisors. The bold numbers in the figure represent the best results among the detection methods, and the underlined numbers represent the suboptimal output results of the detection methods. Compared with other detection methods, the method proposed in this application obtains excellent output results under multiple evaluation indicators and shows the best overall recognition performance.

[0094] According to a method for identifying human tooth cracks based on deep learning provided in the above embodiment, in vitro tooth slice data is obtained, and slices of the whole tooth are sequentially input into a human tooth crack recognition model in the order of slices. The human tooth crack recognition model sequentially selects each slice as a target slice for training. After the trained human tooth crack recognition model is obtained, samples that have not participated in the training are selected from the existing sample training set for testing, the output results of the human tooth crack recognition model are compared with the labels, and evaluation indicators such as accuracy, sensitivity, specificity, positive predictive value and negative predictive value are calculated to judge the recognition effect of the model.

[0095] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for identifying human tooth cracks based on deep learning, characterized in that: The specific steps include: Acquire slice data of ex vivo teeth, and pre-process the slice data to obtain a sample training set; Construct a human tooth crack recognition model and input the sample training set into the human tooth crack recognition model: Perform convolution and downsampling operations on the slice data to obtain the feature map of the slice; Perform position encoding on the feature map, flatten the feature matrix on the encoded feature map, and obtain a one-dimensional feature vector; Perform attention calculation on the one-dimensional feature vector to obtain the feature vector. The specific calculation steps include: ; in, represents the attention calculation of the neighbor perception module, represents the normalized exponential function, The feature map representing the target slice is flattened to obtain a one-dimensional feature vector. The feature map of the previous slice representing the target slice is flattened to obtain a one-dimensional feature vector. The feature map of the next slice representing the target slice is flattened to obtain a one-dimensional feature vector. express The transpose of is a flag used to indicate that the current one-dimensional vector belongs to the target slice. is an identifier used to indicate that the current one-dimensional vector belongs to the previous slice adjacent to the target slice. is an identifier used to indicate that the current one-dimensional vector belongs to the next slice adjacent to the target slice. Represents a one-dimensional feature vector Dimensions; Get the slice before the current slice, concatenate and convolve the feature vector with the data of the previous slice, and obtain a feature map after feature fusion; Take the previous slice as the target slice and get the spatial memory queue; The feature vector, the feature map after feature fusion and the spatial memory queue are subjected to attention calculation, feature extraction and nonlinear transformation to obtain the output feature vector of the current target slice, including: Get the feature fusion matrix size and spatial memory queue size; The feature vector is processed by convolution and SoftMax function, and the processed feature vector is multiplied by the unprocessed feature vector itself; The multiplied feature vector is concatenated with the output feature vector of the current target slice in the channel dimension, and the size of the feature fusion matrix obtained is ; After convolution processing, the feature fusion matrix size is adjusted to ; Get the size of the spatial memory queue ; Based on the feature fusion matrix size and the spatial memory queue size, attention calculation is performed on the acquired feature vector, the feature map after feature fusion, and the spatial memory queue. The specific calculation steps include: ； in, represents the multi-head attention computation in the deep spatial guidance module, Represents the feature map after feature fusion, Indicates the output feature vector after attention calculation, express As the spatial memory queue obtained when slicing the target, is the sequence number of the target slice, Represents the feature vector The dimensions of is an identifier used to indicate the feature vector obtained after attention calculation; Perform maximum pooling convolution and linear transformation on the output feature vector to obtain the crack prediction result of the slice data; A loss function is constructed to optimize the crack prediction results, and an optimized human tooth crack recognition model is obtained; Use the SGD optimizer to set the learning rate, and train the optimized human tooth crack recognition model based on the learning rate; The trained human tooth crack recognition model is obtained, and crack judgment is performed on the slice data of the ex vivo tooth.

2. The method for identifying human tooth cracks based on deep learning according to claim 1, characterized in that: The step of obtaining the sample training set includes: Obtaining ex vivo tooth Micro-CT and CBCT images, slicing the images, and obtaining slice data; Based on the crack location information provided by Micro-CT, the crack distribution information in the corresponding single tooth CBCT data is obtained and converted into labels; The CBCT image of each tooth is integrated with the corresponding label data to obtain a sample training set.

3. The method for identifying human tooth cracks based on deep learning according to claim 1, characterized in that: The constructed human tooth crack recognition model includes: The feature extractor includes the first four downsampling stages of the ConvNext network, where each stage includes a normalization layer, a convolutional layer, a depthwise convolutional layer, and an activation function layer; The neighbor perception module includes multiple hybrid modules containing multi-head attention layers and normalization layers; Deep spatial guidance module, including convolutional layers, multi-head attention layers, normalization layers and fully connected layers.

4. The method for identifying human tooth cracks based on deep learning according to claim 3, characterized in that: The size of the slice is , the feature map size of the slice is , the size of the one-dimensional feature vector is , where C represents the number of output channels of the first downsampling stage in the feature extractor module, and W and H represent the width and height of the feature map, respectively.

5. The method for identifying human tooth cracks based on deep learning according to claim 3, characterized in that: In the process of updating the spatial memory queue, when the queue reaches the optimal queue length, the farthest slice feature information in the current queue is automatically deleted from the queue head. The specific calculation steps include: ； in Represents the updated spatial memory queue, represents the optimal queue length, Indicates the current target slice Serial number, Indicates the sequence number of the previous slice adjacent to the current target slice. Represents the spatial memory queue To update, Integrate into The new spatial memory queue obtained later, ,Right now , used to determine whether a is greater than b. If a is greater than b, execute , otherwise execute D, in It is used to determine whether idx is greater than n. If idx is greater than n, execute , otherwise execute 0, ∃ represents an existential quantifier, Indicates the target slice with serial number idx-n The output feature vector of Indicates the target slice with serial number idx-n. is the exponential decay weight, Indicates the target slice with the serial number idx The output feature vector of .

6. The method for identifying human tooth cracks based on deep learning according to claim 1, characterized in that: The loss function is constructed to optimize the crack prediction result, and the specific steps include: The loss function is constructed using CrossEntropy Loss function and SmoothL1 Loss function; The predicted probability value of the entire slice sequence is compared with the target value of the entire slice sequence, and the predicted probability value of each tooth slice is adjusted based on the loss function. The specific calculation steps include: ； in represents the number of slices of a single tooth, The model predicts the The probability value of a slice having a crack, is an identifier used to indicate that the probability value is the probability value predicted by the network, Indicates the setting The target value of the SmoothL1 Loss function for each slice, is a marker for indicating the target value, Indicates The category number of the slice, where is a label used to indicate the true category.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method for identifying human tooth cracks based on deep learning as described in any one of claims 1 to 6 is implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method for identifying human tooth cracks based on deep learning as described in any one of claims 1 to 6 is implemented.

Images