A segmentation method for CBCT tooth crack images based on convolution kernel coupled update
By using a multi-receptive field feature extraction method with convolution kernel coupling updated in CBCT dental crack image processing, combined with the CGAM jump connection module and ADDC Stack layer, the problem of difficult segmentation at limited resolution is solved, and more accurate teeth are achieved Crack segmentation improves the efficiency of medical image processing.
Patent Information
- Application Number
- CN202411429572.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-14
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-10-14
AI Technical Summary
The prior art is difficult to effectively segment CBCT dental crack images at limited resolution, resulting in reduced dentist identification and diagnostic efficiency.
Using a multi-receptive field feature extraction method based on convolution kernel coupling update, a network framework containing multi-layer encoder and decoder layer is constructed, combining the CGAM jump connection module and the ADDC Stack layer to perform multi-scale fusion and attention on feature information. Calculate, dynamically adjust the convolution kernel expansion rate to extract large receptive field features, and conduct in-depth supervision training through cross entropy and Dice loss.
It significantly improves the segmentation ability of tooth CBCT image cracks, provides a more accurate feature extraction scheme at limited resolution, and enhances the accuracy of medical image segmentation.
Smart Images

Figure CN118939912B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer vision image processing technology, and in particular, relates to a segmentation method for CBCT tooth crack images based on convolution kernel coupling update. Background Art
[0002] Medical image segmentation plays a key role in disease diagnosis and treatment. With the introduction of Transformer in the field of computer vision, many methods combining Transformer with convolutional neural networks have emerged to solve medical image segmentation tasks. However, these methods are challenging in capturing multi-scale features, although multi-scale features contain rich texture and context information, which is crucial for segmentation tasks. In addition, in oral medical image processing, for existing clinical imaging modalities, such as cone beam computed tomography (CBCT), the limited resolution greatly increases the difficulty for dentists to identify and diagnose tooth cracks, and significantly reduces the efficiency of human eyes in identifying cracks. Summary of the invention
[0003] In order to solve the problem in the prior art that the CBCT image has limited resolution and affects the segmentation effect, the present invention proposes a segmentation method for CBCT tooth crack images based on convolution kernel coupling update, which comprises the following steps:
[0004] S1: Construct training data set: From the in vitro teeth collected clinically, evaluate whether there are cracks in the teeth through MicroCT image scanning, and select teeth with cracks to form a tooth crack data set; collect MicroCT images of teeth with cracks, annotate the cracks on the MicroCT images, and obtain crack label maps for each tooth with cracks; in the label map, pixels of objects of different categories are assigned different numbers, where the value of the pixel representing the tooth crack is 1, and the value of the background pixel is 0; then, align the crack label map of the MicroCT image with the tooth CBCT image through the 3D registration method to form a complete tooth CBCT crack data set;
[0005] S2: Build a network framework: The network framework includes multiple layers of corresponding encoder layers and decoder layers, and a jump connection module that transfers feature information between the encoder layer and the decoder layer; each encoder layer includes an ADDCStack layer and a downsampling layer, each decoder layer includes an upsampling layer, an ADDC Stack layer and a segmentation head, and the jump connection module includes a CGAM jump connection layer. The process of one iteration of network training is as follows:
[0006] S2-1: Initialize the dental CBCT image input into the network to obtain the initial feature map , size is ,in , and Represent the initial feature maps length, width and depth, Represents the initial feature map The number of channels of the initial feature map Input into the encoder, pass through four encoder layers in sequence, each encoder layer extracts features and outputs the feature map of each layer ,in Representation feature map It is through The initial feature map obtained by the encoder layer After all encoder layers are processed, the feature map output by the highest encoder is obtained. ,Will Through the double convolution decoder layer processing, the feature map is obtained ,in Representation feature map It is obtained after the fifth decoder layer;
[0007] S2-2: For the encoder layer, decoder layer and CGAM jump connection layer, the processing flow is as follows: Layer encoder layer, respectively get feature maps , for the The CGAM skip connection layer connects the Feature map of the layer encoder output As the The first input of the CGAM skip connection layer is Feature map of the layer decoder output As the The second input of the CGAM skip connection layer, where Representation feature map By The decoder layer obtains The output feature map is obtained after the CGAM skip connection layer is processed ,in Representation feature map It is through The CGAM skip connection layer is obtained for the first The decoder layer transforms the feature map As the The first input of the decoder layer is the feature map As the The second input of the decoder layer is The decoder layer calculates the output feature map ,in Representation feature map By From the fourth layer, for each encoder layer, decoder layer and CGAM jump connection layer, perform the above steps from the fourth layer to the decoder layer. Feature map of the layer encoder output After The CGAM skip connection layer and the layer decoder layer, and finally get the Feature map of the layer decoder output The processing flow is as follows, until the first layer, in the first layer of CGAM jump connection layer, the feature map output by the first layer encoder layer is converted to As the first input of the first CGAM skip connection layer, the feature map output by the second decoder layer is As the second input of the first CGAM skip connection layer, the output feature map is obtained after being processed by the first CGAM skip connection layer. , and finally and As the input of the first decoder layer, the first decoder layer calculates the initial feature map The feature maps of the same size ;
[0008] S2-3: Output feature map of each decoder layer The output is respectively passed through each layer segmentation head for deep supervision and through the feature map Obtain the segmentation result, calculate the loss function with the label map, and perform backpropagation based on the calculated result. This completes one round of iteration.
[0009] Furthermore, in step S2, each ADDC Stack layer includes two ADDC modules, the downsampling layer is composed of a maximum pooling downsampling layer, and the size of the maximum pooling convolution kernel is , with a step size of 2; the upsampling layer consists of a transposed convolution layer, and the size of the transposed convolution kernel is , with a step size of 2.
[0010] Furthermore, in step S2, the input feature map of the ADDC module in each encoder layer and each decoder layer is The size is ,in , and Represents the input feature map length, width and depth, Represents the input feature map The number of channels; for the input feature map , through the convolution kernel size Depthwise convolution Perform feature extraction and obtain the size Feature map , and then The convolution kernel Extracted, where the convolution kernel The size is .
[0011] Furthermore, the convolution kernels in the ADDC modules in each encoder layer and decoder layer are obtained. After that, the attention calculation is performed; the convolution kernel The convolution kernel parameters of the last three dimensions are expanded into a one-dimensional vector form, and the convolution kernel The channel dimension of is placed in the last dimension to obtain the convolution kernel parameters for attention calculation , size is ; Then use PyTorch's nn.Parameter object to initialize 3 and Trainable feature vectors of the same size , , ,Will , , Respectively and Perform multiplication and we get , and ; Then the attention calculation is performed, and the attention calculation formula is as follows:
[0012]
[0013] in Represents the convolution kernel after dynamic adjustment, with a size of ; d represents the convolution kernel parameter The length of the channel dimension; Softmax() represents the probability distribution Softmax function used to calculate the attention score; Express Perform a transpose operation;
[0014] Will and The corresponding positions are added, and then normalized to obtain the convolution kernels for different expansion rates. , and finally the convolution kernel The shape is restored to , the formula is as follows:
[0015]
[0016] Among them, Reshape() means reshaping the size of the convolution kernel; LayerNorm() means layer normalization.
[0017] Furthermore, we get the convolution kernel Afterwards, the convolution kernel As dilated depthwise convolution The convolution kernel is adjusted by The expansion rate is controlled Extract the receptive field of the feature; input feature map By dilating the depthwise convolution Obtain feature maps with large receptive fields , size is ; Then use shared point convolution For feature maps and feature map Calculate and concatenate the obtained feature information in the channel dimension to form multi-receptive field features after BN normalization and Relu activation function processing; finally, use the size of The convolution, BN normalization and Relu activation function operations are used to process the multi-receptive field features to obtain the feature map output by the ADDC module. , size is .
[0018] Further, in step S2-2, for the Layer CGAM jump connection layer, the first input is the feature map , size is ,in , and Represents the feature maps length, width and depth, Feature map representing the input The number of channels; the second input is the feature map , size is ; First, for the feature map , by sampling in height, width and depth with a step size of 2 to extract feature sub-blocks, from the feature map 8 spatially independent feature sub-blocks are obtained, where the elements in each sub-block are The positions in are separated, and then every two feature sub-blocks are spliced in the channel dimension, and then they are respectively passed through a shared weight size of The convolution kernel is used for convolution operation, and then processed by BN normalization and Relu activation function to obtain four low-level fusion feature maps. , express For the The low-level fusion feature map in the group, the low-level fusion feature map Size .
[0019] Furthermore, for the The CGAM skip connection layer obtains the four low-level fusion feature maps. With feature map Perform group-wise parallel attention calculations, where each group contains a and , the low-level fusion feature maps in each group After a shared weight size of The convolution operation is performed with a convolution kernel with a step size of 2, and then processed by BN normalization and Relu activation function to obtain the feature map , size is , the feature map As the Value in each group of attention calculations, the feature map As the key in each group of attention calculations, the low-level fusion feature map As the query in each group, the attention calculation is performed on each group in parallel. The formula is as follows:
[0020]
[0021] in Represents the feature map obtained after multiple group attention calculations, with a size of , , , Represents the linear change layer , and And the corresponding feature map , and The result obtained after multiplication; represents the group attention calculation, The calculation of is as follows:
[0022]
[0023] in express , Softmax() represents the probability distribution Softmax function used to calculate the attention score.
[0024] Further, in step S2-2, for the The CGAM skip connection layer obtains each feature map The concatenation is performed in the channel dimension and then sent to the MLP layer for feature fusion. The MLP layer consists of two linear change layers. The calculation formula is as follows:
[0025]
[0026] in represents a network layer consisting of two linear transformation layers, Represents the feature map obtained by calculating multiple group attention Splicing is done in the channel dimension, Represents the feature map obtained after feature fusion, with a size of ,in , and Represents the feature maps length, width and depth, Representation feature map The number of channels; then Each feature sub-block in is restored according to its original spatial position, and the first Feature map of the output of the CGAM skip connection layer , size is ,in , and Represents the feature maps length, width and depth, Representation feature map The number of channels.
[0027] Further, in step S2-2, for the The decoder layer The first input of the layer decoder Perform upsampling operation and then The second input of the layer decoder layer They are concatenated in the channel dimension and then passed into the ADDC Stack for feature fusion. Finally, the Feature map of the layer decoder output .
[0028] Furthermore, in step S2-3, the segmentation head of each layer consists of a size of The convolution kernel, BN normalization and Relu operation are composed.
[0029] In step S2-3, the final output of the network is obtained after the last segmentation head. In addition, the output of each layer of segmentation head and the label map are used to calculate the cross entropy loss and Dice loss. The calculation formula is as follows:
[0030]
[0031]
[0032] in Represents the cross entropy loss, which is used to measure the difference between the predicted results and the true results. represents a random variable, represents the index of the segmentation category, Representatives The predicted probability of the class is the predicted result, Representative The true label of the class, It is used to measure the overlap between the predicted result and the true label, where TP stands for true positive, FP stands for false positive, and FN stands for false negative. The final loss is obtained by taking a weighted sum of 20% of each layer of segmentation head.
[0033] The beneficial effect of the present invention is that, through the segmentation network design method of the present invention for multi-scale information fusion of CBCT tooth crack images based on the multi-receptive field convolution kernel coupling update mechanism, the segmentation capability of tooth CBCT image cracks in the field of medical image segmentation is effectively improved. This provides an effective feature extraction solution for intelligent medical image processing, which helps to more accurately segment the crack area under limited resolution. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.
[0035] Figure 1 It is a network framework diagram of the present invention;
[0036] Figure 2 It is a framework diagram of the ADDC module of the present invention;
[0037] Figure 3 Schematic diagram of the calculation of the convolution kernel in the ADDC module of the present invention, where the dotted box indicates the corresponding weights involved in calculating the weight of a single convolution kernel element. and part of;
[0038] Figure 4 It is a framework diagram of the CGAM module in the present invention;
[0039] Figure 5 It is a comparison diagram of some output results of the present invention and the existing method. DETAILED DESCRIPTION
[0040] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.
[0041] like Figure 1-Figure 5 As shown, the present invention proposes a segmentation method for CBCT tooth crack images based on convolution kernel coupling update, comprising the following steps:
[0042] S1: Construct training data set: From the in vitro teeth collected clinically, evaluate whether there are cracks in the teeth through MicroCT image scanning, and select teeth with cracks to form a tooth crack data set; collect MicroCT images of teeth with cracks, annotate the cracks on the MicroCT images, and obtain crack label maps for each tooth with cracks; in the label map, pixels of objects of different categories are assigned different numbers, where the value of the pixel representing the tooth crack is 1, and the value of the background pixel is 0; then, align the crack label map of the MicroCT image with the tooth CBCT image through the 3D registration method to form a complete tooth CBCT crack data set;
[0043] S2: Build a network framework: The network framework includes multiple layers of corresponding encoder layers and decoder layers, and a jump connection module that transfers feature information between the encoder layer and the decoder layer; each encoder layer includes an ADDCStack layer and a downsampling layer, each decoder layer includes an upsampling layer, an ADDC Stack layer and a segmentation head, and the jump connection module includes a CGAM jump connection layer. The process of one iteration of network training is as follows:
[0044] S2-1: Initialize the dental CBCT image input into the network to obtain the initial feature map , size is ,in , and Represent the initial feature maps length, width and depth, Represents the initial feature map The number of channels of the initial feature map Input into the encoder, pass through four encoder layers in sequence, each encoder layer extracts features and outputs the feature map of each layer ,in Representation feature map It is through The initial feature map obtained by the encoder layer After all encoder layers are processed, the feature map output by the highest encoder is obtained. ,Will Through the double convolution decoder layer processing, the feature map is obtained ,in Representation feature map It is obtained after the fifth decoder layer;
[0045] S2-2: For the encoder layer, decoder layer and CGAM jump connection layer, the processing flow is as follows: Layer encoder layer, respectively get feature maps , for the The CGAM skip connection layer connects the Feature map of the layer encoder output As the The first input of the CGAM skip connection layer is Feature map of the layer decoder output As the The second input of the CGAM skip connection layer, where Representation feature map By The decoder layer obtains The output feature map is obtained after the CGAM skip connection layer is processed ,in Representation feature map It is through The CGAM skip connection layer is obtained for the first The decoder layer transforms the feature map As the The first input of the decoder layer is the feature map As the The second input of the decoder layer is The decoder layer calculates the output feature map ,in Representation feature map By From the fourth layer, for each encoder layer, decoder layer and CGAM jump connection layer, perform the above steps from the fourth layer to the decoder layer. Feature map of the layer encoder output After The CGAM skip connection layer and the layer decoder layer, and finally get the Feature map of the layer decoder output The processing flow is as follows, until the first layer, in the first layer of CGAM jump connection layer, the feature map output by the first layer encoder layer is converted to As the first input of the first CGAM skip connection layer, the feature map output by the second decoder layer is As the second input of the first CGAM skip connection layer, the output feature map is obtained after being processed by the first CGAM skip connection layer. , and finally and As the input of the first decoder layer, the first decoder layer calculates the initial feature map The feature maps of the same size ;
[0046] S2-3: Output feature map of each decoder layer The output is respectively passed through each layer segmentation head for deep supervision and through the feature map Obtain the segmentation result, calculate the loss function with the label map, and perform backpropagation based on the calculated result. This completes one round of iteration.
[0047] Furthermore, in step S2, each ADDC Stack layer includes two ADDC modules, the downsampling layer is composed of a maximum pooling downsampling layer, and the size of the maximum pooling convolution kernel is , with a step size of 2; the upsampling layer consists of a transposed convolution layer, and the size of the transposed convolution kernel is , with a step size of 2.
[0048] Furthermore, in step S2, the input feature map of the ADDC module in each encoder layer and each decoder layer is The size is ,in , and Represents the input feature map length, width and depth, Represents the input feature map The number of channels; for the input feature map , through the convolution kernel size Depthwise convolution Perform feature extraction and obtain the size Feature map , and then The convolution kernel Extracted, where the convolution kernel The size is .
[0049] Furthermore, the convolution kernels in the ADDC modules in each encoder layer and decoder layer are obtained. After that, the attention calculation is performed; the convolution kernel The convolution kernel parameters of the last three dimensions are expanded into a one-dimensional vector form, and the convolution kernel The channel dimension of is placed in the last dimension to obtain the convolution kernel parameters for attention calculation , size is ; Then use PyTorch's nn.Parameter object to initialize 3 and Trainable feature vectors of the same size , , ,Will , , Respectively and Perform multiplication, and we get , and ; Then the attention calculation is performed, and the attention calculation formula is as follows:
[0050]
[0051] in Represents the convolution kernel after dynamic adjustment, with a size of ; d represents the convolution kernel parameter The length of the channel dimension; Softmax() represents the probability distribution Softmax function used to calculate the attention score; Express Perform a transpose operation;
[0052] Will and The corresponding positions are added, and then normalized to obtain the convolution kernels for different expansion rates. , and finally the convolution kernel The shape is restored to , the formula is as follows:
[0053]
[0054] Among them, Reshape() means reshaping the size of the convolution kernel; LayerNorm() means layer normalization.
[0055] Furthermore, we get the convolution kernel Afterwards, the convolution kernel As a dilated depthwise convolution The convolution kernel is adjusted by The expansion rate is controlled Extract the receptive field of the feature; input feature map By dilating the depthwise convolution Obtain feature maps with large receptive fields , size is ; Then use shared point convolution For feature maps and feature map Calculate and concatenate the obtained feature information in the channel dimension to form multi-receptive field features after BN normalization and Relu activation function processing; finally, use the size of The convolution, BN normalization and Relu activation function operations are used to process the multi-receptive field features to obtain the feature map output by the ADDC module. , size is .
[0056] Further, in step S2-2, for the Layer CGAM jump connection layer, the first input is the feature map , size is ,in , and Represents the feature maps length, width and depth, Feature map representing the input The number of channels; the second input is the feature map , size is ; First, for the feature map , by sampling in height, width and depth with a step size of 2 to extract feature sub-blocks, from the feature map 8 spatially independent feature sub-blocks are obtained, where the elements in each sub-block are The positions in are separated, and then every two feature sub-blocks are spliced in the channel dimension, and then they are respectively passed through a shared weight size of The convolution kernel is used for convolution operation, and then processed by BN normalization and Relu activation function to obtain four low-level fusion feature maps. , express For the The low-level fusion feature map in the group, the low-level fusion feature map Size .
[0057] Furthermore, for the The CGAM skip connection layer obtains the four low-level fusion feature maps. With feature map Perform group-wise parallel attention calculations, where each group contains a and , the low-level fusion feature maps in each group After a shared weight size of The convolution operation is performed with a convolution kernel with a step size of 2, and then processed by BN normalization and Relu activation function to obtain the feature map , size is , the feature map As the Value in each group of attention calculations, the feature map As the key in each group of attention calculations, the low-level fusion feature map As the query in each group, the attention calculation is performed on each group in parallel. The formula is as follows:
[0058]
[0059] in Represents the feature map obtained after multiple group attention calculations, with a size of , , , Represents the linear change layer , and And the corresponding feature map , and The result obtained after multiplication; represents the group attention calculation, The calculation of is as follows:
[0060]
[0061] in express , Softmax() represents the probability distribution Softmax function used to calculate the attention score.
[0062] Further, in step S2-2, for the The CGAM skip connection layer obtains each feature map The concatenation is performed in the channel dimension and then sent to the MLP layer for feature fusion. The MLP layer consists of two linear change layers. The calculation formula is as follows:
[0063]
[0064] in represents a network layer consisting of two linear transformation layers, Represents the feature map obtained by calculating multiple group attention Splicing is done in the channel dimension, Represents the feature map obtained after feature fusion, with a size of ,in , and Represents the feature maps length, width and depth, Representation feature map The number of channels; then Each feature sub-block in is restored according to its original spatial position, and the first Feature map of the output of the CGAM skip connection layer , size is ,in , and Represents the feature maps length, width and depth, Representation feature map The number of channels.
[0065] Further, in step S2-2, for the The decoder layer The first input of the layer decoder Perform upsampling operation and then The second input of the layer decoder layer They are concatenated in the channel dimension and then passed into the ADDC Stack for feature fusion. Finally, the Feature map of the layer decoder output .
[0066] Furthermore, in step S2-3, the segmentation head of each layer consists of a size of The convolution kernel, BN normalization and Relu operation are composed.
[0067] In step S2-3, the final output of the network is obtained after the last segmentation head. In addition, the output of each layer of segmentation head and the label map are used to calculate the cross entropy loss and Dice loss. The calculation formula is as follows:
[0068]
[0069]
[0070] in Represents the cross entropy loss, which is used to measure the difference between the predicted results and the true results. represents a random variable, represents the index of the segmentation category, Representatives The predicted probability of the class is the predicted result, Representative The true label of the class, It is used to measure the overlap between the predicted result and the true label, where TP stands for true positive, FP stands for false positive, and FN stands for false negative. The final loss is obtained by taking a weighted sum of 20% of each layer of segmentation head.
[0071] The segmentation network design method for multi-scale information fusion of CBCT tooth crack images based on a multi-receptive field convolution kernel coupling update mechanism of the present invention effectively improves the segmentation capability of tooth CBCT image cracks in the field of medical image segmentation. This provides an effective feature extraction solution for intelligent medical image processing, which helps to more accurately segment the crack area under limited resolution.
[0072] Example:
[0073] According to step S1, a training data set is constructed. The data is the dental CBCT image obtained by the result scanning and the manually annotated label image containing cracks. Each 3D dental CBCT image forms a pair of samples with the corresponding 3D label.
[0074] According to step S2-1, the trained 3D dental CBCT image is put into the encoder for feature extraction, and then the output result of each encoder layer is retained as the first input of each CGAM module. At the same time, the feature map obtained after multiple downsampling is passed to the decoder through double convolution.
[0075] According to steps S2-2 and S2-3, the output results of each encoder layer and the output results of the decoder one layer higher than this layer are used as the input of the CGAM jump connection layer of this layer, and the output results of the CGAM jump connection layer are passed to the decoder of this layer for feature fusion, and so on. The above process is repeated. Finally, the output result with the same resolution as the input feature map is output through the lowest decoder layer. The output result of each decoder layer is compared with the label map, the loss function is calculated, and back propagation is performed according to the loss function. At this point, a round of training is completed, and the above process is repeated to save the final network model parameters.
[0076] Test: Select sample pairs that have not participated in training for testing, compare the network output results with the labels, and calculate the Dice evaluation index to judge the segmentation effect of the model.
[0077] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A segmentation method for CBCT tooth crack images based on convolution kernel coupled update, characterized in that: The following steps are involved: S1: Construct training data set: From the in vitro teeth collected clinically, evaluate whether there are cracks in the teeth through MicroCT image scanning, and select teeth with cracks to form a tooth crack data set; collect MicroCT images of teeth with cracks, annotate the cracks on the MicroCT images, and obtain crack label maps for each tooth with cracks; in the label map, pixels of objects of different categories are assigned different numbers, where the value of the pixel representing the tooth crack is 1, and the value of the background pixel is 0; then, align the crack label map of the MicroCT image with the tooth CBCT image through the 3D registration method to form a complete tooth CBCT crack data set; S2: Build a network framework: The network framework includes multiple layers of corresponding encoder layers and decoder layers, and a jump connection module that transfers feature information between the encoder layer and the decoder layer; each encoder layer includes an ADDCStack layer and a downsampling layer, each decoder layer includes an upsampling layer, an ADDC Stack layer and a segmentation head, and the jump connection module includes a CGAM jump connection layer. The process of one iteration of network training is as follows: S2-1: Initialize the dental CBCT image input into the network to obtain the initial feature map , size is ,in , and Represent the initial feature maps length, width and depth, Represents the initial feature map The number of channels of the initial feature map Input into the encoder, pass through four encoder layers in sequence, each encoder layer extracts features and outputs the feature map of each layer ,in Representation feature map It is through The initial feature map obtained by the encoder layer After all encoder layers are processed, the feature map output by the highest encoder is obtained. ,Will Through the double convolution decoder layer processing, the feature map is obtained ,in Representation feature map It is obtained after the fifth decoder layer; S2-2: For the encoder layer, decoder layer and CGAM jump connection layer, the processing flow is as follows: Layer encoder layer, respectively get feature maps , for the The CGAM skip connection layer connects the Feature map of the encoder layer output As the The first input of the CGAM skip connection layer is Feature map of the layer decoder output As the The second input of the CGAM skip connection layer, where Representation feature map By The decoder layer obtains The output feature map is obtained after the CGAM skip connection layer is processed ,in Representation feature map It is through The CGAM skip connection layer is obtained. The decoder layer transforms the feature map As the The first input of the decoder layer is the feature map As the The second input of the decoder layer is The decoder layer calculates the output feature map ,in Representation feature map By From the fourth layer, for each encoder layer, decoder layer and CGAM jump connection layer, perform the above steps from the fourth layer to the decoder layer. Feature map of the encoder layer output After The CGAM skip connection layer and the layer decoder layer, and finally get the Feature map of the layer decoder output The processing flow is as follows, until the first layer, in the first layer of CGAM jump connection layer, the feature map output by the first layer encoder layer is converted to As the first input of the first CGAM skip connection layer, the feature map output by the second decoder layer is As the second input of the first CGAM skip connection layer, the output feature map is obtained after being processed by the first CGAM skip connection layer. , and finally and As the input of the first decoder layer, the first decoder layer calculates the initial feature map The feature maps of the same size ; S2-3: Output feature map of each decoder layer The output is respectively passed through each layer segmentation head for deep supervision and through the feature map Obtain the segmentation result, calculate the loss function with the label map, and perform backpropagation based on the calculated result. This completes one round of iteration.
2. The segmentation method for CBCT tooth crack images based on convolution kernel coupled update according to claim 1, characterized in that: In step S2, each ADDC Stack layer includes two ADDC modules, and the downsampling layer is composed of a maximum pooling downsampling layer, and the size of the maximum pooling convolution kernel is , with a step size of 2; the upsampling layer consists of a transposed convolution layer, and the size of the transposed convolution kernel is , with a step size of 2.
3. The segmentation method for CBCT tooth crack images based on convolution kernel coupled update according to claim 2, characterized in that: In step S2, the input feature map of the ADDC module in each encoder layer and each decoder layer is The size is ,in , and Represents the input feature map length, width and depth, Represents the input feature map The number of channels; for the input feature map , through the convolution kernel size Depthwise convolution Perform feature extraction and obtain the size Feature map , and then The convolution kernel Extracted, where the convolution kernel The size is .
4. The segmentation method for CBCT tooth crack images based on convolution kernel coupled update according to claim 3, characterized in that: Get the convolution kernel in the ADDC module in each encoder layer and decoder layer After that, the attention calculation is performed; the convolution kernel The convolution kernel parameters of the last three dimensions are expanded into a one-dimensional vector form, and the convolution kernel The channel dimension of is placed in the last dimension to obtain the convolution kernel parameters for attention calculation , size is ; Then use PyTorch's nn.Parameter object to initialize 3 and Trainable feature vectors of the same size , , ,Will , , Respectively and Perform multiplication and we get , and ; Then the attention calculation is performed, and the attention calculation formula is as follows: in Represents the convolution kernel after dynamic adjustment, with a size of ; d represents the convolution kernel parameter The length of the channel dimension; Softmax() represents the probability distribution Softmax function used to calculate the attention score; Express Perform a transpose operation; Will and The corresponding positions are added, and then normalized to obtain the convolution kernels for different expansion rates. , and finally the convolution kernel The shape is restored to , the formula is as follows: Among them, Reshape() means reshaping the size of the convolution kernel; LayerNorm() means layer normalization.
5. The segmentation method for CBCT tooth crack images based on convolution kernel coupled update according to claim 4, characterized in that: Get the convolution kernel Afterwards, the convolution kernel As dilated depthwise convolution The convolution kernel is adjusted by The expansion rate is controlled Extract the receptive field of the feature; input feature map By dilating the depthwise convolution Obtain feature maps with large receptive fields , size is ; Then use shared point convolution For feature maps and feature map Calculate and concatenate the obtained feature information in the channel dimension to form multi-receptive field features after BN normalization and Relu activation function processing; The final size is The convolution, BN normalization and Relu activation function operations are used to process the multi-receptive field features to obtain the feature map output by the ADDC module. , size is .
6. The segmentation method for CBCT tooth crack images based on convolution kernel coupled update according to claim 1, characterized in that: In step S2-2, for Layer CGAM jump connection layer, the first input is the feature map , size is ,in , and Represents the feature maps length, width and depth, Feature map representing the input The number of channels; the second input is the feature map , size is ; First, for the feature map , by sampling in height, width and depth with a step size of 2 to extract feature sub-blocks, from the feature map 8 spatially independent feature sub-blocks are obtained, where the elements in each sub-block are The positions in are separated, and then every two feature sub-blocks are spliced in the channel dimension, and then they are respectively passed through a shared weight size of The convolution kernel is used for convolution operation, and then processed by BN normalization and Relu activation function to obtain four low-level fusion feature maps. , express For the The low-level fusion feature map in the group, the low-level fusion feature map Size .
7. The segmentation method for CBCT tooth crack images based on convolution kernel coupled update according to claim 6, characterized in that: For The CGAM skip connection layer obtains the four low-level fusion feature maps. With feature map Perform group-wise parallel attention calculations, where each group contains a and , the low-level fusion feature maps in each group After a shared weight size of The convolution operation is performed with a convolution kernel with a step size of 2, and then processed by BN normalization and Relu activation function to obtain the feature map , size is , the feature map As the Value in each group of attention calculations, the feature map As the key in each group of attention calculations, the low-level fusion feature map As the query in each group, the attention calculation is performed on each group in parallel. The formula is as follows: in Represents the feature map obtained after multiple group attention calculations, with a size of , , , Represents the linear change layer , and And the corresponding feature map , and The result obtained after multiplication; represents the group attention calculation, The calculation of is as follows: in express , Softmax() represents the probability distribution Softmax function used to calculate the attention score.
8. The segmentation method for CBCT tooth crack images based on convolution kernel coupled update according to claim 1, characterized in that: In step S2-2, for The CGAM skip connection layer obtains each feature map The concatenation is performed in the channel dimension and then sent to the MLP layer for feature fusion. The MLP layer consists of two linear change layers. The calculation formula is as follows: in represents a network layer consisting of two linear transformation layers, Represents the feature map obtained by calculating multiple group attention Splicing is done in the channel dimension, Represents the feature map obtained after feature fusion, with a size of ,in , and Represents the feature maps length, width and depth, Representation feature map The number of channels; then Each feature sub-block in is restored according to its original spatial position, and the first Feature map of the output of the CGAM skip connection layer , size is ,in , and Represents the feature maps length, width and depth, Representation feature map The number of channels.
9. The segmentation method for CBCT tooth crack images based on convolution kernel coupled update according to claim 1, characterized in that: In step S2-2, for The decoder layer The first input of the layer decoder Perform upsampling operation and then The second input of the layer decoder layer They are concatenated in the channel dimension and then passed into the ADDC Stack for feature fusion. Finally, the Feature map of the layer decoder output .
10. The segmentation method for CBCT tooth crack images based on convolution kernel coupled update according to claim 1, characterized in that: In step S2-3, each layer segmentation head consists of a size of The convolution kernel, BN normalization and Relu operation are composed.
11. The segmentation method for CBCT tooth crack images based on convolution kernel coupled update according to claim 1, characterized in that: In step S2-3, the final output of the network is obtained after the last segmentation head. In addition, the output of each layer of segmentation head and the label map are used to calculate the cross entropy loss and Dice loss. The calculation formula is as follows: in Represents the cross entropy loss, which is used to measure the difference between the predicted results and the true results. represents a random variable, represents the index of the segmentation category, Representatives The predicted probability of the class is the predicted result, Representative The true label of the class, It is used to measure the overlap between the predicted result and the true label, where TP stands for true positive, FP stands for false positive, and FN stands for false negative. The final loss is obtained by taking a weighted sum of 20% of each layer of segmentation head.
