A 3D oral hard palate image segmentation method based on multi-directional state space model

Abstract

The present invention relates to the field of medical image segmentation, and discloses a three-dimensional oral hard palate image segmentation method based on a multi-directional state space model. The method obtains a target image dataset, including a CBCT image dataset and a corresponding delineation label image dataset; constructs a multi-directional state space model; uses the target image dataset as a training set to train the multi-directional state space model; inputs the CBCT image to be segmented into the multi-directional state space model, and outputs the segmentation result of the three-dimensional oral hard palate image. The method can enhance the model's ability to understand the three-dimensional structure, effectively capture the complex spatial dependencies of the oral hard palate region, improve the accuracy and robustness of the segmentation model, and is conducive to improving the computational efficiency and accuracy of existing segmentation methods. Introducing semantics and uncertainty in training for mixed data enhancement can improve the generalization and accuracy of the network, improve delineation efficiency, reduce the delineation burden of dentists, and provide support for achieving efficient, accurate and generalized orthodontic implant planning.

Description

Technical Field

[0001] The present invention relates to the field of medical image segmentation, and in particular to a three-dimensional oral hard palate image segmentation method based on a multi-directional state space model. Background Art

[0002] In orthodontics, the hard palate is the preferred site for orthodontic micro-implants because of its high-quality bone. Therefore, accurate segmentation of the hard palate from oral CBCT (Cone beam Computer Tomography) images is crucial for the implementation of implant surgery. However, the morphology and thickness of the hard palate vary greatly depending on individual age and gender. Therefore, the segmentation accuracy and robustness of the image segmentation model are particularly critical in segmentation execution.

[0003] Traditional methods for oral hard palate segmentation mainly include threshold-based, region growing, shape model and level set methods. The threshold-based method distinguishes the hard palate from other tissues by setting a threshold of the grayscale value, but the effect is poor under noisy and uneven lighting conditions. The region growing method starts from the seed point and gradually expands the region according to the similarity of pixels. It is sensitive to noise and prone to overgrowth. The shape model method guides segmentation by establishing a statistical shape model of the anatomical structure, which requires a large amount of training data. The level set method uses an energy function to evolve the contour line to achieve segmentation. Although it can handle topological changes, it has a high computational complexity.

[0004] The above methods have been successful to a certain extent, but they are easily affected by noise and CBCT slice resolution, have large computational losses, and have poor feature discriminability and robustness. In recent years, with the development of deep learning technology, models such as Convolutional Neural Network (CNN), U-Net and its variants, and Transformer have been widely used in medical image segmentation, and have achieved good results in normal dental cases. However, when faced with the diversity of hard palate anatomical structures and noise in CBCT images, mis-segmentation and missed segmentation still occur. Summary of the invention

[0005] In order to solve the technical problems existing in the prior art, the present invention aims to provide an oral hard palate segmentation method with high segmentation accuracy and high computational efficiency, which can automatically output accurate oral hard palate region results.

[0006] Specifically, the present invention provides a three-dimensional oral hard palate image segmentation method based on a multi-directional state space model, and the technical solution includes the following steps:

[0007] A three-dimensional oral hard palate image segmentation method based on a multi-directional state space model comprises the following steps:

[0008] Step S1: construct a multi-directional state space model;

[0009] Step S2: using the target image dataset as a training set to train the multi-directional state space model; the target image dataset includes a CBCT image dataset and a corresponding delineation label image dataset;

[0010] Step S3: inputting the CBCT image to be segmented into the multi-directional state space model, and outputting the segmentation result of the three-dimensional oral hard palate image;

[0011] The multi-directional state space model includes a multi-directional encoder and a multi-directional decoder;

[0012] Both the multi-directional encoder and the multi-directional decoder include Layer Vision Mamba Layer, =1,2,3…;

[0013] The Vision Mamba layer includes a block embedding module and multiple Mamba encoders;

[0014] The block embedding module includes a linear mapping layer and a normalization layer;

[0015] The Mamba encoder includes the Vision Mamba block and the feed-forward module;

[0016] Among them, both the Vision Mamba block and the feed-forward module include skip connections;

[0017] Step S2 includes:

[0018] For the input original image pair ( , ), using the common mixing method to construct the mixed image And the corresponding mixed label , the formula is as follows:

[0019] ;

[0020] In the formula, is the original image pair ( , ) in the original image and the original image Mixing ratio; , ) is the original image pair ( , ) corresponding label pairs;

[0021] The original image pair is encoded by a multi-directional encoder ( , ) and mixed images , and obtain the feature vector pair containing the image semantic information ( , ) and mixed eigenvectors ;

[0022] Calculate new mixing ratio , the formula is as follows:

[0023] ;

[0024] In the formula, is the Softmax function used for vector normalization, represents the L2 norm;

[0025] Calculating uncertainty , the formula is as follows:

[0026] ;

[0027] In the formula, represents a multi-layer perceptron;

[0028] Reconstruct the feature vector, the formula is as follows:

[0029] ;

[0030] In the formula, is the reconstructed eigenvector, is the mean pixel value of the CBCT image.

[0031] Preferably, before step S2, the method further includes:

[0032] All CBCT images were resized to a uniform size;

[0033] For each CBCT image, a standardization operation is performed, and the formula is as follows:

[0034] ;

[0035] In the formula, is the pixel value of the CBCT image, is the standardized pixel value of the CBCT image, is the mean pixel value of the CBCT image, is the standard deviation of the pixel values ​​of the CBCT image.

[0036] Preferably, through the loss function Assists the convergence of multi-directional state space models, the formula is as follows:

[0037] ;

[0038] In the formula, , and is the weighted adjustment loss function Hyperparameters of The loss function for mixed data enhancement of semantic and uncertainty information, is the binary cross entropy loss function, is the Dice loss function, and the formulas are as follows:

[0039] ;

[0040] ;

[0041] ;

[0042] In the formula, To predict the mixed image Original image The probability of To predict the mixed image Original image The probability of 0≦ , ≦1; The number of pixels to outline the label image; For the The true value of the pixel label. pixels are in the hard palate area, then =1, otherwise 0; To predict the The probability that a pixel belongs to the hard palate area, 0≦ ≦1.

[0043] Preferably, step S3 specifically includes:

[0044] The CBCT image to be segmented X 0 input multi-directional encoder, through The Vision Mamba layer is iteratively processed to obtain the output image X n ;in, is the layer number of the Vision Mamba layer, is the number of Vision Mamba layers, =1,2,…, ;

[0045] The output image X n Enter the multi-directional decoder through The Vision Mamba layer performs iterative processing and outputs an image X 2n , that is, the oral hard palate segmentation result;

[0046] The image processing through the Vision Mamba layer specifically includes the following steps:

[0047] Step T1: The CBCT image to be segmented X 0 Input the block embedding module to embed and obtain the first embedding block;

[0048] Step T2: Input the first embedding block into the Vision Mamba block of the Mamba encoder, and perform reshaping in three directions: depth, height, and width to obtain the corresponding pixel blocks. ;

[0049] Step T3: For each pixel block , pixel position embedding and normalization are performed in turn, the formula is as follows:

[0050] ;

[0051] In the formula, is the normalized pixel block, represents block embedding, represents normalization, It means element-by-element addition;

[0052] Step T4: normalized pixel blocks , extract features through convolution and state space model, and get the corresponding output feature map. The formula is as follows:

[0053] ;

[0054] In the formula, is the output feature map, represents the state space model, represents convolution, represents the activation function, Represents element-wise multiplication;

[0055] Step T5: Restore the size of the output feature map in three directions: depth, height and width through shaping;

[0056] Step T6: superimposing the output feature maps in the three directions of depth, height and width, and performing normalization and feed-forward processing in sequence to generate an output image, i.e., the segmentation result of the three-dimensional oral hard palate image;

[0057] In the multi-directional encoder, +1 layer of Vision Mamba layer to be processed is the first Output image of the Vision Mamba layer X i ; The image to be processed in the first Vision Mamba layer is a CBCT image X 0;

[0058] In the multi-directional decoder, -1 Vision Mamba layer to be processed image is Output image of the Vision Mamba layer X 2n﹣i＋1 and the multi-directional encoder - Output image of 1st Vision Mamba layer X i－1 The superimposed image of The image to be processed in the Vision Mamba layer is the output image. X n ;

[0059] Among them, the layer numbers of the Vision Mamba layer in the multi-directional decoder are arranged in reverse order.

[0060] Preferably, the size of the first embedded block is × × × , the sizes of the pixel blocks corresponding to the depth, height and width directions are × × , × × and × × ;

[0061] in, is the depth of the image to be processed, is the height of the image to be processed, is the width of the image to be processed. is the number of channels of the block embedded module, is the block embedding factor of the block embedding module.

[0062] Preferably, =4.

[0063] Compared with the prior art, the beneficial effects of the present invention are as follows: the technical solution provided by the present invention introduces a multi-directional state space model to extract the features of the input image in different directions, fully utilizes information from different perspectives, enhances the model's ability to understand the three-dimensional structure, and can effectively capture the complex spatial dependencies of the oral hard palate region, improve the accuracy and robustness of the segmentation model, and solve the technical problems of low computational efficiency and poor precision of the existing three-dimensional oral hard palate image segmentation methods. Furthermore, in training, hybrid data enhancement combined with semantic and uncertainty information can solve the problem of shortage of three-dimensional oral hard palate marker data to a certain extent, while improving the generalization and precision of the network; considering uncertainty data enhancement can realize automatic delineation of the oral hard palate region, improve delineation efficiency, reduce the delineation burden of dentists, and provide technical support for the later realization of efficient, accurate and generalized orthodontic implant plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is an overall flow chart of the three-dimensional oral hard palate image segmentation method in the present invention.

[0065] Figure 2 It is a schematic diagram of the multi-directional state space model in the present invention.

[0066] Figure 3 This is a structural diagram of the Vision Mamba layer in the present invention.

[0067] Figure 4 This is a structural diagram of the Vision Mamba block in the present invention.

[0068] Figure 5 This is the result diagram of case segmentation in the present invention. DETAILED DESCRIPTION

[0069] Hereinafter, the technical solution provided by the present invention will be further elaborated in detail with reference to the accompanying drawings.

[0070] In recent years, the Mamba model has achieved great success in the field of language modeling, and many studies have migrated the model from the language field to the visual field. Compared with CNN, U-Net and Transformer, the Mamba model based on the state space model (SSM) shows obvious advantages in processing long sequences, and can still achieve linear complexity without sacrificing the global receptive field. By combining frequency domain features and spatiotemporal selective scanning strategies, the Mamba model can play a greater role in medical image segmentation. Introducing the Mamba model in the three-dimensional oral hard palate segmentation method can effectively combine the frequency domain features and spatial domain features of the image, thereby improving the accuracy and robustness of feature extraction, especially when processing high-noise and low-contrast medical images, it can better achieve accurate segmentation of the hard palate. To this end, the present invention proposes a three-dimensional oral hard palate image segmentation method based on a multi-directional state space model. By using block embedding and multi-directional Mamba blocks to process image sequences, a more accurate and robust segmentation effect of the hard palate area is achieved, and the computational efficiency can be significantly improved, memory usage can be reduced, and the efficiency of dentists' clinical operations can be greatly improved.

[0071] Among them, SSM is usually used to deal with continuous-time systems and is defined as follows:

[0072] ;

[0073] In the formula, is the independent variable of time, for The input sequence of time, for The output sequence at time, for The hidden state of the moment, for The rate of change of the hidden state at time is the state transfer matrix, is the input projection matrix, is the output projection matrix, is a direct transfer matrix.

[0074] In order to enable SSM to handle discrete time systems, the zero-order hold (ZOH) technique is usually used to convert the state transfer matrix and the input projection matrix Discretized, expressed as follows:

[0075] ;

[0076] In the formula, and represent the discretized state transfer matrix and input projection matrix respectively, is the natural base, is the time step, is the identity matrix, Represents the state transfer matrix The inverse matrix of .

[0077] The discretized state transfer matrix and the input projection matrix , the mapping relationship is rewritten as follows:

[0078] ;

[0079] In the formula, and They are The input sequence and output sequence at time, and They are Moment and -The hidden state at time 1, and are the input sequence and hidden state at the initial moment respectively.

[0080] Firstly, oral and hard palate CBCT images were acquired, and after data preprocessing, the images were divided into training set, validation set and test set. Secondly, the training set was input into the constructed multi-directional state-space model, and a data enhancement method that mixed semantic and uncertainty information was added during training. Finally, the prediction effect of the model was quantitatively evaluated on the test set, and the tissue segmentation results were predicted.

[0081] like Figure 1 As shown, the present invention provides a three-dimensional oral hard palate segmentation method based on a multi-directional state space model, and the main steps include: data acquisition, model construction, model training and model testing.

[0082] 1. Data acquisition, including the following steps:

[0083] Step (a): Collect a dataset containing oral CBCT images and oral hard palate labels drawn by doctors.

[0084] Step (b): preprocessing the target image dataset, wherein the target image dataset includes the oral hard palate CBCT image and its delineated label image dataset. The preprocessing includes:

[0085] Adjust the size. Different CBCT images have different sizes. To facilitate subsequent training and testing, all target image datasets need to be resized to 256×256 or other appropriate sizes.

[0086] Standardization. Different CBCTs may have different configurations. In order to eliminate the differences, the oral hard palate CBCT images are standardized using the following formula:

[0087] ;

[0088] In the formula, is the pixel value of the CBCT image, is the standardized pixel value of the CBCT image, is the mean pixel value of the CBCT image, is the standard deviation of the pixel values ​​of the CBCT image.

[0089] Step (c): Dataset division: divide the target image dataset into training set, validation set and test set in a ratio of 8:1:1.

[0090] 2. Model construction. Construct a multi-directional state space model, including a multi-directional encoder and a multi-directional decoder. The main steps include:

[0091] Step (a): Construct a multi-directional encoder and a multi-directional decoder, such as Figure 2 As shown in the figure, each of them contains four VisionMamba layers, and the encoding result and decoding result of each layer are added element by element to form a skip connection; the multi-directional encoder gradually upsamples the three-dimensional oral hard palate CBCT image to obtain a high-dimensional feature map, and the multi-directional decoder gradually downsamples the high-dimensional feature map to obtain the hard palate delineation result;

[0092] Step (b): Build the Vision Mamba layer, such as Figure 3 As shown, each layer consists of a block embedding module and several Mamba encoders to downsample or upsample the input data according to the encoding or decoding process;

[0093] Step (c): Construct a block embedding module, which consists of a linear mapping layer and a normalization layer. The Vision Mamba layer embeds the block into a module with a size of × × × The input data is block embedded to obtain a size of × × × The embedding block of , and Represent the depth, height and width of the input image respectively, and Respectively represent the number of channels and block embedding factor of the previous layer of data, and Respectively represent the number of channels and block embedding factor of the layer data;

[0094] Step (d): Construct a Mamba encoder, which consists of a Vision Mamba block and a feed-forward module. Skip connections are added to both modules to form a residual network structure.

[0095] Step (e): Build the Vision Mamba blocks, such as Figure 4 As shown, first set the size to × × × Output image Reshape in depth, height and width directions respectively × × , × × and × × A block of pixels:

[0096] ;

[0097] in, represents a pixel block in each direction, Represents a shaping operation.

[0098] In order to retain the position information of the original pixel block, the pixel position of the pixel block in each direction is embedded and then normalized:

[0099] ;

[0100] in, is the normalized pixel block, and denote block embedding and normalization, respectively, Represents element-by-element addition.

[0101] The normalized pixel blocks in each direction are convolved and the state space model (SSM) is used to extract features of different dimensions. This feature can be used as the weight coefficient of each pixel, which is multiplied element by element with the normalized pixel block processed by the activation function:

[0102] ;

[0103] in, is the output feature map, , and denote the state space model, convolution and activation function respectively, Represents element-wise multiplication.

[0104] Finally, the feature map in each direction is reshaped back × × × The size of the image is obtained by adding the elements one by one. :

[0105] ;

[0106] Among them, the number of dimensions =3, which means feature extraction from three directions: depth, height and width.

[0107] 3. Model training. Train the multi-directional state space model, input the preprocessed target image data set into the multi-directional state space model, and add a hybrid data enhancement method that combines semantic and uncertainty information during model training. The main steps include:

[0108] Step (a): Use the common hybrid method to construct a hybrid dataset. , ) is the input original image pair, constructing the mixed image And the corresponding mixed label :

[0109] ;

[0110] in, is the original image pair ( , ) in the original image and the original image Mixing ratio, sampled from Beta distribution; ( , ) is the original image pair ( , ) corresponding to the label pair.

[0111] Step (b): Update the mixing ratio. Encode the original image pair ( , ) and mixed images , and obtain the feature vector pair containing the image semantic information ( , ) and mixed eigenvectors ; The formula is as follows:

[0112] ;

[0113] in, is the Softmax function used for vector normalization, represents the L2 norm. When the mixed image and the original image When the semantic information difference between them is large, it indicates that the image is mixed. With less from the original image The characteristic information of tends to 0; on the contrary, when the difference in semantic information between the two is small, the new mixing ratio Tends to 1.

[0114] Step (c): Estimate uncertainty. Using the eigenvector pair ( , ) and mixed eigenvectors , calculate the uncertainty , the formula is as follows:

[0115] ;

[0116] in, Represents a multi-layer perceptron.

[0117] Step (d): Reconstruct the feature vector. After capturing the semantic and uncertainty information of the image, reconstruct the feature vector. The formula is as follows:

[0118] ;

[0119] When there is significant uncertainty or the semantic information is very different, the reconstructed feature vector will receive a smaller gradient; conversely, when there is almost no uncertainty and the semantic information is only slightly different, the reconstructed feature vector will receive a gradient close to 1.

[0120] 4. Model testing. During the multi-directional state space model training process, use As a loss function to help the model converge, it is defined as a loss function for mixed data enhancement of semantic and uncertainty information , binary cross entropy loss function And Dice loss function The sum of , its formula is defined as:

[0121] ;

[0122] in, , and is a hyperparameter used to weight the loss function.

[0123] The loss function for mixed data enhancement of semantic and uncertainty information is defined as:

[0124] ;

[0125] The binary cross entropy loss function is defined as:

[0126] ;

[0127] The Dice loss function is defined as:

[0128] ;

[0129] in, To predict the mixed image Original image The probability of To predict the mixed image Original image The probability of 0≦ , ≦1; The number of pixels to outline the label image; For the The true value of the pixel label, The pixel corresponding to the label is 1 if it is in the hard palate area, otherwise it is 0. The model predicts the probability that the pixel belongs to the hard palate area.

[0130] Loss Function Feedback supervised model prediction probability , and Converges to the true value. As the model is trained, and Converge to their corresponding mixing ratios and 1- , Gradually approaching the true value .

[0131] 5. After the training is completed, the trained model is used to predict the test set, and four indicators, namely Dice score (Dice), average symmetric surface distance (ASSD), Hausdorff distance (HD) and uncertainty error overlap (UEO), are used for quantitative evaluation. The Dice score is defined as:

[0132] ;

[0133] The average symmetric surface distance ASSD is defined as:

[0134] ;

[0135] The Hausdorff distance HD is defined as:

[0136] ;

[0137] The uncertainty-error overlap rate UEO is defined as:

[0138] ;

[0139] in, represents the set of true outline label images, The set of delineated labeled images representing the predictions of the multidirectional state-space model, represents the surface voxel, To truly outline the surface voxels in the labeled image set, delineate surface voxels in the labeled image set predicted by the multi-directional state-space model; To truly draw the label image, Outline label images predicted by multi-directional state-space models; represents the uncertainty of the prediction results of the multi-directional state space model, Indicates a segmentation fault; Indicate point With point set The minimum Euclidean distance of each point in is as follows:

[0140] ;

[0141] In the formula, Point Set The point in .

[0142] After testing, the average values ​​of Dice score Dice, average symmetric surface distance ASSD, Hausdorff distance HD and uncertainty-error overlap rate UEO of the present invention on 5 test sets are 87.81%, 1.336mm, 9.404mm and 88.49% respectively. The example segmentation results of the three-dimensional oral hard palate segmentation method with multi-directional state space model considering uncertainty data enhancement are shown in the figure. Figure 5 shown.

[0143] In summary, the technical solution provided by the present invention introduces a multi-directional state space model to extract features of the input image in different directions, fully utilizes information from different perspectives, enhances the model's ability to understand three-dimensional structures, and can effectively capture the complex spatial dependencies of the oral hard palate area, improve the accuracy and robustness of the segmentation model, and solve the technical problems of low computational efficiency and poor precision of existing three-dimensional oral hard palate image segmentation methods.

[0144] Furthermore, introducing semantics and uncertainty in training for mixed data enhancement can solve the problem of shortage of three-dimensional oral hard palate landmark data to a certain extent, while improving the generalization and accuracy of the network; considering uncertainty data enhancement can realize automatic outlining of the oral hard palate area, improve outlining efficiency, reduce the outlining burden on dentists, and provide technical support for the later realization of efficient, accurate and generalized orthodontic implant plans.

Claims

1. A three-dimensional oral hard palate image segmentation method based on a multi-directional state space model, characterized in that: The following steps are involved: Step S1: construct a multi-directional state space model; Step S2: using the target image dataset as a training set to train the multi-directional state space model; the target image dataset includes a CBCT image dataset and a corresponding delineation label image dataset; Step S3: inputting the CBCT image to be segmented into the multi-directional state space model, and outputting the segmentation result of the three-dimensional oral hard palate image; The multi-directional state space model includes a multi-directional encoder and a multi-directional decoder; Both the multi-directional encoder and the multi-directional decoder include Layer Vision Mamba Layer, =1,2,3…; The Vision Mamba layer includes a block embedding module and multiple Mamba encoders; The block embedding module includes a linear mapping layer and a normalization layer; The Mamba encoder includes the Vision Mamba block and the feed-forward module; Among them, both the Vision Mamba block and the feed-forward module include skip connections; Step S2 includes: For the input original image pair ( , ), using the common mixing method to construct the mixed image And the corresponding mixed label , the formula is as follows: ； In the formula, is the original image pair ( , ) in the original image and the original image Mixing ratio; , ) is the original image pair ( , ) corresponding label pairs; The original image pair is encoded by a multi-directional encoder ( , ) and mixed images , and obtain the feature vector pair containing the image semantic information ( , ) and mixed eigenvectors ; Calculate new mixing ratio , the formula is as follows: ； In the formula, is the Softmax function used for vector normalization, represents the L2 norm; Calculating uncertainty , the formula is as follows: ； In the formula, represents a multi-layer perceptron; Reconstruct the feature vector, the formula is as follows: ； In the formula, is the reconstructed eigenvector, is the mean pixel value of the CBCT image.

2. A three-dimensional oral hard palate image segmentation method based on a multi-directional state space model as claimed in claim 1, characterized in that: Before step S2, the method further includes: All CBCT images were resized to a uniform size; For each CBCT image, a standardization operation is performed, and the formula is as follows: ； In the formula, is the pixel value of the CBCT image, is the standardized pixel value of the CBCT image, is the mean pixel value of the CBCT image, is the standard deviation of the pixel values ​​of the CBCT image.

3. The three-dimensional oral hard palate image segmentation method based on a multi-directional state space model as claimed in claim 1, characterized in that: It also includes the loss function Assists the convergence of multi-directional state space models, the formula is as follows: ； In the formula, , and is the weighted adjustment loss function Hyperparameters of The loss function for mixed data enhancement of semantic and uncertainty information, is the binary cross entropy loss function, is the Dice loss function, and the formulas are as follows: ； ； ； In the formula, To predict the mixed image Original image The probability of To predict the mixed image Original image The probability of 0≦ , ≦1; The number of pixels to outline the label image; For the The true value of the pixel label. pixels are in the hard palate area, then =1, otherwise 0; To predict the The probability that a pixel belongs to the hard palate area, 0≦ ≦1.

4. The three-dimensional oral hard palate image segmentation method based on a multi-directional state space model as claimed in claim 1, characterized in that: Step S3 specifically includes: The CBCT image to be segmented X 0 input multi-directional encoder, through The Vision Mamba layer is iteratively processed to obtain the output image X n ;in, is the layer number of the Vision Mamba layer, is the number of Vision Mamba layers, =1,2,…, ; The output image X n Enter the multi-directional decoder through The Vision Mamba layer performs iterative processing and outputs an image X 2n , that is, the oral hard palate segmentation result; The image processing through the Vision Mamba layer specifically includes the following steps: Step T1: The CBCT image to be segmented X 0 Input the block embedding module to embed and obtain the first embedding block; Step T2: Input the first embedding block into the Vision Mamba block of the Mamba encoder, and perform reshaping in three directions: depth, height, and width to obtain the corresponding pixel blocks. ; Step T3: For each pixel block , pixel position embedding and normalization are performed in turn, the formula is as follows: ； In the formula, is the normalized pixel block, represents block embedding, represents normalization, It means element-by-element addition; Step T4: normalize the pixel block , extract features through convolution and state space model, and get the corresponding output feature map. The formula is as follows: ； In the formula, is the output feature map, represents the state space model, represents convolution, represents the activation function, Represents element-wise multiplication; Step T5: Restore the size of the output feature map in three directions: depth, height and width through shaping; Step T6: superimposing the output feature maps in the three directions of depth, height and width, and performing normalization and feed-forward processing in sequence to generate an output image, i.e., the segmentation result of the three-dimensional oral hard palate image; In the multi-directional encoder, +1 layer of Vision Mamba layer to be processed is the first Output image of the Vision Mamba layer X i ; The image to be processed in the first Vision Mamba layer is a CBCT image X 0; In the multi-directional decoder, -1 Vision Mamba layer to be processed image is Output image of the Vision Mamba layer X 2n﹣i＋1 and the multi-directional encoder - Output image of 1st Vision Mamba layer X i－1 The superimposed image of The image to be processed in the Vision Mamba layer is the output image. X n ; Among them, the layer numbers of the Vision Mamba layer in the multi-directional decoder are arranged in reverse order.

5. A three-dimensional oral hard palate image segmentation method based on a multi-directional state space model as claimed in claim 4, characterized in that: The size of the first embedded block is × × × , the sizes of the pixel blocks corresponding to the depth, height and width directions are × × , × × and × × ; in, is the depth of the image to be processed, is the height of the image to be processed, is the width of the image to be processed. is the number of channels of the block embedded module, is the block embedding factor of the block embedding module.

6. A three-dimensional oral hard palate image segmentation method based on a multi-directional state space model as claimed in claim 4, characterized in that: ＝4。

Images