An image segmentation-based epileptogenic lesion localization method and related device

Abstract

The application discloses a kind of based on image segmentation's epileptogenic focus positioning method and related equipment, the method includes: obtaining target MR sequence image, target MR sequence image is preprocessed, and target data is obtained;FCD epileptogenic focus segmentation model is constructed, the FCD epileptogenic focus segmentation model is trained and tested, and target model is obtained, wherein, target model includes encoder module, double self-attention module and decoder module;Local convolution feature is extracted by inputting target data into the encoder module, and convolution feature is obtained, convolution feature is input into the double self-attention module space, and the fusion of convolution feature and attention is obtained, and output feature is obtained, and output feature is input decoder module and is up-sampled and feature fusion, and the FCD epileptogenic focus segmented from target MR sequence image is obtained.The FCD epileptogenic focus segmentation model proposed in the application has strong feature expression ability and global perception ability, and can realize accurate positioning of epileptogenic focus.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method, system, terminal, and computer-readable storage medium for locating epileptogenic foci based on image segmentation. Background Technology

[0002] Focal cortical dysplasia (FCD) is a common cause of drug-resistant epilepsy. Currently, the most effective treatment is neurosurgical removal of the epileptogenic focus in the brain; therefore, the preoperative localization of the epileptogenic focus directly affects the success of the surgery. Precise localization of the epileptogenic focus using magnetic resonance imaging (MRI) is a commonly used method. Compared to electroencephalography (EEG) and SEEG, MRI offers faster data acquisition, higher spatial resolution, and is non-invasive.

[0003] Precise localization of the epileptogenic focus is crucial for the treatment and prognosis of drug-resistant epilepsy. With advancements in imaging and computer technologies, research on epileptogenic focus localization is increasing, simultaneously driving progress in the diagnosis and treatment of drug-resistant epilepsy.

[0004] However, current methods for locating epileptogenic foci primarily rely on comparing and analyzing low-level features from single-modal or multi-modal sources, such as cortical thickness, sulcus depth, gray-white matter edge contrast, and gray-white matter brightness. These features are susceptible to noise and local deformation, leading to misjudgments, and feature computation is extremely time-consuming. Furthermore, they fail to fully exploit the image features of multi-modal data and lack effective multi-modal feature fusion methods. While deep learning-based epileptogenic focus segmentation and localization methods have been explored, most are currently based on conventional convolutional network models. Due to the limited receptive field of convolutional network models, their feature representation capabilities are insufficient.

[0005] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0006] The main objective of this invention is to provide a method, system, terminal, and computer-readable storage medium for epileptogenic focus localization based on image segmentation. This invention aims to solve the problems in existing technologies where epileptogenic focus localization is limited by the size of the model's receptive field, insufficient model feature expression ability, limited global perception ability, and inaccurate epileptogenic focus localization.

[0007] To achieve the above objectives, the present invention provides a method for localizing epileptogenic foci based on image segmentation, the method comprising the following steps:

[0008] Acquire target MR sequence images, and preprocess the target MR sequence images to obtain target data;

[0009] An FCD epileptogenic focus segmentation model is constructed, and the FCD epileptogenic focus segmentation model is trained and tested to obtain a target model, wherein the target model includes an encoder module, a dual self-attention module, and a decoder module.

[0010] The target data is input into the encoder module for local convolution feature extraction to obtain convolution features. The convolution features are then input into the dual self-attention module space for fusion of convolution features and attention to obtain output features. The output features are then input into the decoder module for upsampling and feature fusion to obtain the FCD epileptogenic focus segmented from the target MR sequence image.

[0011] Optionally, the image segmentation-based method for locating epileptogenic foci, wherein acquiring the target MR sequence image and preprocessing the target MR sequence image to obtain target data specifically includes:

[0012] Acquire the target MR sequence images of the target FCD patient, the target MR sequence images including T1-weighted imaging and FLAIR values;

[0013] The T1-weighted imaging and the FLAIR values ​​are aligned using the FSL tool to obtain aligned data. The aligned data is then standardized using the percentiles of the image grayscale distribution to obtain the target data.

[0014] Optionally, the image segmentation-based epileptogenic focus localization method, wherein training and testing the constructed FCD epileptogenic focus segmentation model to obtain the target model specifically includes:

[0015] MR sequence images and gold standard mask images of epileptogenic foci were acquired from multiple FCD patients. The MR sequence images were preprocessed to obtain sample data.

[0016] The gold standard mask image of the epileptogenic focus is a binary image, which includes the epileptogenic focus region and the background region in the MR sequence image.

[0017] A dataset is obtained based on the sample data and the corresponding epileptogenic focus region and background region images of the sample data. The dataset is then divided into a training set, a validation set, and a test set according to a preset ratio.

[0018] The FCD epileptogenic focus segmentation model is trained using the training set, its performance is validated and its strategy is adjusted using the validation set, and its performance is evaluated using the test set, thus obtaining a target model that meets the requirements.

[0019] Optionally, in the image segmentation-based epileptogenic focus localization method, the encoder module includes six convolutional layers with a kernel size of 3 and a downsampling MaxPooling layer set between every two convolutional layers.

[0020] The step of inputting the target data into the encoder module for local convolutional feature extraction to obtain convolutional features specifically includes:

[0021] The image patch of the target data is input into the initial convolutional layer of the feature channel through the image channel for local convolutional feature extraction. The size of the image patch is... The initial convolutional features are obtained, and then the initial convolutional features are downsampled through the downsampling MaxPooling layer to obtain the downsampled features.

[0022] The downsampled feature f i Local convolutional feature extraction and downsampling operations were performed to obtain six local convolutional features at different scales;

[0023] Where R represents the spatial dimension, c0 represents the number of channels in the input image patch, D, H, and W represent the dimensions of the image patch in the z, y, and x axes, respectively, i represents the ordinal number of the convolutional layer, and c i This represents the number of feature channels in the i-th convolutional layer.

[0024] Optionally, in the image segmentation-based epileptogenic focus localization method, the dual self-attention module includes a reshape unit, a linear mapping layer, and a convolutional layer; the dual self-attention module is provided after the third, fourth, fifth, and sixth convolutional layers.

[0025] A Conv-BN convolutional module is further provided between the encoder module and the dual self-attention module. Before inputting the convolutional features into the space of the dual self-attention module to fuse the convolutional features with attention to obtain the output features, the following steps are also included:

[0026] The four local convolutional features from the 3rd, 4th, 5th, and 6th convolutional layers are input into the Conv-BN convolutional module for feature standardization, resulting in standard convolutional features for each convolutional layer.

[0027] The kernel size of the Conv-BN convolution module is 1.

[0028] Optionally, the method for locating epileptogenic foci based on image segmentation, wherein the step of inputting the convolutional features into the dual self-attention module space to fuse the convolutional features with attention to obtain the output features specifically includes:

[0029] Each of the standard convolutional features f is respectively i The data is input into the reshape unit and flattened to obtain the corresponding two-dimensional sequence data. Where, n i The number of points representing spatial locations, n i =hdw, where h, d, and w represent the standard convolutional feature f, respectively. i 'Dimensions in the z, y, and x axes;'

[0030] For each of the two-dimensional sequence data g i Generate a position embedding vector The embedding vector e i With the corresponding two-dimensional sequence data g i Adding them together yields four feature data points after location embedding;

[0031] Each feature data is input into four linear mapping layers to obtain a self-attention query matrix Q, a self-attention key matrix K, and a spatial self-attention matrix V for each feature data. s Channel self-attention matrix V c The self-attention key matrix K and the spatial self-attention matrix V are... s After dimensionality reduction using a linear mapping layer, we obtain the dimensionality-reduced self-attention key-value matrix. and the spatial self-attention matrix after dimensionality reduction Where p is much smaller than n i The number of spatial location points;

[0032] Based on the self-attention query matrix Q and the dimensionality-reduced self-attention key matrix and the spatial self-attention matrix after dimensionality reduction Calculate the spatial self-attention g for each of the aforementioned feature data. is :

[0033] According to the channel self-attention matrix V c The self-attention key matrix K and the self-attention query matrix Q are used to calculate the channel self-attention g for each of the feature data. ic :

[0034]

[0035] Where Softmax(·) is the normalization exponential function, d is the scaling factor, and T is the transpose operation;

[0036] The spatial self-attention g is and the channel self-attention g ic The result obtained by addition is connected to the standard convolutional feature f via residual concatenation. i 'The features are fused to obtain each of the standard convolutional features f'. i 'First fusion feature z': z i '=f i '+(g is +g ic );

[0037] Each of the first fused features is input into two convolutional models with convolution sizes of 3 and 1, respectively, and the convolution results are residually concatenated with the first fused features to obtain the output feature z of each standard convolutional feature after passing through the dual self-attention module. i :

[0038] z i =z i '+Conv1(Conv3(z i ')),z i =z3, z4, z5, z6;

[0039] Wherein, Conv3(·) represents a convolutional model with a convolutional size of 3, and Conv1(·) represents a convolutional model with a convolutional size of 1.

[0040] Optionally, in the image segmentation-based epileptogenic focus localization method, the decoder module includes an upsampling Deconv2 layer, a convolutional layer, and a feature merging unit.

[0041] The step of upsampling and feature fusion of the output features input to the decoder module to obtain the FCD epileptogenic focus segmented from the target MR sequence image specifically includes:

[0042] The output feature z6 is used as the initial low-resolution feature. The initial low-resolution feature is input into the upsampling Deconv2 layer for upsampling. The initial low-resolution feature is expanded by 2 times through deconvolution or bilinear interpolation to obtain the upsampled feature.

[0043] The number of feature channels in the feature merging unit is transformed into the number of feature channels in the convolutional layer above the convolutional layer corresponding to the output feature z6. The upsampled feature and the output feature corresponding to the previous convolutional layer are merged through the feature merging unit to obtain the merged feature. The merged feature is then fused through the convolutional layer to obtain the second fused feature.

[0044] The second fused feature is used as a low-resolution feature to continue upsampling and feature fusion operations until the output feature and the first two features in the local convolution feature are all fused to obtain the restored image.

[0045] The recovered image is input into a convolutional algorithm with a kernel size of 1 and a channel number of 2 to calculate the target feature map. The target feature map is then normalized using the Softmax function to obtain the FCD epileptogenic focus segmented from the target MR sequence image.

[0046] Furthermore, to achieve the above objectives, the present invention also provides an image segmentation-based epileptogenic focus localization system, wherein the image segmentation-based epileptogenic focus localization system includes:

[0047] The data acquisition and processing module is used to acquire target MR sequence images, preprocess the target MR sequence images, and obtain target data.

[0048] The model building and training module is used to build an FCD epileptogenic focus segmentation model, train and test the FCD epileptogenic focus segmentation model to obtain a target model, wherein the target model includes an encoder module, a dual self-attention module and a decoder module.

[0049] The FCD epileptogenic focus segmentation module is used to input the target data into the encoder module for local convolutional feature extraction to obtain convolutional features, input the convolutional features into the dual self-attention module space for fusion of convolutional features and attention to obtain output features, and input the output features into the decoder module for upsampling and feature fusion to obtain the FCD epileptogenic focus segmented from the target MR sequence image.

[0050] Furthermore, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and an image segmentation-based epileptogenic focus localization program stored in the memory and executable on the processor, wherein when the image segmentation-based epileptogenic focus localization program is executed by the processor, it implements the steps of the image segmentation-based epileptogenic focus localization method as described above.

[0051] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores an image segmentation-based epileptogenic focus localization program, which, when executed by a processor, implements the steps of the image segmentation-based epileptogenic focus localization method as described above.

[0052] In this invention, target MR sequence images are acquired and preprocessed to obtain target data. An FCD epileptogenic focus segmentation model is constructed, trained, and tested to obtain a target model. The target model includes an encoder module, a dual self-attention module, and a decoder module. The target data is input into the encoder module for local convolutional feature extraction to obtain convolutional features. These convolutional features are then input into the dual self-attention module space for fusion of convolutional features and attention, yielding output features. The output features are then input into the decoder module for upsampling and feature fusion to obtain the FCD epileptogenic focus segmented from the target MR sequence image. The proposed FCD epileptogenic focus segmentation model possesses strong feature representation and global perception capabilities, enabling precise localization of epileptogenic foci. Attached Figure Description

[0053] Figure 1 This is a flowchart of a preferred embodiment of the epileptogenic focus localization method based on image segmentation of the present invention;

[0054] Figure 2 This is an architecture diagram of the FCD epileptogenic focus segmentation model in the image segmentation-based epileptogenic focus localization method of the present invention;

[0055] Figure 3 This is a comparison diagram of the epileptogenic focus segmentation effect in the epileptogenic focus localization method based on image segmentation of the present invention.

[0056] Figure 4 This is a schematic diagram illustrating the principle of a preferred embodiment of the epileptogenic focus localization system based on image segmentation of the present invention.

[0057] Figure 5 This is a schematic diagram of the operating environment of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0058] This application provides a method and related equipment for localizing epileptogenic foci based on image segmentation. To make the objectives, technical solutions, and effects of this application clearer and more explicit, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining this application and are not intended to limit this application.

[0059] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0060] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0061] The preferred embodiment of the epileptogenic focus localization method based on image segmentation of the present invention, such as... Figure 1 As shown, the image segmentation-based method for locating epileptogenic foci includes the following steps:

[0062] Step S100: Acquire the target MR sequence image, and preprocess the target MR sequence image to obtain target data.

[0063] Specifically, the target MR sequence images of the target FCD patient are acquired, including T1-weighted imaging and FLAIR values ​​(fluid attenuated inversion recovery).

[0064] FSL (FMRIB's Software Library) is one of the mainstream software programs for processing diffuse brain imaging data. This invention uses the FSL tool to align the T1-weighted imaging data and the FLAIR values ​​to obtain aligned data. The aligned data is then standardized using the percentiles of the image grayscale distribution to obtain the target data.

[0065] Understandably, this invention first uses the FSL toolkit to align intra-individual data, specifically aligning FLAIR and T1 data. Then, the aligned data is registered with the MNI-152 standard brain atlas to achieve inter-individual data alignment. Next, the aligned T1-weighted imaging and FLAIR values ​​are standardized to the [0.5%, 99.5%] range to [0, 1] using the image grayscale distribution percentiles, thereby obtaining the target data. This invention preprocesses the data to ensure consistency. Data typically comes from different sources and may exist in different formats and structures; data preprocessing can give the data a consistent format, making it easier to understand and analyze.

[0066] Step S200: Construct an FCD epileptogenic focus segmentation model, train and test the FCD epileptogenic focus segmentation model to obtain a target model, wherein the target model includes an encoder module, a dual self-attention module and a decoder module.

[0067] Specifically, an FCD epileptogenic focus segmentation model is constructed. The structure of the FCD epileptogenic focus segmentation model is an encoder-decoder structure, including an encoder module, a dual self-attention module, and a decoder module.

[0068] The encoder is a 6-layer convolutional network model used to extract local convolutional features at different scales. Each convolutional layer includes a standard Conv-BN-ReLU convolutional module with a kernel size of 3. The initial number of output channels in the convolutional layers is c1 = 16, and the number of output channels doubles for each subsequent convolutional layer, i.e., c1 = 16. i =2 i-1 • c1,i=1,2,…,6. A downsampling MaxPooling layer is inserted between the two convolutional layers to reduce the feature resolution.

[0069] The dual self-attention module includes a reshape unit, a linear mapping layer, and a convolutional layer; the dual self-attention module is placed after the 3rd, 4th, 5th, and 6th convolutional layers; the decoder module includes an upsampling Deconv2 layer, a convolutional layer, and a feature merging unit. A Conv-BN convolutional module with a kernel size of 1 is also placed between the encoder module and the dual self-attention module to perform feature normalization.

[0070] The process of training and testing the FCD epileptogenic focus segmentation model to obtain the target model specifically includes:

[0071] MR sequence images and gold standard mask images of epileptogenic foci were acquired from multiple FCD patients. The MR sequence images were preprocessed to obtain sample data. The gold standard mask image of the epileptogenic foci was a binary image, comprising the epileptogenic focus region and the background region from the MR sequence images.

[0072] It is understood that the gold standard mask image of the epileptogenic focus is a binary image, corresponding to the region of the epileptogenic focus in the original image, and has the same size as the T1 / FLAIR. If it is an epileptogenic focus region, the label value is 1, while the background region is labeled with a value of 0.

[0073] The process of acquiring MR sequence images from multiple FCD patients and preprocessing the MR sequence images is the same as the process of preprocessing the target MR sequence image in step S100, and will not be repeated here. The MR sequence images from multiple FCD patients will be used as sample data for model training. Preprocessing them can improve the quality of the sample data. Since the original data often contains missing values, outliers, duplicate values, etc., this may lead to inaccurate analysis results. Preprocessing can help identify and handle these problems, thereby improving data quality.

[0074] Further, a dataset is obtained based on the sample data and the corresponding epileptogenic focus area and background area images, and the dataset is divided into a training set, a validation set and a test set according to a preset ratio.

[0075] For example, the dataset can be divided into a training set, a validation set, and a test set in a 7:1:2 ratio. The training set is used for model training, the validation set is used to verify the performance of the model during training and to adjust the training strategy accordingly, and the test set is used to evaluate the performance of the best-performing trained model.

[0076] Furthermore, the FCD epileptogenic focus segmentation model is trained using the training set, the performance of the trained FCD epileptogenic focus segmentation model is validated and the strategy is adjusted using the validation set, and the performance of the trained FCD epileptogenic focus segmentation model is evaluated using the test set to obtain a target model that meets the requirements.

[0077] Specifically, the training set and the validation set are used for model training and validation. The training set is used to update the model parameters, and the validation set is used to verify the model performance during training and save the weights of the best-performing model. During training, each time a positive sample image patch of size [D, H, W] and its corresponding labeled image patch, and a negative sample image patch of size [D, H, W] and its corresponding labeled image patch are randomly sampled from a preprocessed 2-channel image. Here, the central voxel of the positive sample image patch is labeled as the epileptogenic focus (label value 1), and the central voxel of the negative sample image patch is labeled as the background (label value 0).

[0078] To further expand the dataset and increase the diversity of samples, data augmentation was performed on the sampled image patches. The main augmentation methods included: randomly reflecting the x, y, and z axes of the image patches and marker patches with a certain probability (e.g., 0.5); rotating the image within the angle range of [-30, 30] with a certain probability; randomly shifting the gray values ​​of the image patches; and randomly adding Gaussian noise (e.g., with a mean of 0 and a standard deviation of 0.1).

[0079] Organize the augmented image patches into a batch of image data b∈R B×2×D×H×W and labeled batch data s∈R B ×2×D×H×W Where R represents the spatial dimension, B represents the number of image patches, and D, H, and W represent the dimensions of the image patches along the z, y, and x axes, respectively; the image patch size is a cube with D = H = W = 128. The gold standard image is one-hot transformed into 2-channel image patches, meaning that s has 2 channels. The image patches are then input into the model for inference to obtain the prediction result p∈R. B×2×D×H×W (The cube map size of the predicted result p is the same as the input, with C=2 channels, representing the background probability map and the epileptogenic focus probability map, respectively). Finally, the Dice loss function and the CE (CrossEntropy) loss function are used as loss functions to calculate the difference between the predicted segmentation and the gold standard segmentation (Loss), and gradient backpropagation is performed to update the model parameters:

[0080]

[0081] Where j represents the number of categories (output channels), which is 2 categories in this model, i.e., j=0 is the background and j=1 is the lesion; i represents the voxel location; p ij s represents the predicted value of the j-th channel of the i-th voxel in the predicted image p; ij This represents the label value of the j-th channel of the i-th voxel in the labeled batch data s.

[0082] Furthermore, the model parameters are updated by iterating through all images in the training set once. This process is defined as one epoch. Training is performed for 1000 epochs. After every 50 epochs, the model segmentation performance is evaluated using the validation set. Specifically, the validation set images are input into the model to predict the segmented image p. After the argmax operation, the segmented image q = argmax(p) is obtained. Then, the Dice coefficient is used to evaluate the predicted segmentation.

[0083]

[0084] Calculate the average predicted Dice for all data in the validation set and compare it with the previous best Dice. If the current average predicted Dice is better than the previous model performance, save the current model weights as the optimal model.

[0085] Furthermore, after the optimal model is obtained by training the FCD epileptogenic focus segmentation model, the performance of the trained FCD epileptogenic focus segmentation model is evaluated using the test set until the model performance meets the preset requirements, thus obtaining the target model that meets the requirements.

[0086] Step S300: Input the target data into the encoder module to extract local convolutional features and obtain convolutional features. Input the convolutional features into the dual self-attention module space to fuse the convolutional features and attention to obtain output features. Input the output features into the decoder module for upsampling and feature fusion to obtain the FCD epileptogenic focus segmented from the target MR sequence image.

[0087] like Figure 2 As shown, it can be understood that the encoder module includes six convolutional layers with a kernel size of 3 and a downsampling MaxPooling layer placed between every two convolutional layers. The process of inputting the target data into the encoder module for local convolutional feature extraction to obtain convolutional features specifically includes:

[0088] The image patch of the target data is input into the initial convolutional layer of the feature channel through the image channel for local convolutional feature extraction. The size of the image patch is... The initial convolutional features are obtained, and then the initial convolutional features are downsampled through the downsampling MaxPooling layer to obtain the downsampled features.

[0089] The downsampled feature f iLocal convolutional feature extraction and downsampling operations are performed to obtain six local convolutional features at different scales; where R represents the spatial dimension, c0 represents the number of channels in the input image patch (which can be one or more; in this invention, two image channels are involved, namely T1 and FLAIR, so c0 = 2), D, H, and W represent the dimensions of the image patch in the z, y, and x axes, respectively, i represents the ordinal number of the convolutional layer, and c i c represents the number of feature channels in the i-th convolutional layer. In other words, the number of output channels doubles for each subsequent convolutional layer. i =2 i-1 ·c1,i=1,2,…,6.

[0090] Furthermore, in order to further extract global features at different scales, the dual self-attention module is set after the 3rd, 4th, 5th and 6th convolutional layers; the dual self-attention module includes a reshape unit, a linear mapping layer and a convolutional layer.

[0091] A Conv-BN convolutional module is further provided between the encoder module and the dual self-attention module. Before inputting the convolutional features into the space of the dual self-attention module to fuse the convolutional features with attention to obtain the output features, the following steps are also included:

[0092] The four local convolutional features from the 3rd, 4th, 5th, and 6th convolutional layers are input into the Conv-BN convolutional module for feature standardization, resulting in standard convolutional features for each convolutional layer. The kernel size of the Conv-BN convolution module is 1.

[0093] Furthermore, the step of inputting the convolutional features into the dual self-attention module space to fuse the convolutional features and attention to obtain the output features specifically includes:

[0094] Each of the standard convolutional features f is respectively i The data is input into the reshape unit and flattened to obtain the corresponding two-dimensional sequence data. Where, n i The number of points representing spatial locations, n i =hdw, where h, d, and w represent the standard convolutional feature f, respectively. i 'Dimensions in the z, y, and x axes.'

[0095] For each of the two-dimensional sequence data g i Generate a position embedding vector The embedding vector e iWith the corresponding two-dimensional sequence data g i Add them together to get the four feature data after location embedding.

[0096] Each feature data is input into four linear mapping layers to obtain a self-attention query matrix Q, a self-attention key matrix K, and a spatial self-attention matrix V for each feature data. s Channel self-attention matrix V c The self-attention key matrix K and the spatial self-attention matrix V are... s After dimensionality reduction using a linear mapping layer, we obtain the dimensionality-reduced self-attention key-value matrix. and the spatial self-attention matrix after dimensionality reduction Where p is much smaller than n i The number of spatial location points.

[0097] Based on the self-attention query matrix Q and the dimensionality-reduced self-attention key matrix and the spatial self-attention matrix after dimensionality reduction Calculate the spatial self-attention g for each of the aforementioned feature data. is :

[0098] According to the channel self-attention matrix V c The self-attention key matrix K and the self-attention query matrix Q are used to calculate the channel self-attention g for each of the feature data. ic :

[0099]

[0100] Where Softmax(·) is the normalization exponential function, d is the scaling factor, and T is the transpose operation.

[0101] The spatial self-attention g is and the channel self-attention g ic The result obtained by addition is connected to the standard convolutional feature f via residual concatenation. i 'The features are fused to obtain each of the standard convolutional features f'. i 'First fusion feature z': z i '=f i '+(g is +g ic ).

[0102] Each of the first fused features is input into two convolutional models with convolution sizes of 3 and 1, respectively, and the convolution results are residually concatenated with the first fused features to obtain the output feature z of each standard convolutional feature after passing through the dual self-attention module. i:

[0103] z i =z i '+Conv1(Conv3(z i ')),z i =z3, z4, z5, z6;

[0104] Wherein, Conv3(·) represents a convolutional model with a convolutional size of 3, and Conv1(·) represents a convolutional model with a convolutional size of 1.

[0105] As can be seen, the present invention constructs a dual self-attention module, which performs global feature modeling through spatial self-attention and channel self-attention modules respectively, and adopts a Q-K sharing mechanism and a K-V dimensionality reduction method to reduce computational complexity and number of parameters and improve computational efficiency.

[0106] Furthermore, the decoder module includes an upsampling Deconv2 layer, a convolutional layer, and a feature merging unit; the step of upsampling and feature fusion of the output features input to the decoder module to obtain the FCD epileptogenic focus segmented from the target MR sequence image specifically includes:

[0107] The output feature z6 is used as the initial low-resolution feature. The initial low-resolution feature is input into the upsampling Deconv2 layer for upsampling. The initial low-resolution feature is expanded by 2 times through deconvolution or bilinear interpolation to obtain the upsampled feature.

[0108] The feature channel number of the feature merging unit is transformed into the feature channel number of the convolutional layer above the convolutional layer corresponding to the output feature z6. The upsampled feature and the output feature corresponding to the previous convolutional layer are merged through the feature merging unit to obtain the merged feature. The merged feature is then fused through the convolutional layer to obtain the second fused feature.

[0109] The second fused feature is used as a low-resolution feature to continue upsampling and feature fusion operations until the output feature and the first two features in the local convolution feature are all fused together to obtain the restored image.

[0110] The recovered image is input into a convolutional algorithm with a kernel size of 1 and a channel number of 2 to calculate the target feature map. The target feature map is then normalized using the Softmax function to obtain the FCD epileptogenic focus segmented from the target MR sequence image.

[0111] The feasibility of this invention was verified through experimental simulation: Data from 85 patients with focal coronary syndrome (FCD) was used. Each patient's data included T1 and FLAIR sequence data, as well as gold-standard segmentation images of the epileptogenic focus annotated by two radiology experts. The dataset was divided into training, validation, and test sets. After training, the DICE coefficient was used to evaluate the performance on the test set data, and the model was compared with UNETR, 3DAttentionUNet, and 3DResUNet. The comparison showed that the model constructed in this invention had a DICE coefficient of 0.416 ± 0.257, which was superior to the other three models (DICE coefficients were 0.279 ± 0.296, 0.370 ± 0.272, and 0.380 ± 0.316, respectively). This demonstrates more accurate segmentation and localization of the epileptogenic focus.

[0112] like Figure 3 As shown, Figure 3 For a comparison of segmentation results on typical data: The first row shows T1, FLAIR, the fusion of the segmented image of this invention (red) and the gold standard segmentation (yellow), and the 3D model of the gold standard segmentation; the second row shows the 3D models of the segmented image of this invention, the 3D models of the segmented image of 3DResUNet, the 3D models of the segmented image of 3DAttentionUNet, and the 3D models of the segmented image of UNETR. From a 3D perspective, the segmentation result of this invention is closer to the gold standard segmentation.

[0113] As can be seen, this invention provides a method for localizing epileptogenic foci based on image segmentation. It constructs a multi-scale self-attention feature fusion model. First, multi-scale convolutional features are extracted through a multi-layer convolutional network. Then, global feature modeling is performed using a dual self-attention mechanism of channel and space. Finally, the multi-scale features are fused through a decoder network and applied to epileptogenic focus segmentation. Compared with the convolutional network model, this method increases the size of the feature receptive field, enhances the feature representation capability, and is beneficial to improving the segmentation performance of the model, thereby achieving accurate localization of epileptogenic foci.

[0114] Furthermore, such as Figure 4 As shown, based on the above-described image segmentation-based epileptogenic focus localization method, the present invention also provides an image segmentation-based epileptogenic focus localization system, wherein the image segmentation-based epileptogenic focus localization system includes:

[0115] The data acquisition and processing module 51 is used to acquire target MR sequence images and preprocess the target MR sequence images to obtain target data.

[0116] The model building and training module 52 is used to build an FCD epileptogenic focus segmentation model, train and test the FCD epileptogenic focus segmentation model to obtain a target model, wherein the target model includes an encoder module, a dual self-attention module and a decoder module.

[0117] The FCD epileptogenic focus segmentation module 53 is used to input the target data into the encoder module for local convolution feature extraction to obtain convolution features, input the convolution features into the dual self-attention module space for fusion of convolution features and attention to obtain output features, and input the output features into the decoder module for upsampling and feature fusion to obtain the FCD epileptogenic focus segmented from the target MR sequence image.

[0118] Furthermore, such as Figure 5 As shown, based on the above-mentioned image segmentation-based epileptogenic focus localization method and system, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 5 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0119] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card. Further, the memory 20 may include both internal and external storage units. The memory 20 is used to store application software and various types of data installed on the terminal, such as program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores an image segmentation-based epileptogenic focus localization program 40, which can be executed by the processor 10 to implement the image segmentation-based epileptogenic focus localization method of this application.

[0120] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the image segmentation-based epileptogenic focus localization method.

[0121] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The components 10-30 of the terminal communicate with each other via a system bus.

[0122] In one embodiment, when the processor 10 executes the image segmentation-based epileptogenic focus localization program 40 in the memory 20, the following steps are performed:

[0123] Acquire target MR sequence images, and preprocess the target MR sequence images to obtain target data;

[0124] An FCD epileptogenic focus segmentation model is constructed, and the FCD epileptogenic focus segmentation model is trained and tested to obtain a target model, wherein the target model includes an encoder module, a dual self-attention module, and a decoder module.

[0125] The target data is input into the encoder module for local convolution feature extraction to obtain convolution features. The convolution features are then input into the dual self-attention module space for fusion of convolution features and attention to obtain output features. The output features are then input into the decoder module for upsampling and feature fusion to obtain the FCD epileptogenic focus segmented from the target MR sequence image.

[0126] The step of acquiring the target MR sequence image and preprocessing the target MR sequence image to obtain target data specifically includes:

[0127] Acquire the target MR sequence images of the target FCD patient, the target MR sequence images including T1-weighted imaging and FLAIR values;

[0128] The T1-weighted imaging and the FLAIR values ​​are aligned using the FSL tool to obtain aligned data. The aligned data is then standardized using the percentiles of the image grayscale distribution to obtain the target data.

[0129] The step of training and testing the FCD epileptogenic focus segmentation model to obtain the target model specifically includes:

[0130] MR sequence images and gold standard mask images of epileptogenic foci were acquired from multiple FCD patients. The MR sequence images were preprocessed to obtain sample data.

[0131] The gold standard mask image of the epileptogenic focus is a binary image, which includes the epileptogenic focus region and the background region in the MR sequence image.

[0132] A dataset is obtained based on the sample data and the corresponding epileptogenic focus region and background region images of the sample data. The dataset is then divided into a training set, a validation set, and a test set according to a preset ratio.

[0133] The FCD epileptogenic focus segmentation model is trained using the training set, its performance is validated and its strategy is adjusted using the validation set, and its performance is evaluated using the test set, thus obtaining a target model that meets the requirements.

[0134] The encoder module includes six convolutional layers with a kernel size of 3 and a downsampling MaxPooling layer set between every two convolutional layers.

[0135] The step of inputting the target data into the encoder module for local convolutional feature extraction to obtain convolutional features specifically includes:

[0136] The image patch of the target data is input into the initial convolutional layer of the feature channel through the image channel for local convolutional feature extraction. The size of the image patch is... The initial convolutional features are obtained, and then the initial convolutional features are downsampled through the downsampling MaxPooling layer to obtain the downsampled features.

[0137] The downsampled feature f i Local convolutional feature extraction and downsampling operations were performed to obtain six local convolutional features at different scales;

[0138] Where R represents the spatial dimension, c0 represents the number of channels in the input image patch, D, H, and W represent the dimensions of the image patch in the z, y, and x axes, respectively, i represents the ordinal number of the convolutional layer, and c i This represents the number of feature channels in the i-th convolutional layer.

[0139] The dual self-attention module includes a reshape unit, a linear mapping layer, and a convolutional layer; the dual self-attention module is set after the 3rd, 4th, 5th, and 6th convolutional layers;

[0140] A Conv-BN convolutional module is further provided between the encoder module and the dual self-attention module. Before inputting the convolutional features into the space of the dual self-attention module to fuse the convolutional features with attention to obtain the output features, the following steps are also included:

[0141] The four local convolutional features from the 3rd, 4th, 5th, and 6th convolutional layers are input into the Conv-BN convolutional module for feature standardization, resulting in standard convolutional features for each convolutional layer.

[0142] The kernel size of the Conv-BN convolution module is 1.

[0143] Specifically, the step of inputting the convolutional features into the dual self-attention module space to fuse the convolutional features and attention to obtain the output features includes:

[0144] Each of the standard convolutional features f is respectively i The data is input into the reshape unit and flattened to obtain the corresponding two-dimensional sequence data. Where, n i The number of points representing spatial locations, n i =hdw, where h, d, and w represent the standard convolutional feature f, respectively. i 'Dimensions in the z, y, and x axes;'

[0145] For each of the two-dimensional sequence data g i Generate a position embedding vector The embedding vector e i With the corresponding two-dimensional sequence data g i Adding them together yields four feature data points after location embedding;

[0146] Each feature data is input into four linear mapping layers to obtain a self-attention query matrix Q, a self-attention key matrix K, and a spatial self-attention matrix V for each feature data. s Channel self-attention matrix V c The self-attention key matrix K and the spatial self-attention matrix V are... s After dimensionality reduction using a linear mapping layer, we obtain the dimensionality-reduced self-attention key-value matrix. and the spatial self-attention matrix after dimensionality reduction Where p is much smaller than n i The number of spatial location points;

[0147] Based on the self-attention query matrix Q and the dimensionality-reduced self-attention key matrix and the spatial self-attention matrix after dimensionality reduction Calculate the spatial self-attention g for each of the aforementioned feature data. is :

[0148] According to the channel self-attention matrix V c The self-attention key matrix K and the self-attention query matrix Q are used to calculate the channel self-attention g for each of the feature data. ic :

[0149]

[0150] Where Softmax(·) is the normalization exponential function, d is the scaling factor, and T is the transpose operation;

[0151] The spatial self-attention g is and the channel self-attention g ic The result obtained by addition is connected to the standard convolutional feature f via residual concatenation. i 'The features are fused to obtain each of the standard convolutional features f'. i 'First fusion feature z': z i '=f i '+(g is +g ic );

[0152] Each of the first fused features is input into two convolutional models with convolution sizes of 3 and 1, respectively, and the convolution results are residually concatenated with the first fused features to obtain the output feature z of each standard convolutional feature after passing through the dual self-attention module. i :

[0153] z i =z i '+Conv1(Conv3(z i ')),z i =z3, z4, z5, z6;

[0154] Wherein, Conv3(·) represents a convolutional model with a convolutional size of 3, and Conv1(·) represents a convolutional model with a convolutional size of 1.

[0155] The decoder module includes an upsampling Deconv2 layer, a convolutional layer, and a feature merging unit.

[0156] The step of upsampling and feature fusion of the output features input to the decoder module to obtain the FCD epileptogenic focus segmented from the target MR sequence image specifically includes:

[0157] The output feature z6 is used as the initial low-resolution feature. The initial low-resolution feature is input into the upsampling Deconv2 layer for upsampling. The initial low-resolution feature is expanded by 2 times through deconvolution or bilinear interpolation to obtain the upsampled feature.

[0158] The number of feature channels in the feature merging unit is transformed into the number of feature channels in the convolutional layer above the convolutional layer corresponding to the output feature z6. The upsampled feature and the output feature corresponding to the previous convolutional layer are merged through the feature merging unit to obtain the merged feature. The merged feature is then fused through the convolutional layer to obtain the second fused feature.

[0159] The second fused feature is used as a low-resolution feature to continue upsampling and feature fusion operations until the output feature and the first two features in the local convolution feature are all fused to obtain the restored image.

[0160] The recovered image is input into a convolutional algorithm with a kernel size of 1 and a channel number of 2 to calculate the target feature map. The target feature map is then normalized using the Softmax function to obtain the FCD epileptogenic focus segmented from the target MR sequence image.

[0161] In summary, this invention provides a method and related equipment for epileptogenic focus localization based on image segmentation. The method includes: acquiring a target MR sequence image; preprocessing the target MR sequence image to obtain target data; constructing an FCD epileptogenic focus segmentation model; training and testing the FCD epileptogenic focus segmentation model to obtain a target model, wherein the target model includes an encoder module, a dual self-attention module, and a decoder module; inputting the target data into the encoder module for local convolutional feature extraction to obtain convolutional features; inputting the convolutional features into the space of the dual self-attention module for fusion of convolutional features and attention to obtain output features; and inputting the output features into the decoder module for upsampling and feature fusion to obtain the FCD epileptogenic focus segmented from the target MR sequence image. The FCD epileptogenic focus segmentation model proposed in this invention has strong feature representation and global perception capabilities, improves the feature receptive field size, and can achieve accurate localization of the epileptogenic focus.

[0162] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0163] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0164] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for locating epileptogenic foci based on image segmentation, characterized in that, The image segmentation-based method for locating epileptogenic foci includes: Acquire target MR sequence images, and preprocess the target MR sequence images to obtain target data; An FCD epileptogenic focus segmentation model is constructed, and the FCD epileptogenic focus segmentation model is trained and tested to obtain a target model, wherein the target model includes an encoder module, a dual self-attention module, and a decoder module. The target data is input into the encoder module for local convolution feature extraction to obtain convolution features. The convolution features are then input into the dual self-attention module space for fusion of convolution features and attention to obtain output features. The output features are then input into the decoder module for upsampling and feature fusion to obtain the FCD epileptogenic focus segmented from the target MR sequence image. The encoder module includes six convolutional layers with a kernel size of 3 and a downsampling MaxPooling layer set between every two convolutional layers; The step of inputting the target data into the encoder module for local convolutional feature extraction to obtain convolutional features specifically includes: The image patch of the target data is input into the initial convolutional layer of the feature channel through the image channel for local convolutional feature extraction. The size of the image patch is... The initial convolutional features are obtained, and then the initial convolutional features are downsampled through the MaxPooling layer to obtain the downsampled features. ; The downsampled features Local convolutional feature extraction and downsampling operations were performed to obtain six local convolutional features at different scales; in, Represents the dimension of space. This indicates the number of channels in the input image block. , , These represent the dimensions of the image patch along the z, y, and x axes, respectively. Indicates the ordinal number of the convolutional layer. Indicates the first The number of feature channels in each convolutional layer; The dual self-attention module includes a reshape unit, a linear mapping layer, and a convolutional layer; the dual self-attention module is set after the 3rd, 4th, 5th, and 6th convolutional layers; A Conv-BN convolutional module is further provided between the encoder module and the dual self-attention module. Before inputting the convolutional features into the space of the dual self-attention module to fuse the convolutional features with the attention to obtain the output features, the following steps are also included: The four local convolutional features from the 3rd, 4th, 5th, and 6th convolutional layers are input into the Conv-BN convolutional module for feature standardization, resulting in standard convolutional features for each convolutional layer. ; The kernel size of the Conv-BN convolution module is 1. The step of inputting the convolutional features into the dual self-attention module space to fuse the convolutional features and attention to obtain the output features specifically includes: Each of the standard convolutional features is respectively The data is input into the reshape unit for flattening, yielding the corresponding two-dimensional sequence data. ,in, The number of points representing spatial locations. , These respectively represent the standard convolutional features. Dimensions in the z, y, and x axes; For each of the two-dimensional sequence data Generate a position embedding vector , the embedding vector With the corresponding two-dimensional sequence data Adding them together yields four feature data points after location embedding; Each feature data is input into four linear mapping layers to obtain the self-attention query matrix Q, the self-attention key matrix K, and the spatial self-attention matrix for each feature data. Channel self-attention matrix The self-attention key matrix K and the spatial self-attention matrix are... After dimensionality reduction using a linear mapping layer, we obtain the dimensionality-reduced self-attention key-value matrix. Spatial self-attention matrix ,in, much smaller ; Based on the self-attention query matrix Q and the dimensionality-reduced self-attention key matrix Inter-attention matrix Calculate the spatial self-attention for each of the aforementioned feature data. : ; According to the channel self-attention matrix The self-attention key matrix K and the self-attention query matrix Q are used to calculate the channel self-attention for each of the feature data. : ； in, Normalized exponential function Scaling factor This is a transpose operation; The spatial self-attention and the channel self-attention The result obtained by addition is connected to the standard convolutional features via residual concatenation. The features are then fused to obtain each of the standard convolutional features. First fusion feature : ; Each of the first fused features is input into two convolutional models with convolution sizes of 3 and 1, respectively, and the convolution results are residually concatenated with the first fused features to obtain the output features of each standard convolutional feature after passing through the dual self-attention module. : ， = 、 、 、 ； in, This represents a convolutional model with a convolution size of 3. This represents a convolutional model with a convolution size of 1.

2. The method for localizing epileptogenic foci based on image segmentation according to claim 1, characterized in that, The acquisition of the target MR sequence image and the preprocessing of the target MR sequence image to obtain target data specifically include: Acquire the target MR sequence images of the target FCD patient, the target MR sequence images including T1-weighted imaging and FLAIR values; The T1-weighted imaging and the FLAIR values ​​are aligned using the FSL tool to obtain aligned data. The aligned data is then standardized using the percentiles of the image grayscale distribution to obtain the target data.

3. The method for localizing epileptogenic foci based on image segmentation according to claim 1, characterized in that, The process of training and testing the FCD epileptogenic focus segmentation model to obtain the target model specifically includes: MR sequence images and gold standard mask images of epileptogenic foci were acquired from multiple FCD patients. The MR sequence images were preprocessed to obtain sample data. The gold standard mask image of the epileptogenic focus is a binary image, which includes the epileptogenic focus region and the background region in the MR sequence image. A dataset is obtained based on the sample data and the corresponding epileptogenic focus region and background region images of the sample data. The dataset is then divided into a training set, a validation set, and a test set according to a preset ratio. The FCD epileptogenic focus segmentation model is trained using the training set, its performance is validated and its strategy is adjusted using the validation set, and its performance is evaluated using the test set, thus obtaining a target model that meets the requirements.

4. The method for localizing epileptogenic foci based on image segmentation according to claim 1, characterized in that, The decoder module includes an upsampling Deconv2 layer, a convolutional layer, and a feature merging unit; The step of upsampling and feature fusion of the output features input to the decoder module to obtain the FCD epileptogenic focus segmented from the target MR sequence image specifically includes: Output features As the initial low-resolution feature, the initial low-resolution feature is input into the upsampling Deconv2 layer for upsampling. The initial low-resolution feature is expanded by 2 times through deconvolution or bilinear interpolation to obtain the upsampled feature. The number of feature channels in the feature merging unit is transformed into output features. The number of feature channels of the previous convolutional layer corresponding to the convolutional layer is used to merge the upsampled feature with the output feature corresponding to the previous convolutional layer through the feature merging unit to obtain the merged feature. The merged feature is then fused through the convolutional layer to obtain the second fused feature. The second fused feature is used as a low-resolution feature to continue upsampling and feature fusion operations until the output feature and the first two features in the local convolution feature are all fused to obtain the restored image. The recovered image is input into a convolutional algorithm with a kernel size of 1 and 2 channels to calculate the target feature map. The target feature map is then normalized using the Softmax function to obtain... FCD epileptogenic focus segmented from the target MR sequence image.

5. A system for locating epileptogenic foci based on image segmentation, characterized in that, The image segmentation-based epileptogenic focus localization system is used to implement the image segmentation-based epileptogenic focus localization method according to any one of claims 1-4, wherein the image segmentation-based epileptogenic focus localization system comprises: The data acquisition and processing module is used to acquire target MR sequence images and preprocess the target MR sequence images to obtain target data. The model building and training module is used to build an FCD epileptogenic focus segmentation model, train and test the FCD epileptogenic focus segmentation model to obtain a target model, wherein the target model includes an encoder module, a dual self-attention module and a decoder module. The FCD epileptogenic focus segmentation module is used to input the target data into the encoder module for local convolutional feature extraction to obtain convolutional features, input the convolutional features into the dual self-attention module space for fusion of convolutional features and attention to obtain output features, and input the output features into the decoder module for upsampling and feature fusion to obtain the FCD epileptogenic focus segmented from the target MR sequence image.

6. A terminal, characterized in that, The terminal includes: a memory, a processor, and an image segmentation-based epileptogenic focus localization program stored in the memory and executable on the processor. When the image segmentation-based epileptogenic focus localization program is executed by the processor, it implements the steps of the image segmentation-based epileptogenic focus localization method as described in any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an image segmentation-based epileptogenic focus localization program, which, when executed by a processor, implements the steps of the image segmentation-based epileptogenic focus localization method as described in any one of claims 1-4.

Images