A medical image segmentation system and method based on frequency-space dual-flow decoupling and global graph relationship reasoning

Abstract

A medical image segmentation system and a segmentation method, the method comprising the steps of: constructing an encoder-decoder segmentation network, introducing a frequency-space dual-flow decoupling module DualWaveNet in the shallow coding stage, extracting local anatomical structure features through the spatial branch, and using fixed Haar wavelet decomposition to construct the frequency branch to explicitly separate and recombine the low-frequency structure information and the high-frequency boundary texture in the horizontal, vertical and diagonal directions. In the deep bottleneck stage, a global graph relationship reasoning module GraphRM is introduced, the deep channel response is constructed into a potential semantic node, the long-range topological dependence between the lesion area and the similar background tissue is modeled and calibrated by combining the local graph perception branch and the global semantic relationship branch. After the decoder is gradually up-sampled, the segmentation result of the prostate cancer MRI image is output. The system comprises a processor which can run the image segmentation method. The present application can improve the ability of micro-lesion positioning, fuzzy boundary recovery and background mis-segmentation suppression under the condition of single modality MRI input.

Description

Technical Field

[0001] This invention relates to the field of medical image segmentation technology, specifically to a medical image segmentation system and method based on frequency-spatial dual-stream decoupling and global graph relation reasoning. Background Technology

[0002] In recent years, the incidence and mortality rates of prostate cancer have been on the rise, making it one of the most common malignant tumors in men. Magnetic resonance imaging (MRI), due to its high soft tissue contrast, is widely used for prostate cancer screening, lesion localization, and clinical auxiliary diagnosis. However, in actual clinical settings, prostate cancer lesions often exhibit characteristics such as small size, low contrast, blurred boundaries, significant morphological variations, and similar signals to surrounding normal tissues. This makes manual interpretation time-consuming and reliant on physician experience, highlighting the significant challenges of automated and precise segmentation.

[0003] Existing medical image segmentation methods typically improve model performance through multi-scale feature fusion, attention mechanisms, Transformers, Mamba, or frequency domain enhancement. Some methods utilize frequency domain transformations to enhance boundary textures, preserving some shallow high-frequency information; others improve global semantic representation through long-range dependency modeling; and still others enhance spatial connections between different regions through graph structure inference. However, these methods still have shortcomings in prostate cancer single-modal MRI lesion segmentation tasks: directly fusing frequency features at deep or global levels easily amplifies background artifacts; conventional long-range modeling lacks explicit suppression of similar background tissues; and graph inference methods based on pixel space or coarse segmentation region construction easily introduce a large number of irrelevant background nodes, increasing computational burden and weakening the expression of key lesion features. Summary of the Invention

[0004] To address the aforementioned technical problems, this technical solution provides a medical image segmentation system and method based on frequency-space dual-stream decoupling and global graph relation reasoning. Under single-modal MRI input conditions, it can simultaneously preserve shallow high-frequency boundary details, suppress redundant background interference, and perform global topological relation calibration in deep semantic space, thereby improving segmentation accuracy and robustness in low-contrast, small lesion, and blurred boundary scenes; effectively solving the aforementioned technical problems.

[0005] This invention is achieved through the following technical solution:

[0006] A medical image segmentation method based on frequency-spatial dual-stream decoupling and global graph relation reasoning includes the following steps:

[0007] A DWGNet encoder-decoder prostate cancer MRI image segmentation network was constructed, comprising an encoder, a bottleneck layer, a decoder, and a skip connection layer.

[0008] The frequency-space dual-stream decoupling module DualWaveNet is embedded in the first two layers of the encoder to perform spatial domain structure modeling and frequency domain boundary modeling on shallow features. Local anatomical features are extracted through spatial domain branches, and low-frequency structural information and multi-directional high-frequency boundary texture information are extracted through frequency domain wavelet branches.

[0009] A global graph relation reasoning module GraphRM is embedded in the network bottleneck layer. GraphRM includes a local graph-aware branch and a global semantic graph reasoning branch, which are used to model local anatomical continuity and global semantic topological dependency, respectively.

[0010] Furthermore, the operation process of the frequency-space dual-stream decoupling module DualWaveNet is as follows:

[0011] 1) Input feature tensor Input the shared convolutional backbone to obtain intermediate feature representations. Its expression is:

[0012] ;

[0013] in, , Represents the ReLU activation function. This represents the batch normalization function. This indicates a convolution operation with a kernel size of 3×3;

[0014] 2) Input the spatial domain branch and the frequency domain branch separately; the spatial domain branch uses 3×3 convolution to extract local structural features. Its expression is:

[0015] ;

[0016] in, The intermediate feature map representing the input spatial domain branch; Indicates spatial location on the feature map; A set of two-dimensional local coordinate offsets representing a standard 3×3 convolution; Indicates relative to the center position Local coordinate offset; Indicates the first Each input channel is located in the neighborhood. Eigenvalues ​​at; This represents the learnable convolutional kernel weight parameters at the corresponding local offset position; Indicates a learnable scaling factor Translation factor Batch normalization function;

[0017] 3) In the frequency domain branch, a fixed db1 / Haar wavelet is used for two-dimensional discrete wavelet decomposition. The one-dimensional low-pass filter and high-pass filter of the db1 wavelet are as follows: Four two-dimensional filters are constructed using tensor outer products, and their expressions are as follows:

[0018] ;

[0019] in, Used to extract low-frequency approximate structures. These are used to capture high-frequency detail responses in the horizontal, vertical, and diagonal directions, respectively; ⊗ represents the outer product.

[0020] 4) Let the first... The low-frequency characteristics of the input level are In the initial state Wavelet decomposition is performed using a two-dimensional discrete convolution operation with a stride of 2, for any output coordinates. and corresponding frequency band The calculation is as follows:

[0021] ;

[0022] in, Indicates the first Frequency subband after level decomposition Output spatial coordinates The characteristic response at the location; These correspond to the low-frequency approximation, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency sub-bands, respectively. In order to match the target frequency band Matched two-dimensional fixed wavelet filter;

[0023] 5) The decomposition results Subbands are spliced ​​along the channel dimension to obtain a multi-scale frequency domain feature tensor. Its expression is:

[0024] ;

[0025] in, , This represents a tensor splicing operation performed along the channel dimension; Indicates the first The low-frequency approximate component is obtained after two-dimensional discrete wavelet decomposition. These represent the high-frequency detail components in the horizontal, vertical, and diagonal directions, respectively.

[0026] 6) To Frequency domain modulation is introduced, and depthwise separable group convolution is used to achieve independent spatial aggregation of each frequency channel, and a learnable frequency scaling tensor is used. The expression for global recalibration of different subbands is as follows:

[0027] ;

[0028] in, This represents the modulated multi-scale frequency domain feature tensor; Scaling the tensor to learnable frequencies; ⊙ represents a depthwise convolution with a kernel size of 5×5; ⊙ represents element-wise multiplication.

[0029] 7) After completing frequency domain modulation, By segmenting along the channel dimension, low-frequency approximate components are obtained. and high-frequency detail component set Its expression is:

[0030] , ;

[0031] in, As input to the next level of wavelet decomposition, it is used to recursively model structural information at a larger scale; Used for boundary detail restoration in the subsequent reverse reconstruction stage;

[0032] 8) After completing the frequency domain feature evolution at all levels, the multi-scale frequency sub-bands are restored to the original spatial resolution through inverse discrete wavelet transform, and a system based on... The basic spatial residual path;

[0033] ;

[0034] in, This is the basic space convolution projection operation. Its learnable parameters, This represents the learnable space scaling factor in the residual path. This represents the discrete wavelet inverse transform operator. Indicates the maximum decomposition level. This indicates the sum of the residuals;

[0035] 9) The spatial domain branching obtained Obtained from frequency domain branch Perform cross-channel fusion to obtain the output features of DualWaveNet. Its expression is:

[0036] ;

[0037] ;

[0038] in, , This indicates a splicing operation along the channel dimension. express Pointwise convolution.

[0039] Furthermore, the specific process of the local graph-aware branch of the GraphRM module is as follows:

[0040] 1) Output features of the deep encoder Input the local graph convolutional network GraphCN, which consists of three concatenated 3×3 depthwise convolutions with a stride of 2 and a batch normalization operation:

[0041] ;

[0042] in, , ; Indicates the first Assigned to the first layer in the layer The depthwise convolutional kernel weights for each channel; Represents the compressed local feature basis; Indicates the first The layer output feature map is at the 1st individual channels, spatial locations The characteristic response value at that location; Indicates the first The batch normalization function for layers; Indicates the spatial coordinate index on the feature map; Indicates the previous level (the first) (Layer) Input features in the first layer Each channel, combined with a step size of 2 and local offset The characteristic response value at that location; This indicates that the final feature map output after all three cascaded operations is assigned to... ;

[0043] 2) The channel responses in the data are treated as a local semantic set, and are represented by three linear projection matrices. , and The features are projected into a compressed latent space and then the spatial flattening operator is used. Get the query tensor Key tensors Sum tensor Its expression is:

[0044] ;

[0045] in, , ⊛ represents the discrete spatial convolution operation, Represents the space flattening operator;

[0046] 3) Calculate the adjacency matrix of the local semantic graph along the channel dimension. , used to measure the correlation of semantic responses of different channels in the compressed space, where the line, number Column elements The calculation is as follows:

[0047] ;

[0048] in, Represents the adjacency matrix of the channel semantic graph Elements in; and Represent the query tensor Q and the value tensor respectively. In the The first spatial node and the One channel response; Value tensor In the The first spatial node One channel response; Represented by natural constant An exponential function with base 0;

[0049] 4) Utilizing the adjacency matrix of the local semantic graph Key tensor Perform global graph feature aggregation to obtain higher-order semantic information. Its expression is:

[0050] ;

[0051] in, Key tensor The transpose of the matrix; The dimension is The real tensor space ( For batch size, The number of spatial nodes. (Number of channels after compression).

[0052] 5) By inverse flattening operator Restored to the spatial dimension, and then convolved using a one-dimensional graph. Batch normalization and channel mapping operators Execution graph signal smoothing yields local graph-aware features. Its expression is:

[0053] ;

[0054] in, This represents a parameterized one-dimensional graph convolution operator. express convolution, Represents the inverse flattening operator; This represents the Batch Normalization function;

[0055] 6) Utilizing upsampling Will The original spatial scale is restored, and the upsampled graph features are used as a soft topological attention mask on the original deep features. Local graph relationship enhancement features are obtained. Its expression is:

[0056] ;

[0057] in, ∠ represents the bilinear space upsampling mapping, and ⊙ represents element-wise multiplication.

[0058] Furthermore, the specific process of the global semantic graph reasoning branch of the GraphRM module is as follows:

[0059] 1) Deep features The channel response is treated as a channelized semantic node, utilizing two independent nonlinear mapping functions. and Input features Projected onto latent spaces with different compression ratios, and then processed by the spatial flattening operator. Generate feature basis tensors and transformation mapping matrix Its expression is:

[0060] ;

[0061] in, , Set the channel compression factor. ,but ;

[0062] 2) Construct a global adjacency affinity matrix between channelized semantic nodes using matrix multiplication. Its expression is:

[0063] ;

[0064] in, The covariance matrix represents the relationship between global semantic nodes, used to measure the coactivation relationship and topological dependency of different channelized semantic nodes throughout the entire space.

[0065] 3) Apply two sets of parameterized one-dimensional graph convolution operators along the channel node dimension. and For the global adjacency affinity matrix Message passing and feature smoothing are performed, and residual skip connections are introduced to obtain updated graph node features. Its expression is:

[0066] ;

[0067] Where Zout∈R^(B×Cs×Cm), ⊛1D represents a one-dimensional convolution operation. and Represents the normalization function;

[0068] 4) Features of graph nodes used for completing relational reasoning With transformation mapping matrix Perform right multiplication and use the inverse flattening operator. Restore the spatial dimension, then utilize the channel mapping function. The number of channels to be restored, compared with the initial input. By forming residual connections, global semantic reasoning features are obtained. Its expression is:

[0069] ;

[0070] in, For inclusion Feature recovery mapping with convolution and batch normalization;

[0071] 5) Enhance the features of local graph relationships With global semantic reasoning features Concatenating along the feature channel dimension and then dynamically reducing and fusing them through a final cross-channel pointwise convolutional layer to obtain the GraphRM output features. Its expression is:

[0072] ;

[0073] in, , This indicates splicing along the feature channel axis. Indicates the kernel size as The pointwise convolution operator, This represents the batch normalization function; Through upsampling operation Will The original spatial scale is restored, and the upsampled graph features are used as a soft topological attention mask on the original deep features. The characteristics obtained.

[0074] Furthermore, before the step of constructing the encoder-decoder prostate cancer MRI image segmentation network DWGNet, there is a step of obtaining the prostate cancer MRI image dataset to be processed, dividing the dataset into training set and test set according to patient level, and performing uniform preprocessing, and uniformly adjusting the input image size to 256×256.

[0075] Furthermore, the datasets include the PI-CAI dataset, the HY prostate cancer MRI dataset, and the PROMISE12 dataset; among them, the PI-CAI dataset is subjected to five-fold cross-validation according to patient level, the HY prostate cancer MRI dataset is divided into training and testing sets in a 4:1 ratio according to patient level, and the PROMISE12 dataset is used for training and testing according to the official patient level division to avoid the same patient slice appearing in both the training and testing sets at the same time.

[0076] Furthermore, after the step of constructing the encoder-decoder prostate cancer MRI image segmentation network DWGNet, there is a step of inputting the multi-scale features processed by DualWaveNet and GraphRM into the decoder. The decoder upsamples the multi-scale features step by step and fuses them with skip connections to restore the spatial resolution and outputs a segmentation mask for the prostate cancer MRI image.

[0077] A medical image segmentation system based on frequency-spatial dual-stream decoupling and global graph relation reasoning includes a memory, a processor, and a computer program stored in the memory and executable on the processor; when the computer program is loaded onto the processor, it can implement the above-mentioned prostate cancer MRI image segmentation method.

[0078] Beneficial effects

[0079] The medical image segmentation system and method proposed in this invention, based on frequency-space dual-stream decoupling and global graph relation reasoning, has the following advantages compared with existing technologies:

[0080] (1) The present invention designs a frequency-space dual-stream decoupling module (DualWaveNet), which explicitly decouples shallow features through spatial domain branches and frequency domain wavelet branches. While preserving local anatomical information, it enhances the high-frequency texture of lesion boundaries, effectively solving the problems of low contrast, blurred boundaries and difficulty in locating small lesions in prostate cancer MRI images, and providing more discriminative shallow boundary priors for subsequent deep semantic modeling.

[0081] (2) In the frequency domain branch, the present invention uses a fixed db1 / Haar wavelet filter for two-dimensional discrete wavelet decomposition, which decomposes the shallow features into low-frequency approximate structures and high-frequency detail responses in the horizontal, vertical and diagonal directions. Since the fixed wavelet filter does not introduce additional learnable parameters, it can reduce the instability and overfitting risk in the training process of small sample medical images, while improving the model's sensitivity to edge mutations, texture changes and small lesion contours.

[0082] (3) This invention proposes a global graph relation reasoning module (GraphRM), which models local anatomical continuity and global semantic topological relationships through a local graph perception branch and a global semantic graph reasoning branch, respectively. The local graph perception branch can constrain the spatial continuity of lesion boundaries, while the global semantic graph reasoning branch can construct deep channel responses into channelized semantic nodes and calibrate lesion responses with similar background tissue responses through global covariance relationships, thereby effectively suppressing false positives and missegmentation.

[0083] (4) This invention achieves shallow boundary detail enhancement, background redundancy suppression and deep semantic topology calibration under single-modal MRI input conditions through the synergistic effect of DualWaveNet and GraphRM. It does not rely on multimodal image registration or complex interactive prompts and is suitable for routine clinical prostate MRI-assisted screening and computer-aided diagnosis scenarios.

[0084] (5) The present invention achieved good segmentation performance on the PROMISE12, HY prostate cancer MRI dataset and PI-CAI dataset. Experimental results show that the present invention can effectively improve the regional overlap, boundary localization stability and anti-interference ability of prostate cancer MRI image segmentation, and has good clinical application value and technical advantages. Attached Figure Description

[0085] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0086] Figure 2 This is a schematic diagram of the overall architecture of the image segmentation network in this invention.

[0087] Figure 3 This is a schematic diagram of the structure of the DualWaveNet frequency-space dual-stream decoupling module in this invention.

[0088] Figure 4 This is a schematic diagram of the structure of the GraphRM global graph relationship reasoning module in this invention.

[0089] Figures 5(a) to 5(d) This is a distribution chart of the Dice, IoU, Precision, and Recall metrics in the five-fold cross-validation of the PI-CAI dataset in this invention.

[0090] Figure 6 This is a visualization comparison of the segmentation results of this invention and other models on different prostate MRI datasets.

[0091] Figure 7 These are visualization results of the characteristic responses or heatmaps at different stages of the present invention. Detailed Implementation

[0092] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. The described embodiments are merely some embodiments of the present invention, and not all embodiments. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the design concept of the present invention should fall within the protection scope of the present invention.

[0093] Example 1:

[0094] like Figure 1 As shown, a medical image segmentation method based on frequency-spatial dual-stream decoupling and global graph relation reasoning includes the following steps:

[0095] Step 1: Obtain the prostate cancer MRI image segmentation dataset and divide the sample set.

[0096] Step 1.1: Collect relevant prostate cancer MRI image segmentation datasets.

[0097] (1) PI-CAI Dataset: This dataset contains multi-parameter prostate MRI images. This embodiment focuses on 220 clinically significant prostate cancer cases with artificial lesion annotations and performs five-fold cross-validation at the patient level. Patient ID is used as the basic unit for partitioning, and a fixed random seed of 42 is used to divide patients into 5 mutually exclusive subsets. In each fold experiment, approximately 80% of patients are used for training and approximately 20% of patients are used for testing, ensuring that all two-dimensional slices of the same patient do not appear in both the training and testing sets at the same time.

[0098] (2) HY Prostate Cancer MRI Dataset: This dataset was constructed by the First People's Hospital of Huai'an, Jiangsu Province, China, and includes 398 MRI images of prostate cancer from 169 patients. In this embodiment, the training set and the test set are divided into a 4:1 ratio according to the patient level. The training set contains 320 images from 136 patients, and the test set contains 78 images from 33 patients. Different slice images of the same patient appear in only one subset of the dataset.

[0099] (3) PROMISE12 Dataset: This dataset comes from the 2012 Prostate MRI Image Public Segmentation Challenge and is used to verify the model's generalization ability on the public prostate MRI dataset. This embodiment follows the official patient-level segmentation method, where the training set contains 1277 MRI images of 50 patients and the test set contains 795 MRI images of 30 patients. The patients in the training and testing phases do not overlap.

[0100] Step 1.2: Standardize and preprocess the prostate cancer MRI image dataset. First, all input images are uniformly adjusted to a scale of 256×256. Then, training and test sets are constructed according to the dataset partitioning principles to ensure that all slices from the same patient appear in only the same subset of data, thus avoiding leakage of patient-level information.

[0101] Step 2: Construct the encoder-decoder prostate cancer MRI image segmentation network DWGNet;

[0102] like Figure 2 As shown, the network includes an encoder, a bottleneck layer, a decoder, and a jump connection layer; the encoder includes a shallow encoder and a mid-to-high-level encoder, the shallow encoder includes a first encoder layer and a second encoder layer, and the mid-to-high-level encoder includes a third encoder layer and a fourth encoder layer.

[0103] DualWaveNet modules are embedded in the first and second layers of the encoder for shallow frequency-spatial feature decoupling; the third and fourth layers of the encoder adopt a standard double convolutional structure for extracting stable high-level semantic features; a GraphRM module is embedded in the bottleneck layer for deep global graph relation reasoning; finally, the decoder upsamples step by step and fuses skip connection features to output the final segmentation mask.

[0104] Step 2.1: Embed the frequency-spatial dual-stream decoupling module DualWaveNet in the first two layers of the encoder to perform spatial domain structure modeling and frequency domain boundary modeling on shallow features. Local anatomical structural features are extracted through the spatial domain branch, and low-frequency structural information and multi-directional high-frequency boundary texture information are extracted through the frequency domain wavelet branch. The operation process of the frequency-spatial dual-stream decoupling module DualWaveNet is as follows:

[0105] 1) Input feature tensor Input the shared convolutional backbone to obtain intermediate feature representations. Its expression is:

[0106] ;

[0107] in, , Represents the ReLU activation function. This represents the batch normalization function. This indicates a convolution operation with a kernel size of 3×3;

[0108] 2) Input the spatial domain branch and the frequency domain branch separately; the spatial domain branch uses 3×3 convolution to extract local structural features. Its expression is:

[0109] ;

[0110] in, The intermediate feature map representing the input spatial domain branch; Indicates spatial location on the feature map; A set of two-dimensional local coordinate offsets representing a standard 3×3 convolution; Indicates relative to the center position Local coordinate offset; Indicates the first Each input channel is located in the neighborhood. Eigenvalues ​​at; This represents the learnable convolutional kernel weight parameters at the corresponding local offset position; Indicates a learnable scaling factor Translation factor Batch normalization function;

[0111] 3) In the frequency domain branch, a fixed db1 / Haar wavelet is used for two-dimensional discrete wavelet decomposition. The one-dimensional low-pass filter and high-pass filter of the db1 wavelet are as follows: Four two-dimensional filters are constructed using tensor outer products, and their expressions are as follows:

[0112] ;

[0113] in, Used to extract low-frequency approximate structures. These are used to capture high-frequency detail responses in the horizontal, vertical, and diagonal directions, respectively; ⊗ represents the outer product.

[0114] 4) Let the first... The low-frequency characteristics of the input level are In the initial state Wavelet decomposition is performed using a two-dimensional discrete convolution operation with a stride of 2, for any output coordinates. and corresponding frequency band The calculation is as follows:

[0115] ;

[0116] in, Indicates the first Frequency subband after level decomposition Output spatial coordinates The characteristic response at the location; These correspond to the low-frequency approximation, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency sub-bands, respectively. In order to match the target frequency band Matched two-dimensional fixed wavelet filter;

[0117] 5) The decomposition results Subbands are spliced ​​along the channel dimension to obtain a multi-scale frequency domain feature tensor. Its expression is:

[0118] ;

[0119] in, , This represents a tensor splicing operation performed along the channel dimension; Indicates the first The low-frequency approximate component is obtained after two-dimensional discrete wavelet decomposition. These represent the high-frequency detail components in the horizontal, vertical, and diagonal directions, respectively.

[0120] 6) To Frequency domain modulation is introduced, and depthwise separable group convolution is used to achieve independent spatial aggregation of each frequency channel, and a learnable frequency scaling tensor is used. The expression for global recalibration of different subbands is as follows:

[0121] ;

[0122] in, This represents the modulated multi-scale frequency domain feature tensor; Scaling the tensor to learnable frequencies; ⊙ represents a depthwise convolution with a kernel size of 5×5; ⊙ represents element-wise multiplication.

[0123] 7) After completing frequency domain modulation, By segmenting along the channel dimension, low-frequency approximate components are obtained. and high-frequency detail component set Its expression is:

[0124] , ;

[0125] in, As input to the next level of wavelet decomposition, it is used to recursively model structural information at a larger scale; Used for boundary detail restoration in the subsequent reverse reconstruction stage;

[0126] 8) After completing the frequency domain feature evolution at all levels, the multi-scale frequency sub-bands are restored to the original spatial resolution through inverse discrete wavelet transform, and a system based on... The basic spatial residual path;

[0127] ;

[0128] in, This is the basic space convolution projection operation. Its learnable parameters, This represents the learnable space scaling factor in the residual path. This represents the discrete wavelet inverse transform operator. Indicates the maximum decomposition level. This indicates the sum of the residuals;

[0129] 9) The spatial domain branching obtained Obtained from frequency domain branch Perform cross-channel fusion to obtain the output features of DualWaveNet. Its expression is:

[0130] ;

[0131] ;

[0132] in, , This indicates a splicing operation along the channel dimension. express Pointwise convolution.

[0133] Step 2.2: Construct the global graph relation reasoning module GraphRM.

[0134] like Figure 4 As shown, this module consists of a local graph perception branch and a global semantic graph reasoning branch, which are used to perform local topological constraints and global semantic calibration on deep semantic features at the bottleneck layer.

[0135] A global graph relation reasoning module GraphRM is embedded in the network bottleneck layer. GraphRM includes a local graph-aware branch and a global semantic graph reasoning branch, which are used to model local anatomical continuity and global semantic topological dependency, respectively.

[0136] Step 2.21: The specific process of constructing local graph-aware branches is as follows:

[0137] 1) Output features of the deep encoder Input the local graph convolutional network GraphCN, which consists of three concatenated 3×3 depthwise convolutions with a stride of 2 and a batch normalization operation:

[0138] ;

[0139] in, , ; Indicates the first Assigned to the first layer in the layer The depthwise convolutional kernel weights for each channel; Represents the compressed local feature basis; Indicates the first The layer output feature map is at the 1st individual channels, spatial locations The characteristic response value at that location; Indicates the first The batch normalization function for layers; Indicates the spatial coordinate index on the feature map; Indicates the previous level (the first) (Layer) Input features in the first layer Each channel, combined with a step size of 2 and local offset The characteristic response value at that location; This indicates that the final feature map output after all three cascaded operations is assigned to... ;

[0140] 2) The channel responses in the data are treated as a local semantic set, and are represented by three linear projection matrices. , and The features are projected into a compressed latent space and then the spatial flattening operator is used. Get the query tensor Key tensors Sum tensor Its expression is:

[0141] ;

[0142] in, , ⊛ represents the discrete spatial convolution operation, Represents the space flattening operator;

[0143] 3) Calculate the adjacency matrix of the local semantic graph along the channel dimension. , used to measure the correlation of semantic responses of different channels in the compressed space, where the line, number Column elements The calculation is as follows:

[0144] ;

[0145] in, Represents the adjacency matrix of the channel semantic graph Elements in; and Represent the query tensor Q and the value tensor respectively. In the The first spatial node and the One channel response; Value tensor In the The first spatial node One channel response; Represented by natural constant An exponential function with base 0;

[0146] 4) Utilizing the adjacency matrix of the local semantic graph Key tensor Perform global graph feature aggregation to obtain higher-order semantic information. Its expression is:

[0147] ;

[0148] in, Key tensor The transpose of the matrix; The dimension is The real tensor space ( For batch size, The number of spatial nodes. (Number of channels after compression).

[0149] 5) By inverse flattening operator Restored to the spatial dimension, and then convolved using a one-dimensional graph. Batch normalization and channel mapping operators Execution graph signal smoothing yields local graph-aware features. Its expression is:

[0150] ;

[0151] in, This represents a parameterized one-dimensional graph convolution operator. express convolution, Represents the inverse flattening operator; This represents the Batch Normalization function;

[0152] 6) Utilizing upsampling Will The original spatial scale is restored, and the upsampled graph features are used as a soft topological attention mask on the original deep features. Local graph relationship enhancement features are obtained. Its expression is:

[0153] ;

[0154] in, ∠ represents the bilinear space upsampling mapping, and ⊙ represents element-wise multiplication.

[0155] Step 2.22: The specific process of constructing the global semantic graph reasoning branch is as follows:

[0156] 1) Deep features The channel response is treated as a channelized semantic node, utilizing two independent nonlinear mapping functions. and Input features Projected onto latent spaces with different compression ratios, and then processed by the spatial flattening operator. Generate feature basis tensors and transformation mapping matrix Its expression is:

[0157] ;

[0158] in, , Set the channel compression factor. ,but ;

[0159] 2) Construct a global adjacency affinity matrix between channelized semantic nodes using matrix multiplication. Its expression is:

[0160] ;

[0161] in, The covariance matrix represents the relationship between global semantic nodes, used to measure the coactivation relationship and topological dependency of different channelized semantic nodes throughout the entire space.

[0162] 3) Apply two sets of parameterized one-dimensional graph convolution operators along the channel node dimension. and For the global adjacency affinity matrix Message passing and feature smoothing are performed, and residual skip connections are introduced to obtain updated graph node features. Its expression is:

[0163] ;

[0164] Where Zout∈R^(B×Cs×Cm), ⊛1D represents a one-dimensional convolution operation. and Represents the normalization function;

[0165] 4) Features of graph nodes used for completing relational reasoning With transformation mapping matrix Perform right multiplication and use the inverse flattening operator. Restore the spatial dimension, then utilize the channel mapping function. The number of channels to be restored, compared with the initial input. By forming residual connections, global semantic reasoning features are obtained. Its expression is:

[0166] ;

[0167] in, For inclusion Feature recovery mapping with convolution and batch normalization;

[0168] 5) Enhance the features of local graph relationships With global semantic reasoning features Concatenating along the feature channel dimension and then dynamically reducing and fusing them through a final cross-channel pointwise convolutional layer to obtain the GraphRM output features. Its expression is:

[0169] ;

[0170] in, , This indicates splicing along the feature channel axis. Indicates the kernel size as The pointwise convolution operator, This represents the batch normalization function; Through upsampling operation Will The original spatial scale is restored, and the upsampled graph features are used as a soft topological attention mask on the original deep features. The characteristics obtained.

[0171] Step 3: Train the DWGNet network using the training set and validate the trained model using the test set. After training, input the prostate cancer MRI image to be segmented into the trained DWGNet network. The DWGNet network will segment the input image and output the corresponding prostate cancer lesion or prostate structure segmentation result.

[0172] The DWGNet network performs the following segmentation process on the input image:

[0173] Step 3.1: Obtain the prostate MRI image to be processed, adjust all input images to a uniform scale of 256×256, and divide the training set and test set according to patient level to ensure that all slices of the same patient do not appear in the training set and test set at the same time.

[0174] Step 3.2: Input the preprocessed MRI image into the encoder. Embed a DualWaveNet module in the first two layers of the encoder. Extract local anatomical features through spatial domain branching and explicitly decouple low-frequency structures and high-frequency boundary textures of shallow features through frequency domain wavelet branching, thereby enhancing lesion boundary expression and suppressing redundant background responses.

[0175] Step 3.3: The third and fourth layers of the encoder adopt a standard double convolutional structure to further semantically abstract the features after shallow frequency-space decoupling, extract stable high-level semantic features, and provide deep feature input for subsequent bottleneck layer graph relation reasoning.

[0176] Step 3.4: Input the deepest layer output features of the encoder into the GraphRM module. GraphRM models the continuity of local anatomy through a local graph perception branch and constructs the global covariance relationship between channelized semantic nodes through a global semantic graph inference branch, thereby performing topological calibration on the lesion response and the response to similar background.

[0177] Step 3.5: Input the deep enhancement features output by GraphRM into the decoder, and combine them with the encoder skip connection features to restore the spatial resolution step by step, finally generating a pixel-level segmentation mask for prostate cancer MRI images.

[0178] The DWGNet model described in this invention will be compared with current advanced medical image segmentation models, and the feasibility and effectiveness of the method will be verified through ablation experiments and visualization analysis.

[0179] The DWGNet model described in this invention was compared with other models such as TransUNet, AttentionUNet, UNet++, CATNet, H2former, Hiformer, MDSAUNet, DAMAF, PMFSNet, and MAUNet on the HY prostate cancer MRI dataset, the PROMISE12 dataset, and the PI-CAI dataset. The segmentation results of each model on the HY dataset and the PROMISE12 dataset are shown in Tables 1 and 2, respectively. The five-fold cross-validation results on the PI-CAI dataset are shown in Tables 1 and 2, respectively. Figures 5(a)-5(d) As shown.

[0180] Table 1. Comparison results of the HY prostate cancer image segmentation dataset with other models.

[0181] TransUnet 0.5772 0.4057 55.6917 0.5428 0.6162 AttentionUnet 0.6003 0.4288 43.4560 0.5072 0.7352 Unet++ 0.6128 0.4418 32.0605 0.5628 0.6727 CATNet 0.6008 0.4294 42.2291 0.5575 0.6515 H2former 0.4989 0.3323 42.5718 0.4514 0.5575 Hiformer 0.4505 0.2907 38.8074 0.4828 0.4222 MDSAUNet 0.5860 0.4144 41.8127 0.5249 0.6632 DAMA 0.4989 0.3323 38.2232 0.5581 0.4510 PMFSNet 0.4686 0.3060 57.5306 0.3816 0.6070 MAUNet 0.5743 0.4028 45.3057 0.5059 0.6639 DWGNet This Invention 0.6514 0.4830 45.0961 0.5929 0.7226

[0182] As shown in Table 1, on the HY prostate cancer MRI dataset, which features low contrast, blurred boundaries, and strong background interference, the proposed DWGNet achieves superior region overlap performance, with a DSC of 0.6514 and an IoU of 0.4830. Compared to models such as TransUNet, AttentionUNet, UNet++, and CATNet, this invention can more stably preserve lesion boundary details and reduce missegmentation caused by similar background tissues. This fully demonstrates that DualWaveNet's enhancement of shallow boundary textures and GraphRM's calibration of deep semantic topological relationships are complementary.

[0183] Table 2. Comparison results of the PROMISE12 prostate image segmentation dataset with other models

[0184] TransUnet 0.8580 0.7513 32.0593 0.8809 0.8363 AttentionUnet 0.8932 0.8071 20.4641 0.9068 0.8800 Unet++ 0.8946 0.8093 27.7194 0.8866 0.8933 CATNet 0.8899 0.8017 31.6395 0.8866 0.8933 H2former 0.8327 0.7133 38.8425 0.8353 0.8301 Hiformer 0.8459 0.7329 39.4061 0.8922 0.8041 MDSAUNet 0.8785 0.7833 39.5247 0.8754 0.8816 DAMA 0.8373 0.7201 38.5186 0.8326 0.8420 PMFSNet 0.8783 0.7831 31.6309 0.8875 0.8693 MAUNet 0.8793 0.7846 28.4060 0.8877 0.8710 DWGNet (this invention) 0.8951 0.8100 25.9061 0.9030 0.8873

[0185] As shown in Table 2, most models have achieved high segmentation accuracy on the PROMISE12 dataset, but the DWGNet proposed in this invention still achieves competitive results, with a DSC of 0.8951, an IoU of 0.8100, and a Precision of 0.9030. This effectively demonstrates that this invention is not only applicable to low-contrast prostate cancer lesion segmentation scenarios, but also maintains stable segmentation performance on publicly available datasets with relatively clear prostate structural boundaries.

[0186] The PI-CAI dataset presents significant segmentation challenges due to issues such as minute lesions, severe imbalances in foreground and background ratios, and substantial morphological differences between various cases. (Refer to...) Figures 5(a)-5(d) The present invention exhibits relatively stable distributions of Dice, IoU, Precision, and Recall in five-fold cross-validation. Compared with some hybrid attention models and Transformer-like models, DWGNet's metric distribution is more concentrated, indicating that it has better stability and generalization ability in complex clinical scenarios.

[0187] Visual analysis:

[0188] Reference Figure 6The method of this invention is visually compared with medical image segmentation models such as TransUNet, AttentionUNet, UNet++, CATNet, H2former, Hiformer, MDSAUNet, DAMAF, PMFSNet, and MAUNet. In the PROMISE12 dataset, due to the relatively clear boundaries of the prostate structure, most methods can obtain a relatively complete organ outline, but some models still exhibit outline breaks or morphological deviations in local boundary regions. In contrast, the segmentation results generated by this invention are closer to the real annotations, with smoother boundaries and more stable spatial connectivity.

[0189] In the HY and PI-CAI datasets, due to the smaller size, lower contrast, and similar grayscale distribution of lesions to surrounding tissues, some contrastive models are prone to missed detections, boundary drift, or missegmentation of adjacent normal tissue as foreground. This invention uses DualWaveNet to preserve high-frequency boundary details in the shallow layer and GraphRM for semantic topology calibration in the deep layer, resulting in a more compact predicted region, more continuous lesion boundaries, and a significant reduction in background pseudo-activations. Therefore, this invention can better maintain the connectivity and boundary integrity of lesion regions under complex background interference.

[0190] Reference Figure 7 The feature responses at different stages of this invention were visualized and analyzed. Overall, the feature responses showed a trend of gradually converging from scattered background responses to the lesion area. After inputting the original image, there were still many background texture responses in the shallow features; after introducing DualWaveNet in the first layer of the encoder, the model began to capture the macroscopic structure of the prostate and some high-frequency boundary information; after continuing to introduce DualWaveNet in the second layer of the encoder, the low-frequency anatomical structure and high-frequency boundary texture were further separated, the redundant background responses were gradually weakened, and the feature activation was more concentrated in the candidate lesion area.

[0191] When GraphRM is introduced into the bottleneck layer, the feature responses further exhibit an enhanced trend of structural consistency. The local graph-aware branch helps maintain the spatial continuity of lesion boundaries, while the global semantic graph inference branch can utilize the global relationships between channelized semantic nodes to suppress false activations in similar background regions. The predicted region in the final output stage is more compact, has a more coherent morphology, and is closer to the actual lesion region. The visualization results above demonstrate that DualWaveNet and GraphRM play roles in shallow boundary enhancement and deep semantic calibration, respectively, and their combined efforts improve the localization accuracy and anti-interference capability of prostate cancer MRI image segmentation.

[0192] Ablation experiment:

[0193] To systematically verify the effectiveness of the DualWaveNet and GraphRM modules, this embodiment uses a basic encoder-decoder network as the baseline, and introduces the DualWaveNet and GraphRM modules respectively to construct Baseline+WAVE, Baseline+GRA, and the complete DWGNet model. The ablation experiment results of each model on the HY, PI-CAI, and PROMISE12 datasets are shown in Tables 3, 4, and 5, respectively.

[0194] Table 3. Ablation experiments on the HY prostate cancer image segmentation dataset

[0195] Baseline 0.5971 0.4265 49.4635 0.4851 0.7792 Baseline+WAVE 0.6451 0.4761 81.7463 0.6430 0.6472 Baseline+GRA 0.6443 0.4752 44.7874 0.5923 0.7063 Baseline+WAVE+GRA (DWGNet) 0.6514 0.4830 45.0961 0.5929 0.7226

[0196] As shown in Table 3, on the HY prostate cancer MRI dataset, the baseline DSC is 0.5971 and IoU is 0.4265. After introducing DualWaveNet, the DSC improves to 0.6451 and the IoU improves to 0.4761, indicating that frequency-spatial decoupling can effectively enhance shallow boundary representation and improve regional overlap performance. After introducing GraphRM, the model's DSC reaches 0.6443 and IoU reaches 0.4752, indicating that deep graph relation reasoning can utilize semantic topological information to improve lesion localization. The complete DWGNet, combined with DualWaveNet and GraphRM, achieves the best overall performance with a DSC of 0.6514 and an IoU of 0.4830, proving that shallow frequency-spatial decoupling and deep graph relation reasoning have complementary gains.

[0197] Table 4. Ablation experiments on the PI-CAI prostate cancer image segmentation dataset

[0198] Baseline 0.4571 0.2963 128.9189 0.4530 0.4614 Baseline+WAVE 0.4844 0.3196 124.8960 0.4698 0.4999 Baseline+GRA 0.4659 0.3037 102.5229 0.3860 0.5874 Baseline+WAVE+GRA (DWGNet) 0.4866 0.3216 119.0032 0.4988 0.4750

[0199] As shown in Table 4, the segmentation task on the PI-CAI dataset is more challenging due to the small lesion size, unbalanced foreground-background ratio, and significant case variability. The baseline DSC is 0.4571 and IoU is 0.2963; after introducing DualWaveNet, the DSC improves to 0.4844 and the IoU to 0.3196; the introduction of GraphRM enhances the semantic calibration capability in complex backgrounds through deep semantic graph relationship modeling. The complete DWGNet achieves a DSC of 0.4866 and an IoU of 0.3216, indicating that the proposed method can alleviate the problems of insufficient region overlap and false background responses in small lesion segmentation.

[0200] Table 5. Ablation experiments on the PROMISE12 prostate image segmentation dataset

[0201] Baseline 0.8896 0.8012 30.2923 0.8875 0.8917 Baseline+WAVE 0.8942 0.8087 22.9661 0.9140 0.8753 Baseline+GRA 0.8927 0.8063 26.3484 0.9064 0.8795 Baseline+WAVE+GRA (DWGNet) 0.8951 0.8100 23.3712 0.9030 0.8873

[0202] As shown in Table 5, the baseline already achieved high segmentation accuracy on the PROMISE12 dataset, but the model was further optimized after adding the proposed module. The full DWGNet achieved a DSC of 0.8951 and an IoU of 0.8100, indicating that in prostate structure segmentation tasks with relatively clear boundaries, DualWaveNet can still improve the ability to model boundary details, and GraphRM can still enhance global structural consistency.

[0203] In practical applications, this invention can be embedded as a core algorithm in a medical imaging computer-aided diagnostic system. After receiving a prostate MRI image, the system first performs size normalization processing, then inputs it into a trained DWGNet network, outputting a pixel-level segmentation mask for prostate cancer lesions or prostate structures. Doctors can combine this segmentation result for lesion localization, contour confirmation, radiomics feature extraction, and subsequent diagnostic and treatment decision analysis. This method can reduce the burden of manual annotation, improve the consistency and stability of prostate cancer MRI image segmentation, and has good clinical auxiliary application value.

[0204] The prostate cancer MRI image segmentation network based on frequency-space dual-stream decoupling and global graph relation reasoning of the present invention can also be set in a computer program. When the computer program is loaded into the processor, it implements the above-mentioned prostate cancer MRI image segmentation method including the DualWaveNet module and the GraphRM module.

[0205] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent transformations or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A medical image segmentation method based on frequency-spatial dual-stream decoupling and global graph relation reasoning, characterized in that, Including the following steps: A DWGNet encoder-decoder prostate cancer MRI image segmentation network was constructed, comprising an encoder, a bottleneck layer, a decoder, and a skip connection layer. The frequency-space dual-stream decoupling module DualWaveNet is embedded in the first two layers of the encoder to perform spatial domain structure modeling and frequency domain boundary modeling on shallow features. Local anatomical features are extracted through spatial domain branches, and low-frequency structural information and multi-directional high-frequency boundary texture information are extracted through frequency domain wavelet branches. A global graph relation reasoning module GraphRM is embedded in the network bottleneck layer. GraphRM includes a local graph-aware branch and a global semantic graph reasoning branch, which are used to model local anatomical continuity and global semantic topological dependency, respectively.

2. The medical image segmentation method based on frequency-spatial dual-stream decoupling and global graph relation reasoning according to claim 1, characterized in that: The operation process of the frequency-space dual-stream decoupling module DualWaveNet is as follows: 1) Input feature tensor Input the shared convolutional backbone to obtain intermediate feature representations. ; 2) Input the spatial domain branch and the frequency domain branch separately; the spatial domain branch uses 3×3 convolution to extract local structural features. ; 3) In the frequency domain branch, a fixed db1 / Haar wavelet is used for two-dimensional discrete wavelet decomposition. The one-dimensional low-pass filter and high-pass filter of the db1 wavelet are as follows: Four two-dimensional filters are constructed using tensor outer products, and their expressions are as follows: ； in, Used to extract low-frequency approximate structures. These are used to capture high-frequency detail responses in the horizontal, vertical, and diagonal directions, respectively; ⊗ represents the outer product. 4) Let the first... The low-frequency characteristics of the input level are In the initial state Wavelet decomposition is performed using a two-dimensional discrete convolution operation with a stride of 2, for any output coordinates. and corresponding frequency band The calculation is as follows: ； in, Indicates the first Frequency subband after level decomposition Output spatial coordinates The characteristic response at the location; These correspond to the low-frequency approximation, horizontal high-frequency, vertical high-frequency, and diagonal high-frequency sub-bands, respectively. In order to match the target frequency band Matched two-dimensional fixed wavelet filter; 5) The decomposition results Subbands are spliced ​​along the channel dimension to obtain a multi-scale frequency domain feature tensor. ; 6) To Frequency domain modulation is introduced, and depthwise separable group convolution is used to achieve independent spatial aggregation of each frequency channel, and a learnable frequency scaling tensor is used. The expression for global recalibration of different subbands is as follows: ； in, Represents the modulated multi-scale frequency domain feature tensor; Scaling the tensor to learnable frequencies; ⊙ represents a depthwise convolution with a kernel size of 5×5; ⊙ represents element-wise multiplication. 7) After completing frequency domain modulation, By segmenting along the channel dimension, low-frequency approximate components are obtained. and high-frequency detail component set Its expression is: ， ； in, As input to the next level of wavelet decomposition, it is used to recursively model structural information at a larger scale; Used for boundary detail restoration in the subsequent reverse reconstruction stage; 8) After completing the frequency domain feature evolution at all levels, the multi-scale frequency sub-bands are restored to the original spatial resolution through inverse discrete wavelet transform, and a system based on... The basic spatial residual path; ； in, This is the basic space convolution projection operation. Its learnable parameters, This represents the learnable space scaling factor in the residual path. This represents the discrete wavelet inverse transform operator. Indicates the maximum decomposition level. This indicates the sum of the residuals; 9) The spatial domain branching obtained Obtained from frequency domain branch Cross-channel fusion is performed to obtain the output features of DualWaveNet. Its expression is: ； ； in, , This indicates a splicing operation along the channel dimension. express Pointwise convolution.

3. The medical image segmentation method based on frequency-spatial dual-stream decoupling and global graph relation reasoning according to claim 1, characterized in that: The specific process of the local graph-aware branch of the GraphRM module is as follows: 1) Output features of the deep encoder The input is a local graph convolutional network, GraphCN, which consists of three concatenated 3×3 depthwise convolutions with a stride of 2 and a batch normalization operation, resulting in a compressed local feature basis. ; 2) The channel responses in the data are treated as a local semantic set, and are represented by three linear projection matrices. , and The features are projected into a compressed latent space and then the spatial flattening operator is used. Get the query tensor Key tensors Sum tensor ; 3) Calculate the adjacency matrix of the local semantic graph along the channel dimension. , used to measure the correlation of semantic responses of different channels in the compressed space, where the line, number Column elements The calculation is as follows: ； in, Represents the adjacency matrix of the channel semantic graph Elements in; and Represent the query tensor Q and the value tensor respectively. In the The first spatial node and the One channel response; 4) Utilizing the adjacency matrix of the local semantic graph Key tensor Perform global graph feature aggregation to obtain higher-order semantic information. ; 5) By inverse flattening operator Restored to the spatial dimension, and then convolved using a one-dimensional graph. Batch normalization and channel mapping operators Execution graph signal smoothing yields local graph-aware features. ; 6) Utilizing upsampling Will The original spatial scale is restored, and the upsampled graph features are used as a soft topological attention mask on the original deep features. Local graph relationship enhancement features are obtained. .

4. A medical image segmentation method based on frequency-spatial dual-stream decoupling and global graph relation reasoning according to claim 1 or 3, characterized in that: The specific process of the global semantic graph reasoning branch of the GraphRM module is as follows: 1) Deep features The channel response is treated as a channelized semantic node, utilizing two independent nonlinear mapping functions. and Input features Projected onto latent spaces with different compression ratios, and then processed by the spatial flattening operator. Generate feature basis tensors and transformation mapping matrix Its expression is: ； in, , Set the channel compression factor. ,but ; 2) Construct a global adjacency affinity matrix between channelized semantic nodes using matrix multiplication. Its expression is: ； in, The covariance matrix represents the relationship between global semantic nodes, used to measure the coactivation relationship and topological dependency of different channelized semantic nodes throughout the entire space. 3) Apply two sets of parameterized one-dimensional graph convolution operators along the channel node dimension. and For the global adjacency affinity matrix Message passing and feature smoothing are performed, and residual skip connections are introduced to obtain updated graph node features. ; 4) Features of graph nodes used for completing relational reasoning With transformation mapping matrix Perform right multiplication and use the inverse flattening operator. Restore the spatial dimension, then utilize the channel mapping function. The number of channels to be restored, compared with the initial input. By forming residual connections, global semantic reasoning features are obtained. Its expression is: ； in, For inclusion Feature recovery mapping with convolution and batch normalization; 5) Enhance the features of local graph relationships With global semantic reasoning features Concatenating along the feature channel dimension and then dynamically reducing and fusing them through a final cross-channel pointwise convolutional layer to obtain the GraphRM output features. Its expression is: ； in, , This indicates splicing along the feature channel axis. Indicates the kernel size as The pointwise convolution operator, This represents the batch normalization function; Through upsampling operation Will The original spatial scale is restored, and the upsampled graph features are used as a soft topological attention mask on the original deep features. The characteristics obtained.

5. The medical image segmentation method based on frequency-spatial dual-stream decoupling and global graph relation reasoning according to claim 1, characterized in that: Before the step of constructing the encoder-decoder prostate cancer MRI image segmentation network DWGNet, there is a step of obtaining the prostate cancer MRI image dataset to be processed. The dataset is divided into training set and test set according to patient level, and uniform preprocessing is performed, and the input image size is uniformly adjusted to 256×256.

6. The medical image segmentation method based on frequency-spatial dual-stream decoupling and global graph relation reasoning according to claim 5, characterized in that: The datasets include the PI-CAI dataset, the HY prostate cancer MRI dataset, and the PROMISE12 dataset. The PI-CAI dataset is subjected to five-fold cross-validation according to patient level, the HY prostate cancer MRI dataset is divided into training and testing sets in a 4:1 ratio according to patient level, and the PROMISE12 dataset is used for training and testing according to the official patient level division.

7. The medical image segmentation method based on frequency-spatial dual-stream decoupling and global graph relation reasoning according to claim 1, characterized in that: After the step of constructing the encoder-decoder prostate cancer MRI image segmentation network DWGNet, there is a step of inputting the multi-scale features processed by DualWaveNet and GraphRM into the decoder. The decoder upsamples the multi-scale features step by step and fuses them with skip connections to restore the spatial resolution and outputs a segmentation mask for the prostate cancer MRI image.

8. A medical image segmentation system based on frequency-spatial dual-stream decoupling and global graph relation reasoning, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor; characterized in that: When the computer program is loaded into the processor, it can implement the prostate cancer MRI image segmentation method according to any one of claims 1-7.

Images