Medical image segmentation method based on bidirectional scanning Mama and dual-scale constraint deformation convolution

Abstract

The invention discloses a medical image segmentation method based on bidirectional scanning Mama and dual-scale constraint deformation convolution. The method comprises the following steps: firstly, acquiring a medical image and preprocessing the medical image; then constructing a segmentation model; the segmentation model adopts an encoder-decoder architecture; the encoder is formed by stacking a plurality of bidirectional scanning Mamba modules and convolution attention modules, and a dual-scale constraint deformation convolution module is arranged at the bottom of the encoder; training, testing and verifying the segmentation model by using the preprocessed medical image; and finally, processing the medical image to be segmented by using the verified segmentation model, and segmenting an organ part. According to the method, the deformation convolution and the Mama are effectively combined, the function that the Mama can remotely depend on and capture global information is combined with the function that the deformation convolution can capture more detailed local information, and the organ part is automatically segmented from the medical image.

Description

Technical Field

[0001] This invention belongs to the field of image segmentation technology and relates to a medical image segmentation method based on bidirectional scanning Mamba and dual-scale constrained deformable convolution. Background Technology

[0002] Medical image segmentation plays a crucial role in clinical diagnosis and treatment planning, aiding in the identification of tumors, lesions, and organ structures. However, compared to natural images, medical images such as CT (computed tomography) and MR (magnetic resonance imaging) often suffer from higher noise levels, lower contrast, and blurred boundaries, making it difficult to extract structural features from target regions. Furthermore, significant differences in organ morphology among patients, and the small size and susceptibility to occlusion by surrounding tissues, further complicate segmentation.

[0003] Current deep learning methods mainly include Convolutional Neural Networks (CNNs) and Transformer architectures. CNNs excel in local feature extraction, but their limited receptive field restricts their ability to model long-range dependencies, making them ill-suited for handling the complex spatial relationships between abdominal organs. While Transformers can model global dependencies, their self-attention mechanism is computationally complex, resulting in significant resource consumption in high-resolution medical images and hindering practical deployment. Hybrid architectures combining both approaches have emerged in recent years, but they still have shortcomings in resolving blurred boundaries, enhancing small organ features, and improving computational efficiency.

[0004] On the other hand, the task of segmenting multiple organs in the abdomen is also significantly challenging: organs have similar gray-scale distributions and unclear boundaries; organs such as the pancreas are long and irregular in shape; and small organs such as the adrenal glands are small in size and easily blend with surrounding tissues. These characteristics mean that the model needs to have both strong long-range dependency modeling capabilities and high-precision local structure capture capabilities.

[0005] In summary, existing methods generally struggle to balance global modeling efficiency with the ability to depict local details, resulting in insufficient segmentation accuracy for abdominal organs, especially small organs. Therefore, a novel segmentation method is needed that can simultaneously achieve efficient long-range dependency modeling, enhance local detail representation, and adapt to organ structures at different scales. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing a medical image segmentation method based on bidirectional scanning Mamba and dual-scale constrained deformable convolution, which can accurately segment organ parts in medical images.

[0007] The specific implementation steps of this invention are as follows:

[0008] In a first aspect, the present invention provides a medical image segmentation method based on bidirectional scanning Mamba and dual-scale constrained deformable convolution, comprising the following steps:

[0009] Medical images containing organs are acquired and preprocessed to construct a dataset;

[0010] A segmentation model, DDMamba, is constructed. The DDMamba model employs an encoder-decoder architecture. The encoder consists of multiple bidirectional scanning Mamba modules (BSM) and convolutional attention modules (CSCA). The BSM module performs bidirectional scanning state space modeling on the input feature map, using forward and backward SSM to model forward long-range dependencies and capture reverse contextual information, respectively, resulting in feature maps with enhanced sequence dependencies and contextual features. The CSCA module captures the global semantic information of the feature map output by the BSM through channel attention and uses spatial attention to locate salient regions, obtaining enhanced feature representations. These enhanced feature representations are input into a dual-scale constrained deformable convolutional module (DSCDC), which uses different branches to capture wide-area contextual information and structural details of the medical image, and then inputs them into the decoder. The decoder consists of four upsampling modules and a final output projection layer, used to progressively restore spatial resolution and generate organ segmentation results.

[0011] The dataset is divided into training, testing, and validation sets to train, test, and validate the segmentation model DDMamba.

[0012] The validated segmentation model DDMamba was used to process the medical image to be segmented and segment out the organ parts.

[0013] Furthermore, the specific processing steps of the BSM module include:

[0014] The input feature map is flattened and then split into two parallel branches. The first branch passes through a linear layer, a 1D convolutional layer, and a SiLU activation function in sequence; the second branch passes through a linear layer and a SiLU activation function in sequence.

[0015] The outputs of the first and second branches are fed into the forward SSM and backward SSM respectively for scanning; the outputs of the forward SSM and backward SSM are multiplied element-wise with the corresponding gated signals, and the results are added together element-wise.

[0016] The fused features are projected through a linear layer and restored to spatial feature maps through a reshape operation. Finally, a residual connection is performed with the original input feature map of the BSM module to obtain feature maps with sequence dependency enhancement and context feature enhancement.

[0017] Furthermore, the CSCA module includes a cascaded channel attention module and a spatial attention module, wherein:

[0018] The channel attention module includes a parallel global average pooling layer and a global max pooling layer. The outputs of the two layers are respectively passed through a 1×1 convolutional layer, a ReLU activation function, and a 1×1 convolutional layer. Then the two outputs are added together, and channel weights are generated by passing through a Sigmoid activation function. These weights are then multiplied element-wise with the input features of the channel attention module and input to the spatial attention module.

[0019] The spatial attention module performs average pooling and max pooling on the input features along the channel dimension, concatenates the two feature maps along the channel dimension, and then generates spatial weights through a 7×7 convolutional layer and a sigmoid activation function. These weights are then multiplied element-wise with the input features to obtain the enhanced feature representation.

[0020] Furthermore, the DSCDC module includes a large-scale offset branch, a small-scale offset branch, and a feature fusion weight generation branch. The large-scale offset branch uses deformable convolutions with large offsets to generate large-scale offset features; the small-scale offset branch uses deformable convolutions with small offsets to generate small-scale offset features. In the feature fusion branch, the input feature map is sequentially passed through a 1×1 convolution, a 3×3 convolution, a 1×1 convolution, and a sigmoid activation function to generate feature fusion weights. Finally, the large-scale offset features and the small-scale offset features are fused according to the feature fusion weights to obtain the final output.

[0021] In a second aspect, the present invention provides a medical image segmentation system for implementing the above method, comprising:

[0022] The medical image processing module is used to acquire and preprocess medical images containing organs.

[0023] The medical image segmentation module is used to process preprocessed medical images using the validated segmentation model DDMamba to segment organ parts.

[0024] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the above-described method.

[0025] Fourthly, the present invention provides a computing device, including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the above-described method.

[0026] Compared with the prior art, the present invention has the following technical effects:

[0027] This invention effectively combines bidirectional scanning Mamba, convolutional attention, and dual-scale constrained deformable convolution. It combines Mamba's ability to capture global information remotely with deformable convolution's ability to capture more detailed local information, thereby enhancing the functionality and flexibility of the traditional encoder-decoder architecture. This invention has been applied to the field of medical image segmentation, enabling the automatic segmentation of organ parts from medical images.

[0028] This invention introduces an additional State-Space Model (SSM) branch on top of the existing Mamba architecture, enabling simultaneous scanning and feature fusion of the input image in both forward and reverse directions. This design aims to enhance the model's ability to extract features from the central region of the image and improve spatial awareness, thereby more effectively capturing key information in tasks such as multi-organ segmentation. Furthermore, a convolutional attention module is introduced after the bidirectional Mamba module to achieve effective communication between different sub-windows of Mamba, enhancing the model's ability to capture local features while simultaneously focusing on global information, thus improving the overall feature representation. This mechanism helps the model to more comprehensively understand spatial relationships and contextual information in the image. Finally, a dual-scale constrained deformable convolutional module (DSCDC) is introduced at the bottom of the encoder, effectively compensating for the shortcomings of traditional methods in multi-scale organ modeling and achieving collaborative capture of global information from large organs and local details from small organs. While ensuring controllable computational overhead, this mechanism significantly improves the model's feature representation ability and robustness for abdominal multi-organ segmentation tasks. Attached Figure Description

[0029] Figure 1 This is a structural diagram of the DDMamba network of the present invention, wherein (a) is the overall architecture diagram, (b) is the BSM module, (c) is the CSCA module, and (d) is the DSCDC module.

[0030] Figure 2 It is a true 2D medical image (Ground Truth).

[0031] Figure 3 It is a true 3D medical image (Ground Truth).

[0032] Figure 4 This is the result image after organ segmentation.

[0033] Figure 5 This is the result image after organ segmentation. Detailed Implementation

[0034] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0035] This invention provides a medical image segmentation method based on bidirectional scanning Mamba and dual-scale constrained deformable convolution, comprising the following steps:

[0036] Step 1: Preprocessing and enhancing medical image data:

[0037] Collect medical images, preprocess these images, expand the sample size, and normalize the sample images;

[0038] The medical image dataset used in this embodiment of the invention comes from the MICCAI 2022 FLARE Challenge and includes segmentation of 13 abdominal organs, including the liver, right kidney, spleen, pancreas, aorta, inferior vena cava, right adrenal gland, left adrenal gland, gallbladder, esophagus, stomach, duodenum, and left kidney. The training set contains 50 CT scans from the MSD pancreas dataset, with annotations from AbdomenCT-1K. An additional 50 cases from different medical centers were used for evaluation, with annotations provided by the challenge organizers. These images were then subjected to rotation, cropping, scaling, grayscale conversion, and normalization to expand the dataset and increase the sample size for the experiment. Finally, the images were normalized.

[0039] Step 2: Construct the segmentation model DDMamba:

[0040] like Figure 1 As shown in (a), the segmentation model DDMamba is constructed based on the U-Net encoder-decoder structure, which includes three parts: bidirectional scanning Mamba module (BSM), convolutional attention module (CSCA), and dual-scale constrained deformable convolutional module (DSCDC).

[0041] First, a CT image is input. If the image is single-channel, it is copied twice using the repeat function to expand it to three channels. Then, Input Projection is used to transform the input medical image into an input feature map. .

[0042] In the encoder:

[0043] The input feature map is processed through four layers of modules consisting of BSM and CSCA blocks, and then input into the bottom-level DSCDC module for further processing.

[0044] like Figure 1As shown in (b), in the BSM module, the input feature map is first flattened to transform the spatial dimension into a sequence form, facilitating subsequent modeling along the sequence dimension. The flattened features are divided into two parallel branches. The left branch passes through a Linear layer to complete the feature dimension mapping, a 1D convolutional layer to capture local sequence relationships, and introduces nonlinearity through the SiLU activation function. The right branch directly passes through a Linear layer to obtain another linearly transformed sequence representation, and is also activated by SiLU. The two branches generate different initial sequence features, providing input for subsequent bidirectional state space modeling.

[0045] Subsequently, sequence features from the two branches are fed into the forward SSM and backward SSM, respectively. The forward SSM scans the sequence from front to back to model forward long-range dependencies, while the backward SSM scans from back to front to capture reverse contextual information. Through bidirectional scanning, the module can simultaneously obtain sequence structure features in both forward and reverse directions.

[0046] After bidirectional modeling, the outputs of the forward and backward SSMs are multiplied element-wise with their corresponding input sequences, and the results are then added and fused to achieve sequence dependency enhancement, bidirectional information interaction, and contextual feature enhancement. The fused sequence features are then projected through a Linear layer to further adjust the feature expressive power and dimensionality.

[0047] Next, the linearly transformed sequence is restored to a spatial structure through a Reshape operation, ensuring it matches the size of the input feature map. Finally, the module performs a residual concatenation between this output and the original input feature map to obtain the final enhanced feature map, which helps stabilize training and improve feature representation capabilities.

[0048] The SSM calculation process can be summarized as a linear ordinary differential equation (ODE):

[0049]

[0050]

[0051] in Represents the state matrix, while and All of these represent projection parameters. Represented as an intermediate state, used to process the input of a one-dimensional input function or sequence. R mapped to the final output R, intermediate state The instantaneous rate of change over time.

[0052] To make SSM more suitable for deep learning tasks, it was discretized. Discretization transforms a continuous time axis into a series of discrete time points, allowing the model to be solved efficiently using numerical methods. This step significantly improves the computational efficiency of SSM when dealing with large-scale datasets. An overview of discretization is as follows:

[0053]

[0054]

[0055] The discrete process is as follows:

[0056]

[0057]

[0058] in This is a selection factor with the same shape as the input, used to control the degree of state update at each time step; The state matrix is ​​the discretized state matrix. These are the discretized projection parameters; Represents the identity matrix, whose size is the same as the state matrix. Consistency, used in matrix operations; and These represent the discrete hidden states at the current time and the previous time, respectively.

[0059] In the CSCA module, the input feature map is used. As input, the importance of different channels is first modeled using a channel attention module. Then, the weighted feature map is fed into a spatial attention module to obtain salient spatial regions, and finally, the enhanced feature representation is output. Specifically, the input features first pass through the channel attention module to generate channel weight vectors. The enhanced feature map is obtained by fusing it with the original feature map through channel-by-channel multiplication. Subsequently, the feature map enters the spatial attention module to generate a spatial weight map. The final output features are obtained by fusing them with the input features through element-wise multiplication. The entire process can be summarized as follows:

[0060]

[0061]

[0062] Among them, symbols This represents element-wise multiplication. During multiplication, attention weights are broadcast (copied) as needed. Channel attention weights are broadcast along the spatial dimension, while spatial attention weights are broadcast along the channel dimension. This represents the final optimized output.

[0063] In the channel attention module, its main function is to capture global semantic information to determine which channels are more important. This is achieved by performing global average pooling and global max pooling on the input features in the spatial dimension, resulting in two pools of size [value missing]. The vectors represent the average response and the strongest response, respectively. Then, the two vectors are passed through two layers with shared weights. Feature transformation is performed using convolution and ReLU activation functions. Next, the two outputs are added element-wise along the channel dimension, and channel attention weights are generated using a Sigmoid activation function. Ultimately, The feature map is multiplied channel by channel with the original feature map to obtain the enhanced feature map. The entire process can be summarized as follows:

[0064]

[0065] in This represents the Sigmoid activation function. This indicates a two-layer fully connected layer (used in the implementation). (using convolution instead), with shared weights in both MLP paths. and Let $\mathbf$ and $\mathbf$ represent global average pooling and global max pooling, respectively, in the spatial dimension.

[0066] In the spatial attention module, its main function is to locate salient regions of features in two-dimensional space. Its implementation involves performing average pooling and max pooling on the input features along the channel dimension, resulting in two images of size [missing information]. The two feature maps are then concatenated along the channel dimension to form a single-channel feature map. The feature map. Passed through a convolutional kernel of size... Feature fusion is performed using convolutional layers, and spatial attention weights are generated using the Sigmoid activation function. Finally, The feature map is obtained by multiplying the input features position by position to obtain the spatially augmented feature map. The entire process can be summarized as follows:

[0067]

[0068] in This represents the Sigmoid activation function. This indicates that the kernel size is Convolution operation, and Let $\mathbf$ and $\mathbf$ represent global average pooling and global max pooling, respectively, along the channel dimension.

[0069] In the DSCDC module, there are two independent offset convolution branches with different offset ranges. The large offset range branch focuses on capturing the broad contextual information of large organs, thereby obtaining a more complete structural outline and spatial relationships; the small offset range branch focuses on extracting the fine structural details of small organs, preserving their boundary and morphological features. Its convolution operation is as follows:

[0070]

[0071] in This is the center position of the current convolution kernel. For the nth sampling point in the grid, This represents the convolution weight coefficients at the corresponding point positions. The offset is calculated from the input feature map through an additional convolutional branch.

[0072] Subsequently, the feature maps are processed by a feature fusion module, and then element-wise multiplied with the outputs of the corresponding dual-scale deformable convolution branches to achieve scale-specific feature enhancement. Finally, the two feature paths are fused into the final output using a weighted summation method. The mathematical formula for this method is as follows:

[0073]

[0074] in The normalized probability value, ranging from [0, 1], is generated after the feature fusion module applies the Sigmoid function. This probability represents the confidence level of the input feature point x corresponding to a specific organ. Subsequently, it is used to guide feature allocation and weighted fusion, thereby facilitating adaptive feature extraction of organs at different scales. and These represent the output feature maps of the two constrained deformable convolution branches, respectively.

[0075] In the decoder:

[0076] It consists of four identical upsampling blocks, and its processing steps are as follows: the input feature map x is bilinearly upsampled, and its size is doubled along two dimensions. Then, the feature map skip provided by the skip connection is concatenated with the upsampled feature map x along the channel dimension. Then, the concatenated feature map x is passed to two convolutional layers for convolution operation to obtain the returned feature map.

[0077] Finally, after another upsampling, the feature map is restored to the size of the input model, and then a 3×3 convolutional layer (Output Projection) is used to transform the feature map into the output image.

[0078] Step 3: Set the training strategy and loss function, and train the model;

[0079] The preprocessed dataset is divided into training, testing, and validation sets; a combination of Dice loss and BCE loss is used to train the model, while regularization is employed to prevent overfitting; the backpropagation algorithm is used in the DDMamba network model to update the weights and biases in the network; and the loss function is used to update the parameters during the training iteration.

[0080] The Dice loss function is a commonly used metric for evaluating the performance of image segmentation models. Its core idea is to measure the spatial overlap between the predicted results and the ground truth labels, making it particularly suitable for object segmentation tasks. This function is based on the Dice coefficient (Dice similarity), which measures the similarity between two sets. Specifically, it reflects the overlap by calculating the ratio of twice the intersection of the predicted result and the ground truth label to the sum of the pixels of both. The loss value is typically defined as... The coefficient is important because a higher Dice coefficient indicates a greater overlap between the prediction and the actual result, resulting in a smaller loss value.

[0081] The optimization objective of the Dice loss function is to minimize this loss, thereby making the model output more spatially consistent with the true label. This optimization method not only improves segmentation accuracy but also preserves the integrity of object boundaries to a certain extent. Its mathematical definition is as follows:

[0082]

[0083] Where X is the predicted output and Y is the true label.

[0084] Binary Cross Entropy (BCE) loss is a widely used loss function in medical image segmentation tasks. Its principle originates from the concept of cross-entropy in information theory, used to quantify the difference between the model's prediction and the ground truth label. This method typically converts the network output into probability values ​​and compares them with the corresponding ground truth labels to calculate the loss. Specifically, in binary classification scenarios, BCE loss is processed differently depending on the label value: when the label is 1, it measures the log probability that the model predicts the sample as positive; when the label is 0, it calculates the log probability that the sample is predicted as negative.

[0085] Finally, the overall BCE loss is obtained by averaging the loss values ​​of all samples. The optimization objective of this function is to minimize the loss, so that the probability distribution predicted by the model is as close as possible to the true label, thereby improving the accuracy of the segmentation task. Its mathematical expression is as follows:

[0086]

[0087] in It's a real label. It is the probability predicted by the model.

[0088] The final loss function of the model is a combination of Dice loss and BCE loss:

[0089]

[0090] Step 4: Validate the trained network model:

[0091] The pre-segmented validation set is input into the trained segmentation model DDMamba. After the model segments the medical image, the lesion part is segmented out. The segmented image is compared with the lesion area judged by experts to evaluate and validate the network model.

[0092] Four widely adopted evaluation metrics were used to measure the performance of the EMA-TransUNet model; the evaluation metrics are shown below:

[0093] Dice similarity coefficient (DSC): DSC is a metric for measuring classification performance, typically used to calculate the similarity between two samples. The value ranges from 0 to 1, with a value of 1 for the best segmentation and 0 for the worst. The DSC metric is commonly used in fields such as medical image analysis. Its formula is:

[0094]

[0095] Its denominator includes the truly positive samples (FP+TP) and all samples that are actually positive (TP+FN). TP (True Positive) is the number of samples correctly predicted as positive, FP (False Positive) is the number of samples incorrectly predicted as positive (actually negative), and FN (False Negative) is the number of samples incorrectly predicted as negative (actually positive).

[0096] Normalized Surface Dice (NSD) is a metric used to evaluate surface similarity in medical image segmentation tasks, particularly suitable for assessing the boundary accuracy of 3D structures. Unlike traditional DSC, NSD focuses on whether the boundaries or surfaces of the segmented result accurately match the boundaries of the actual anatomical structures. Similar to DSC, NSD is normalized to the [0,1] interval, where 1 represents perfect similarity and 0 represents no overlap. Its formula is:

[0097]

[0098] in It is the surface that predicts the segmentation results. It is the surface referenced to the standard (true ground value). Indicates the allowable error Inside, predict the surface With reference surface Intersecting parts, and These represent the sizes (usually surface areas) of the predicted surface and the reference surface, respectively.

[0099] Through these comprehensive evaluation criteria, this paper can not only accurately evaluate the performance of the proposed method in medical image segmentation tasks, but also effectively compare it with other state-of-the-art methods, providing strong data support and theoretical basis for further optimization of the model.

[0100] Step 5: Input any medical image into the validated model and segment out the organ parts.

[0101] Select as Figure 2 and Figure 3 The medical images shown were tested, and the results were as follows: Figure 4 and Figure 5 As shown, the DDMamba model can accurately segment organs, providing powerful assistance to doctors in diagnosis and treatment.

[0102] Compared to traditional medical image segmentation methods, the DDMamba model has better performance, can segment organ regions more accurately, and is more suitable for different types of medical image datasets.

[0103] The accompanying drawings of the embodiments disclosed in this invention only involve structures relevant to the embodiments of this disclosure. However, the above description is merely a preferred embodiment of the present invention, and it can be fully applied to various fields suitable for the present invention. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A medical image segmentation method based on bidirectional scanning Mamba and dual-scale constrained deformable convolution, characterized in that, Includes the following steps: Medical images containing organs are acquired and preprocessed to construct a dataset; A segmentation model, DDMamba, is constructed. The DDMamba model employs an encoder-decoder architecture. The encoder consists of multiple bidirectional scanning Mamba modules and convolutional attention modules stacked together. The bidirectional scanning Mamba modules perform bidirectional scanning state space modeling on the input feature map, modeling forward long-range dependencies and capturing reverse contextual information through forward and backward SSMs, respectively, to obtain a feature-enhanced feature map. The convolutional attention module captures the global semantic information of the feature-enhanced feature map through channel attention and uses spatial attention to locate salient regions, obtaining an enhanced feature representation. This enhanced feature representation is input into a dual-scale constrained deformable convolutional module, which uses different branches to capture the wide-area contextual information and structural details of the medical image, and then inputs it into the decoder. The decoder is used to progressively restore spatial resolution and generate organ segmentation results. The dataset is divided into training, testing, and validation sets to train, test, and validate the segmentation model DDMamba. The validated segmentation model DDMamba is used to process the medical image to be segmented and output the organ segmentation results.

2. The medical image segmentation method according to claim 1, characterized in that, The specific processing steps for the bidirectional scanning Mamba module include: The input feature map is flattened and then divided into two parallel branches; the first branch passes through a linear layer, a 1D convolutional layer and a SiLU activation function in sequence; the second branch passes through a linear layer and a SiLU activation function in sequence. The outputs of the first and second branches are respectively fed into the forward SSM and the backward SSM for scanning; the outputs of the forward SSM and the backward SSM are multiplied element-wise with the corresponding gating signals, and the results are added element-wise and fused together. The fused features are projected through a linear layer and restored to spatial feature maps through a Reshape operation. Finally, a residual connection is made with the original input feature map of the bidirectional scanning Mamba module to obtain feature maps with sequence dependency enhancement and context feature enhancement.

3. The medical image segmentation method according to claim 1, characterized in that, The convolutional attention module includes a cascaded channel attention module and a spatial attention module.

4. The medical image segmentation method according to claim 3, characterized in that, The channel attention module includes a parallel global average pooling layer and a global max pooling layer. The outputs of the two layers are respectively passed through a 1×1 convolutional layer, a ReLU activation function, and a 1×1 convolutional layer. Then the two outputs are added together, and channel weights are generated by passing through a Sigmoid activation function. These weights are then multiplied element-wise with the input features of the channel attention module and input to the spatial attention module. The spatial attention module performs average pooling and max pooling on the input features along the channel dimension, concatenates the two feature maps along the channel dimension, and then generates spatial weights through a 7×7 convolutional layer and a sigmoid activation function. These weights are then multiplied element-wise with the input features to obtain the enhanced feature representation.

5. The medical image segmentation method according to claim 1, characterized in that, The dual-scale constrained deformable convolution module includes a large-scale offset branch, a small-scale offset branch, and a feature fusion weight generation branch; the large-scale offset branch uses deformable convolution with a large offset to generate large-scale offset features; the small-scale offset branch uses deformable convolution with a small offset to generate small-scale offset features. In the feature fusion branch, the input feature map is sequentially processed by 1×1 convolution, 3×3 convolution, 1×1 convolution and Sigmoid activation function to generate feature fusion weights; finally, the large-scale offset features and small-scale offset features are fused according to the feature fusion weights to obtain the final output.

6. The medical image segmentation method according to claim 1, characterized in that, The decoder consists of four upsampling modules and a final output projection layer.

7. A medical image segmentation system implementing the method according to any one of claims 1-6, characterized in that, include: The medical image processing module is used to acquire medical images and perform preprocessing. The medical image segmentation module is used to process preprocessed medical images using the validated segmentation model DDMamba to segment organ parts.

8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed in a computer, causes the computer to perform the method as described in any one of claims 1-6.

9. A computing device, characterized in that, The method includes a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the method as described in any one of claims 1-6.

Images