Medical image segmentation method based on multi-scale convolution bidirectional Mamba

Abstract

The present application relates to the technical field of image processing, in particular to a medical image segmentation method based on multi-scale convolution bidirectional Mamba. In the present application, a medical image segmentation model including an encoder, a decoder and a skip connection module is constructed; the medical image segmentation model is trained using a medical image, wherein local features are extracted through a CNN branch, and long-range context dependency relationships are captured through a multi-scale bidirectional Mamba branch; the parameters of the medical image segmentation model are updated according to a target loss, and the medical image segmentation is performed through the trained medical image segmentation model, thereby improving the effect of medical image segmentation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a medical image segmentation method based on multi-scale convolutional bidirectional Mamba. BACKGROUND

[0002] In the field of medical image segmentation, state space models (SSMs) and their variants Mamba provide novel and effective technical solutions. SSMs can model long-range dependencies with linear computational complexity by encoding input signals into latent states and relying on linear dynamical systems for their evolution, significantly reducing computational requirements while maintaining strong modeling performance. Compared with the Transformer architecture, SSMs are more efficient in high-resolution medical image processing, can effectively capture relationships between anatomical regions that are far apart in space, and can overcome the limitations of convolutional neural networks (CNNs) in capturing global context dependencies. Mamba combines a selective scanning mechanism and hardware-aware optimization to achieve linear-time processing of long sequences, combining the sequence modeling capabilities of recurrent neural networks (RNNs) with the parallelism of Transformers, providing a practical and scalable solution for long-range dependency modeling in high-resolution imaging tasks. Researchers have also proposed a variety of architectures that integrate SSMs, such as U-Mamba, which combines the global modeling of SSMs with the local feature extraction capabilities of CNNs in a U-shaped architecture, VMamba, which enhances spatial dependency modeling through a two-dimensional state space mechanism, and VM-Unet, which constructs a complete encoder-decoder framework based on SSM principles.

[0003] However, existing technical solutions based on SSMs have key limitations that hinder their practical clinical application in medical image segmentation. First, the traditional Mamba model uses a unidirectional scanning strategy, which causes information to gradually decay when modeling long-range dependencies, greatly limiting the ability to fully capture spatial relationships in complex medical images. Second, existing technical solutions struggle to balance fine-grained local feature extraction and global context information modeling, lacking a robust multi-scale feature fusion mechanism that reduces segmentation accuracy, especially when dealing with anatomical structures and pathological regions with complex morphological variations. Third, most Mamba-based architectures focus primarily on spatial domain features, with insufficient attention to key texture and boundary information encoded in the frequency domain. SUMMARY

[0004] Therefore, the present application provides a medical image segmentation method based on multi-scale convolutional bidirectional Mamba, which at least solves one problem in the prior art.

[0005] In a first aspect, the present application provides a medical image segmentation model training method based on multi-scale convolutional bidirectional Mamba, which includes the following steps:

[0006] constructing a medical image segmentation model comprising an encoder, a decoder and a skip connection module; wherein the decoder comprises a multi-scale convolutional bidirectional Mamba module, and the multi-scale convolutional bidirectional Mamba module comprises a CNN branch and a multi-scale bidirectional Mamba branch;

[0007] training the medical image segmentation model using medical images, wherein local features are extracted by the CNN branch and long-range context dependency relationships are captured by the multi-scale bidirectional Mamba branch;

[0008] updating parameters of the medical image segmentation model according to a target loss.

[0009] In some optional embodiments, the CNN branch comprises at least two convolutional layers, and a batch normalization layer and an activation layer are sequentially connected after each convolutional layer; features are sequentially convolved by the convolutional layers, batch-normalized by the batch normalization layers, and nonlinearly activated by the activation layers.

[0010] In some optional embodiments, the multi-scale bidirectional Mamba branch comprises a first layer normalization layer, a bidirectional Mamba module, a second layer normalization layer and a multi-scale feature fusion module; features are sequentially layer-normalized by the first layer normalization layer, learn feature representations by the bidirectional Mamba module, and are residual-connected, and then are sequentially layer-normalized by the second layer normalization layer, fused by the multi-scale feature fusion module, and are residual-connected again.

[0011] In some optional embodiments, operations in the bidirectional Mamba module are as shown in the following formula:

[0012] ;

[0013] ;

[0014] ;

[0015] ;

[0016] ;

[0017] wherein, represents input features of the bidirectional Mamba module, represents layer normalization, represents linear projection, represents SiLU function activation; represents a processing procedure of a forward state space model (SSM), represents an operation of reversing the order of elements in a sequence, represents a backward convolution operation, represents a process of a reverse state space model (SSM); represents a matrix element point multiplication operation, represents output features of a bidirectional Mamba module.

[0018] In some optional embodiments, operations in the multi-scale feature fusion module are shown in the following formula:

[0019] ;

[0020] ;

[0021] ;

[0022] wherein, represents input features of the multi-scale feature fusion module, represents a 3x3 convolution with a dilation rate of 2, represents a 3x3 convolution with a dilation rate of 3, represents a 1x1 convolution, represents a 3D convolution with a kernel size of 3x3x3, represents a nonlinear activation function (GeLU) activation, represents output features of the multi-scale feature fusion module.

[0023] In some optional embodiments, the skip connection module includes a multi-frequency attention (MFA) module, wherein operations in the multi-frequency attention (MFA) module include:

[0024] fusing features from the encoder and features from the decoder to obtain fused features ;

[0025] performing adaptive pooling and discrete cosine transform (DCT) on the fused features to obtain a plurality of frequency components;

[0026] performing global average pooling, global maximum pooling, and global minimum pooling on each frequency component, combining by element-wise addition, and dividing by the number of frequency components to respectively generate aggregated features;

[0027] aggregating all aggregated features, and then processing through a fully connected layer and activating through an activation function to obtain a channel attention map ;

[0028] fusing features from the encoder and features from the decoder according to the channel attention map to generate a feature map ;

[0029] The feature map is processed through a spatial attention mechanism to obtain output features of a multi-frequency attention (MFA) module.

[0030] In some optional embodiments, the target loss is calculated by the following formula:

[0031] ;

[0032] wherein, denotes the target loss; and denote hyperparameters, respectively; denotes a Dice loss function, denotes a cross-entropy loss function.

[0033] In a second aspect, the present application provides a medical image segmentation method based on a multi-scale convolutional bidirectional Mamba, which comprises the following steps:

[0034] inputting the medical image into a medical image segmentation model to output a segmentation result from the medical image segmentation model;

[0035] wherein, the medical image segmentation model is obtained by the medical image segmentation method based on the multi-scale convolutional bidirectional Mamba as described above.

[0036] In a third aspect, the present application provides a medical image segmentation device, which comprises:

[0037] at least one processor;

[0038] and a memory in communication connection with the at least one processor;

[0039] wherein, the memory stores instructions, and the instructions, when executed by the at least one processor, implement the medical image segmentation method based on the multi-scale convolutional bidirectional Mamba as described above.

[0040] In a fourth aspect, the present application provides a computer-readable storage medium, which stores instructions, and the instructions, when executed by a processor, implement the medical image segmentation method based on the multi-scale convolutional bidirectional Mamba as described above.

[0041] Due to the adoption of the above technical solutions, the embodiments of the present application at least have the following beneficial effects:

[0042] (1) The CNN branch, the multi-scale bidirectional Mamba branch and the multi-scale feature fusion module are integrated in the medical image segmentation model, which breaks through the key limitations of the existing methods in local feature extraction, global context modeling and multi-scale feature integration;

[0043] (2) The multi-scale convolutional bidirectional Mamba module integrates local detail preservation and global semantic understanding functions, and through cooperative local-global feature extraction, realizes special multi-scale aggregation of complex anatomical structures;

[0044] (3) The multi-frequency attention (MFA) module uses frequency domain representation through discrete cosine transform (DCT) and combines spatial attention to improve the feature fusion of the encoder-decoder, thereby significantly enhancing the model's attention to key areas of anatomy and diagnostics. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 It is a network structure schematic diagram of a medical image segmentation model in an embodiment of the application.

[0046] Figure 2 It is a network structure schematic diagram of a multi-scale convolutional bidirectional Mamba module in an embodiment of the application.

[0047] Figure 3 It is a network structure schematic diagram of a bidirectional Mamba module in an embodiment of the application.

[0048] Figure 4 It is a network structure schematic diagram of a multi-scale feature fusion module in an embodiment of the application.

[0049] Figure 5 It is a network structure schematic diagram of a multi-frequency attention module in an embodiment of the application.

[0050] Figure 6 It is a part of the visual segmentation results of the MCBM-MFA model and the existing medical image segmentation model on the ACDC dataset in an embodiment of the application.

[0051] Figure 7 It is a part of the visual segmentation results of the MCBM-MFA model and the existing medical image segmentation model on the ISIC2018 dataset in an embodiment of the application.

[0052] Figure 8 It is a part of the visual segmentation results of the MCBM-MFA model and the existing medical image segmentation model on the GlaS dataset in an embodiment of the application.

[0053] Figure 9 It is a part of the visual segmentation results of the MSFF module using 3D convolution and 2D convolution on the ACDC, GlaS and ISIC2018 datasets in an embodiment of the application. DETAILED DESCRIPTION

[0054] The concept of the present application and the technical effects produced will be described clearly and completely below to fully set forth the purpose, scheme and effects of the present application.

[0055] Figure 1 The network structure of a medical image segmentation model of one embodiment of the present application is shown, and the medical image segmentation model is abbreviated as an MCBM-MFA model for ease of representation.

[0056] Specifically, the MCBM-MFA model includes an encoder, a skip connection module and a decoder. The encoder can adopt a pre-trained PVTv2 (Pyramid Vision Transformer v2) backbone network, which is divided into four hierarchical stages with channel dimensions increasing in turn [64, 128, 320, 512]. Through the stacking of convolution and step operation, the encoder realizes effective multi-scale feature extraction, while reducing the spatial resolution by half at each stage to obtain feature maps with input resolutions of [1 / 4, 1 / 8, 1 / 16, 1 / 32].

[0057] The decoder includes an up-sampling layer and a multi-scale convolutional bidirectional Mamba (MCBM) module. Each MCBM module contains two parallel branches: a CNN-based branch (CNN branch) for local detailed feature extraction and a multi-scale bidirectional Mamba branch for capturing long-range contextual dependencies. The decoder contains four stages with channel dimensions decreasing in turn [512, 320, 128, 64], gradually recovering the spatial resolution. At each decoding stage, the features from the encoder and the skip connection module are independently processed by the two branches, and then fused to realize the combination of fine-grained local features and high-level semantic context. This collaborative design enhances the semantic consistency under multi-scale.

[0058] To further strengthen the interaction between the encoder and the decoder, a multi-frequency attention (MFA) module is introduced as an enhanced skip connection mechanism. That is, the skip connection module includes the MFA module. The MFA module converts spatial features to the frequency domain through discrete cosine transform (DCT) and combines with the spatial attention mechanism to realize more effective feature alignment and texture preservation. This design significantly improves the segmentation accuracy and boundary outlining effect while maintaining computational efficiency.

[0059] As Figure 2As shown, the MCBM module includes a CNN branch and a multi-scale bidirectional Mamba branch, which synergistically integrates the local modeling advantage of CNN with the global dependency modeling capability of state-space models. Among them, the CNN branch adopts a double-layer 3x3 convolution architecture, and a batch normalization (BN) layer and an activation layer are connected in turn after each convolution layer. The activation layer can be activated using the ReLU function. This design gradually extracts hierarchical local features, which can effectively capture low-level texture patterns and high-level edge contours while preserving morphological structures. The stacked convolution layers with nonlinear activation generate multi-scale receptive fields, producing fine-grained visual representations, which are particularly critical for accurate medical image interpretation and analysis.

[0060] The multi-scale bidirectional Mamba branch includes a bidirectional Mamba (BM) module and a multi-scale feature fusion (MSFF) module, and a layer normalization layer (LayerNorm) is arranged before the bidirectional Mamba module and the multi-scale feature fusion module. The BM module learns sequential feature representations, significantly enhancing the model's ability to capture long-range dependencies and global context information. At the same time, the MSFF module aggregates features of multiple receptive field scales, achieving robust semantic integration and improving adaptability to different target sizes and complex anatomical structures.

[0061] The processing process of the MCBM module at the i-th stage is shown in the following formula:

[0062] ;

[0063] ;

[0064] ;

[0065] ;

[0066] wherein, represents the input feature of the i-th stage MCBM module, and i is 1, 2, 3 or 4 because the decoder contains four stages; represents sequentially performing 3x3 convolution, batch normalization and ReLU function activation; , , respectively represent bidirectional Mamba module processing, layer normalization and multi-scale feature fusion module processing; represents the output feature of the i-th stage MCBM module.

[0067] As shown in Figure 3 , the bidirectional Mamba module adopts a double-branch architecture. The first branch linearly projects the input feature, and then uses the SiLU function for nonlinear activation, as shown in the following formula:

[0068] ;

[0069] wherein, represents the input feature of the bidirectional Mamba module, , H, W, C represent the height, width, and channel number of the feature map, respectively; represents layer normalization; represents linear projection; represents SiLU function activation.

[0070] The second branch first obtains a sequence feature by linear projection, and then the sequence is processed by forward and reverse state space model (SSM) modules, respectively, as shown in the following equations:

[0071] ;

[0072] ;

[0073] ;

[0074] wherein, represents the processing procedure of the forward state space model (SSM), represents the operation of reversing the order of elements in the sequence, represents backward convolution operation, represents the processing procedure of the reverse state space model (SSM).

[0075] Finally, the output of the bidirectional Mamba module is calculated by the following equation:

[0076] ;

[0077] wherein, represents matrix element point multiplication operation, represents the output feature of the bidirectional Mamba module.

[0078] The bidirectional Mamba module simultaneously models the forward and reverse context dependencies, thereby being able to learn richer and more comprehensive semantic representations. This bidirectional scanning mode significantly enhances the ability of the model to learn spatially sensitive features, and has higher representation fidelity compared with unidirectional models.

[0079] As shown in Figure 4 , the MSFF module adopts a double-branch architecture. The main branch first expands the channel dimension through a 1x1 convolutional layer, and then performs a series of multi-scale dilated convolutions to achieve comprehensive feature extraction. These dilated convolutions adopt a consistent 3x3 kernel size but have different dilation rates in order to extract fine-grained local features at different spatial resolutions. For the input feature The operation of the first branch is shown in the following equation:

[0080] ;

[0081] where, and denote 3x3 convolutions with dilation rates of 2 and 3, respectively, denotes a standard 1x1 convolution.

[0082] In the second branch, 3D convolutional layers are adopted to capture and model the contextual information and spatial structure, followed by a nonlinear activation function (GeLU) activation, as shown in the following equation:

[0083] ;

[0084] where, denotes a 3D convolution operation with a kernel size of 3x3x3.

[0085] To obtain the final output of the MSFF module, the features of the two branches are fused by element-wise multiplication, as shown in the following equation:

[0086] ;

[0087] where, denotes the output feature of the MSFF module.

[0088] The MSFF module aggregates features from multiple receptive field scales to extract common complex multi-scale contextual dependencies in medical images, enhancing the network's nonlinear representation ability and spatial modeling ability.

[0089] As shown in Figure 5 , the MFA module first applies a discrete cosine transform (DCT) to the fused features of the encoder and decoder. This frequency domain decomposition can more effectively model multi-scale anatomical structures, covering a range from small pathological lesions to large organ systems. Let and denote the output features of the encoder and decoder, respectively, they are first concatenated to obtain the fused feature , which is then adjusted to a fixed size by adaptive pooling, as shown in the following equation:

[0090] ;

[0091]

[0092] where, denotes concatenation, denotes adaptive pooling.

[0093] Next, frequency features are extracted using a set of predefined DCT filters, each corresponding to a specific coordinate in the frequency domain As shown in the following equation:

[0094] ;

[0095] wherein, u and v represent horizontal and vertical frequency components in the frequency domain, respectively; represents the j th DCT filter, which is defined as shown in the following equation:

[0096] ;

[0097] ;

[0098] ;

[0099] wherein, and represent coordinate positions in the spatial domain.

[0100] For each frequency component, global average pooling, global max pooling and global min pooling operations are applied, and then all the obtained pooling features are combined by element-wise addition and divided by the number of frequency branches, respectively, to generate aggregated features , and , as shown in the following equations:

[0101] ;

[0102] ;

[0103] ;

[0104] wherein, n represents the number of frequency branches, represents the global average pooling operation, represents the global max pooling operation, represents the global min pooling operation.

[0105] Then, , and are aggregated, input to a fully connected layer and processed by an activation function to generate a channel attention map M, as shown in the following equation:

[0106] ;

[0107] ;

[0108] where, , and denote Sigmoid function activation, 1x1 convolution and ReLU function activation, respectively.

[0109] The channel attention map M is combined with the encoder output features to generate the feature map as shown in the following equation:

[0110] ;

[0111] ;

[0112] where, a denotes an adaptive adjustment factor. The adaptive adjustment factor is a learnable parameter, which dynamically adjusts the channel attention distribution, thereby realizing adaptive and context-aware feature fusion, achieving optimal balance between different frequency feature components, and enhancing the adaptability of the model in various medical image segmentation tasks.

[0113] Subsequently, to further address the imbalance of spatial information after frequency domain fusion and enhance the sensitivity to key anatomical regions, the MFA module integrates a spatial attention module. This spatial attention module improves the contour delineation of lesion boundaries and other salient structures by adaptively re-weighting the spatial features. Specifically, it first computes the max-pooling and average-pooling in the channel dimension in parallel to extract complementary spatial information. Then, by applying a 7x7 convolution to the concatenated features, the model can encode fine-grained spatial interactions. Next, the Sigmoid activation function generates normalized attention weights in the range [0, 1], and the resulting attention map highlights regions with spatial information, guiding the model to focus on anatomically relevant regions and thus improving segmentation performance. The calculation of the spatial attention weight is shown in the following equation:

[0114] ;

[0115] where, denotes the input of the spatial attention module, denotes the Sigmoid activation function, denotes the 7x7 convolution, denotes the max-pooling in the channel dimension, denotes the average-pooling in the channel dimension.

[0116] Finally, the spatial attention weight is combined with the feature map and the encoder output feature are multiplied to generate their respective weighted representations, which are fused by element-wise addition as shown in the following formula:

[0117]

[0118] wherein, represents the output feature of the MFA module.

[0119] The parameters of the medical image segmentation model are updated according to a composite loss (target loss), and the composite loss is defined as shown in the following formula:

[0120]

[0121] wherein, represents the composite loss; and respectively represent hyperparameters; represents a Dice loss function, represents a cross-entropy loss function.

[0122] In order to verify the effect of the medical image segmentation model in the embodiment of the application, experiments are performed using the GlaS dataset, the ISIC2018 dataset and the ACDC dataset, and compared with a plurality of existing medical image segmentation models.

[0123] The gland segmentation (GlaS) dataset is a widely recognized benchmark dataset in the field of colon gland segmentation, and is derived from the MICCAI gland segmentation challenge in 2015. The dataset contains 165 hematoxylin-eosin (H&E) stained microscopic images of colon tissue sections, covering benign and malignant samples. Each image is annotated by a professional pathologist pixel by pixel to outline the gland structure. In order to ensure consistency in training and evaluation, all images are adjusted to 224x224 pixels and divided into 85 training samples and 80 test samples. The experiment adopts a five-fold cross-validation scheme to comprehensively evaluate the segmentation performance and generalization ability.

[0124] The ISIC2018 dataset is compiled by the International Skin Imaging Collaboration (International Skin Imaging Collaboration) and is a benchmark dataset for skin lesion segmentation and melanoma detection. The dataset contains 2594 images, covering clinical photos and dermatoscope images, and the lesion masks are annotated by dermatologists. In order to ensure consistency, each image is adjusted to 224x224 pixels. The dataset is divided into training set, validation set and test set in the ratio of 70 / 10 / 20. The experiment adopts a five-fold cross-validation to comprehensively evaluate the generalization ability of the model.

[0125] ​The ACDC dataset was released at MICCAI 2017 conference, providing comprehensive cardiac MRI labeled images for segmentation tasks. The dataset contains short-axis cine MRI sequences of 100 subjects, covering healthy individuals and patients with right ventricular (RV), left ventricular (LV), and myocardial (Myo) related diseases. To standardize the input dimension, all slices are adjusted to 224x224 pixels. The dataset is divided into 70 training samples, 10 validation samples, and 20 test samples.

[0126] The image segmentation quality is evaluated by the Dice coefficient, mIoU, and HD95 key indicators. Among them, the Dice coefficient is used to quantify the similarity between the segmentation results and the corresponding true labels. mIoU (mean intersection over union) is used to measure the average overlap between the predicted segmentation mask and the true mask. The values of Dice coefficient and mIoU are between 0 and 1, and the higher the value, the closer the segmentation result to the true label, indicating that the model has higher accuracy and reliability. 95% Hausdorff distance (HD95) is used to quantify the maximum boundary difference between the predicted segmentation and the true label, and the smaller the value, the higher the boundary fitting degree. The specific calculation of these indicators is shown in the following formula:

[0127] ;

[0128] ;

[0129] ;

[0130] where TP represents the number of pixels correctly predicted by the model as target regions, FP represents the number of pixels incorrectly predicted by the model as target regions, and FN represents the number of pixels incorrectly predicted by the model as background; y represents the true value, represents the predicted segmentation result, and respectively represent the one-way Hausdorff distance from the true value to the prediction result and the predicted segmentation result to the true value.

[0131] The experiment is based on the PyTorch framework, and all training and testing are completed on an NVIDIA RTX4080 GPU with 16GB of video memory. As shown in Table 1, the input image is standardized to a resolution of 224x224; in terms of optimization, the AdamW algorithm is used, with an initial learning rate of 0.001 to speed up convergence and improve generalization performance.

[0132] Table 1 Hyperparameters used in the experiment

[0133]

[0134] Table 2 shows the Dice coefficient results of the MCBM-MFA model and existing medical image segmentation models on the ACDC dataset in embodiments of the present application. The Dice scores of the MCBM-MFA model on the right ventricle (RV), myocardium (Myo), and left ventricle (LV) are 89.96%, 89.58%, and 95.99%, respectively, with an average Dice score of 91.83%. Compared with the powerful Transformer-based models (such as PVT-CASCADE and CSWin-UNet, both achieving 91.46%), the MCBM-MFA model achieves a relative improvement of 0.37%, which indicates its stronger ability in accurately segmenting cardiac substructures with diverse morphologies and scales. In addition, the MCBM-MFA model outperforms various hybrid CNN-Transformer architectures, including TransUNet (89.71%) and MISSFormer (90.86%), highlighting the effectiveness of the proposed bidirectional Mamba module and multi-scale feature fusion module in jointly capturing local texture and global contextual dependencies. Furthermore, although pure Transformer-based models (such as Swin-UNet, CSWin-UNet, and PVT-CASCADE) generally have better global receptive fields than traditional CNNs, they still have limitations in modeling fine-grained anatomical details, especially in regions with ambiguous boundaries or low contrast. In contrast, the MCBM-MFA model effectively compensates for this gap by integrating frequency domain attention mechanisms and spatial context modulation, enabling accurate boundary delineation and better discrimination of subtle structural changes. These results collectively demonstrate the robustness and generalization ability of the MCBM-MFA model in the cardiac MRI segmentation task.

[0135] Table 2 Dice coefficient results of various medical image segmentation models on the ACDC dataset

[0136]

[0137] Figure 6 Table 2 shows the Dice coefficient results of the MCBM-MFA model and existing medical image segmentation models on the ACDC dataset in embodiments of the present application. The Dice scores of the MCBM-MFA model on the right ventricle (RV), myocardium (Myo), and left ventricle (LV) are 89.96%, 89.58%, and 95.99%, respectively, with an average Dice score of 91.83%. Compared with the powerful Transformer-based models (such as PVT-CASCADE and CSWin-UNet, both achieving 91.46%), the MCBM-MFA model achieves a relative improvement of 0.37%, which indicates its stronger ability in accurately segmenting cardiac substructures with diverse morphologies and scales. In addition, the MCBM-MFA model outperforms various hybrid CNN-Transformer architectures, including TransUNet (89.71%) and MISSFormer (90.86%), highlighting the effectiveness of the proposed bidirectional Mamba module and multi-scale feature fusion module in jointly capturing local texture and global contextual dependencies. Furthermore, although pure Transformer-based models (such as Swin-UNet, CSWin-UNet, and PVT-CASCADE) generally have better global receptive fields than traditional CNNs, they still have limitations in modeling fine-grained anatomical details, especially in regions with ambiguous boundaries or low contrast. In contrast, the MCBM-MFA model effectively compensates for this gap by integrating frequency domain attention mechanisms and spatial context modulation, enabling accurate boundary delineation and better discrimination of subtle structural changes. These results collectively demonstrate the robustness and generalization ability of the MCBM-MFA model in the cardiac MRI segmentation task.

[0138] The evaluation on the ISIC2018 dataset adopts five-fold cross-validation, and the results are summarized in Table 3. It can be seen that the MCBM-MFA model performs strong and stable performance in the skin lesion segmentation task, and is superior to many existing medical image segmentation models in both mIoU and Dice indicators. Specifically, in terms of mIoU, the MCBM-MFA model achieves a score of 84.95%, which is 0.49% higher than that of CFA TransUnet; compared with the classic U-Net, the improvement reaches 3.72%, which indicates that the MCBM-MFA model has significant effect in accurately outlining the lesion boundary and reducing background interference. In terms of Dice, the MCBM-MFA model ranks first with a score of 91.08%, which is significantly higher than the second-ranked CFA TransUnet; compared with I2U-Net and Swin-Unet, the MCBM-MFA model achieves performance improvement of 0.98% and 1.47% respectively, which indicates that the MCBM-MFA model has more excellent segmentation reliability and consistency when facing various lesion morphologies and contrast changes.

[0139] Table 3 mIoU and Dice coefficient results of each model on the ISIC2018 dataset

[0140]

[0141] Figure 7 The qualitative segmentation results of each model on the ISIC2018 skin lesion dataset are shown, highlighting the superior performance of the MCBM-MFA model compared to models such as U-Net, UNet++, MFCNet and HSH-UNet. In the five representative cases with different lesion complexity, the MCBM-MFA model always shows significant advantages in boundary accuracy and morphological fidelity.

[0142] The comprehensive evaluation results of each model on the GlaS dataset are shown in Table 4, which shows that the MCBM-MFA model is superior to many existing CNN-based, Transformer-based and hybrid architecture-based models in segmentation performance. Specifically, the MCBM-MFA model achieves the highest Dice coefficient (91.50%) and the lowest HD95 (6.08mm). Compared with the powerful MMUNet, the Dice coefficient of the MCBM-MFA model is increased by 0.33% and the HD95 is reduced by 0.26mm, which confirms its enhanced ability to accurately outline the glandular structure in complex histopathology images.

[0143] Table 4 HD95 and Dice coefficient results of each model on the GlaS dataset

[0144]

[0145] Figure 8 Qualitative results of gland segmentation on GlaS dataset are demonstrated, highlighting the superior performance of the MCBM-MFA model compared to several existing medical image segmentation models, especially in handling complex pathological tissue structures.

[0146] To explore the individual effects of each component in the MCBM-MFA model, ablation experiments were conducted on the ACDC and GlaS datasets. Specifically, six model variants were designed to systematically evaluate the individual and combined effects of the MFA module and the MCBM module. In addition, the MCBM module was split into two independent components (the MSFF module and the bidirectional Mamba module) to evaluate their respective effects. Based on the MCBM-MFA model, six variant models were obtained by selectively removing, replacing, or modifying modules to systematically test the influence of each module:

[0147] Model I: Replace the MSFF module with a traditional MLP module to test the effectiveness of the multi-scale feature fusion mechanism;

[0148] Model II: Replace the bidirectional Mamba module with a unidirectional Mamba to evaluate the impact of bidirectional modeling capability;

[0149] Model III: Replace the MFA module with simple feature addition (Add) to evaluate the role of multi-level feature aggregation;

[0150] Model IV: Replace the MFA and MSFF modules with simple feature addition (Add) and MLP modules, respectively, to verify the joint effect of feature aggregation and multi-scale fusion;

[0151] Model V: Replace the MFA and bidirectional Mamba modules with simple feature addition (Add) and unidirectional Mamba, respectively, to test the comprehensive impact on feature aggregation and sequence modeling;

[0152] Model VI: Replace the MSFF and bidirectional Mamba modules with MLP and unidirectional Mamba, respectively, to evaluate the joint effect of multi-scale feature fusion and bidirectional sequence modeling.

[0153] As shown in Table 5, on the ACDC dataset, the complete MCBM-MFA model achieved the most advanced performance in all evaluation indicators, with an overall Dice coefficient of 91.83%, and specific scores for each chamber: right ventricle (RV) 89.96%, myocardium (Myo) 89.58%, and left ventricle (LV) 95.99%. Compared with the six variant models, the Dice score of the complete model increased by 0.16%, 0.46%, 0.39%, 0.23%, 0.15%, and 0.86%, respectively. Similarly, the evaluation on the GlaS dataset further confirmed the robustness and effectiveness of the proposed architecture; the MCBM-MFA model achieved the highest Dice coefficient of 91.50% and the lowest HD95 of 6.08mm, which exceeded the corresponding variant models by 0.36%, 0.54%, 0.25%, 1.14%, 0.78%, and 0.72%, respectively. The consistent quantitative improvement on the two datasets collectively proves the indispensability of MFA, MSFF, and bidirectional Mamba modules and their synergistic effect.

[0154] Table 5 Ablation experiment results

[0155]

[0156] To explore the impact of different encoders on the performance of the MCBM-MFA model, comprehensive experiments were conducted on the GlaS and ISIC2018 datasets using five existing pre-trained encoders, and the results are summarized in Table 6. PVTv2 achieved the highest performance, achieving a Dice coefficient of 91.50% and an HD95 of 6.08mm on the GlaS dataset, and a Dice coefficient of 91.08% and an mIoU of 84.95% on the ISIC2018 dataset. PVTv2 significantly reduces computational complexity while efficiently preserving multi-scale feature extraction, enabling better segmentation of multi-scale targets in medical imaging tasks and precise boundary positioning.

[0157] Table 6 Impact of different encoders on MCBM-MFA performance

[0158]

[0159] On the ACDC, GlaS, and ISIC2018 datasets, the performance difference between 3D convolution and 2D convolution in the MSFF module was evaluated. As shown in Table 7 and Figure 9 3D convolution consistently outperformed 2D convolution in all evaluation indicators, indicating that the 3D convolution architecture can better capture multi-dimensional spatial context and improve the segmentation accuracy of volumetric medical imaging tasks and two-dimensional-based medical imaging tasks.

[0160] Table 7 Comparison of 2D convolution and 3D convolution in MSFF module

[0161]

[0162] To evaluate the computational efficiency of the MCBM-MFA model, the system compared the trainable parameters and floating-point operations per second (FLOPs) of several representative medical image segmentation models, and the detailed results are shown in Table 8. The MCBM-MFA model uses 48.62M parameters and requires 10.96 GFLOPs. Although it is not the lightest model, it effectively balances efficiency and strong representation ability.

[0163] Table 8 Trainable parameters and floating-point operations per second of each model

[0164]

[0165] The above is only a preferred embodiment of the present application, and the present application is not limited to the above-mentioned embodiments, as long as the same or equivalent means achieve the technical effects of the present application, which should belong to the protection scope of the present application. Within the protection scope of the present application, the technical solutions and / or embodiments can have various modifications and changes.

Claims

1. A training method for a medical image segmentation model based on multi-scale convolutional bidirectional Mamba, characterized in that, Includes the following steps: A medical image segmentation model is constructed, comprising an encoder, a decoder, and a skip connection module; wherein the decoder includes a multi-scale convolutional bidirectional Mamba module, and the multi-scale convolutional bidirectional Mamba module includes a CNN branch and a multi-scale bidirectional Mamba branch; The medical image segmentation model is trained using medical images, where local features are extracted through CNN branches and long-range contextual dependencies are captured through multi-scale bidirectional Mamba branches. The parameters of the medical image segmentation model are updated based on the target loss; The multi-scale bidirectional Mamba branch includes a first normalization layer, a bidirectional Mamba module, a second normalization layer, and a multi-scale feature fusion module. Features are sequentially normalized through the first normalization layer, learned through the bidirectional Mamba module, and then connected via residuals. They are then sequentially normalized through the second normalization layer, fused through the multi-scale feature fusion module, and finally connected via residuals. The operations in a bidirectional Mamba module are shown in the following formula: ； ； ； ； ； in, This indicates the input characteristics of a bidirectional Mamba module. Representation layer normalization, Indicates linear projection. Indicates SiLU function activation; This represents the processing procedure for the forward state-space model. This indicates an operation that reverses the order of elements in a sequence. This represents the processing procedure of the inverse state-space model; This represents the element-wise multiplication operation of matrices. This indicates the output characteristics of a bidirectional Mamba module; The skip connection module includes a multi-frequency attention module, wherein the operations in the multi-frequency attention module include: Features from the encoder and decoder features By performing fusion, fusion characteristics are obtained. ; Fusion features Adaptive pooling and discrete cosine transform are performed to obtain multiple frequency components; For each frequency component, global average pooling, global max pooling, and global min pooling are performed. The components are then combined by adding them element by element and divided by the number of frequency components to generate aggregated features. All aggregated features are aggregated, then processed through a fully connected layer and activated by an activation function to obtain the channel attention map. ; According to the channel attention map Fuse features from the encoder and decoder features Generate feature maps ; Processing feature maps using spatial attention mechanism The output features of the multi-frequency attention module are obtained.

2. The method according to claim 1, characterized in that, The CNN branch includes at least two convolutional layers, and each convolutional layer is followed by a batch normalization layer and an activation layer in sequence; features are sequentially convolved by the convolutional layer, batch normalized by the batch normalization layer, and non-linearly activated by the activation layer.

3. The method according to claim 1, characterized in that, The operations in the multi-scale feature fusion module are shown in the following equation: ； ； ； in, This represents the input features of the multi-scale feature fusion module. This represents a 3×3 convolution with a hole ratio of 2. This represents a 3×3 convolution with a hole rate of 3. Represents a 1×1 convolution. This represents a 3D convolution with a kernel size of 3×3×3. This indicates activation by a non-linear activation function. This represents the output features of the multi-scale feature fusion module.

4. The method according to claim 1, characterized in that, The target loss is calculated by the following formula: ； in, Indicates target loss; and These represent hyperparameters; Represents the Dice loss function. This represents the cross-entropy loss function.

5. A medical image segmentation method based on multi-scale convolutional bidirectional Mamba, characterized in that, Includes the following steps: The medical image is input into the medical image segmentation model, and the medical image segmentation model outputs the segmentation result; The medical image segmentation model is obtained by the medical image segmentation model training method based on multi-scale convolutional bidirectional Mamba as described in any one of claims 1-4.

6. A medical image segmentation device, characterized in that, include: At least one processor; and memory that is communicatively connected to at least one processor; The memory stores instructions that, when executed by at least one processor, implement the medical image segmentation method based on multi-scale convolutional bidirectional Mamba as described in claim 5.

7. A computer-readable storage medium, characterized in that, The system stores instructions that, when executed by a processor, implement the medical image segmentation method based on multi-scale convolutional bidirectional Mamba as described in claim 5.

Images