A medical image segmentation model and a segmentation method

Abstract

The application discloses a medical image segmentation model and a segmentation method, and relates to the field of medical image processing.The model comprises an encoder, a decoder and N jump connection modules.The encoder comprises N image splicing modules connected in sequence, and each image splicing module comprises an HCM block and a patch merging layer connected in sequence.The decoder comprises N image expansion modules connected in sequence, and the i-th image expansion module comprises a patch expansion layer and an HCM block connected through the i-th jump connection module.The HCM block in the i-th image expansion module is connected with the HCM block in the N+1-i-th image splicing module through the i-th jump connection module.The HCM block comprises a VSS block and a CNN model connected in parallel.The patch merging layer in the last image splicing module in the encoder is connected with the patch expansion layer in the first image expansion module in the decoder.The application can improve segmentation accuracy and reduce computational complexity.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical image processing, in particular to a medical image segmentation model and a segmentation method. BACKGROUND

[0002] Medical image segmentation is usually implemented using deep learning models represented by convolutional neural networks and Vision Transformers. Since the UNet model was introduced in 2015, CNN has attracted widespread attention from researchers in the field of image segmentation due to its excellent image feature extraction capability. However, the convolution operation itself has some limitations, such as only being able to capture local feature information when processing global dependencies, and the size of the receptive field being limited. In contrast, segmentation models based on the Vision Transformer architecture can effectively capture non-local dependencies between objects in an image through global self-attention mechanisms, thereby significantly improving segmentation accuracy. However, at the same time, the Vision Transformer model also brings higher computational complexity due to its self-attention characteristics. Therefore, it is very important to design a medical image segmentation model that can improve segmentation accuracy and reduce computational complexity. SUMMARY

[0003] The purpose of the present application is to provide a medical image segmentation model and a segmentation method that can improve segmentation accuracy and reduce computational complexity.

[0004] To achieve the above-mentioned purpose, the present application provides the following solutions: In a first aspect, the present application provides a medical image segmentation model, comprising an encoder, a decoder and N jump connection modules; the encoder comprises N image stitching modules connected in turn, and each image stitching module comprises an HCM block and a patch merging layer connected in turn; the decoder comprises N image expansion modules connected in turn, and the i-th image expansion module comprises a patch expansion layer and an HCM block connected through the i-th jump connection module; the HCM block in the i-th image expansion module is connected with the HCM block in the N+1-i-th image stitching module through the i-th jump connection module; the HCM block comprises a VSS block and a CNN model connected in parallel; the patch merging layer in the last image stitching module in the encoder is connected with the patch expansion layer in the first image expansion module in the decoder.

[0005] In a second aspect, the present application provides a medical image segmentation method, comprising: obtaining a medical image; inputting the medical image into the medical image segmentation model described above to obtain a segmentation result.

[0006] According to the specific embodiments provided in the application, the application has the following technical effects: the application provides a medical image segmentation model and a segmentation method, the HCM block in the medical image segmentation model provided by the application is composed of a VSS block that can extract global context information and a CNN model that focuses on capturing local features, both context information and local features can be focused on, the purpose of improving segmentation accuracy is achieved, and the processing efficiency is improved by connecting the VSS block and the CNN model in parallel, and the computational complexity can be reduced. BRIEF DESCRIPTION OF DRAWINGS

[0007] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed to be used in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0008] Figure 1 The structural diagram of the medical image segmentation model provided by an embodiment of the application.

[0009] Figure 2 The HCM block structural diagram provided by an embodiment of the application.

[0010] Figure 3 The skip connection module structural diagram provided by an embodiment of the application.

[0011] Figure 4 The actual segmentation diagram of each model on the ISIC2018 dataset.

[0012] Figure 5 The actual segmentation diagram of each model on the ACDC dataset. DETAILED DESCRIPTION

[0013] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only some of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0014] In order to make the above-mentioned purposes, features and advantages of the application more obvious and easy to understand, the application will be further described in detail below with reference to the drawings and specific embodiments.

[0015] In an exemplary embodiment, a brand-new U-shaped hybrid CNN-Mamba network model, HCMUNet, is proposed, which combines the advantages of CNN in extracting local features and the ability of Mamba in capturing long-range dependencies, thereby efficiently learning local and global information and significantly reducing computational complexity. As shown in Figure 1 A medical image segmentation model is provided, comprising an encoder, a decoder and N skip connection modules (SCDA); the encoder comprises N image stitching modules connected in sequence, and each image stitching module comprises an HCM block (HCM Block) and a patch merging layer (Patch Merging) connected in sequence; the decoder comprises N image expansion modules connected in sequence, and the i th image expansion module comprises a patch expanding layer (Patch Expanding) and an HCM block connected through the i th skip connection module; the HCM block in the i th image expansion module is connected with the HCM block in the N+1-i th image stitching module through the i th skip connection module; the HCM block comprises a VSS block and a CNN model connected in parallel; The patch merging layer in the last image stitching module in the encoder is connected with the patch expanding layer in the first image expansion module in the decoder. The serial numbers of the image stitching modules are numbered according to the connection order in the encoder, and the serial numbers of the image expansion modules are numbered according to the connection order in the decoder.

[0016] The HCM block in the medical image segmentation model provided by the present application is composed of a VSS block that can extract global context information and a CNN model that focuses on capturing local features, which can focus on both context information and local features, achieve the purpose of improving segmentation accuracy, and improve processing efficiency by connecting the VSS block and the CNN model in parallel, which can reduce the computational complexity.

[0017] In another exemplary embodiment of the present application, the encoder further comprises a patch embedding module (Patch Embedding), and the output end of the patch embedding module is connected with the input end of the HCM block in the first image stitching module.

[0018] In another exemplary embodiment of the present application, as Figure 2As shown, the HCM block further comprises a Split module (Split), a concatenation layer (Concat), a channel shuffle mechanism, and a first concatenation operation; an input end of the Split module and a first input end of the first concatenation operation are input ends of the HCM block, output ends of the Split module are connected with an input end of a VSS block and an input end of a CNN model respectively, an output end of the VSS block and an output end of the CNN model are connected with an input end of the concatenation layer, an output end of the concatenation layer is connected with an input end of the channel shuffle mechanism, an output end of the channel shuffle mechanism is connected with a second input end of the first concatenation operation, and an output end of the first concatenation operation is an output end of the HCM block. The HCM block is a high-efficiency double-branch module. The working process of the HCM block is generally as follows: the input feature map is divided into two groups according to the channel dimension, and is processed through a convolution branch (CNN module) and a structured branch (VSS module) respectively. The convolution branch focuses on capturing local features, while the structured branch is responsible for extracting global context information. The processing results of the two branches are then recombined through channel merging to restore the original channel dimension of the feature map. In order to further improve the information flow and reduce the loss of features caused by channel segmentation, the present application introduces a channel shuffle mechanism, which rearranges the feature map in the channel dimension, effectively promoting the fusion of information between different groups and avoiding the isolation of information between channels caused by grouped convolution.

[0019] In another exemplary embodiment of the present application, in the design of the HCM block, the present application carefully selects the activation functions in the convolution branch and the structured branch. Specifically, ReLU is used as the activation function in the convolution branch to enhance the non-linear expression ability of the network; and SiLU activation function is used in the structured branch to improve the training stability and performance of the network with its smooth and non-monotonic characteristics. Such configuration enables the HCM block to effectively extract and fuse multi-scale features while maintaining lightweight, providing a solid foundation for the overall performance of the medical image segmentation model.

[0020] Suppose a feature input is and the feature output after processing by the HCM block is First, divide X into X1 and X2 according to the number of channels to enter different branches for feature extraction. In order to ensure compatibility with convolution operations, the original feature map is rearranged for subsequent convolution processing. The algorithm flow of the HCM block is as follows, where BN represents Batch Normalization and LN represents Layer Normalization.

[0021]

[0022] In another example embodiment of the present application, the medical image segmentation model further comprises: a bottleneck layer; and a patch merging layer in the last image patching module of the encoder is connected to a patch expanding layer in the first image expanding module of the decoder through the bottleneck layer. The resolution and dimension of the feature map in the bottleneck layer remain unchanged to focus on the extraction of global features.

[0023] In another example embodiment of the present application, the bottleneck layer comprises a plurality of HCM blocks connected in sequence.

[0024] In another example embodiment of the present application, the decoder further comprises: a first patch expanding layer; an output end of the first patch expanding layer is connected to an output end of the HCM block in the Nth image expanding module, and an output end of the first patch expanding layer is an output end of the decoder.

[0025] In another example embodiment of the present application, the medical image segmentation model further comprises: a linear projection module; and an input end of the linear projection module is connected to an output end of the decoder.

[0026] Channel attention focuses on strengthening important feature channels in the feature map while suppressing unimportant channels. This is achieved by assigning weights to each channel of the feature map, allowing the network to focus on more critical information for the task. For example, SENET effectively improves the discriminative ability of the network by adaptively recalibrating channel feature responses. Spatial attention, on the other hand, focuses on spatial positions in the feature map by learning the importance of different regions in the image to enhance the network's sensitivity to spatial information. Spatial attention can help the network focus on key regions in the image and ignore irrelevant background information, thereby improving the accuracy of image recognition. For example, STN can convert spatial deformation data and automatically capture features of important regions. CBAM attention mechanism fuses channel attention and spatial attention in a concatenated manner to capture the complex dependency between channel and spatial position, improving the model's ability to capture important features without significantly increasing computational burden. BAM processes channel attention and spatial attention in parallel to further improve the richness of feature expression. ECA-Net proposes a channel attention module without dimension reduction, which realizes efficient channel feature recalibration through simple convolution operations. Compared with using a single attention mechanism, the combination of channel attention and spatial attention can enhance the model's ability to capture contextual features. For example, Figure 3As shown, in another example embodiment of the present application, the skip connection module is a skip connection module with double attention, including a second concatenation operation, a CBAM attention mechanism and a linear layer (Linear) connected in sequence. This module can effectively make up for the loss of spatial information caused by downsampling due to the use of the CBAM attention mechanism to fuse multi-scale features, and enhance the global dimensional interaction between the layers of the decoder, improving the ability of cross-dimensional feature interaction, thereby better recovering the detailed information of the image while generating the output mask.

[0027] In actual applications, the activation function in the skip connection module uses ReLU.

[0028] The working process of the skip connection module provided by the present application is as follows: first, the skip connection module obtains the fused feature representation by concatenating the encoder output and the features of the corresponding layer in the decoder in the channel dimension, thereby realizing effective aggregation of multi-scale features. Second, the channel attention submodule (ChannelAttention) in the CBAM attention mechanism captures the global dependency across channels, reduces the data dimension by introducing a linear layer to reduce the amount of calculation, better learns the dependency between different channels, and emphasizes important features and suppresses unimportant features by weighting the channels, thereby improving the expression ability of the network. The spatial attention submodule (SpatialAttention) in the CBAM attention mechanism focuses on capturing the context information in the feature space. In actual applications, the spatial attention submodule of the present application uses two 7x7 large convolution kernels to focus on the global dependencies in the space through a larger receptive field. At the same time, in order to be consistent with the channel attention submodule, the spatial attention submodule uses the same hyperparameters for dimension reduction operation, thereby realizing information filtering and efficient calculation. Finally, the outputs of the channel attention submodule and the spatial attention submodule are weighted and fused to generate enhanced features containing rich context information. These features are fed back to the decoder to make up for the loss of detailed information caused by downsampling and improve the fine-grained performance of the generated mask.

[0029] In another example embodiment of the present application, the output end of the HCM block in the (N+1-i)th image concatenation module and the output end of the patch expansion layer in the i th image expansion module are connected to the input end of the second concatenation operation in the i th skip connection module; the output end of the linear layer in the i th skip connection module is connected to the input end of the HCM block in the i th image expansion module.

[0030] The working process of the encoder provided in the application is generally as follows: for an input medical image with a size of HxWx3, the input medical image is first divided into non-overlapping patches, and then patch embedding operation is performed on each patch by a patch embedding module to project the features of each patch to a preset dimension. The converted patches are subjected to patch merging processing via a plurality of HCM blocks and a patch merging layer to generate hierarchical feature representations. The patch merging layer is used to reduce the resolution of the feature map while increasing the feature dimension, and the HCM block is responsible for learning feature representation.

[0031] The working process of the decoder provided in the application is generally as follows: the decoder part is composed of an HCM block and a patch expansion layer. The patch expansion layer is responsible for up-sampling operation on the feature map while reducing the dimension. In this process, the extracted context features are fused with features of different scales in the encoder through a skip connection module to compensate for the loss of spatial information in the down-sampling process and enhance the interaction of global dimensions. Finally, the original resolution of the feature map is restored by 4 times up-sampling through the first patch expansion layer, and a pixel-level segmentation prediction is finally generated using a linear projection layer.

[0032] In another exemplary embodiment of the application, the ISIC2018 dataset, the Synapse dataset and the ACDC dataset can be used to train the medical image segmentation model provided in the application. In the training process, for the ISIC2018 dataset, the basic binary cross-entropy loss and the Dice loss are used as the loss function , and the specific formula is as follows.

[0033] .

[0034] .

[0035] .

[0036] In the above formula, N represents the total number of samples, is the weight of the loss function , and the default is 0.6. In the segmentation task, the BCE loss function compares the real label and the predicted label (output result of the medical image segmentation model) pixel by pixel, which can be understood as a microscopic label value; respectively represent the real situation and the output prediction value (output result of the medical image segmentation model), and the Dice loss function is a kind of overall collection similarity measure index, and are used to measure the global overlap, which can be understood as a macroscopic statistical quantity.

[0037] For the Synapse dataset and the ACDC dataset, a hybrid loss function is adopted , which combines a cross-entropy loss function and a Dice loss to address the class imbalance related issues, and the specific formula is: and .

[0038] wherein C represents the total number of classes. is an indicator, and if sample i belongs to class C, the value is 1, otherwise it is equal to 0. is the probability of the model predicting that sample i belongs to class C. is the weight of the hybrid loss function, and the default is 1.

[0039] The application also provides a specific embodiment, which is different from the above-mentioned embodiments in that in the embodiment, N = 3, the HCM block in the third image expansion module is connected to the HCM block in the first image splicing module through the third jump connection module, the HCM block in the second image expansion module is connected to the HCM block in the second image splicing module through the second jump connection module, and the HCM block in the first image expansion module is connected to the HCM block in the third image splicing module through the first jump connection module.

[0040] The application also provides an embodiment which comprehensively analyzes the performance of the medical image segmentation model HCMUNet provided above and other advanced models, and verifies its superiority through multiple datasets. The model is tested and compared on ISIC2018, Synapse and ACDC datasets. In addition, in order to further evaluate the contribution of each module of the model, the application designs a series of ablation experiments on the above datasets, and analyzes the role and influence of each component in the model in depth, and the specific steps are steps 4.1 to 4.5.

[0041] Step 4.1: Select the dataset.

[0042] ISIC18 dataset: The ISIC2018 dataset is a challenge dataset provided by the International Skin Imaging Collaboration in 2018, which is a publicly available skin lesion segmentation dataset, containing 2694 dermoscopy images with segmentation mask labels. In this experiment, the dataset is divided into training set and test set according to the ratio of 8:2.

[0043] Synapse Dataset: The Synapse dataset is a classic multi-organ segmentation dataset widely known in the field of medical image analysis. It was provided by the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). This dataset contains 30 abdominal computed tomography (CT) cases, totaling 3,779 axial abdominal clinical CT images, covering eight abdominal organs: aorta, gallbladder, left and right kidneys, liver, pancreas, spleen, and stomach. In this experiment, the Synapse dataset was partitioned, with 21 cases used for model training and the remaining 9 for model testing.

[0044] ACDC Dataset: The ACDC dataset focuses on cardiac magnetic resonance imaging analysis and is widely used to evaluate the performance of automatic cardiac structure segmentation and diagnostic algorithms. This dataset contains cardiac MRI scans from 100 patients, including expert-annotated labels for three key structures: the left ventricle, right ventricle, and myocardium. This application uses 80 samples for training and 20 for testing.

[0045] Step 4.2: Set evaluation metrics.

[0046] To evaluate the performance of the proposed HCMUNet on various datasets, this application uses mean intersection over union (mIoU), Dice Similarity Coefficient (DSC), accuracy (Acc), sensitivity (Sen), and specificity (Spe) as primary performance metrics on the ISIC2018 dataset. For the Synapse dataset, the Dice Similarity Coefficient (DSC) is used as the primary performance metric to assess segmentation performance on different organs. HD95 is used as an auxiliary metric to quantify the degree of shape matching between the predicted segmentation results and the ground truth (GroundTruth), with a particular focus on edge accuracy. For the ACDC dataset, the Dice Similarity Coefficient is used as the primary performance metric. The metric calculation formulas are as follows.

[0047] .

[0048] .

[0049] .

[0050] .

[0051] .

[0052] .

[0053] where FN represents false negative, specifically, being judged as negative sample but actually positive sample. TN represents true negative, specifically, being judged as negative sample and actually negative sample. FP represents false positive, specifically, being judged as positive sample when it is negative sample. TP represents true positive, specifically, being judged as positive sample and actually positive sample. For medical image segmentation, positive sample and negative sample represent pixels in lesion area and non-lesion area respectively. h represents the boundary point of predicted segmentation, S represents the boundary point of real segmentation, and d(h, S) represents the distance between point h and point S.

[0054] Step 4.3: Designing experimental details.

[0055] In terms of parameter configuration, the images of all data sets are preprocessed in this embodiment, and their sizes are uniformly adjusted to 256x256 pixels. In order to effectively suppress the overfitting phenomenon, data augmentation strategies such as random flipping and random rotation are introduced, which effectively increase the diversity of training samples. In terms of running parameters of model training, the batch size is set to 32 in this application, and AdamW is selected as the optimization algorithm, which combines Adam and weight decay strategy, and can effectively prevent gradient explosion and overfitting problem; the initial learning rate is set to 0.001, and is dynamically adjusted according to the training progress to accelerate the convergence; training is performed in 300 training cycles (epochs) to ensure that the model can fully learn the feature distribution of the data. In the whole experiment, no pre-trained model is used to verify the learning ability of the model on the data without external assistance. All experiments are run on V100 GPU.

[0056] Step 4.4: Experimental results and analysis.

[0057] Table 1 Experimental results of the model on ISIC2018 data set

[0058]

[0059] Automatic segmentation of skin lesions is of great significance in improving the efficiency of clinical workflows such as disease diagnosis, prognosis analysis, and treatment plan formulation. Therefore, this application selected the ISIC2018 dataset to evaluate the performance of various models. The results are shown in Table 1 (the best results are marked in bold). Compared with the baseline model, HCMUNet has achieved certain performance improvements in multiple evaluation indicators, and has achieved the best performance in mIoU, DSC and Sen; compared with the Transformer-based model, it has also improved by about two points. This shows that the U-shaped structure designed with a dual-branch structure can effectively capture local and global information, and show relatively outstanding performance in segmentation tasks. From Figure 4 It can be seen that the model proposed in this application can capture the edge information of the target more accurately than the segmentation results of other models, and effectively avoid the edge blurring phenomenon that occurs in other models; for areas of different sizes, the model can achieve accurate segmentation without obvious breaks or missing phenomena, showing good scale invariance and shape adaptability.

[0060]

[0061] To validate the model's multi-segmentation performance, we selected the Synapse multi-organ abdominal CT segmentation dataset to evaluate its performance. Table 2 shows the segmentation results of the HCMUNet model and other models on this dataset (the best results are shown in bold). Compared to the baseline model, the model achieves significant improvements in segmentation accuracy. Compared to mainstream Transformer-based segmentation methods (such as TransUNet and Swin-Unet), HCMUNet achieves improvements of 4.34% and 2.69% on DSC, respectively, validating the effectiveness of CNN and Mamba as foundational models in medical image segmentation. The results in Table 2 demonstrate that the proposed model performs well on most organ segmentation tasks, particularly improving on the segmentation of the gallbladder, left kidney, and liver, and significantly outperforming other models in spleen segmentation. HCMUNet achieves exceptionally low scores on the HD95 metric, demonstrating its strong ability to learn edge information and handle complex boundaries.

[0062] Table 3 Experimental results of the model on the ACDC dataset

[0063]

[0064] To verify the effect of the model on other data sets, the commonly used ACDC data set is selected to verify the segmentation ability of the model. Compared with some classic models and the latest models, the model has relatively excellent results. The experimental results are shown in Table 3. Although the model has certain differences in the performance of the left ventricle compared with other models, it is obviously superior to other models in comprehensive performance. Compared with the classic model UNet, the DSC is increased by 4.57 percentage points; compared with the baseline model, it is increased by 1.23 percentage points, and the segmentation effect is further improved. From the actual segmentation diagram Figure 5 It can be seen that the model still shows excellent segmentation performance on the ACDC data set, and the segmentation result is closer to the mask of the original image than other models.

[0065] Through the above experiments, it can be observed that the model can maintain good performance on different data sets, which indicates that the model has strong robustness and generalization ability. Whether it is a simple or complex segmentation region, the model can accurately identify and segment the target, and the segmentation result not only completely covers the target region, but also accurately depicts the details and structure inside the target. There is no mis-segmentation or missed segmentation. In terms of edge definition, the performance of the model is particularly outstanding. The edge part of the segmentation result is delicate and coherent, and there is no edge blur phenomenon commonly seen in other models. The edge contour closely follows the actual boundary of the target region. Even in the case of complex organs or slight changes in the edge, the model can maintain high edge positioning accuracy.

[0066] Step 4.5: ablation experiment.

[0067] The present application carries out a series of ablation experiments on three data sets respectively to verify the effectiveness of the key components of the model. Table 4 details the performance changes of the model using different components. In Table 4, SSM refers to using only the baseline model, Conv refers to adding a convolution branch to the original basis, paras represents the parameter size, and FLOPs represents the calculation amount. Specifically, when only using SSM, the evaluation indicators of the model are slightly insufficient. When integrating the Conv branch on this basis, the parameters and calculation amount of the model are reduced compared with the basic model, and the evaluation indicators are improved compared with the basic model. After adding the skip connection module, the use of channel and spatial attention can enhance the cross-dimensional interaction in channel and space, which can help the model generate more accurate segmentation results. The results show that compared with not adding this module, the DSC of the model is improved to a certain extent. However, the introduction of the skip connection module also leads to a significant increase in the parameter amount of the model, and a possible main reason for analysis is that combining CNN features with global semantic features learned by Mamba structure cannot significantly benefit the segmentation task.

[0068] Table 4 Performance of different components on HCMU Net

[0069]

[0070] The present application conducts experimental verification on the model on three commonly used medical image data sets, and the DICE similarity coefficient (DSC) is improved by 0.64%, 0.74% and 1.23% compared with the baseline model. The results show that the model proposed in the present application exhibits excellent performance on multiple public medical image segmentation data sets, especially in the processing of edge features, and exhibits good robustness and generalization ability, and exhibits excellent performance in segmentation accuracy and generalization ability. The model is tested on three public medical image data sets (ISIC2018, Synapse multi-organ segmentation and ACDC heart segmentation data set). The experimental results show that whether in different types of medical images or in various segmentation tasks, the model exhibits stable and efficient segmentation effect, verifying its practicability and reliability in the field of medical image segmentation.

[0071] Accurate medical image segmentation plays a crucial role in clinical diagnosis, disease treatment, and many other applications. In recent years, CNN-based and Transformer-based architectures have been extensively studied and explored for medical image segmentation tasks. However, CNNs have limitations in capturing global image information, while Transformers can effectively model global information but their computational complexity grows quadratically with image size. Recently, state-space models (SSMs) have shown linear complexity in terms of input size and memory usage, making them ideal choices for lightweight base models. Additionally, SSMs' ability to capture long-range dependencies is crucial for addressing long-distance information extraction problems. Based on traditional SSMs, the Mamba model was proposed by introducing time-varying parameters. Research results show that compared to Transformers, Mamba models require fewer parameters when processing text information. In the field of computer vision, the introduction of Vision Mamba further expands the application scenarios of this model. Vision Mamba has lower computational complexity while maintaining the same modeling capability as Vision Transformer. VMamba proposes a two-dimensional selection scan, bridging the gap between one-dimensional array scanning and two-dimensional plane traversal. U-Mamba creatively combines convolutional neural networks with state-space models, becoming the first Mamba model used in the field of medical image segmentation. VMUNet introduces a visual state-space block as a basic block to capture extensive contextual information and constructs an asymmetric encoder-decoder structure, establishing the first medical image segmentation model based on pure SSMs. Compared with VMUNet, the proposed medical image segmentation model first integrates convolutional neural networks into the VSS block in the encoder part to construct a hybrid basic block called HCM. The HCM block can capture both local features and long-range dependencies in the image and achieve downsampling through a patch merging strategy, improving the efficiency and accuracy of feature extraction. The decoder part consists of HCM blocks and patch expansion operations, responsible for gradually restoring the spatial resolution of the segmentation result, thus generating more accurate segmentation boundaries. This model innovatively combines the advantages of CNNs in local feature extraction with the ability of Mamba to capture long-range dependencies, effectively improving feature extraction and expression while reducing computational complexity. Furthermore, the bottleneck layer consists of two HCM blocks, not only reducing the model's parameter quantity but also enhancing the model's expression ability, enabling the model to more effectively learn the features in the image.Then for the jump connection part, the medical image segmentation model redesigns the jump connection part, combines attention mechanism with the jump connection, compensates the spatial information loss caused by the down-sampling process by integrating multi-scale features, further enhances the global interaction ability across dimensions, enhances the information interaction between the encoder and the decoder, minimizes the information loss, and expands the global dimensional interaction of multi-scale features, thereby improving the efficiency of feature transmission and the overall performance of the model.

[0072] In an exemplary embodiment, a medical image segmentation method is provided, comprising: obtaining a medical image. Inputting the medical image into the medical image segmentation model described in the above embodiments to obtain a segmentation result.

[0073] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0074] The technical features of the above embodiments can be combined arbitrarily, and in order to make the description concise, not all possible combinations of the technical features in the above embodiments are described, however, as long as the combination of the technical features does not exist contradictory, it should be considered as the scope of the present application.

[0075] In the present application, specific examples are applied to explain the principles and implementation modes of the present application, and the above embodiment descriptions are only used to help understand the method and its core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed. In conclusion, the content of the present application should not be understood as a limitation.

Claims

1. A medical image segmentation model, characterized in that: The medical image segmentation model includes: A bottleneck layer, an encoder, a decoder, and N skip connection modules; the encoder includes N sequentially connected image splicing modules, each of which includes an HCM block and a patch merging layer connected in sequence; the decoder includes N sequentially connected image expansion modules, the i-th image expansion module includes a patch expansion layer and an HCM block connected via the i-th skip connection module; the HCM block in the i-th image expansion module is connected to the HCM block in the (N+1-i)-th image splicing module via the i-th skip connection module; the HCM block includes a VSS block and a CNN model connected in parallel; The CNN model includes batch normalization, convolutional layer, RELU activation function, batch normalization, convolutional layer, RELU activation function, batch normalization, point-by-point convolution and RELU activation function connected in sequence. The patch merging layer in the last image splicing module in the encoder is connected to the patch expansion layer in the first image expansion module in the decoder; The jump connection module includes a second splicing operation, a CBAM attention mechanism and a linear layer connected in sequence; the patch merging layer in the last image splicing module in the encoder is connected to the patch expansion layer in the first image expansion module in the decoder through the bottleneck layer; the bottleneck layer includes multiple HCM blocks connected in sequence.

2. The medical image segmentation model according to claim 1, characterized in that The HCM block also includes: Split module, connection layer, channel shuffling mechanism and first splicing operation; the input end of the Split module and the first input end of the first splicing operation are the input end of the HCM block, the output end of the Split module is respectively connected to the input end of the VSS block and the input end of the CNN model, the output end of the VSS block and the output end of the CNN model are connected to the input end of the connection layer, the output end of the connection layer is connected to the input end of the channel shuffling mechanism, the output end of the channel shuffling mechanism is connected to the second input end of the first splicing operation, and the output end of the first splicing operation is the output end of the HCM block.

3. The medical image segmentation model according to claim 1, characterized in that The output end of the HCM block in the N+1-i th image splicing module and the output end of the patch expansion layer in the i th image expansion module are connected to the input end of the second splicing operation in the i th jump connection module; the output end of the linear layer in the i th jump connection module is connected to the input end of the HCM block in the i th image expansion module.

4. The medical image segmentation model according to claim 1, characterized in that The encoder further includes a patch embedding module, an output end of which is connected to an input end of the HCM block in the first image splicing module.

5. The medical image segmentation model according to claim 1, characterized in that The decoder further includes: a first patch extension layer; an input end of the first patch extension layer is connected to an output end of the HCM block in the Nth image extension module, and an output end of the first patch extension layer is an output end of the decoder.

6. The medical image segmentation model according to claim 1, characterized in that The medical image segmentation model further includes: a linear projection module; an input end of the linear projection module is connected to an output end of the decoder.

7. A medical image segmentation method, characterized in that: The medical image segmentation method comprises: Acquiring medical images; The medical image is input into the medical image segmentation model described in any one of claims 1 to 6 to obtain a segmentation result.

Images