Medical image segmentation model and segmentation method

Abstract

The invention discloses a medical image segmentation model and a segmentation method, and relates to the field of medical image processing, and the model comprises an encoder, a decoder and N jump connection modules. The encoder comprises N image splicing modules which are connected in sequence. Each image splicing module comprises an HCM block and a patch merging layer which are connected in sequence. The decoder comprises N image expansion modules which are connected in sequence, and the ith image expansion module comprises a patch expansion layer and an HCM block which are connected through the ith jump connection module; the HCM block in the ith image expansion module is connected with the HCM block in the (N + 1-i) th image splicing module through the ith jump connection module; the HCM block comprises a VSS block and a CNN model which are connected in parallel; and 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. According to the application, the segmentation precision can be improved, and the calculation complexity can be reduced.

Description

Technical Field

[0001] This application relates to the field of medical image processing, and particularly to a medical image segmentation model and a segmentation method. Background Art

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

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

[0004] To achieve the above purpose, this application provides the following solutions: In the first aspect, this application provides a medical image segmentation model, including: an encoder, a decoder, and N skip connection modules; the encoder includes N sequentially connected image splicing modules, and the image splicing module includes an HCM block and a patch merging layer connected in sequence; the decoder includes N sequentially connected image expansion modules, and the i-th image expansion module includes a patch expansion layer and an HCM block connected through 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 through the i-th skip connection module; the HCM block includes a parallel VSS block and a CNN model; ; the patch merging layer in the last image splicing module of the encoder is connected to the patch expansion layer in the first image expansion module of the decoder.

[0005] In the second aspect, this application provides a medical image segmentation method, including: obtaining a medical image; inputting the medical image into the above-mentioned medical image segmentation model to obtain a segmentation result.

[0006] According to the specific embodiments provided by the present application, the present application has the following technical effects: The present application provides a medical image segmentation model and a segmentation method. 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. It can not only pay attention to context information but also local features, achieving the purpose of improving segmentation accuracy. And by connecting the VSS block and the CNN model in parallel, the processing efficiency can be improved, and the computational complexity can be reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0008] Figure 1 It is a structural diagram of the medical image segmentation model provided by an embodiment of the present application.

[0009] Figure 2 It is a structural diagram of the HCM block provided by an embodiment of the present application.

[0010] Figure 3 It is a structural diagram of the skip connection module provided by an embodiment of the present application.

[0011] Figure 4 It is the actual segmentation map of each model on the ISIC2018 dataset.

[0012] Figure 5 It is the actual segmentation map of each model on the ACDC dataset. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0013] The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all of them. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0014] To make the above objects, features, and advantages of the present application more obvious and understandable, the present 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. This model 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 the computational complexity. As Figure 1 shown, a medical image segmentation model is provided, including: an encoder, a decoder, and N skip connection modules (SCDA); the encoder includes N sequentially connected image stitching modules, and each image stitching module includes an HCM block (HCM Block) and a patch merging layer connected in sequence; the decoder includes N sequentially connected image expansion modules, and the i-th image expansion module includes a patch expansion layer (PatchExpanding) and an HCM block connected through 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 stitching module through the i-th skip connection module; the HCM block includes a VSS block and a CNN model connected in parallel; The patch merging layer in the last image stitching module of the encoder is connected to the patch expansion layer in the first image expansion module of 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 in this application consists of a VSS block that can extract global context information and a CNN model that focuses on capturing local features. It can not only pay attention to context information but also local features, achieving the purpose of improving the segmentation accuracy. And by connecting the VSS block and the CNN model in parallel to improve the processing efficiency, the computational complexity can be reduced.

[0017] In another exemplary embodiment of this application, the encoder further includes a patch embedding module (PatchEmbedding), and the output end of the patch embedding module is connected to the input end of the HCM block in the first image stitching module.

[0018] In another exemplary embodiment of this application, as Figure 2As shown, the HCM block further includes: a Split module (Split), a Concat layer, a channel shuffle mechanism (shuffle), and a first concatenation operation; the input end of the Split module and the first input end of the first concatenation operation are the input ends of the HCM block, the output ends of the Split module are respectively connected to the input ends of the VSS block and the CNN model, the output ends of the VSS block and the CNN model are connected to the input end of the Concat layer, the output end of the Concat layer is connected to the input end of the channel shuffle mechanism, the output end of the channel shuffle mechanism is connected to the second input end of the first concatenation operation, and the output end of the first concatenation operation is the output end of the HCM block. The HCM block is an efficient dual-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 processed through a convolutional branch (CNN module) and a structured branch (VSS module) respectively. The convolutional branch focuses on capturing local features, while the structured branch is responsible for extracting global context information. The processing results of these two branches are then recombined through channel merging to restore the original channel dimension of the feature map. To further improve information flow and reduce feature loss that may be caused by channel splitting, 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 information isolation between channels that may be 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 convolutional branch and the structured branch. Specifically, the ReLU is used as the activation function in the convolutional branch to enhance the non-linear expression ability of the network; while the 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 a 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] Assume that a given feature input is , and the feature output after being processed by the HCM block is . First, X is evenly divided into X1 and X2 according to the number of channels and enters different branches for feature extraction. To ensure compatibility with the convolution operation, 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 exemplary embodiment of the present application, the medical image segmentation model further includes: a Bottleneck layer; the patch merging layer in the last image stitching module in the encoder is connected to the patch expansion layer in the first image expansion module in the decoder through the Bottleneck layer. In the Bottleneck layer, the resolution and dimension of the feature map remain unchanged to focus on the extraction of global features.

[0023] In another exemplary embodiment of the present application, the Bottleneck layer includes a plurality of sequentially connected HCM blocks.

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

[0025] In another exemplary embodiment of the present application, the medical image segmentation model further includes: a Linear Projection module; the input end of the Linear Projection module is connected to the output end of the decoder.

[0026] Channel attention focuses on strengthening important feature channels in the feature map while suppressing unimportant channels. It is achieved by assigning weights to each channel of the feature map, enabling the network to focus on information more critical to 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, enhancing the network's sensitivity to spatial information by learning the importance of different regions in the image. 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. Taking STN as an example, it can transform deformed data in space and automatically capture the features of important regions. The CBAM attention mechanism fuses channel attention and spatial attention in a cascaded manner to jointly capture the complex dependencies between channels and spatial positions, improving the model's ability to capture important features without significantly increasing the computational burden. BAM processes channel attention and spatial attention in parallel, further enhancing the richness of feature representation. The Efficient Channel Attention Network (ECA-Net) proposes a channel attention module without dimensionality reduction, achieving efficient channel feature recalibration through simple convolutional operations. Compared with using a single attention mechanism, the combination of channel attention and spatial attention can enhance the model's ability to capture context features. Such as Figure 3As shown, in another exemplary embodiment of the present application, the skip connection module is a skip connection module with dual attention, including a second splicing operation, a CBAM attention mechanism, and a linear layer connected in sequence. Since this module uses the CBAM attention mechanism to fuse multi-scale features, it can effectively compensate for the loss of spatial information caused by downsampling, enhance the global dimension interaction of multi-scale features between layers of the decoder, and improve the ability of cross-dimensional feature interaction. Therefore, while generating the output mask, it can better restore the detailed information of the image.

[0027] In practical 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 generally as follows: First, the skip connection module obtains a fused feature representation by splicing the encoder output and the features of the corresponding layer in the decoder in the channel dimension, thereby achieving effective aggregation of multi-scale features. Second, the channel attention sub-module (ChannelAttention) in the CBAM attention mechanism captures the global dependencies across channels. By introducing a linear layer to reduce the data dimension and computational complexity, it can better learn the dependencies between different channels, and emphasize important features by weighting the channels and suppressing unimportant features, thereby improving the expressive power of the network. The spatial attention sub-module (SpatialAttention) in the CBAM attention mechanism focuses on capturing the context information of the feature space. In practical applications, the spatial attention sub-module of the present application uses two 7×7 large convolutional kernels to focus on the global dependencies in space through a larger receptive field. At the same time, in order to be consistent with the channel attention sub-module, the same hyperparameters are used for dimensionality reduction operations in the spatial attention sub-module, thereby achieving information filtering and efficient computation. Finally, the outputs of the channel attention sub-module and the spatial attention sub-module are weighted and fused to generate enhanced features containing rich context information. These features are fed back to the decoder to compensate for the loss of detailed information caused by downsampling and improve the fine-grained performance of the generated mask.

[0029] In another exemplary embodiment of the present application, 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 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 this application is generally as follows: For an input medical image with a size of H×W×3, it is first segmented into non-overlapping patches, and then patch embedding operations are performed through a patch embedding module to project the features of each patch into a preset dimension. The transformed patches are subjected to patch merging processing through multiple HCM blocks and a patch merging layer to generate a hierarchical feature representation. Among them, the patch merging layer is used to reduce the resolution of the feature map and simultaneously increase the feature dimension, while the HCM block is responsible for learning feature representations.

[0031] The working process of the decoder provided in this application is generally as follows: The decoder part consists of HCM blocks and a patch expansion layer. The patch expansion layer is responsible for performing upsampling operations on the feature map while reducing the dimension. During 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 lost spatial information during downsampling and enhance the interaction of the global dimension. Finally, the original resolution of the feature map is restored through a 4-fold upsampling by the first patch expansion layer, and finally a pixel-level segmentation prediction is generated using a linear projection layer.

[0032] In another exemplary embodiment of this application, the ISIC2018 dataset, the Synapse dataset, and the ACDC dataset can be used to train the medical image segmentation model proposed in this application. During the training process, for the ISIC2018 dataset, this application uses the basic binary cross-entropy loss and the Dice loss 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 , with a default value of 0.6. In the segmentation task, the BCE loss function compares the true label and the predicted label (the output result of the medical image segmentation model) pixel by pixel, which can be understood as microscopic label values; respectively represent the true situation and the predicted value of the output (the output result of the medical image segmentation model). The Dice loss function is a holistic set similarity metric, and are used to measure the global overlap degree, which can be understood as macroscopic statistics.

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

[0038] Among them, C represents the total number of classes. is an indicator. If sample i belongs to class C, its value is 1; otherwise, it is equal to 0. is the probability that the model predicts that sample i belongs to class C. is the weight of the hybrid loss function, with default values of 1 for both.

[0039] This application also provides a specific embodiment. Different from the above embodiment, in this 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 skip 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 skip connection module. The HCM block in the first image expansion module is connected to the HCM block in the third image splicing module through the first skip connection module.

[0040] This application also provides an embodiment that comprehensively analyzes the performance of the above-provided medical image segmentation model HCMUNet and other advanced models, and verifies its superiority through multiple datasets. The model was tested and compared on the ISIC2018, Synapse, and ACDC datasets. In addition, to further evaluate the contributions of each module of the model, this application designed a series of ablation experiments on the above datasets, and deeply analyzed the roles and impacts of each component in the model. The specific steps are as follows from Step 4.1 to Step 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. It is a publicly available skin lesion segmentation dataset, containing 2,694 dermoscopic images with segmentation mask labels. In this experiment, the dataset was divided into a training set and a test set according to a ratio of 8:2.

[0043] Synapse Dataset: The Synapse dataset is a well-known classic multi-organ segmentation dataset in the field of medical image analysis, provided by the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI). This dataset contains 30 cases of abdominal computed tomography (CT) scans, totaling 3,779 axial abdominal clinical CT images, covering 8 abdominal organs: aorta, gallbladder, left kidney, right kidney, liver, pancreas, spleen, and stomach. In this experiment, the Synapse dataset was partitioned, with 21 cases used for model training and the remaining 9 cases used for model testing.

[0044] ACDC Dataset: The ACDC dataset is a dataset focused on cardiac magnetic resonance imaging analysis, widely used to evaluate the performance of automatic segmentation and diagnostic algorithms for cardiac structures. This dataset covers cardiac MRI scan images of 100 patients, including expert manual annotation labels for three important structures: left ventricle, right ventricle, and myocardium. In this application, 80 samples were used for training and 20 samples were used for testing.

[0045] Step 4.2: Set the evaluation metrics.

[0046] To evaluate the performance of the proposed HCMUNet on different datasets, on the ISIC2018 dataset, this application uses the Mean Intersection over Union (mIoU), Dice Similarity Coefficient (DSC), Accuracy (Acc), Sensitivity (Sen), and Specificity (Spe) as the main evaluation metrics for performance evaluation; for the Synapse dataset, the Dice Similarity Coefficient (DSC) is used as the main evaluation metric for performance evaluation to assess the segmentation performance on different organs; HD95 is used as an auxiliary metric to quantify the shape matching degree between the predicted segmentation result and the ground truth, especially focusing on the accuracy of the edges. For the ACDC dataset, the Dice Similarity Coefficient is used as the main evaluation metric for performance evaluation. The formula for the metrics is as follows.

[0047] 。

[0048] 。

[0049] 。

[0050] 。

[0051] 。

[0052] 。

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

[0054] Step 4.3: Design the experimental details.

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

[0056] Step 4.4: Experimental results and analysis.

[0057] Table 1 Experimental results table of the model on the ISIC2018 dataset

[0058] The 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 selects the ISIC2018 dataset to evaluate the performance of various models. The results are shown in Table 1 (where the optimal results are marked in bold). Compared with the baseline model, HCMUNet has achieved a certain performance improvement in multiple evaluation metrics and has achieved the best performance in terms of mIoU, DSC, and Sen; there is also an improvement of about two points compared with the models mainly based on Transformer. This shows that the U-shaped structure designed with a dual-branch structure can effectively capture local and global information and demonstrate excellent performance in the segmentation task. From Figure 4 It can be seen that compared with the segmentation results of other models, the model proposed in this application can capture the edge information of the target more accurately and effectively avoid the edge blur phenomenon that appears in other models; for regions of different sizes, the model can achieve accurate segmentation without obvious breaks or omissions, demonstrating good scale invariance and shape adaptability.

[0059]

[0060] To verify the multi-segmentation effect of the model, this application selects the Synapse multi-organ abdominal CT segmentation dataset to evaluate the performance of the model. Table 2 shows the specific segmentation results of the HCMUNet model and other models on this dataset (where the optimal results are shown in bold). Compared with the baseline model, the segmentation accuracy of the model has a certain improvement. Compared with the mainstream Transformer-based segmentation methods (such as TransUNet and Swin-Unet), HCMUNet has improved by 4.34% and 2.69% respectively in terms of DSC, verifying the effectiveness of using CNN and Mamba as the basic models in the field of medical image segmentation. From the results in Table 2, it can be seen that the model proposed in this application performs excellently in most organ segmentation tasks, especially in the segmentation of the gallbladder, left kidney, and liver, showing a certain improvement compared with other models, and far exceeding other models in the segmentation of the spleen. In terms of the HD95 index, HCMUNet has obtained a very low value, indicating that the model also has strong capabilities in learning edge information and processing complex boundaries.

[0061] Table 3 Experimental results of the model on the ACDC dataset

[0062] To verify the performance of the model on other datasets, the common ACDC dataset was selected to validate the segmentation ability of the model. Compared with some classical models and the latest models, this model has relatively excellent results. Multiple experiments were conducted on the ACDC dataset, and the average of the results was taken. 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, its comprehensive performance is significantly better than other models. Compared with the classical model UNet, the DSC increased by 4.57 percentage points; compared with the baseline model, it increased by 1.23 percentage points, and the segmentation effect was further improved. From the actual segmentation map Figure 5 It can be seen that this model still demonstrates excellent segmentation performance on the ACDC dataset, and its segmentation results are closer to the mask of the original image compared with other models.

[0063] Through the above experiments, it can be observed that the model can maintain good performance on different datasets, which indicates that the model has strong robustness and generalization ability. Whether it is a simple or complex segmentation area, the model can accurately identify and segment the target. The segmentation results not only completely cover the target area, but also accurately depict the details and structures inside the target, without missegmentation or missed segmentation. In terms of edge sharpness, the performance of the model is particularly prominent. The edge part of the segmentation result is delicate and coherent, without the common edge blur phenomenon in other models. The edge contour closely adheres to the actual boundary of the target area. Even in the case of complex organs or subtle changes in the edge, the model can maintain a high edge localization accuracy.

[0064] Step 4.5: Ablation experiment.

[0065] A series of ablation experiments were conducted on three datasets in this application 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 convolutional branch on the original basis, paras represents the parameter size, and FLOPs represents the computational volume. 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 computational volume of the model are reduced compared with the basic model, and at the same time, its evaluation indicators are improved compared with the basic model. After adding the skip connection module, by using channel and spatial attention to enhance the cross-dimensional interaction in the channel and spatial aspects, it can help the model generate more accurate segmentation results. The results show that the DSC of the model has a certain degree of improvement compared with that without this module. However, the introduction of the skip connection module also led to a significant increase in the number of model parameters. One main reason for the analysis may be that combining CNN features with the global semantic features learned by the Mamba structure does not significantly benefit the segmentation task.

[0066] Table 4 Performance of Different Components on HCMUNet

[0067] In this application, the model was experimentally verified on three commonly used medical image datasets. Compared with the baseline model, the DICE similarity coefficient (DSC) increased by 0.64%, 0.74% and 1.23% respectively. The results show that the model proposed in this application has demonstrated excellent performance on multiple publicly available medical image segmentation datasets, especially in the processing of edge features, showing good robustness and generalization ability, and excellent performance in both segmentation accuracy and generalization ability. The model was experimented on three publicly available medical image datasets (ISIC2018, Synapse multi-organ segmentation, and ACDC cardiac segmentation datasets). The experimental results show that the model has demonstrated stable and efficient segmentation results in both different types of medical images and diverse segmentation tasks, verifying its practicability and reliability in the field of medical image segmentation.

[0068] Accurate medical image segmentation plays a crucial role in clinical diagnosis, disease treatment, and many application fields. In recent years, CNN-based and Transformer-based architectures have been deeply studied and widely explored in medical image segmentation tasks. However, CNN still has limitations in capturing global image information, while although Transformers can effectively perform global modeling, their computational complexity grows quadratically with the increase in image size. Recently, state space models (SSMs) have shown linear complexity in terms of input size and memory footprint, making them an ideal choice for lightweight backbone models. At the same time, the ability of SSMs to capture long-range dependencies is crucial for solving the problem of long-distance information extraction. Based on traditional SSMs, the Mamba model was proposed by introducing time-varying parameters. The research results show that compared with Transformers, the Mamba model has fewer parameter requirements when processing text information. In the field of computer vision, the proposal of Vision Mamba further expands the application scenarios of this model. Vision Mamba has the same modeling ability as Vision Transformer while having lower computational complexity. VMamba proposed two-dimensional selective scanning, bridging the gap between one-dimensional array scanning and two-dimensional plane traversal. U-Mamba creatively combines convolutional neural networks with state space models and is the first Mamba model for the field of medical image segmentation. VMUNet captures extensive context information by introducing visual state space blocks as basic blocks and constructs an asymmetric encoder-decoder structure to establish the first medical image segmentation model built based on a pure SSMs model. Compared with VMUNet, the medical image segmentation model proposed in this application, firstly, in the encoder part of the medical image segmentation model, a convolutional neural network is incorporated on the basis of the VSS block to construct a hybrid basic block called HCM. The HCM block can capture both local features and long-range dependencies of the image and perform downsampling through a patch merging strategy to improve the efficiency and accuracy of feature extraction. The decoder part consists of HCM blocks and patch expansion operations, which are responsible for gradually restoring the spatial resolution of the segmentation result, thereby generating a more accurate segmentation boundary. It innovatively integrates the advantages of CNN in local feature extraction and the ability of Mamba in capturing long-range dependencies, effectively improving the feature extraction and expression ability while reducing the computational complexity. In addition, the bottleneck layer is composed of two HCM blocks, which not only reduces the number of parameters of the model but also enhances the expression ability of the model, enabling the model to more effectively learn the features in the image.Then for the skip connection part, the medical image segmentation model redesigned the skip connection part, combined the attention mechanism with the skip connection, effectively compensated for the spatial information loss caused by the downsampling process by integrating multi-scale features, and further enhanced the global interaction ability across dimensions, enhanced the information interaction between the encoder and the decoder, minimized information loss, and extended the global dimensional interaction of multi-scale features, thereby improving the efficiency of feature transmission and the overall performance of the model.

[0069] In an exemplary embodiment, a medical image segmentation method is provided, including: obtaining a medical image. Inputting the medical image into the medical image segmentation model described in the above embodiment to obtain a segmentation result.

[0070] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.

[0071] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0072] Specific examples are used in this application to elaborate on the principle and implementation manner of this application. The description of the above embodiments is only used to help understand the method and its core idea of this application; at the same time, for those of ordinary skill in the art, based on the idea of this application, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to this application.

Claims

1. A medical image segmentation model, characterized in that, The medical image segmentation model includes: An encoder, a decoder, and N skip connection modules; the encoder includes N sequentially connected image stitching modules, and each image stitching module includes an HCM block and a patch merging layer that are sequentially connected; the decoder includes N sequentially connected image expansion modules, and the i-th image expansion module includes a patch expansion layer and an HCM block connected through 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 stitching module through the i-th skip connection module; the HCM block includes 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 to the patch expansion layer in the first image expansion module in the decoder.

2. The medical image segmentation model according to claim 1, wherein The HCM block further includes: A Split module, a connection layer, a channel shuffling mechanism, and a first splicing operation; the input end of the Split module and the first input end of the first splicing operation are the input ends 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 ends of the VSS block and 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, wherein The skip connection module includes a second splicing operation, a CBAM attention mechanism, and a linear layer connected in sequence.

4. The medical image segmentation model according to claim 1, characterized in that, The medical image segmentation model further includes: a bottleneck layer; 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.

5. The medical image segmentation model according to claim 4, wherein The bottleneck layer includes a plurality of HCM blocks connected in sequence.

6. The medical image segmentation model according to claim 3, wherein 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 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.

7. The medical image segmentation model according to claim 1, wherein The encoder further includes a patch embedding module, and the output end of the patch embedding module is connected to the input end of the HCM block in the first image splicing module.

8. The medical image segmentation model according to claim 1, wherein The decoder further includes: a first patch expansion layer; the input end of the first patch expansion layer is connected to the output end of the HCM block in the N-th image expansion module, and the output end of the first patch expansion layer is the output end of the decoder.

9. The medical image segmentation model according to claim 1, wherein, The medical image segmentation model further includes: a linear projection module; the input end of the linear projection module is connected to the output end of the decoder.

10. A medical image segmentation method, characterized in that, The medical image segmentation method includes: Obtaining a medical image; Inputting the medical image into the medical image segmentation model according to any one of claims 1 to 9 to obtain a segmentation result.

Images