Heart MRI image segmentation method and system based on space channel reconstruction

By introducing multi-scale extended Transformer's spatial channel convolution block and sliding window expansion attention operation in cardiac MRI image segmentation, the problems of receptive field expansion and calculation complexity in the prior art are solved, and efficient and robust hearts are achieved MRI image segmentation significantly improves segmentation accuracy and performance.

CN118941559AInactive Publication Date: 2024-11-12QILU UNIVERSITY OF TECHNOLOGY (SHANDONG ACADEMY OF SCIENCES) +2
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411418789.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-12
Publication Date
2024-11-12
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing cardiac MRI image segmentation method has insufficient efficiency and robustness. The local dependence of local operations and convolution limits the model receptive field, making it difficult to capture the position and shape characteristics of the ventricles, and the numerous model parameters increase the computational complexity. Spend.

Method used

The space channel convolution block with multi-scale extended Transformer was introduced, combined with local sparse convolution blocks and sliding window expansion attention operation, space channel reconstruction was carried out, multi-scale context information was captured and receptive fields were expanded, and the ES-Net architecture was proposed. and Scc-MSWA module.

Benefits of technology

It significantly improves the efficiency and robustness of cardiac MRI image segmentation, reduces computational complexity, improves the segmentation ability of complex cardiac anatomy, and shows excellent performance on multiple cardiac datasets.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118941559A_ABST
    Figure CN118941559A_ABST
Patent Text Reader

Abstract

The invention provides a heart MRI image segmentation method and system based on space channel reconstruction, and relates to the field of heart MRI image segmentation, and the specific scheme comprises the steps: obtaining a to-be-segmented heart MRI image; inputting the heart MRI image into the trained segmentation model to obtain a segmented heart MRI image; wherein the segmentation model is characterized in that on a U-Net structure, a space channel convolution block with a multi-scale expansion Transform is introduced, a local sparse convolution block and a sliding window expansion attention operation are combined, space channel reconstruction is carried out, multi-scale context information is captured, and a receptive field is expanded; according to the method, the space channel convolution block with the multi-scale expansion Transform is introduced, multi-scale context information is captured, the receptive field is expanded, and the efficiency and robustness of heart MRI image segmentation are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of cardiac MRI image segmentation, and in particular relates to a cardiac MRI image segmentation method and system based on spatial channel reconstruction. Background Art

[0002] In the field of medical image analysis, especially in the application of cardiac magnetic resonance imaging (MRI), accurate image segmentation is crucial for the diagnosis and management of cardiovascular diseases. Detailed segmentation of key anatomical structures such as the left ventricle (LV) and right ventricle (RV) not only provides clinicians with a powerful tool to evaluate cardiac function, morphology, and potential pathology, but also helps develop early intervention strategies, thereby significantly improving patient outcomes.

[0003] Cardiac MRI image segmentation faces multiple challenges, including low image contrast, partial volume effects, and significant anatomical differences between patients, all of which make it difficult to achieve high-fidelity segmentation; traditional cardiac MRI image segmentation methods mainly rely on manual delineation by experts, a process that is not only labor-intensive and time-consuming, but also susceptible to inter-observer variability, thus limiting its applicability in large-scale applications; the emergence of deep learning, especially the introduction of convolutional neural networks (CNNs), has revolutionized this field by automating the cardiac MRI segmentation process with unprecedented accuracy and efficiency; among various CNN architectures, U-Net and its variants have become the first choice in medical image segmentation tasks due to their ability to capture local and global contextual information.

[0004] Although deep learning methods are widely used in cardiac medical image segmentation and have significant effects and can provide accurate results, they still face the following problems: 1. The local dependency operation and the locality of convolution limit the expansion of the model's receptive field, making it difficult to capture the location and shape characteristics of the ventricle. Although increasing the network depth can expand the receptive field, it also increases the computational complexity and training difficulty.

[0005] 2. Medical image segmentation models often need to process complex details of high-resolution images, resulting in a large number of model parameters and increased demand for computing resources. This limitation is particularly evident in the deep layers of the network, making it difficult for the receptive field to cover the entire input image.

[0006] Therefore, existing cardiac MRI image segmentation schemes are insufficient in terms of efficiency and robustness. Summary of the invention

[0007] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a cardiac MRI image segmentation method and system based on spatial channel reconstruction, introduces a spatial channel convolution block with a multi-scale extended Transformer, captures multi-scale contextual information and expands the receptive field, and improves the efficiency and robustness of cardiac MRI image segmentation.

[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: A first aspect of the present invention provides a cardiac MRI image segmentation method based on spatial channel reconstruction.

[0009] A cardiac MRI image segmentation method based on spatial channel reconstruction, comprising: Acquire a cardiac MRI image to be segmented; Input the cardiac MRI image into the trained segmentation model to obtain a segmented cardiac MRI image; Among them, the segmentation model introduces a spatial channel convolution block with a multi-scale extended Transformer on the U-Net structure, combines local sparse convolution blocks and sliding window expansion attention operations, performs spatial channel reconstruction, captures multi-scale contextual information and expands the receptive field.

[0010] Further, the segmentation model includes an encoder, a bottleneck and a decoder; The encoder is composed of a plurality of encoding blocks, and the encoding block is composed of a residual U block and a pooling layer; The bottleneck consists of spatial channel convolution blocks and pooling layers with multi-scale dilated Transformers; The decoder is composed of a plurality of decoding blocks, and the decoding block is composed of a residual U block and a pooling layer.

[0011] Furthermore, the residual U block first extracts local and global features by gradually reducing the sampling rate, and then fuses them into a high-resolution image through upsampling, feature cascading and convolution to obtain multi-scale features.

[0012] Furthermore, the spatial channel convolution block with the multi-scale extended Transformer is composed of sequentially connected spatial and channel convolution blocks and a multi-scale window attention layer; The spatial and channel convolution blocks are used to compress and refine the multi-scale features to obtain reconstructed multi-scale features; The multi-scale window attention layer is used to model the long-range dependencies in the multi-scale features to obtain enhanced multi-scale features.

[0013] Furthermore, the spatial and channel convolution blocks include spatial reconstruction convolution and channel reconstruction convolution; The spatial reconstruction convolution uses separation and reconstruction operations to distinguish two feature maps with rich information and poor information according to spatial content characteristics, and the two feature maps are merged into a spatial refinement feature map through a cross reconstruction operation; The channel reconstruction convolution uses segmentation and transformation operations to divide the input feature map into a feature map rich in representative features and a feature map supplementing detail information, and performs weighted fusion on the two feature maps to obtain reconstructed multi-scale features.

[0014] Furthermore, the multi-scale window attention layer selects representative queries, keys, and values ​​in sliding windows of different scales and performs self-attention operations.

[0015] Furthermore, the multi-scale window attention layer specifically operates as follows: Given a feature graph, obtain the corresponding query, key, and value through linear projection; The channels of the feature map are divided into multiple different heads, and each head uses a different dilation rate to perform multi-scale sliding window dilation attention.

[0016] A second aspect of the present invention provides a cardiac MRI image segmentation system based on spatial channel reconstruction.

[0017] A cardiac MRI image segmentation system based on spatial channel reconstruction includes an acquisition module and a segmentation module: An acquisition module is configured to: acquire a cardiac MRI image to be segmented; The segmentation module is configured to: input the cardiac MRI image into the trained segmentation model to obtain a segmented cardiac MRI image; Among them, the segmentation model introduces a spatial channel convolution block with a multi-scale extended Transformer on the U-Net structure, combines local sparse convolution blocks and sliding window expansion attention operations, performs spatial channel reconstruction, captures multi-scale contextual information and expands the receptive field.

[0018] The third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in a cardiac MRI image segmentation method based on spatial channel reconstruction as described in the first aspect of the present invention.

[0019] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in a cardiac MRI image segmentation method based on spatial channel reconstruction as described in the first aspect of the present invention are implemented.

[0020] One or more of the above technical solutions have the following beneficial effects: This paper proposes a spatial channel convolution block (Scc-MSWA) based on multi-scale dilated Transformer, which combines local sparse convolution blocks and sliding window dilated attention operations. It not only effectively captures multi-scale contextual information and expands the receptive field, but also significantly reduces the computational complexity and improves the efficiency and robustness of cardiac MRI image segmentation.

[0021] The present invention introduces a novel residual U-block (RSU) technology to capture multi-scale features and extract rich contextual information at different stages, enhance feature extraction and nonlinear mapping, and significantly improve the segmentation capability of complex cardiac anatomical structures.

[0022] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0024] Figure 1 This is the ES-Net structure diagram of the first embodiment.

[0025] Figure 2 This is a diagram of the RSU structure of the first embodiment.

[0026] Figure 3 FIG. 4 is a structural diagram of Scc-MSWA of the first embodiment.

[0027] Figure 4 This is the spatial reconstruction convolution (SRC) structure diagram of the first embodiment.

[0028] Figure 5 This is a diagram of the channel reconstruction convolution (CRC) structure of the first embodiment.

[0029] Figure 6 This is a diagram of the multi-scale window attention structure of the first embodiment. DETAILED DESCRIPTION

[0030] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0031] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0032] Embodiment 1 In one embodiment of the present disclosure, a cardiac MRI image segmentation method based on spatial channel reconstruction is provided, comprising the following steps: Step S1: Acquire a cardiac MRI image to be segmented; Step S2: inputting the cardiac MRI image into the trained segmentation model to obtain a segmented cardiac MRI image; Among them, the segmentation model introduces a spatial channel convolution block with a multi-scale extended Transformer on the U-Net structure, combines local sparse convolution blocks and sliding window expansion attention operations, performs spatial channel reconstruction, captures multi-scale contextual information and expands the receptive field.

[0033] Further, the segmentation model includes an encoder, a bottleneck and a decoder; The encoder is composed of a plurality of encoding blocks, and the encoding block is composed of a residual U block and a pooling layer; The bottleneck consists of spatial channel convolution blocks and pooling layers with multi-scale dilated Transformers; The decoder is composed of a plurality of decoding blocks, and the decoding block is composed of a residual U block and a pooling layer.

[0034] Furthermore, the residual U block first extracts local and global features by gradually reducing the sampling rate, and then fuses them into a high-resolution image through upsampling, feature cascading and convolution to obtain multi-scale features.

[0035] Furthermore, the spatial channel convolution block with the multi-scale extended Transformer is composed of sequentially connected spatial and channel convolution blocks and a multi-scale window attention layer; The spatial and channel convolution blocks are used to compress and refine the multi-scale features to obtain reconstructed multi-scale features; The multi-scale window attention layer is used to model the long-range dependencies in the multi-scale features to obtain enhanced multi-scale features.

[0036] Furthermore, the spatial and channel convolution blocks include spatial reconstruction convolution and channel reconstruction convolution; The spatial reconstruction convolution uses separation and reconstruction operations to distinguish two feature maps with rich information and poor information according to spatial content characteristics, and the two feature maps are merged into a spatial refinement feature map through a cross reconstruction operation; The channel reconstruction convolution uses segmentation and transformation operations to divide the input feature map into a feature map rich in representative features and a feature map supplementing detail information, and performs weighted fusion on the two feature maps to obtain reconstructed multi-scale features.

[0037] Furthermore, the multi-scale window attention layer selects representative queries, keys, and values ​​in sliding windows of different scales and performs self-attention operations.

[0038] Furthermore, the multi-scale window attention layer specifically operates as follows: Given a feature graph, obtain the corresponding query, key, and value through linear projection; The channels of the feature map are divided into multiple different heads, and each head uses a different dilation rate to perform multi-scale sliding window dilation attention.

[0039] The above segmentation model adopts the ES-Net architecture. The structure of the ES-Net architecture is described in detail below.

[0040] In order to take advantage of the advantages of CNN and Transformers, the ES-Net architecture was developed with a U-shaped design to improve the accuracy and efficiency of cardiac MRI segmentation. ES-Net is based on the traditional U-Net structure, such as Figure 1 As shown, it includes encoder, bottleneck, decoder and skip connections, with multiple scales (n) ranging from n=[1,2,3,4,5,6].

[0041] There are two improvements in ES-Net: 1) The encoder-decoder architecture is improved by replacing the traditional convolutional layers in the layer (n=1 to 4) with residual U blocks (RSU), which enhances feature extraction and nonlinear mapping; 2) In order to expand the receptive field and feature representation of the model, the spatial channel convolution block with multi-scale dilated Transformer (Scc-MSWA) is introduced, which combines local sparse convolution blocks and sliding window dilated attention (SWDA) operations to reduce computational complexity and improve the performance of visual tasks to ensure accurate segmentation through effective multi-level information utilization; How Scc-MSWA is embedded in the U-shaped architecture and the reasons behind these choices are described in detail below.

[0042] 1. Residual U Block (RSU) In order to improve model efficiency and reduce parameters, RSU is introduced to capture multi-scale features and extract rich contextual information at different stages.

[0043] The structure of RSU is as follows Figure 2 As shown in the figure, the architecture processes the input image and extracts features through multiple layers of local sparse convolution blocks and jump connections; from left to right, the image depicts multiple convolution layers, where the output feature dimensions of each layer are marked in the form of "HxWxC", where H and W represent height and width, respectively, and C represents the number of channels; the network structure also contains jump connections, which allow feature maps from early layers to be directly transmitted to subsequent layers, helping to preserve image detail information and gradient flow; in addition, upsampling and downsampling operations are used, which are used to increase and reduce the size of feature maps, respectively, so as to capture image features at different scales. From Figure 2 From the structure shown, we can see that as the number of encoders increases, the RSU is deepened through more pooling operations, thereby expanding the receptive field to capture richer local and global features; local and global features are extracted by gradually reducing the sampling rate, and then fused into high-resolution images, i.e., multi-scale features, through upsampling, feature cascading and convolution, thereby minimizing detail loss; this method significantly improves the segmentation accuracy, and the weight transformation and local feature fusion of RSU enhance the original feature representation; ES-Net uses RSU blocks to efficiently extract multi-scale features, greatly reducing computational overhead while maintaining excellent performance in cardiac MRI segmentation.

[0044] 2. Spatial Channel Convolution Block with Multi-Scale Expanded Transformer (Scc-MSWA) like Figure 3 As shown, Scc-MSWA consists of sequentially connected spatial and channel convolution blocks and multi-scale window attention layers; the spatial and channel convolution blocks are used to compress and refine multi-scale features to obtain reconstructed multi-scale features; the multi-scale window attention layer is used to model the long-range dependencies in multi-scale features to obtain enhanced multi-scale features.

[0045] 1. Spatial and Channel Convolution Block (SCC) The spatial and channel convolutional block (SCC) module is constructed by sequentially combining two sub-modules: spatial reconstruction convolution (SRC) and channel reconstruction convolution (CRC). For the input features in the bottleneck residual block (i.e., residual U block), The SCC module first applies the SRC submodule to reduce the redundancy of the spatial dimension and generate spatially refined features. Subsequently, the CRC submodule processes these spatially refined features , to reduce the redundancy of the channel dimension, thereby obtaining the refined channel features By effectively utilizing the redundancy between features, the SCC module achieves effective compression and refinement of features, significantly enhancing the feature representation capability while reducing redundancy.

[0046] (1) Spatial Reconstruction Convolution (SRC) The structure of spatial reconstruction convolution (SRC) is as follows Figure 4 As shown, starting from the top input features (X), these features undergo a series of processing and transformations to extract key information; the input features are followed by a mini-batch normalization operation to obtain This step is crucial to improving the training efficiency and stability of the network; then, the data is obtained by using formula (2) to obtain the matrix , , A linear transformation is performed. These weights are learned by the model during training and are used to adjust the distribution of the input data. A Sigmoid activation function is applied, which maps the input data to the interval (0,1), increasing the nonlinear expression ability of the network. The output of the Sigmoid activation function is then sent to a "spatial reconstruction" operation, which involves rearranging or transforming the data in the spatial dimension to better suit the needs of subsequent tasks. Finally, the network outputs a feature map , this feature map contains the key information representation of the input data after being processed by the entire network, and is the basis for the model to make predictions or classifications.

[0047] It can be seen that SRC can cleverly combine separation and reconstruction operations. The purpose of the separation operation is to distinguish information-rich and information-poor feature maps based on spatial content characteristics. Apply group normalization as shown in formula (1) ( ), calculate the channel weight , using the normalized scaling factor ; The relevant channel weights , reflects the importance of each feature map and is calculated by formula (2), where and yes The mean and standard deviation of η and yes The learnable affine transformation parameters in the layer, is a small constant used for numerical stability.

[0048] (1) (2) Next, the channel weights are mapped to binary values ​​in the range (0,1) through the Sigmoid function and the gating mechanism T, and then thresholded: (3) These binary weights are used to transform the input feature map Divided into two parts: Information-rich components and components with less information , the formulas to obtain these components are as follows:

[0049] (4) in, is element-wise multiplication.

[0050] Finally, these two feature maps are merged into a spatially refined feature map through a cross reconstruction operation. .

[0051] (2) Channel Reconstruction Convolution (CRC) CRC architecture, such as Figure 5 As shown in Figure 5, the architecture is basically divided into two key stages: segmentation and transformation. In the segmentation stage, as shown in formula (5), the CRC method calculates the segmentation ratio according to the predetermined segmentation ratio. The feature map of spatial refinement It is divided into two parts, left and right. In these two parts, 1×1 convolution kernel is used to reduce the dimension and reduce the computational complexity to obtain and .

[0052] (5) In the transformation stage, it is further subdivided into two branches: the left branch uses efficient convolution operations to deeply explore and extract high-level features, while the right branch cleverly uses less computationally demanding 1×1 convolutions to enhance the representation of shallow detail features; these carefully designed adjustments not only optimize the model architecture, but also significantly improve the clarity and accuracy of feature representation.

[0053] Specifically, after the 1×1 convolution operation, the left part Apply 1×1 convolution and k×k convolution operations to generate a feature map rich in representative features ; At the same time, the right part After point-by-point convolution and feature reuse, a feature map with supplementary detail information is generated , get and The formula is as follows:

[0054] (6) (7)

[0055] In this framework, is the predetermined segmentation ratio, c is the number of channels, r is the squeeze ratio, which is usually set to 2. In this embodiment, the group size g = 2 is set, ∪ represents the connection operation, and Represents the learnable weight matrix associated with it, and in the left part is associated with the k × k and 1 × 1 convolution operations respectively. and is the learnable weight matrix associated with the 1 × 1 convolution operation on the right side is also a learnable weight matrix.

[0056] Finally, we get and Weight and , these weights are used to and Perform weighted fusion to generate a channel-refined feature map The specific formula for this process is as follows: (8) (9) (10) in, represents the global average pooling operation, They are the global spatial information on the left and the spatial global information on the right. is the feature importance vector calculated by the softmax function, i.e., the weight.

[0057] (3) Multi-Scale Window Attention (MSWA) The sliding window extended attention mechanism, i.e., multi-scale window attention, is innovatively introduced to effectively model the long-range dependencies in feature maps. Figure 6 As shown in the figure, MSWA selects representative keys and values ​​in sliding windows of different scales and performs self-attention operations. This method captures and integrates multi-scale contextual semantic dependencies and utilizes the rich information in the data blocks.

[0058] In the sliding window centered on the query patch, representative keys and values ​​are sparsely selected and then self-attention is performed as follows: (11)

[0059] in, and Represent the height and width of the feature map respectively, and Respectively represent the feature maps and The keys and values ​​selected in , calculate the dot product of the query with all the keys, and divide each dot product result by , then apply the softmax function to get the weights on the values, and finally, concatenate the features from different heads and feed them into a linear layer for further processing.

[0060] For different attention heads ( ), using different expansion rates The sliding window dilated attention (SWDA) is used to effectively capture multi-scale information; each attention head independently calculates its corresponding attention output ,in , and is from The query, key, and value matrices separated from the matrix are as follows: (12) here, Indicates The output of an attention head.

[0061] These outputs are then concatenated and aggregated via a linear transformation to generate the feature map , this process can be formalized into the following formula: (13) Where n represents the number of attention heads, Rather, it is the intermediate result after connection and linear transformation.

[0062] The MSWA module utilizes a multi-head partitioning method and combines it with a multi-scale self-attention mechanism, where different heads use different dilation rates to effectively extract and integrate multi-scale semantic information, which enhances the model's ability to handle complex visual tasks while maintaining relatively low computational complexity.

[0063] like Figure 6 As shown, given the feature map , the corresponding query, key and value are obtained through linear projection. Then, the channels of the feature map are divided into n different heads, and each head uses a different expansion rate for multi-scale SWDA (sliding window dilation attention). For the blue points with different expansion rates around the yellow points, the model uses a multi-scale extended Transformer block to capture the complex relationship between these points, and promotes the interaction between features of different scales through the attention mechanism, sliding window and dilation attention, thereby constructing a more comprehensive and multi-level feature representation.

[0064] In a convolutional neural network, the input feature map The dimension is and , where B represents the batch size, Represents the number of channels, and are the height and width of the feature map, respectively. For each convolutional layer, a different dilation rate is used ( ), these expansion rates are usually positive integers chosen based on network design and specific task requirements.

[0065] Specifically, by assigning different dilation rates to different heads, MSWA effectively aggregates semantic information of various scales within the receptive field, reducing the redundancy in the self-attention mechanism while avoiding complex operations or additional computational overhead and computational cost. The output of MSWA is expressed as: (14) The main contributions of this example are threefold: (1) The ES-Net model is proposed: a novel network architecture tailored for cardiac MRI image analysis; (2) The Scc-MSWA module is proposed: a spatial channel convolution block with a multi-scale extended Transformer, which is designed to capture multi-scale contextual information and long-range dependencies while effectively expanding the receptive field by reducing module complexity; (3) A comprehensive evaluation of ES-Net is performed: ES-Net is compared in detail with baseline and state-of-the-art segmentation architectures, demonstrating its superior performance in both in-database and cross-database evaluation scenarios.

[0066] Its main design goal is to significantly reduce model parameters while improving global feature analysis capabilities. In order to improve global analysis performance, a spatial channel multi-scale self-attention mechanism (Scc-MSWA) that integrates Scc and MSWA modules is developed and integrated with the Scc-MSWA mechanism. This mechanism effectively captures and integrates comprehensive global dependencies across spatial and channel dimensions; in addition, by replacing the traditional U-Net architecture with the RSU module, efficient integration of receptive fields at different scales is achieved, thereby enhancing the model's multi-scale feature extraction capabilities; while maintaining a low model complexity, the core dilated convolution characteristics are retained.

[0067] Through extensive experiments on multiple cardiac datasets, including ACDC, Sunnybrook MRI, and RVSC, the ES-Net model proposed in this embodiment shows excellent performance in cardiac image segmentation tasks; on the ACDC dataset, ES-Net achieves the highest average Dice score of 90.68%, significantly better than existing methods. Its efficient architecture and enhanced receptive field capability not only reduce the number of parameters, but also improve the overall performance, showing a clear advantage over competitors; on the Sunnybrook MRI dataset, ES-Net surpasses the current state-of-the-art methods with the highest average Dice score of 94.62%; it is worth noting that its excellent segmentation performance under a variety of cardiac pathological conditions further verifies its efficiency and accuracy; in experiments on the RVSC dataset, ES-Net achieves state-of-the-art segmentation accuracy and precision, fully demonstrating its wide applicability and excellent stability.

[0068] By optimizing the original U-Net architecture and integrating the efficient Scc-MSWA module, ES-Net not only achieved a significant improvement in the Dice coefficient, but also reduced the model parameters by up to 43.88%. In addition, the computational complexity of ES-Net is only 24.89G FLOPs, which is much lower than other leading methods such as UNet++ (103.67G FLOPs) and MT-UNet (44.75G FLOPs), which highlights its high efficiency.

[0069] Embodiment 2 In one embodiment of the present disclosure, a cardiac MRI image segmentation system based on spatial channel reconstruction is provided, including an acquisition module and a segmentation module: An acquisition module is configured to: acquire a cardiac MRI image to be segmented; The segmentation module is configured to: input the cardiac MRI image into the trained segmentation model to obtain a segmented cardiac MRI image; Among them, the segmentation model introduces a spatial channel convolution block with a multi-scale extended Transformer on the U-Net structure, combines local sparse convolution blocks and sliding window expansion attention operations, performs spatial channel reconstruction, captures multi-scale contextual information and expands the receptive field.

[0070] Embodiment 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0071] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in a cardiac MRI image segmentation method based on spatial channel reconstruction as described in the first embodiment of the present disclosure.

[0072] Embodiment 4 The purpose of this embodiment is to provide an electronic device.

[0073] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in a cardiac MRI image segmentation method based on spatial channel reconstruction as described in Embodiment 1 of the present disclosure are implemented.

[0074] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A cardiac MRI image segmentation method based on spatial channel reconstruction, It is characterized in that include: Acquire a cardiac MRI image to be segmented; Input the cardiac MRI image into the trained segmentation model to obtain a segmented cardiac MRI image; Among them, the segmentation model introduces a spatial channel convolution block with a multi-scale extended Transformer on the U-Net structure, combines local sparse convolution blocks and sliding window expansion attention operations, performs spatial channel reconstruction, captures multi-scale contextual information and expands the receptive field.

2. A cardiac MRI image segmentation method based on spatial channel reconstruction as claimed in claim 1, It is characterized in that The segmentation model includes an encoder, a bottleneck and a decoder; The encoder is composed of a plurality of encoding blocks, and the encoding block is composed of a residual U block and a pooling layer; The bottleneck consists of spatial channel convolution blocks and pooling layers with multi-scale dilated Transformers; The decoder is composed of a plurality of decoding blocks, and the decoding block is composed of a residual U block and a pooling layer.

3. A cardiac MRI image segmentation method based on spatial channel reconstruction as claimed in claim 2, It is characterized in that The residual U block first extracts local and global features by gradually reducing the sampling rate, and then fuses them into a high-resolution image through upsampling, feature cascading and convolution to obtain multi-scale features.

4. A cardiac MRI image segmentation method based on spatial channel reconstruction as claimed in claim 2, It is characterized in that The spatial channel convolution block with multi-scale extended Transformer is composed of sequentially connected spatial and channel convolution blocks and multi-scale window attention layers; The spatial and channel convolution blocks are used to compress and refine the multi-scale features to obtain reconstructed multi-scale features; The multi-scale window attention layer is used to model the long-range dependencies in the multi-scale features to obtain enhanced multi-scale features.

5. A cardiac MRI image segmentation method based on spatial channel reconstruction as claimed in claim 4, It is characterized in that The spatial and channel convolution blocks include spatial reconstruction convolution and channel reconstruction convolution; The spatial reconstruction convolution uses separation and reconstruction operations to distinguish two feature maps with rich information and poor information according to spatial content characteristics, and the two feature maps are merged into a spatial refinement feature map through a cross reconstruction operation; The channel reconstruction convolution uses segmentation and transformation operations to divide the input feature map into a feature map rich in representative features and a feature map supplementing detail information, and performs weighted fusion on the two feature maps to obtain reconstructed multi-scale features.

6. A cardiac MRI image segmentation method based on spatial channel reconstruction as claimed in claim 5, It is characterized in that The multi-scale window attention layer selects representative queries, keys, and values ​​in sliding windows of different scales and performs self-attention operations.

7. A cardiac MRI image segmentation method based on spatial channel reconstruction as claimed in claim 6, It is characterized in that The multi-scale window attention layer specifically operates as follows: Given a feature graph, obtain the corresponding query, key, and value through linear projection; The channels of the feature map are divided into multiple different heads, and each head uses a different dilation rate to perform multi-scale sliding window dilation attention.

8. A cardiac MRI image segmentation system based on spatial channel reconstruction, It is characterized in that Including acquisition module and segmentation module: An acquisition module is configured to: acquire a cardiac MRI image to be segmented; The segmentation module is configured to: input the cardiac MRI image into the trained segmentation model to obtain a segmented cardiac MRI image; Among them, the segmentation model introduces a spatial channel convolution block with a multi-scale extended Transformer on the U-Net structure, combines local sparse convolution blocks and sliding window expansion attention operations, performs spatial channel reconstruction, captures multi-scale contextual information and expands the receptive field.

9. An electronic device, Its characteristics include: a memory for non-transitory storage of computer readable instructions; a processor for executing the computer-readable instructions; Wherein, when the computer-readable instructions are executed by the processor, the method according to any one of claims 1 to 7 is executed.

10. A storage medium, Its characteristics are: The computer-readable instructions are non-transitorily stored, wherein when the computer-readable instructions are executed by a computer, the method of any one of claims 1 to 7 is performed.

Citation Information

Patent Citations

  • Choroid sublayer and great blood vessel joint segmentation method based on OCT (Optical Coherence Tomography) image

    CN115631207A

  • Cardiac MRI segmentation method and system based on U-Net and Transform fusion improvement

    CN116823850A

  • HL-UNet image segmentation model and heart dynamic magnetic resonance imaging segmentation method

    CN118470036A

  • Character wheel type mechanical water meter reading identification method based on Meter-YOLO model

    CN118587722A