A medical image segmentation method based on dynamic region aggregation convolution
This medical image segmentation method, which utilizes dynamic region aggregation convolution, addresses the limitation of local receptive fields in convolutional neural networks for medical image segmentation. It achieves efficient global context modeling and accuracy improvement, making it suitable for various medical image segmentation tasks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- UNIV OF ELECTRONICS SCI & TECH OF CHINA
- Filing Date
- 2025-07-29
- Publication Date
- 2026-06-09
Smart Images
Figure CN121121094B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image segmentation technology, specifically to a medical image segmentation method based on dynamic region aggregation convolution. Background Technology
[0002] Medical image segmentation is a core task in computer-aided diagnosis and treatment planning, widely used for the automatic identification of anatomical structures or lesion regions in various imaging modalities such as CT, MRI, and ultrasound. In recent years, convolutional neural networks (CNNs) have become the mainstream method in this field due to their excellent spatial modeling capabilities. Models such as UNet and nnUNet are widely used in multi-organ and lesion segmentation tasks.
[0003] However, traditional convolutional neural networks, limited by their local receptive field design, have an inherent disadvantage in modeling long-range dependencies and capturing the global context of images. Medical images, on the other hand, exhibit highly regular structural locations and significant contextual constraints; for example, brain structures do not appear on the abdomen, and lungs do not appear near the thyroid gland. Therefore, effectively modeling long-range dependency information is crucial for improving segmentation accuracy.
[0004] To compensate for the limitations of convolutional neural networks (CNNs) in long-range modeling, some works employ large convolutional kernels, dilated convolutions, or introduce Transformer modules to achieve long-distance modeling between features. However, these methods often incur significant computational overhead or parameter redundancy, limiting their practical application in medical images. PeLK is a parameter-efficient CNN architecture based on maximally ...
[0005] Although PeLK has certain advantages in expanding the receptive field and reducing the number of parameters, its core method still has several limitations, which are precisely the problems that this invention aims to solve:
[0006] Still limited by convolutional kernel structure design, PeLK's computation is still strictly constrained by the sliding window mechanism, lacking a full-image-level interaction mechanism between display pixels. Each convolution only acts on a local region, which, although covering a wide range, is essentially a local stacking model and cannot achieve efficient global dependency modeling within a single layer.
[0007] Contextual modeling lacks directional structure and spatial explicitness: PeLK uses uniform shared parameters for aggregation in the outer region, which cannot distinguish the different semantic contributions of different directions (such as upper left, lower right, etc.) to the central pixel, making it difficult for the model to capture important spatial layout information in medical images.
[0008] Fixed receptive field configuration and lack of input adaptability: The size of the PeLK convolution kernel is preset during the training phase and does not have the ability to dynamically and adaptively adjust according to the input image resolution, regional structure or organ size, which is not conducive to maintaining robustness in highly heterogeneous medical image data. Summary of the Invention
[0009] To address the shortcomings of existing technologies, this invention provides a medical image segmentation method based on dynamic region aggregation convolution, which solves the problems mentioned in the background section.
[0010] To achieve the above objectives, the present invention provides the following technical solution: a medical image segmentation method based on dynamic region aggregation convolution, comprising the following steps:
[0011] S1. Construct an image segmentation network architecture based on full-region convolution (ORConv). The network architecture includes a full-region encoder architecture (Orea), a lightweight upsampling decoder architecture (Luna), and a multi-scale full-region skip enhancement module (Mosa).
[0012] S2. Using the Orea as an encoder, hierarchical feature extraction is performed on the input medical image. The Orea contains multiple Orea modules, and each Orea module integrates ORConv to realize the modeling of local and global contextual information.
[0013] S3. Using the Luna as a decoder, spatial resolution is restored to the features output by the encoder. The Luna adopts a simplified structure to maintain low computational overhead.
[0014] S4. By enhancing the skip connection features through the Mosa, the multi-scale features of the encoder are effectively transmitted to the decoder. The Mosa enhances the feature representation capability through multi-scale region perception fusion.
[0015] S5. Use the network architecture to segment the medical image and output the segmentation result.
[0016] Optionally, the implementation steps of the full-region convolution (ORConv) include:
[0017] S101, For the input feature map At each sliding window position (h, w), the feature map is divided into 9 regions: 1 central region. and 8 surrounding areas The central area For a fixed size, the shape of the surrounding area changes dynamically with (h,w);
[0018] S102, Calculate the output Y of the central region center(c,h,w) uses independent convolution parameters to extract local fine features through depthwise convolution;
[0019] S103, Calculate the output Y of the surrounding area. surround (c,h,w) uses shared weights to sum the pixels in each surrounding region and introduces a region suppression factor to balance the contributions of different regions;
[0020] S104. Merge the outputs from the central region and the surrounding region to obtain the final output Y(c,h,w) of ORConv. center (c,h,w)+λ·Y surround (c,h,w), where λ is the weight parameter with a value of 0.25.
[0021] Optionally, the central region outputs Y center The formula for calculating (c,h,w) is:
[0022]
[0023] in, Here are the learnable convolution kernel parameters for the central region, and K×K is the size of the central convolution window. The pixel value of the central region at position (i,j) in channel c.
[0024] Optionally, the surrounding area outputs Y surround The calculation steps for (c,h,w) include:
[0025] (1) For each surrounding area Within each channel, pixel summation is performed within the region to obtain the region aggregation feature;
[0026] (2) Introducing a shared weight tensor The aggregation features of the eight surrounding regions were weighted separately;
[0027] (3) Based on the number of pixels in the region Calculate the region inhibition factor And apply it to the weighted aggregated features;
[0028] (4) Calculate the output of the suppressed surrounding region using the following formula:
[0029]
[0030] Among them, H i W i The surrounding areas Height and width.
[0031] Optionally, the region inhibition factor The formula for calculation is:
[0032]
[0033] Where α is a hyperparameter with a value of 0.5. For the region The number of pixels (excluding channels).
[0034] Optionally, the Orea module implementation steps in step S2 include:
[0035] S201. Expand the channel dimension of the input features through the first linear layer;
[0036] S202. Input the expanded channel features into ORConv to perform local and global context modeling.
[0037] S203. The number of channels output by ORConv is reduced back to its original size through the second linear layer;
[0038] S204. Two residual paths are introduced: one is a module-level residual connection (direct connection between input and output), and the other is a residual path outside the ORConv layer to enhance training stability.
[0039] Optionally, the Luna implementation steps in step S3 include:
[0040] S301. Channel adjustment of input features is performed through a linear mapping layer;
[0041] S302. Use 9×9 depthwise convolution to refine the spatial features of the adjusted features;
[0042] S303, further process the features through another linear layer;
[0043] S304. Perform bilinear upsampling to restore the spatial resolution of the features.
[0044] Optionally, the Mosa implementation step in step S4 includes:
[0045] S401. Perform hierarchical scaling modulation on the input skip connection features by scaling and shifting the normalized features and the original features element by element using two learnable vectors.
[0046] S402. Perform channel dimensionality reduction on the scaled features through a linear layer;
[0047] S403. Input the dimensionality-reduced features into three parallel branches. Each branch contains an ORConv module and an SE attention module. The ORConv modules of the three branches use 3×3, 5×5, and 7×7 convolution kernels respectively to capture the spatial dependence of different receptive fields.
[0048] S404. After averaging the outputs of the three branches, perform residual fusion with the results of the input branches, process them with the activation function, project them back to the original channel dimension, and fuse them again with the module input to obtain the enhanced skip connection features.
[0049] This invention provides a medical image segmentation method based on dynamic region aggregation convolution, which has the following beneficial effects:
[0050] This medical image segmentation method based on dynamic region aggregation convolution, with its proposed full-region convolution and derived image segmentation network architecture, demonstrates significant performance improvements and efficiency advantages in several representative medical image segmentation tasks. Specifically, it has the following notable effects:
[0051] The method of this invention has been comprehensively compared with mainstream CNN and Transformer architectures on multiple publicly available medical image segmentation benchmarks, achieving leading performance in all cases. In the Synapse multi-organ CT segmentation task, this method achieves an average Dice coefficient of 84.13% and an HD95 of 10.44, outperforming strong baseline models such as PVT-EMCAD-B2 (83.63%, HD95 of 15.68). In the ACDC cardiac MRI segmentation, it achieves an average Dice score of 92.15%, surpassing representative architectures such as TransUNet, SwinUNet, MISSFormer, and TransCASCADE. In multiple lesion binary segmentation tasks (Kvasir-SEG, CVC-ClinicDB, CVC-ColonDB, TN3K), this method achieves the highest Dice score on all datasets, with a maximum improvement of 1.2%.
[0052] The method of this invention effectively reduces the overall number of parameters and computational cost of the model while maintaining or improving segmentation accuracy; the entire network contains only 3.05M parameters and 3.85GFLOPs, which is significantly lower than mainstream Transformer architectures (such as SwinUNet and TransUNet) and high-parameter convolutional models (such as RepLKNet and PeLK); this indicates that the invention can be efficiently deployed in resource-constrained environments such as edge devices and clinical real-time systems.
[0053] This invention does not rely on specific task data or assumed structures, but completes region context modeling only through convolution operations, thus possessing good task versatility. Experimental results show that the structure can be stably applied to multiple task scenarios with blurred organ boundaries, large differences in structural scale, and different imaging modalities, and exhibits strong cross-dataset generalization ability. Attached Figure Description
[0054] Figure 1 This is a schematic diagram of the invention process;
[0055] Figure 2 A schematic diagram of the full-area convolution of the invention (left) and a schematic diagram of a medical image (right);
[0056] Figure 3 The diagram shows the framework of the invention, where (a) the overall architecture of the network consists of an omni-region encoder architecture (Orea), a lightweight upsample decoder architecture (Luna), and a multi-scale omni-region skipping enhancement module (Mosa); (b) the design of the Orea module, which integrates ORConv to achieve omni-region encoding; (c) the design of the Luna module, which uses depthwise convolution to maintain a lightweight structure; and (d) the design of the Mosa module, which enhances skip connections through multi-scale region-aware fusion. Detailed Implementation
[0057] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0058] In the description of this invention, unless otherwise stated, "a plurality of" means two or more; the terms "upper," "lower," "left," "right," "inner," "outer," "front end," "rear end," "head," "tail," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0059] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0060] Please see Figures 1 to 3 A medical image segmentation method based on dynamic region aggregation convolution includes the following steps:
[0061] S1. Construct an image segmentation network architecture based on full-region convolution (ORConv). The network architecture includes a full-region encoder architecture (Orea), a lightweight upsampling decoder architecture (Luna), and a multi-scale full-region skip enhancement module (Mosa).
[0062] The implementation steps of full-region convolution (ORConv) include:
[0063] S101, For the input feature map At each sliding window position (h, w), the feature map is divided into 9 regions: 1 central region. and 8 surrounding areas The central area For a fixed size, the shape of the surrounding area changes dynamically with (h,w);
[0064] S102, Calculate the output Y of the central region center (c,h,w) uses independent convolution parameters to extract local fine features through depthwise convolution;
[0065] Central area output Y center The formula for calculating (c,h,w) is:
[0066]
[0067] in, Here are the learnable convolution kernel parameters for the central region, and K×K is the size of the central convolution window. The pixel value of the central region at position (i,j) in channel c;
[0068] S103, Calculate the output Y of the surrounding area. surround (c,h,w) uses shared weights to sum the pixels in each surrounding region and introduces a region suppression factor to balance the contributions of different regions;
[0069] Output Y in the surrounding area surround The calculation steps for (c,h,w) include:
[0070] (1) For each surrounding area Within each channel, pixel summation is performed within the region to obtain the region aggregation feature;
[0071] (2) Introducing a shared weight tensor The aggregation features of the eight surrounding regions were weighted separately;
[0072] (3) Based on the number of pixels in the region Calculate the region inhibition factor And apply it to the weighted aggregated features;
[0073] (4) Calculate the output of the suppressed surrounding region using the following formula:
[0074]
[0075] Among them, H i W i The surrounding areas Height and width;
[0076] Restrictive factors The formula for calculation is:
[0077]
[0078] Where α is a hyperparameter with a value of 0.5. For the region The number of pixels (excluding channels);
[0079] S104. Merge the outputs from the central region and the surrounding region to obtain the final output Y(c,h,w) of ORConv. center (c,h,w)+λ·Y surround (c,h,w), where λ is the weight parameter with a value of 0.25;
[0080] S2. Using the Orea as an encoder, hierarchical feature extraction is performed on the input medical image. The Orea contains multiple Orea modules, and each Orea module integrates ORConv to realize the modeling of local and global contextual information.
[0081] The steps to implement the Orea module include:
[0082] S201. Expand the channel dimension of the input features through the first linear layer;
[0083] S202. Input the expanded channel features into ORConv to perform local and global context modeling.
[0084] S203. The number of channels output by ORConv is reduced back to its original size through the second linear layer;
[0085] S204. Two residual paths are introduced: one is a module-level residual connection (direct connection between input and output), and the other is a residual path outside the ORConv layer to enhance training stability.
[0086] S3. Using the Luna as a decoder, spatial resolution is restored to the features output by the encoder. The Luna adopts a simplified structure to maintain low computational overhead.
[0087] The steps to implement Luna include:
[0088] S301. Channel adjustment of input features is performed through a linear mapping layer;
[0089] S302. Use 9×9 depthwise convolution to refine the spatial features of the adjusted features;
[0090] S303, further process the features through another linear layer;
[0091] S304. Perform bilinear upsampling to restore the spatial resolution of the features;
[0092] S4. By enhancing the skip connection features through the Mosa, the multi-scale features of the encoder are effectively transmitted to the decoder. The Mosa enhances the feature representation capability through multi-scale region perception fusion.
[0093] The steps involved in implementing Mosa include:
[0094] S401. Perform hierarchical scaling modulation on the input skip connection features by scaling and shifting the normalized features and the original features element by element using two learnable vectors.
[0095] S402. Perform channel dimensionality reduction on the scaled features through a linear layer;
[0096] S403. Input the dimensionality-reduced features into three parallel branches. Each branch contains an ORConv module and an SE attention module. The ORConv modules of the three branches use 3×3, 5×5, and 7×7 convolution kernels respectively to capture the spatial dependence of different receptive fields.
[0097] S404. After averaging the outputs of the three branches, the residual is fused with the results of the input branches. After processing by the activation function, the residual is projected back to the original channel dimension and fused with the module input again to obtain the enhanced skip connection feature.
[0098] S5. Use the network architecture to segment the medical image and output the segmentation result.
[0099] Table 1 provides a quantitative comparison of segmentation performance on the Synapse dataset, with evaluation metrics including mDice and mHD95.
[0100]
[0101]
[0102] Table 2 provides a quantitative comparison of segmentation performance on the ACDC dataset, using mDice as the evaluation metric.
[0103]
[0104] Table 3 compares the performance on the binary lesion segmentation dataset, using mDice as the evaluation metric; results are reported as mean ± standard deviation of three runs.
[0105]
[0106]
[0107] The model is lightweight, significantly reducing the number of parameters and computational complexity.
[0108] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A medical image segmentation method based on dynamic region aggregation convolution, characterized in that, Includes the following steps: S1. Construct an image segmentation network architecture based on full-region convolution ORConv. The network architecture includes a full-region encoder architecture Orea, a lightweight upsampling decoder architecture Luna, and a multi-scale full-region skip enhancement module Mosa. The steps to implement full-region convolution include: S101, For the input feature map At each sliding window position The feature map is divided into 9 regions: 1 central region. and 8 surrounding areas The central area For a fixed size, the shape of the surrounding area varies. Dynamic changes; S102, Calculate the output of the central region. It uses independent convolution parameters to extract local fine features through depthwise convolution; S103, Calculate the output of the surrounding area. A shared weight is used to sum the pixels in each surrounding region, and a region suppression factor is introduced to balance the contributions of different regions. S104. Merge the output from the central region with the output from the surrounding region to obtain the final output of ORConv. ,in This is a weighting parameter with a value of 0.25; S2. Using the Orea as an encoder, hierarchical feature extraction is performed on the input medical image. The Orea contains multiple Orea modules, and each Orea module integrates ORConv to realize the modeling of local and global contextual information. S3. Using the Luna as a decoder, spatial resolution is restored to the features output by the encoder. The Luna adopts a simplified structure to maintain low computational overhead. S4. By enhancing the skip connection features through the Mosa, the multi-scale features of the encoder are effectively transmitted to the decoder. The Mosa enhances the feature representation capability through multi-scale region perception fusion. S5. Use the network architecture to segment the medical image and output the segmentation result.
2. The medical image segmentation method based on dynamic region aggregation convolution as described in claim 1, characterized in that, The central area output The formula for calculation is: ； in, The learnable convolutional kernel parameters for the central region, The size of the center convolution window, For the central area in the passage ,Location The pixel value.
3. The medical image segmentation method based on dynamic region aggregation convolution as described in claim 1, characterized in that, The surrounding area output The calculation steps include: (1) For each surrounding area Within each channel, pixel summation within the region is performed to obtain the region aggregation feature; (2) Introduce shared weight tensors The aggregation features of the eight surrounding areas are weighted separately; (3) Based on the number of pixels in the region Calculate the region inhibition factor And apply it to the weighted aggregated features; (4) Calculate the output of the suppressed surrounding region, using the following formula: ； in, , The surrounding areas The height and width.
4. The medical image segmentation method based on dynamic region aggregation convolution as described in claim 3, characterized in that, The region inhibitory factor The formula for calculation is: ； in, This is a hyperparameter with a value of 0.
5. For the region The number of pixels (excluding channels).
5. The medical image segmentation method based on dynamic region aggregation convolution as described in claim 1, characterized in that, The Orea module implementation steps in step S2 include: S201. Expand the channel dimension of the input features through the first linear layer; S202. Input the expanded channel features into ORConv to perform local and global context modeling. S203. The number of channels output by ORConv is reduced back to its original size through the second linear layer; S204. Two residual paths are introduced: one is a module-level residual connection (direct connection between input and output), and the other is a residual path outside the ORConv layer to enhance training stability.
6. The medical image segmentation method based on dynamic region aggregation convolution as described in claim 1, characterized in that, The Luna implementation steps in step S3 include: S301. Channel adjustment of input features is performed through a linear mapping layer; S302. Use 9×9 depthwise convolution to refine the spatial features of the adjusted features; S303, further process the features through another linear layer; S304. Perform bilinear upsampling to restore the spatial resolution of the features.
7. The medical image segmentation method based on dynamic region aggregation convolution as described in claim 1, characterized in that, The Mosa implementation steps in step S4 include: S401. Perform hierarchical scaling modulation on the input skip connection features by scaling and shifting the normalized features and the original features element by element using two learnable vectors. S402. Perform channel dimensionality reduction on the scaled features through a linear layer; S403. Input the dimensionality-reduced features into three parallel branches. Each branch contains an ORConv module and an SE attention module. The ORConv modules of the three branches use 3×3, 5×5, and 7×7 convolution kernels respectively to capture the spatial dependence of different receptive fields. S404. After averaging the outputs of the three branches, perform residual fusion with the results of the input branches, process them with the activation function, project them back to the original channel dimension, and fuse them again with the module input to obtain the enhanced skip connection features.