Medical image lesion segmentation method and system based on frequency domain perception structure propagation
Patent Information
- Application Number
- CN202611051399.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-15
- Publication Date
- 2026-10-20
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision, medical image processing and deep learning, specifically to a method and system for medical image lesion segmentation based on frequency domain-aware structure propagation. Background Technology
[0002] Lesion segmentation in medical images is a core component of computer-aided diagnosis, directly impacting tumor staging, target delineation, and treatment efficacy evaluation. Clinically common tumor lesions often exhibit characteristics such as blurred boundaries, low contrast, and invasive growth. Traditional manual delineation is inefficient and suffers from poor inter-observer consistency, necessitating the support of high-precision automated segmentation methods.
[0003] Deep learning-based encoder-decoder segmentation frameworks, such as U-Net, have been widely used in medical image segmentation tasks. nnU-Net further enhances the generalization ability of segmentation methods by adaptively configuring preprocessing schemes, network architectures, and training strategies.
[0004] In recent years, researchers have combined CNNs with Transformers to capture both local details and global dependencies simultaneously. Hybrid architectures such as nnWNet have achieved significant progress in various medical image segmentation tasks. However, during multiple downsampling processes, the high-frequency structural information of lesion boundaries is easily attenuated and lost, leading to fragmented segmentation and incomplete contours in weak boundary regions.
[0005] To address this issue, existing frequency domain methods, while introducing discrete wavelet transform to extract high-frequency structural information, suffer from two shortcomings: First, they lack a targeted noise suppression mechanism, as directly enhancing high-frequency information can simultaneously amplify background texture and imaging artifacts; second, they lack a long-distance structural propagation mechanism, as isolated local high-frequency enhancement cannot solve the modeling problem of continuous boundaries in the immersion region. Existing technologies cannot simultaneously solve the problems of reliable extraction, denoising, and continuous propagation across regions of high-frequency structural information, resulting in limited segmentation accuracy in weak boundary regions. Summary of the Invention
[0006] To address the aforementioned technical problems, this application provides a method and system for medical image lesion segmentation based on frequency domain-aware structure propagation.
[0007] In a first aspect, this application provides a method for lesion segmentation in medical images based on frequency domain-aware structure propagation, comprising the following steps:
[0008] S1: Input the medical image into the encoder, downsample layer by layer to extract multi-scale features, and obtain a deep feature map;
[0009] S2: Input the deep feature map into the frequency domain perceptual structure propagation module to perform frequency domain decomposition and extract low-frequency semantic features and high-frequency structural features;
[0010] S3: Perform gating screening on the high-frequency structural features to obtain the screened high-frequency features;
[0011] S4: Input the selected high-frequency features into the bidirectional visual state space model unit for multi-directional state space serialization scanning and fusion to obtain enhanced high-frequency features;
[0012] S5: Reconstruct and fuse the low-frequency semantic features with the enhanced high-frequency features, and then inject them back into the deep feature map through an adaptive residual gating method to obtain the enhanced deep features;
[0013] S6: Input the enhanced deep features into the decoder for upsampling, combine them with the skip connection features of each layer of the encoder, and output the lesion segmentation result.
[0014] Secondly, this application provides a medical image lesion segmentation system based on frequency domain-aware structure propagation, comprising:
[0015] The feature extraction module is used to input medical images into the encoder, downsample layer by layer to extract multi-scale features, and obtain a deep feature map;
[0016] The frequency domain decomposition module is used to input the deep feature map into the frequency domain perceptual structure propagation module to perform frequency domain decomposition and extract low-frequency semantic features and high-frequency structural features.
[0017] The feature filtering module is used to perform gated filtering on the high-frequency structural features to obtain the filtered high-frequency features.
[0018] The multi-directional structure propagation module is used to input the selected high-frequency features into the bidirectional visual state space model unit for multi-directional state space serialization scanning and fusion to obtain enhanced high-frequency features.
[0019] The reconstruction and injection module is used to reconstruct and fuse the low-frequency semantic features with the enhanced high-frequency features, and then inject them back into the deep feature map through an adaptive residual gating method to obtain the enhanced deep features.
[0020] The decoding module is used to input the enhanced deep features into the decoder for upsampling, and combine them with the skip connection features of each layer of the encoder to output the lesion segmentation result.
[0021] Thirdly, this application provides an electronic device, comprising: at least one processor and a memory; the memory storing computer execution instructions; the at least one processor executing the computer execution instructions stored in the memory, causing the at least one processor to perform the medical image lesion segmentation method based on frequency domain-aware structure propagation as described in the first aspect and various possible designs of the first aspect.
[0022] The beneficial effects of this application are:
[0023] This application explicitly extracts and filters high-frequency information of lesion boundaries by embedding a frequency-domain perceptual structure propagation module; it uses spatial attention gating to suppress noise and introduces learnable weight coefficients for high-frequency subbands in each direction; it introduces a bidirectional visual state space model to establish global long-range dependencies with linear complexity, enabling scattered and broken local boundaries to form a continuous structural expression; and it significantly improves the segmentation accuracy of weak boundary regions through an adaptive residual-gated back-injection backbone network. Attached Figure Description
[0024] The technical solution and other beneficial effects of this application will become apparent from the following detailed description of specific embodiments in conjunction with the accompanying drawings.
[0025] Figure 1 A flowchart illustrating the overall process of the medical image lesion segmentation method based on frequency domain sensing structure propagation provided in this application embodiment;
[0026] Figure 2 A schematic diagram of the overall architecture of the medical image lesion segmentation model based on frequency domain sensing structure propagation and the internal structure of the frequency domain sensing structure propagation module provided in the embodiments of this application;
[0027] Figure 3 This is a schematic diagram of multi-directional sequential scanning of a bidirectional visual state space model provided in an embodiment of this application;
[0028] Figure 4 A schematic diagram comparing the results of lesion segmentation in medical images provided in the embodiments of this application. Detailed Implementation
[0029] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.
[0030] It should be noted that this embodiment uses the segmentation of gastric cancer CT lesions and the use of nnWNet as the backbone segmentation network as an example for illustration, but this does not constitute a limitation on the scope of protection of this application.
[0031] Reference Figure 1 This application provides a method for lesion segmentation in medical images based on frequency domain-aware structure propagation, comprising the following steps:
[0032] Step S1: Feature Extraction
[0033] Please refer to Figure 2In this embodiment, a medical image (in this embodiment, a 512×512 single-channel enhanced abdominal CT two-dimensional slice image) is input into the encoder of the nnWNet backbone segmentation network with an encoder-decoder structure. The nnWNet encoder contains four downsampling blocks, each consisting of two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function, downsampled through a 2×2 max-pooling layer. The input CT image undergoes four layer-by-layer downsampling steps to extract multi-scale features, resulting in a deep feature map. The output of each downsampling stage is retained as skip connection features for fusion in the corresponding layer of the decoder.
[0034] Furthermore, in this embodiment, the deep feature map is the bottleneck layer feature map of the backbone network, with a size of 32×32×256, which is the feature representation at 1 / 16 resolution of the input image.
[0035] Step S2, Frequency Domain Decomposition
[0036] Please refer to Figure 2 and Figure 3 In this embodiment of the application, the deep feature map obtained in step S1 is input to the discrete wavelet transform unit in the frequency domain sensing structure propagation module to perform a two-dimensional discrete wavelet transform.
[0037] Furthermore, in this embodiment, the Haar wavelet basis is used to explicitly decompose the deep feature map into four sub-bands: the low-frequency semantic sub-band. And three high-frequency structural sub-bands, corresponding to the horizontal sub-bands respectively. Vertical sub-band and diagonal sub-band The three high-frequency structural subbands are collectively referred to as This embodiment reduces the feature space resolution to half of its original value and increases the number of channels to four times its original value through this transformation, achieving lossless information decomposition.
[0038] Step S3: High-frequency gating screening
[0039] Further, please refer to Figure 2 The specific steps of step S3 are as follows:
[0040] First, the three high-frequency structural sub-bands , , The stitching is performed along the channel dimension, and the stitching formula is as follows:
[0041]
[0042] in, The corresponding number of characteristic channels is 3C, where C is the number of channels in a single high-frequency subband.
[0043] Secondly, average pooling and max pooling are performed on the spliced high-frequency features along the channel dimension to obtain two high-frequency response maps with preserved space dimensions.
[0044] Next, the two high-frequency response maps are concatenated along the channel dimension and then input into the spatial convolutional layer, and then... The activation function generates a single-channel spatial attention weight map with the same spatial size as the high-frequency subband. (The range is (0,1)), the calculation formula is as follows:
[0045]
[0046]
[0047]
[0048] in, This represents the average pooling operation along the channel dimension. This represents the max pooling operation along the channel dimension. This represents the spatial convolution operation. for Activation function. Weight graph. Medium- to high-response locations correspond to the actual lesion boundaries, while low-response locations correspond to background textures and imaging artifacts.
[0049] Finally, the spatial attention weight map Element-wise multiplication with the three original high-frequency structure subbands yields the filtered high-frequency features. (Include , , The formula is as follows:
[0050]
[0051] in, This represents the element-wise multiplication of matrices (Hadamard product).
[0052] Step S4, Multidirectional structural propagation
[0053] Please refer to Figure 2 and Figure 3 In this embodiment, corresponding learnable directional weight coefficients are applied to the selected high-frequency features in the horizontal, vertical, and diagonal directions, and then fed into the bidirectional visual state space model unit for multi-directional state space sequential scanning. The specific steps are as follows:
[0054] First, the three high-frequency features after screening , , Apply learnable directional weights respectively , , The network adaptively adjusts the contribution weights of high-frequency components in the horizontal, vertical, and diagonal directions to structural propagation, enabling it to differentiate and emphasize structural information in different directions based on the directional distribution characteristics of the lesion boundary. All weight coefficients are initialized to 1.0 and are jointly optimized with the network parameters during training. The weighted high-frequency features are denoted as... , , Collectively referred to as ;
[0055] Secondly, the weighted high-frequency features The sequence is expanded into a one-dimensional sequence and then independently scanned in the state space serialization along the four directions: top to bottom, bottom to top, left to right, and right to left. Each direction uses independent state space model parameters to establish long-range context dependencies for each position in the sequence.
[0056] Finally, the scan output features from the four directions are weighted and fused, and reconstructed into a two-dimensional feature map to obtain enhanced high-frequency features. .
[0057] Step S5: Reconstructing the injection
[0058] Please refer to Figure 2 The enhanced high-frequency features obtained in step S4 Compared with the low-frequency semantic subband obtained in step S2 The features are fed into a two-dimensional inverse discrete wavelet transform unit for inverse transformation and reconstruction, resulting in context fusion features with the same size as the original deep feature map X. Simultaneously, based on the three high-frequency structural sub-bands obtained in step S2... , , Calculate the high-frequency energy map The calculation formula is as follows:
[0059]
[0060] in, To prevent numerical instability, a minimal constant is introduced, which is taken in this embodiment. .
[0061] Furthermore, the high-frequency energy map After 1×1 convolution mapping and The activation function is used to obtain the spatial residual gating coefficient S, which is calculated using the following formula:
[0062]
[0063] in, Indicates a high-frequency structure subband , , The calculated high-frequency energy map, This represents the spatial residual gating coefficient used to control the intensity of context fusion feature backinjection.
[0064] Subsequently, the spatial residual gating coefficient was used. Context fusion features Perform element-wise gating and use a learnable scaling factor. Adaptively back-inject the gated context fusion features into the original deep feature map. The enhanced deep features are obtained. The calculation formula is as follows:
[0065]
[0066] in, This indicates element-wise multiplication. This represents the learnable scaling factor. This represents the original deep feature map of the input frequency domain sensing structure propagation module. This represents the context fusion feature reconstructed by the two-dimensional inverse discrete wavelet transform. This represents the enhanced deep features.
[0067] Step S6, Decoding Output
[0068] Please refer to Figure 2 The enhanced deep features obtained in step S5 The image is fed into a decoder containing four upsampling blocks. Each upsampling block restores the spatial resolution through bilinear interpolation or transposed convolution, and is then concatenated and fused with the skip connection features of the corresponding layer in the encoder to gradually restore the input image resolution. Finally, it is processed by 1×1 convolution and... The activation function outputs a segmentation prediction map.
[0069] Furthermore, the embodiments of this application will be specifically described from two dimensions: model training and performance evaluation.
[0070] Furthermore, the backbone segmentation network in this embodiment is trained as follows: training set data is input into the backbone network containing the frequency domain-aware structure propagation module for end-to-end joint training. All learnable parameters in the network participate in gradient updates, including backbone network parameters, convolution parameters in the high-frequency spatial attention gating unit, and learnable direction weight coefficients. , , and adaptive fusion scaling factor Preprocessing normalizes CT values to [-1,1] and resamples them to a resolution of 1mm×1mm×1mm.
[0071] To accurately guide network convergence, this embodiment employs a joint supervision model training method using the Dice loss function and the Cross-Entropy (CE) loss function. Let... This is a probability map for lesion segmentation prediction output by the network. Manually draw the gold standard (Ground Truth) for the corresponding experts. This represents the total number of pixels in the image.
[0072] Furthermore, the formula for calculating the Dice coefficient used in the model evaluation in this application embodiment is as follows:
[0073]
[0074] in, Indicates the first The predicted probability that a pixel belongs to a lesion. Indicates the first The real label of each pixel For the smoothing term, this embodiment takes... To ensure numerical stability.
[0075] Based on the Dice coefficients mentioned above, the Dice loss function used for network training is... Defined as:
[0076]
[0077] Meanwhile, the classic binary classification cross-entropy loss function The expression is:
[0078]
[0079] Finally, the total loss function used for end-to-end joint training The configuration is a weighted sum with each component having a weight of 0.5, and the formula is:
[0080]
[0081] Furthermore, in this embodiment, the optimizer is AdamW, and the initial learning rate is set to... The weight decays to Training employs 5-fold cross-validation with a batch size of 8 and a maximum training epoch configuration of 200. An early stopping mechanism is triggered when the Dice coefficient on the validation set no longer improves for 10 consecutive epochs, and the model parameters with the best performance on the validation set are used as the final segmentation model.
[0082] In summary, this application embeds a frequency domain-aware structure propagation module into the backbone network in a plug-and-play manner through steps S1 to S6, thereby realizing explicit extraction of high-frequency structural information (step S2), adaptive noise suppression (step S3), directional adaptive weighting and cross-regional continuous propagation (step S4), and residual enhancement back injection (step S5), which can effectively solve the problem of weak boundary region contour breakage in lesion segmentation of medical images.
[0083] Five-fold cross-validation experiments on a gastric cancer CT dataset (containing 251 enhanced CT scans and corresponding lesion annotations) showed that, compared with the baseline nnWNet, the method of this invention improved the average Dice coefficient by 2.1% (from 59.3% to 61.4%) and decreased HD95 by 10.67 mm (from 50.09 mm to 39.42 mm), significantly improving the contour integrity and structural continuity of invasive lesion segmentation.
[0084] To further demonstrate the technical advantages of this application, please refer to... Figure 4 The figure shows a comparative diagram of lesion segmentation results in medical images. Figure 4 From left to right, the images show: the original CT image, the gold standard (GT) manually drawn by the doctor, the segmentation result of the baseline method (nnWNet), and the segmentation result of the method of this invention. Through the comparison of these four columns, it is clearly observed that when segmenting gastric cancer lesions with relatively blurred boundaries, the traditional baseline model is prone to undersegmentation, edge breaks, or false positive areas in weak boundary regions. In contrast, the segmentation result output by the method of this invention significantly improves the continuity, integrity, and accuracy of the boundaries, and has a higher degree of fit with the doctor's manually drawn gold standard. It effectively suppresses oversegmentation and undersegmentation problems, verifying the technical improvement effect of this application in the structural representation of weak boundary regions.
[0085] In one exemplary embodiment, a medical image lesion segmentation system based on frequency domain-aware structure propagation is provided, characterized in that it includes:
[0086] The feature extraction module is used to input medical images into the encoder, downsample layer by layer to extract multi-scale features, and obtain a deep feature map;
[0087] The frequency domain decomposition module is used to input the deep feature map into the frequency domain perceptual structure propagation module to perform frequency domain decomposition and extract low-frequency semantic features and high-frequency structural features.
[0088] The feature filtering module is used to perform gated filtering on the high-frequency structural features to obtain the filtered high-frequency features.
[0089] The multi-directional structure propagation module is used to input the selected high-frequency features into the bidirectional visual state space model unit for multi-directional state space serialization scanning and fusion to obtain enhanced high-frequency features.
[0090] The reconstruction and injection module is used to reconstruct and fuse the low-frequency semantic features with the enhanced high-frequency features, and then inject them back into the deep feature map through an adaptive residual gating method to obtain the enhanced deep features.
[0091] The decoding module is used to input the enhanced deep features into the decoder for upsampling, and combine them with the skip connection features of each layer of the encoder to output the lesion segmentation result.
[0092] In one exemplary embodiment, an electronic device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above-described method for medical image lesion segmentation based on frequency domain-aware structure propagation.
[0093] The above provides a detailed description of a medical image lesion segmentation method and system based on frequency domain sensing structure propagation provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the technical solutions and core ideas of this application. Those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.
Claims
1. A method for lesion segmentation in medical images based on frequency domain-aware structure propagation, characterized in that, Includes the following steps: S1: Input the medical image into the encoder, downsample layer by layer to extract multi-scale features, and obtain a deep feature map; S2: Input the deep feature map into the frequency domain perceptual structure propagation module to perform frequency domain decomposition and extract low-frequency semantic features and high-frequency structural features; S3: Perform gating screening on the high-frequency structural features to obtain the screened high-frequency features; S4: Input the selected high-frequency features into the bidirectional visual state space model unit for multi-directional state space serialization scanning and fusion to obtain enhanced high-frequency features; S5: Reconstruct and fuse the low-frequency semantic features with the enhanced high-frequency features, and then inject them back into the deep feature map through an adaptive residual gating method to obtain the enhanced deep features; S6: Input the enhanced deep features into the decoder for upsampling, combine them with the skip connection features of each layer of the encoder, and output the lesion segmentation result.
2. The method for medical image lesion segmentation based on frequency domain sensing structure propagation according to claim 1, characterized in that, The encoder contains four downsampling blocks, each consisting of two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function. It uses 2×2 max pooling for downsampling while retaining the output of each downsampling stage as skip connection features.
3. A method for lesion segmentation in medical images based on frequency domain-aware structure propagation according to claim 1 or 2, characterized in that, In the frequency domain sensing structure propagation module, a discrete wavelet transform unit is used to explicitly decompose the deep feature map into one low-frequency semantic sub-band and three high-frequency structure sub-bands based on the Haar wavelet basis.
4. The method for lesion segmentation in medical images based on frequency domain-aware structure propagation according to claim 3, characterized in that, The three high-frequency structural subbands include a horizontal subband, a vertical subband, and a diagonal subband.
5. The method for lesion segmentation in medical images based on frequency domain-aware structure propagation according to claim 4, characterized in that, In step S3, when gating and filtering high-frequency structural features to obtain the filtered high-frequency features, the following steps are performed: The three high-frequency structural sub-bands are spliced together along the channel dimension; The high-frequency features after splicing are averaged and max-pooled respectively to obtain high-frequency response maps with two retained space sizes; The two high-frequency response maps are concatenated along the channel dimension and then input into a spatial convolutional layer, and then... The activation function generates a single-channel spatial attention weight map with the same spatial size as the high-frequency subband. The spatial attention weight map is multiplied element-wise with the three original high-frequency structural subbands to obtain the filtered high-frequency features.
6. The method for lesion segmentation in medical images based on frequency domain-aware structure propagation according to claim 1, characterized in that, When performing S4 to obtain enhanced high-frequency features, the following steps are performed: Learnable directional weight coefficients are applied to the selected high-frequency features; The weighted high-frequency features are unfolded into a one-dimensional sequence, and state space serialization scanning is performed independently in four directions: top to bottom, bottom to top, left to right, and right to left, with independent state space model parameters used in each direction; The scan output features from the four directions are weighted, fused, and reshaped to obtain enhanced high-frequency features.
7. A method for lesion segmentation in medical images based on frequency domain-aware structure propagation according to claim 1 or 6, characterized in that, When performing S5 to obtain the enhanced deep features, the following steps are performed: The enhanced high-frequency features and low-frequency semantic subbands are input into a two-dimensional inverse discrete wavelet transform unit, and then reconstructed and fused through inverse transform to obtain context fusion features; Calculate the high-frequency energy map based on three high-frequency structural sub-bands; The high-frequency energy map is processed by 1×1 convolution mapping and Sigmoid activation function to obtain the spatial residual gating coefficients; The context fusion features are gated element-wise using spatial residual gating coefficients. The gated context fusion features are then adaptively back-injected into the original deep feature map using a learnable scaling factor to obtain the enhanced deep features.
8. The method for lesion segmentation in medical images based on frequency domain-aware structure propagation according to claim 7, characterized in that, The decoder contains four upsampling blocks. Each upsampling block restores the spatial resolution through bilinear interpolation or transposed convolution, and is then spliced and fused with the skip connection features of the corresponding layer of the encoder to gradually restore the resolution to the input image.
9. A medical image lesion segmentation system based on frequency domain sensing structure propagation, characterized in that, include: The feature extraction module is used to input medical images into the encoder, downsample layer by layer to extract multi-scale features, and obtain a deep feature map; The frequency domain decomposition module is used to input the deep feature map into the frequency domain perceptual structure propagation module to perform frequency domain decomposition and extract low-frequency semantic features and high-frequency structural features. The feature filtering module is used to perform gated filtering on the high-frequency structural features to obtain the filtered high-frequency features. The multi-directional structure propagation module is used to input the selected high-frequency features into the bidirectional visual state space model unit for multi-directional state space serialization scanning and fusion to obtain enhanced high-frequency features. The reconstruction and injection module is used to reconstruct and fuse the low-frequency semantic features with the enhanced high-frequency features, and inject them back into the deep feature map through an adaptive residual gating method to obtain the enhanced deep features; The decoding module is used to input the enhanced deep features into the decoder for upsampling, and combine them with the skip connection features of each layer of the encoder to output the lesion segmentation result.
10. An electronic device, characterized in that, It includes a memory and a processor; the memory stores an application program, and the processor is used to run the application program in the memory to perform the steps in the medical image lesion segmentation method according to any one of claims 1 to 8.