Soft supervision medical image segmentation method and system based on pseudo boundary suppression

Abstract

The invention provides a soft supervision medical image segmentation method and system based on pseudo boundary suppression, and relates to the technical field of medical image processing and analysis, and the method comprises the steps: obtaining a to-be-segmented medical image; a medical image is input into a pseudo boundary suppression network of a U-Net structure, firstly, an encoder extracts multi-scale features from the medical image, then, a decoder maps the multi-scale features back to a pixel space to obtain a prediction probability graph, and finally, a lesion segmentation result is generated through thresholding operation; wherein a soft supervision strategy is introduced in the training process of the pseudo boundary suppression network, a boundary loss item constructed based on edge complexity is added in a loss function, and the weight of the boundary loss item is adaptively modulated; according to the invention, the pseudo-boundary suppression network is constructed based on the soft supervision mechanism and the frequency selective representation, the generation of pseudo-boundaries is avoided while the fuzzy boundaries are processed, and the precision and robustness of medical image segmentation are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and analysis technology, specifically to a soft-supervised medical image segmentation method and system based on pseudo-boundary suppression. Background Technology

[0002] In the field of medical image segmentation, especially lesion segmentation in medical images, accurate identification of structural boundaries in the image is crucial. Currently, many medical image segmentation methods rely on hard labels (such as binarized masks) as supervisory signals. This hard labeling method often forces blurred transition regions to sharp edges when annotating images. While this method can provide efficient segmentation results in some cases, it ignores the complexity and blurred boundaries of many anatomical structures in medical images, leading to the generation of false boundaries.

[0003] Traditional image segmentation methods predict lesion regions in images through pixel-level classification or region segmentation. These methods typically optimize the model by minimizing loss functions, such as cross-entropy loss or Dice loss, to minimize the difference between the predicted segmentation map and the ground truth label. However, these methods perform poorly for regions with blurred boundaries, especially in breast ultrasound images, where the similarity of pixel intensity distribution between the lesion and surrounding tissue fluctuates greatly, making it difficult for traditional methods to accurately capture details and boundaries.

[0004] With the development of deep learning technology, Convolutional Neural Networks (CNNs) have made significant progress in medical image segmentation. Early models based on shallow CNNs could train quickly and provide basic segmentation results with small datasets. However, as the complexity of medical images increases, especially when dealing with regions with blurred boundaries and unclear structures, the feature extraction capabilities of shallow networks gradually become insufficient. To improve segmentation accuracy, U-Net and its variants (such as U-Net++ and Attention U-Net) have effectively improved multi-scale feature extraction capabilities and adaptability to blurred boundaries by introducing encoder-decoder structures and skip connections, but problems such as over-sharpening of boundaries and the formation of false boundaries still exist.

[0005] To address the issues of false and blurred boundaries, researchers have proposed various methods to enhance model performance. In recent years, frequency domain-based modeling methods have achieved initial success, particularly when combined with traditional spatial feature extraction. Frequency domain enhancement techniques can improve the sensitivity of segmentation models to boundaries, especially by effectively suppressing high-frequency noise in images. However, most existing methods still have limitations when dealing with complex boundaries in medical images, especially in areas with unclear boundaries or noise, where false boundaries are easily generated, leading to over-segmentation or under-segmentation. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes a soft-supervised medical image segmentation method and system based on pseudo-boundary suppression. Based on a soft-supervised mechanism and frequency-selective representation, a pseudo-boundary suppression network (PBS-Net) is constructed. This network handles blurred boundaries while avoiding the generation of pseudo-boundaries, significantly improving the accuracy and robustness of medical image segmentation.

[0007] According to some embodiments, the present invention adopts the following technical solution: A soft-supervised medical image segmentation method based on pseudo-boundary suppression includes: Acquire the medical image to be segmented; The medical image is input into the pseudo-boundary suppression network with a U-Net structure. First, the encoder extracts multi-scale features from the medical image. Then, the decoder maps the multi-scale features back to the pixel space to obtain the predicted probability map. Finally, the lesion segmentation result is generated through thresholding. In the training process of the pseudo-boundary suppression network, a soft supervision strategy is introduced, which adds a boundary loss term based on edge complexity to the loss function and adaptively modulates the weight of the boundary loss term.

[0008] According to some embodiments, the present invention adopts the following technical solution: A soft-supervised medical image segmentation system based on pseudo-boundary suppression, comprising: The image acquisition module is configured to acquire medical images to be segmented; The image segmentation module is configured to: input medical images into a pseudo-boundary suppression network with a U-Net structure; firstly, the encoder extracts multi-scale features from the medical images; then the decoder maps the multi-scale features back to the pixel space to obtain a predicted probability map; and finally, the lesion segmentation result is generated through a thresholding operation. In the training process of the pseudo-boundary suppression network, a soft supervision strategy is introduced, which adds a boundary loss term based on edge complexity to the loss function and adaptively modulates the weight of the boundary loss term.

[0009] According to some embodiments, the present invention adopts the following technical solution: A computer program product includes a computer program that, when executed by a processor, implements the aforementioned soft-supervised medical image segmentation method based on pseudo-boundary suppression.

[0010] According to some embodiments, the present invention adopts the following technical solution: A non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the aforementioned soft-supervised medical image segmentation method based on pseudo-boundary suppression.

[0011] According to some embodiments, the present invention adopts the following technical solution: An electronic device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the aforementioned soft-supervised medical image segmentation method based on pseudo-boundary suppression.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a pseudo-boundary suppression network (PBS-Net) based on soft supervision and frequency-selective representation to achieve high-precision segmentation of medical images. Through the dual mechanism of soft supervision and frequency-selective representation, it avoids the generation of pseudo-boundaries while handling blurred boundaries, significantly improving the accuracy and robustness of medical image segmentation.

[0013] This method uses convolutional neural networks to extract multi-scale features from images and combines soft supervision mechanisms to strengthen supervision in regions with high boundary complexity and reduce supervision in blurred boundary regions, thereby effectively reducing the generation of false boundaries. The method calculates the boundary complexity of the image through Sobel gradient and dynamically adjusts the structural loss according to the complexity, enhancing the flexibility of the model in blurred transition regions.

[0014] A frequency-selective representation strategy is adopted to perform frequency decomposition at different stages of the network encoder. Attention-guided two-dimensional discrete wavelet transform (DWT) is used to enhance high-frequency structural information and suppress noise interference. At the same time, a frequency context block (FCB) is introduced to restore the consistency between frequency and semantics, ensuring frequency-semantic alignment in the decoding stage. Attached Figure Description

[0015] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0016] Figure 1 This is a flowchart of the method in Example 1. Figure 2 This is a schematic diagram of the soft supervision mechanism structure in Example 1.

[0017] Figure 3 This is a schematic diagram of the pseudo-boundary suppression network structure in Example 1.

[0018] Figure 4 This is a schematic diagram of the APB module structure in Example 1.

[0019] Figure 5 This is a schematic diagram of the SFD module structure in Example 1.

[0020] Figure 6 This is a schematic diagram of the FCB module structure in Example 1. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0023] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0024] Example 1 One embodiment of the present invention provides a soft-supervised medical image segmentation method based on pseudo-boundary suppression, comprising: Step S1: Obtain the medical image to be segmented; Step S2: Input the medical image into the pseudo-boundary suppression network of the U-Net structure. First, the encoder extracts multi-scale features from the medical image, and then the decoder maps the multi-scale features back to the pixel space to obtain the predicted probability map. Finally, the lesion segmentation result is generated through thresholding. In the training process of the pseudo-boundary suppression network, a soft supervision strategy is introduced, which adds a boundary loss term based on edge complexity to the loss function and adaptively modulates the weight of the boundary loss term.

[0025] As one embodiment, this invention provides a soft-supervised medical image segmentation method based on pseudo-boundary suppression. Based on a soft-supervised mechanism and frequency-selective representation, a pseudo-boundary suppression network is constructed. This method handles blurred boundaries while avoiding the generation of pseudo-boundaries, significantly improving the accuracy and robustness of medical image segmentation. The specific implementation process is described below from the perspectives of model training and model inference. Figure 1 As shown: Step 1: Obtain and preprocess the medical image dataset, specifically: Step 1.1: Obtain publicly available medical image datasets BUSI, Dataset B, STU, PH2, ISIC2017, and COVID19-1 as training datasets. Simultaneously, perform preliminary processing on the training datasets by reading the data according to the image acquisition order and dividing the original training dataset. and the original test dataset ; Step 1.2: Preprocess the training dataset: Perform standardization processing on the original training dataset obtained in step 1.1. and the original test dataset Preprocessing is performed to obtain the preprocessed training dataset. and the preprocessed test dataset .

[0026] Step 2: Construct a pseudo-boundary suppression network (PBS-Net), introducing a soft supervision mechanism and a frequency-selective representation strategy into the U-Net structure, specifically as follows: Step 2.1: In the loss function, a soft-supervision strategy is introduced, adding a boundary loss term constructed based on edge complexity, and adaptively modulating the weights of the boundary loss term, such as... Figure 2 As shown: Step 2.1.1: Sobel Edge Complexity Estimation To accommodate the intensity of supervision across regions, boundary sharpness is evaluated pixel-by-pixel. For this purpose, the Sobel operator is used to evaluate the ground truth labels. The edge magnitude map is calculated as follows: Configure two Sobel kernels to capture horizontal and vertical gradients:

[0027] truth label and and After convolution, the gradient magnitude is calculated to obtain the edge magnitude map, which can be expressed by the formula:

[0028] The resulting edge amplitude map reflects the sharpness of the boundary. High values ​​indicate sharp edges that require strong supervision, while low values ​​correspond to smooth transitions where strict supervision may induce false boundaries.

[0029] Step 2.1.2: Weighting the adaptive modulation boundary loss term To ensure that the supervision signal remains consistent with the boundary complexity, structural constraints are incorporated into the conventional loss function. This conventional loss function typically consists of pixel-level classification and region-level consistency, which are achieved through binary cross-entropy. and dice loss To enforce:

[0030]

[0031] in, express The prediction result on the b-th sample, i.e., the final predicted probability map output by the decoder, Prevent division by zero.

[0032] These traditional losses treat all foreground-background transitions indiscriminately, enforcing strict alignment using hard labels. To allow for more flexible boundary modeling, a Sobel-based boundary loss is incorporated, which captures structural consistency beyond binary contours, expressed as:

[0033] This loss encourages predictions to be made within the bounds of the truth label. Aligning geometrically, rather than simply matching binary categories, improves boundary sensitivity and reduces the illusion of sharp edges in anatomically smooth areas.

[0034] To transform this structural constraint into adaptive strength, we first normalize the overall edge complexity of the label:

[0035] in, , , , These represent batch size, number of channels, space height, and space width, respectively. As a global boundary complexity metric for the entire sample, all dims represents the sum of the absolute values ​​of gradients calculated for all image dimensions (all pixels and channels of each sample), thus evaluating the edge complexity of the entire image.

[0036] To avoid overly strict or lenient supervision in extreme cases, adaptive supervision weights are introduced to constrain the model's degrees of freedom, defined as:

[0037] here, It is a global adjustment factor. , It is the clipping boundary that ensures training stability. Reflecting the global marginal complexity of each sample, the boundary graph derived from Sobel does not enforce boundary constraints independently, but provides a structural prior that dynamically adjusts the strength of supervision.

[0038] The previously defined loss terms are ultimately integrated into a unified objective with adaptive weights to adapt to different boundary features in different regions. The final loss function is:

[0039] Here, and The contributions of pixel-level classification and region-level consistency were controlled separately, and the boundary loss was determined by... Adaptive modulation.

[0040] Step 2.2: In the structure of the Pseudo-Boundary Suppression Network (PBS-Net), a frequency-selective representation strategy is introduced, including Selective Frequency Downsampling (SFD) and Frequency Context Block (FCB). The Pseudo-Boundary Suppression Network (PBS-Net) adopts a U-Net structure, such as... Figure 3 As shown, it mainly consists of an encoder, a decoder, and skip connections. The encoder includes an APB module and a selective frequency downsampling module, employing a frequency-selective representation strategy to perform frequency decomposition at different stages of the encoder. It uses attention-guided two-dimensional discrete wavelet transform to enhance high-frequency structural information and generate multi-scale feature maps. The decoder introduces a frequency context module (FCB) to restore the consistency between frequency and semantics, ensuring frequency-semantic alignment during the decoding stage, and obtaining a predicted probability map, specifically: I. Encoder In this embodiment, PBS-Net includes a five-level encoder. Each encoder stage consists of two components: , , The original input image, For APB from the input image Extract local attention propagation features from them. for The output multi-scale feature maps can be flexibly instantiated with any effective module by the encoder.

[0041] 1. APB module To better support soft supervision of local boundaries, an attention propagation module (APB) is designed to process the input image. Perform local attention propagation and feature enhancement. For example... Figure 4 As shown, the APB module comprises two stages: in the first stage (attention propagation stage), local descriptors are extracted from the input feature map through convolution and average pooling. And generate a spatial attention map. It is used to propagate boundary information within the spatial neighborhood, thereby enhancing the edge features of blurred regions and outputting an enhanced feature map modulated by attention weighting and gating mechanisms. In the second stage (feature extraction stage), the enhanced feature map is... Multi-layer depthwise separable convolution and channel convolution operations are performed, combined with GELU activation and normalization, to further extract boundary-sensitive features and suppress false boundary noise, outputting a structure-enhanced feature map. As the multi-scale feature input of the encoder, specifically: In the first stage, a smooth local descriptor is first computed:

[0042] in, For average pooling, This represents element-wise multiplication. We apply two stacked 3×3 convolutions to generate an attention map:

[0043] This attention map diffuses local edge responses into the spatial neighborhood, allowing boundary cues to propagate smoothly within ambiguous areas.

[0044] In the After upsampling, a gating mechanism is applied, in which The first channel (activated by Sigmoid) acts as a space gate, ultimately enhancing the output as follows:

[0045] In the second stage, stacked depthwise separable convolutions were applied to refine the structure. Begin, calculate:

[0046]

[0047] in, and This indicates that GELU and BatchNorm are used for depthwise convolution and channel convolution, with a kernel size of [missing information]. and depth With each stage And the changes:

[0048] 2. Selective Frequency Downsampling (SFD) SFD operates during encoder downsampling, propagating the extracted local attention features. The system performs structure-aware frequency domain decomposition, generates low-frequency and high-frequency sub-bands using two-dimensional discrete wavelet transform (DWT), then enhances effective high-frequency features through spatial attention mechanisms and performs noise suppression to generate an enhanced frequency feature map, such as... Figure 5 As shown, specifically: (1) Using two-dimensional discrete wavelet transform (DWT) to transform features Decomposed into four frequency sub-bands:

[0049] in, Capture low-frequency semantic content, and These represent the high-frequency edge responses along the vertical, horizontal, and diagonal directions, respectively. 2D indicates two-dimensionality, and 2D DWT is the two-dimensional discrete wavelet transform.

[0050] (2) These high-frequency subbands are connected to form a unified spectrum representation. .because Incorporating meaningful edge cues and noisy activations, through 1×1 convolution and Softmax. Generate spatial attention diagram And apply it to suppress irrelevant responses:

[0051]

[0052] (3) To Averaging is performed to derive the semantic boundary representation:

[0053] (4) and They are connected and projected back to the original channel size through a 1×1 convolution:

[0054] in, It is a low-frequency sub-band. These are high-frequency features enhanced and averaged through spatial attention mechanisms. The low-frequency and enhanced high-frequency subbands are merged using 1×1 convolution.

[0055] II. Decoder PBS-Net consists of four decoders, each stage including a Frequency Context Module (FCB) and an improved FusionConv that propagates the local attention features from the APB output. Obtained through FCB processing in the jump connection ,right Transposed convolution is performed, with upsampling layer by layer. Then, an improved FusionConv is used to process the concatenated tensor to obtain... ( Finally, the output of the last stage decoder will be... As a prediction probability map.

[0056] 1. Frequency Context Block (FCB) Despite spectral suppression in the encoder, frequency semantic misalignment may still occur in the decoder. Therefore, FCB is introduced and applied to each hop connection before decoding, such as... Figure 6 As shown, FCB first uses orthogonal fast Fourier transform (FFT) to... Projecting onto the frequency domain, as shown below:

[0057] here, Represents frequency domain signal, and These represent the indices for horizontal and vertical frequencies, respectively. Represents the original spatial domain signal. , Indicates the pixel position.

[0058] Then, The real and imaginary parts are separated and represented as follows: and And connect them to form a spectral representation:

[0059] To recalibrate the frequency response, the spectral representation... Attention weights are generated by applying a depthwise convolution with residual connections, adaptive average pooling (AAP), and a 1×1 convolution with softmax:

[0060] in, Output a Each channel attention vector is used to reweight the spectral components. Represents a learnable spectral kernel. It is a weighted output of the frequency response. AAP is adaptive average pooling, which dynamically adjusts the size of the pooling window according to the size of the input feature map, mapping the size of the output feature map to a specific target size, thus helping to extract global frequency information.

[0061] from Reconstructing the spectrum and apply inverse FFT ( Recover spatial features. Then apply frequency-guided channel attention:

[0062]

[0063] in, It is a real value output. Represents a two-dimensional Fourier transform. This represents the amplitude spectrum after frequency centering. It is a two-layer channel attention module with ReLU and Sigmoid activation.

[0064] 2. Improved FusionConv The decoding process is defined as follows:

[0065] here, Represented as an improved FusionConv, it uses grouped convolutions with different kernel sizes (3×3, 5×5, 7×7) for enhancement to enrich multi-scale context and better preserve boundary structure. The final output is a predicted probability map. .

[0066] The main improvement here is that the original FusionConv only used 3×3 grouped convolutions, while this embodiment uses grouped convolutions with different kernel sizes (3×3, 5×5, 7×7) to enrich the multi-scale context and better preserve the boundary structure.

[0067] Step 3: Train the Pseudo-Boundary Suppression Network (PBS-Net) model, including the following steps: Step 3.1: Prepare the input training dataset Input medical images , H is the batch size of the training dataset, representing the number of samples in the same batch of input. H and W represent the height and width of the image, respectively. C is the number of channels, representing the feature dimension of each pixel. Step 3.2: Initialize encoder and decoder parameters, including the weights of convolutional and fully connected layers in the Selective Frequency Downsampling (SFD), Frequency Context Block (FCB), and APB modules. A random initialization method is used, with weight values ​​following a Gaussian distribution with a mean of 0 and a standard deviation of 0.01. ; Verify the initialization parameters through forward propagation, and input Multi-scale feature maps are generated using an encoder. ( This ensures that the input dimension matches the decoder's input dimension.

[0068] Step 3.3 Initialize the stochastic gradient descent (SGD) optimizer and set the initial learning rate. Learning rate decay rate Used to adaptively adjust the gradient update step size; input Feature maps are generated using the encoder. The decoder outputs a predicted probability map. Through the loss function defined earlier Calculate the joint loss and backpropagate to update the model parameters.

[0069] Step 4: Use the trained model to perform lesion segmentation and prediction in medical images, including the following steps: Step 4.1: Obtain the medical image to be segmented ; Step 4.2: Use the encoder in the model to extract multi-scale features from the input image.

[0070] Specifically, the input tensor Input the model trained in step 3, and first extract features using the APB module. The encoder then downsamples layer by layer through the Selective Frequency Downsampling (SFD) module to generate multi-scale feature maps. ( ),in , Number of channels The numbers are 16, 32, 128, 160, and 256 respectively. In the APB module, the input features To perform local attention propagation feature extraction, firstly, an initial feature map is generated through convolution operations, and then average pooling is applied to generate local descriptors. Next, an attention map is generated through multiple convolution operations. Then, element-wise multiplication, gating mechanisms, and depthwise separable convolutions generate the final enhanced feature map. .

[0071] In the SFD module, for The system performs structure-aware frequency domain decomposition, generates low-frequency and high-frequency sub-bands using two-dimensional discrete wavelet transform (DWT), then enhances effective high-frequency features through spatial attention mechanism, and performs noise suppression to generate an enhanced frequency feature map. That is, multi-scale feature maps.

[0072] Step 4.3: Process the multi-scale feature map using the Frequency Context Module (FCB) and the improved FusionConv in the decoder to generate a predicted probability map.

[0073] Specifically, the multi-scale feature map output by APB Obtained through FCB processing in the jump connection The data is passed to the decoder, which upsamples layer by layer through transposed convolution, and then processes the concatenated tensor using an improved FusionConv algorithm to obtain... ( ).

[0074] In the FCB module, for To perform frequency domain transformation, the input feature map is first transformed using a Fast Fourier Transform (FFT). The data is then transformed to the frequency domain, followed by the integration of frequency position awareness, frequency dynamic convolution, and frequency-guided channel attention to adjust the frequency response, thereby restoring the consistency between frequency and semantics, and finally outputting the final feature map. .

[0075] Step 4.4: Map the predicted probability map to the final lesion segmentation result and generate a segmentation mask. .

[0076] Specifically, the predicted probability map generated in step 4.3 Post-processing is performed, and a thresholding operation is used to generate a binary segmentation mask. The process is represented as follows:

[0077] in, ,like ,but (lesion area), otherwise (Background area); Output the final segmentation mask This represents the lesion area in each input image, used for subsequent evaluation or clinical applications.

[0078] Example 2 One embodiment of the present invention provides a soft-supervised medical image segmentation system based on pseudo-boundary suppression, comprising: The image acquisition module is configured to acquire medical images to be segmented; The image segmentation module is configured to: input medical images into a pseudo-boundary suppression network with a U-Net structure; firstly, the encoder extracts multi-scale features from the medical images; then the decoder maps the multi-scale features back to the pixel space to obtain a predicted probability map; and finally, the lesion segmentation result is generated through a thresholding operation. In the training process of the pseudo-boundary suppression network, a soft supervision strategy is introduced, which adds a boundary loss term based on edge complexity to the loss function and adaptively modulates the weight of the boundary loss term.

[0079] Example 3 One embodiment of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned soft-supervised medical image segmentation method based on pseudo-boundary suppression.

[0080] Example 4 In one embodiment of the present invention, a non-transitory computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the aforementioned soft-supervised medical image segmentation method based on pseudo-boundary suppression.

[0081] Example 5 One embodiment of the present invention provides an electronic device, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the soft-supervised medical image segmentation method based on pseudo-boundary suppression.

[0082] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0083] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0084] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A soft-supervised medical image segmentation method based on pseudo-boundary suppression, characterized in that, include: Acquire the medical image to be segmented; The medical image is input into the pseudo-boundary suppression network with a U-Net structure. First, the encoder extracts multi-scale features from the medical image. Then, the decoder maps the multi-scale features back to the pixel space to obtain the predicted probability map. Finally, the lesion segmentation result is generated through thresholding. In the training process of the pseudo-boundary suppression network, a soft supervision strategy is introduced, which adds a boundary loss term based on edge complexity to the loss function and adaptively modulates the weight of the boundary loss term.

2. The soft-supervised medical image segmentation method based on pseudo-boundary suppression as described in claim 1, characterized in that, The encoder employs a frequency-selective representation strategy, performing frequency decomposition at different stages of the encoder and using attention-guided two-dimensional discrete wavelet transform to enhance high-frequency structural information, thereby generating multi-scale feature maps.

3. The soft-supervised medical image segmentation method based on pseudo-boundary suppression as described in claim 1, characterized in that, The decoder introduces a frequency context block to restore the consistency between frequency and semantics, ensuring frequency-semantic alignment during the decoding stage and obtaining a predicted probability map.

4. The soft-supervised medical image segmentation method based on pseudo-boundary suppression as described in claim 1, characterized in that, The boundary loss term constructed based on edge complexity is expressed by the formula: in, An edge magnitude map used to quantify edge complexity. , These are the prediction probability map and the truth label map, respectively. The total number of samples.

5. The soft-supervised medical image segmentation method based on pseudo-boundary suppression as described in claim 4, characterized in that, The edge magnitude map, calculated using the Sobel operator, is expressed by the following formula: in, For truth labeling, , These are the Sobel kernels for capturing horizontal and vertical gradients, respectively.

6. The soft-supervised medical image segmentation method based on pseudo-boundary suppression as described in claim 1, characterized in that, The weights of the adaptive modulation boundary loss term are expressed by the following formula: in, The weights of the boundary loss terms in the loss function. It is a global adjustment factor. , It is the clipping boundary that ensures training stability. It is a global boundary complexity metric based on all samples in the training dataset. , , , These represent batch size, number of channels, space height, and space width, respectively.

7. A soft-supervised medical image segmentation system based on pseudo-boundary suppression, characterized in that, include: The image acquisition module is configured to acquire medical images to be segmented; The image segmentation module is configured to: input medical images into a pseudo-boundary suppression network with a U-Net structure; firstly, the encoder extracts multi-scale features from the medical images; then the decoder maps the multi-scale features back to the pixel space to obtain a predicted probability map; and finally, the lesion segmentation result is generated through a thresholding operation. In the training process of the pseudo-boundary suppression network, a soft supervision strategy is introduced, which adds a boundary loss term based on edge complexity to the loss function and adaptively modulates the weight of the boundary loss term.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the soft-supervised medical image segmentation method based on pseudo-boundary suppression as described in any one of claims 1-6.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement a soft-supervised medical image segmentation method based on pseudo-boundary suppression as described in any one of claims 1-6.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to perform a soft-supervised medical image segmentation method based on pseudo-boundary suppression as described in any one of claims 1-6.

Images