A medical image segmentation method and device based on spatial perception and frequency domain information

Abstract

The present invention provides a medical image segmentation method and device based on spatial perception and frequency domain information. By combining frequency domain information guidance and multi-head state space perception technology, the accuracy and robustness of medical image segmentation are effectively improved. The medical image is transformed in the frequency domain, and the high-frequency and low-frequency components of the image are separated by multi-scale decomposition technology, and low-frequency features are extracted to obtain global information. By introducing a learnable noise filtering mechanism, the noise and irrelevant background information in the frequency domain are suppressed, so that the model can focus on the lesion area more accurately. A multi-head perception visual state space module is designed at the bottleneck layer, and the lesion features of different scales are captured through a multi-scale adaptive feature fusion mechanism, thereby enhancing the segmentation capability of small-sized lesions and complex structures. A contextual focus attention mechanism is introduced in the jump connection to further strengthen the fusion of global information and local details, ensuring the accuracy of the segmentation results; finally, the high-resolution segmented image is restored through the decoder.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision and pattern recognition, and in particular to a medical image segmentation method and device based on spatial perception and frequency domain information. Background Art

[0002] In the field of medical image processing, image segmentation technology is a key component in achieving precision medicine. By accurately demarcating different tissues, organs, or lesion areas in medical images, it provides a crucial basis for disease diagnosis, treatment planning, and efficacy evaluation. In recent years, with the rise of deep learning technology, convolutional neural network (CNN)-based segmentation methods have made significant progress. U-shaped network architectures, such as UNet, have become a mainstream choice for medical image segmentation tasks due to their unique encoder-decoder structure and skip connection mechanism.

[0003] However, existing technologies still face many challenges when processing complex medical images. On the one hand, traditional CNN encoders perform well in extracting local features, but their fixed convolution kernels are insufficient in capturing global semantic information and long-range spatial dependencies. Although some improved methods (such as dilated spatial pyramid pooling and pyramid pooling modules) alleviate the problem of information loss through multi-scale feature fusion, they often ignore the complementarity between features at different scales, resulting in insufficient learning of global and detailed features. On the other hand, low-frequency information in medical images (such as the macroscopic structure of organs) is highly robust to noise, but existing methods fail to fully utilize frequency domain features, and low-frequency information is easily interfered with by irrelevant background, which affects segmentation accuracy.

[0004] Furthermore, the bottleneck layer of the U-shaped network, connecting the encoder and decoder, must strike a balance between global context modeling and computational efficiency. While the Transformer-based bottleneck layer enhances global perception through self-attention, its high computational complexity limits its application in high-resolution medical images. While some lightweight models (such as Mamba, which combines state-space models) reduce computational overhead, they lack the ability to extract features for multi-scale lesions (especially tiny lesions), making them difficult to adapt to the complex structure of medical images.

[0005] In general, existing medical image segmentation techniques still have shortcomings in global and local feature fusion, multi-scale feature extraction, and computational efficiency. This is particularly true when processing complex medical images, which can easily lead to problems such as loss of edge details and omission of lesion areas. Therefore, developing a medical image segmentation method that can effectively fuse frequency domain information, enhance multi-scale perception capabilities, and improve computational efficiency is crucial for improving the accuracy and reliability of medical image segmentation.

[0006] In view of this, this application is filed. Summary of the Invention

[0007] The present invention provides a medical image segmentation method and device based on spatial perception and frequency domain information, which can at least partially improve the above problems.

[0008] To achieve the above object, the present invention adopts the following technical solutions:

[0009] A medical image segmentation method based on spatial perception and frequency domain information, comprising:

[0010] Acquiring a medical image to be processed, performing data preprocessing on the medical image, and performing quality assessment on the preprocessed medical image to calculate the signal-to-noise ratio and peak signal-to-noise ratio of the image, assess the contrast and clarity of the image, detect artifacts and noise levels in the image, and if the signal-to-noise ratio is lower than a preset threshold, performing adaptive histogram equalization based on the image contrast, selectively applying gamma correction to locally enhance the detected artifact area;

[0011] Call the pre-trained FDMUNet network model and use the context feature extraction module focusing on low-frequency information to perform feature extraction on the pre-processed medical image to obtain the output feature map;

[0012] The output feature map is extracted and processed using the multi-head perception visual state space module in the FDMUNet network model to obtain global information features of the output feature map at different scales;

[0013] Use the focused context attention module in the FDMUNet network model to perform boundary enhancement on the feature map to obtain enhanced features;

[0014] The decoder is used to restore and preprocess the global information features, and the processed global information features are fused with the enhanced features to obtain the segmentation result. This operation is repeated until the preset number of times is reached to generate a high-resolution feature map.

[0015] The present invention also provides a medical image segmentation device based on spatial perception and frequency domain information, which includes:

[0016] a preprocessing unit, configured to acquire a medical image to be processed, perform data preprocessing on the medical image, and perform quality assessment on the preprocessed medical image, calculate the signal-to-noise ratio and peak signal-to-noise ratio of the image, assess the contrast and clarity of the image, detect artifacts and noise levels in the image, and if the signal-to-noise ratio is lower than a preset threshold, perform adaptive histogram equalization based on the image contrast, selectively apply gamma correction to perform local enhancement processing on the detected artifact area;

[0017] The first extraction unit is used to call the pre-trained FDMUNet network model and use the context feature extraction module focusing on low-frequency information to perform feature extraction on the pre-processed medical image to obtain an output feature map;

[0018] A second extraction unit is used to extract the output feature map using a multi-head perception visual state space module in the FDMUNet network model to obtain global information features of the output feature map at different scales;

[0019] The enhancement unit is used to perform boundary enhancement processing on the feature map using the focused context attention module in the FDMUNet network model to obtain enhanced features;

[0020] The fusion unit is used to use the decoder to restore the global information features and fuse the processed global information features with the enhanced features to obtain the segmentation result. This operation is repeated until a preset number of times is reached to generate a high-resolution feature map.

[0021] In summary, this medical image segmentation method based on spatial and frequency-domain information significantly improves segmentation accuracy and robustness by integrating frequency-domain information with multi-head state-space perception. The core of this method lies in decomposing the image into different frequency components through frequency-domain transformation, focusing on extracting low-frequency features that contain global information. A learnable noise filtering mechanism effectively suppresses background noise and enhances the ability to identify lesions. In terms of model architecture, a multi-head perceptual visual state-space module is introduced, which extracts image features at multiple scales, effectively capturing multi-scale information from subtle lesions to overall anatomical structures, addressing the shortcomings of traditional methods in processing complex medical images. Furthermore, this method incorporates a context-focused attention mechanism within skip connections, further optimizing the fusion of global and local information and improving the model's adaptability to complex scenes. Finally, a decoder is used to restore high-resolution segmentation results, ensuring segmentation accuracy. This method is not only theoretically innovative but also demonstrates significant performance improvements in practical applications, providing a new technical solution for the field of medical image segmentation with important clinical application value.

[0022] Compared with the existing technology, the medical image segmentation method based on spatial perception and frequency domain information has the following beneficial effects: (1) Through the multi-scale decomposition capability of frequency domain transformation and a learnable frequency domain filtering module for low-frequency components, the features of different frequencies in medical images are efficiently captured and the noise in low-frequency information is effectively suppressed. (2) Through the multi-head perception visual state space module, the model's understanding ability is enhanced by extracting information from different scales, which can effectively capture fine-grained (such as small lesions) and coarse-grained (such as overall anatomical structure) features, ensuring that the network has sufficient feature expression capabilities at different resolutions. (3) It effectively demonstrates the important influence of detail features on the segmentation performance of the model, and detail features also have an important influence on the performance of semantic segmentation tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 1 is a flow chart of a medical image segmentation method based on spatial perception and frequency domain information provided by the first embodiment of the present invention;

[0024] Figure 2 This is a diagram showing the overall model structure of the medical image segmentation method based on spatial perception and frequency domain information provided by an embodiment of the present invention;

[0025] Figure 3 is a structural diagram of a FLICEB module provided in an embodiment of the present invention;

[0026] Figure 4 is a structural diagram of the MPVSS module provided by an embodiment of the present invention;

[0027] Figure 5 is a structural diagram of a CFA module provided by an embodiment of the present invention;

[0028] Figure 6 This is a module diagram of a medical image segmentation device based on spatial perception and frequency domain information provided by the second embodiment of the present invention. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0030] refer to Figure 1 、 Figure 2 As shown, the first embodiment of the present invention discloses a medical image segmentation method based on spatial perception and frequency domain information, which can be performed by a medical image segmentation device based on spatial perception and frequency domain information (hereinafter referred to as a segmentation device), and in particular, by one or more processors in the segmentation device to implement the following method:

[0031] S1, obtaining a medical image to be processed, performing data preprocessing on the medical image, and performing quality assessment on the preprocessed medical image, calculating the signal-to-noise ratio and peak signal-to-noise ratio of the image, evaluating the contrast and clarity of the image, detecting artifacts and noise levels in the image, and if the signal-to-noise ratio is lower than a preset threshold, performing adaptive histogram equalization based on the image contrast, selectively applying gamma correction to perform local enhancement processing on the detected artifact area;

[0032] Specifically, step S1 includes: calculating the signal-to-noise ratio and peak signal-to-noise ratio of the image to quantitatively evaluate the image quality, wherein the signal-to-noise ratio measures the image quality by calculating the ratio of the average power of the image signal to the average power of the noise, and the peak signal-to-noise ratio evaluates the image fidelity by the logarithmic relationship between the maximum possible pixel value of the image and the mean square error;

[0033] The root mean square contrast method is used to evaluate the global contrast of the image, and the second-order derivative of the image is calculated using the Laplacian operator to evaluate the clarity;

[0034] To address the artifacts and noise in images, a noise detection mechanism based on high-frequency component analysis is designed, combining directional gradient histogram features and Radon transform to identify existing striping artifacts and motion artifacts.

[0035] When the signal-to-noise ratio is detected to be lower than a preset threshold, contrast-limited adaptive histogram equalization is used for enhancement, where the cropping limit parameters are dynamically adjusted according to the signal-to-noise ratio value;

[0036] For areas with insufficient contrast, adaptive gamma correction based on image statistical characteristics is introduced to automatically adjust the gamma value by analyzing the average brightness of the image;

[0037] For the detected artifact area, local enhancement processing based on guided filtering is adopted, and Poisson fusion is used to ensure a natural transition between the processed area and the surrounding images.

[0038] In this embodiment, the preprocessed image undergoes image quality assessment and an adaptive enhancement mechanism. This mechanism first quantitatively assesses image quality by calculating the image's signal-to-noise ratio (SNR) and peak signal-to-noise ratio (PSNR). The SNR measures image quality by calculating the ratio of the image's average signal power to the average noise power, while the PSNR assesses image fidelity by comparing the logarithmic relationship between the image's maximum possible pixel value and the mean square error. Furthermore, the RMS (root mean square) contrast method is used to assess global image contrast, and the Laplacian operator is used to calculate the image's second-order derivative to assess clarity. To address potential artifacts and noise in the image, this method designs a noise detection mechanism based on high-frequency component analysis, combining the histogram of oriented gradients (HOG) feature with the Radon transform to identify possible banding and motion artifacts. Based on these evaluation metrics, this method further implements an adaptive image enhancement strategy: when the SNR is detected to be lower than a preset threshold (usually set to 15dB), contrast-limited adaptive histogram equalization (CLAHE) is used for enhancement, where the cropping limit parameters are dynamically adjusted according to the SNR value; for areas with insufficient contrast, adaptive gamma correction based on image statistics is introduced to automatically adjust the gamma value by analyzing the average brightness of the image; for detected artifact areas, local enhancement processing based on guided filtering is adopted, and Poisson fusion is used to ensure a natural transition between the processed area and the surrounding image.

[0039] S2, calling the pre-trained FDMUNet network model, using the context feature extraction module focusing on low-frequency information to perform feature extraction on the pre-processed medical image to obtain the output feature map;

[0040] Specifically, step S2 includes: for the input pre-processed medical feature map , H, W, C represent the height, width and number of channels of the image respectively. The context feature extraction module focusing on low-frequency information of the FDMUNet network model is used to perform feature decomposition processing on the medical feature map X at different preset frequency scales to capture different frequency features in the medical image. The formula is: , , T , ,in, is the frequency domain transform operation, is the low-frequency component of the medical image, are all high-frequency components of medical images. is the output feature map obtained after the noise filtering mechanism, is the sigmoid function, is the maximum pooling operation, is the average pooling operation, Indicates connection by channel number, is a 1x1 convolution operation, is the batch normalization operation, is the ReLU activation function, is the characteristic information after high-frequency information and low-frequency information are connected through channels, is the extracted feature map;

[0041] After three consecutive feature extractions, we get The output feature map of the three-times downsampling feature extraction is expressed as ,in, , , ;

[0042] , , .

[0043] Preferably, the data preprocessing includes: size scaling, random cropping, horizontal mirroring, and random cropping.

[0044] In this embodiment, first, medical images to be processed are acquired. These images can be original images from hospital CT scans, MRI imaging, or other medical imaging equipment. To improve the model's generalization ability and segmentation accuracy, these medical images are subjected to data preprocessing. The specific steps of data preprocessing include resizing the images to meet the model's input requirements; performing random cropping to increase data diversity; performing horizontal mirroring to further enrich the dataset; and performing random cropping again to ensure that the model can learn different parts of the image. These preprocessing steps not only enhance the model's robustness but also improve its adaptability to different image sizes and orientations.

[0045] Because existing segmentation schemes often use CNN encoders that ignore the complementary information between features at different scales, hindering the learning of global and detailed features, this approach, taking into account the complementarity between frequency-domain and spatial features, uses a pretrained FDMUNet network model to extract features from preprocessed medical images after preprocessing. The core of the FDMUNet network model lies in its low-frequency information-focused context feature extraction block (FLICEB). This block extracts key information from the input data, thereby improving model performance. Leveraging the multi-scale analysis capabilities of frequency-domain transforms, it performs feature decomposition on medical images at pre-defined frequency scales. Specifically, this block uses frequency-domain transforms to decompose the input feature map into low-frequency and high-frequency components. The low-frequency component primarily contains global image information, while the high-frequency component contains detailed information. This decomposition enables the model to more comprehensively capture various features in medical images. This low-frequency information-focused context feature extraction block not only effectively extracts key features from medical images but also enhances the model's focus on lesion areas through noise filtering.

[0046] During the feature decomposition process, special attention is paid to the processing of low-frequency components. Because frequency domain transformation may introduce background noise and low-frequency components in areas unrelated to the target, a learnable noise filtering mechanism is introduced to address this problem. This mechanism can automatically identify and suppress noise or irrelevant background information in the low-frequency part, allowing the model to focus more on areas with higher semantic significance, such as lesions. This process not only effectively suppresses noise and irrelevant background information in low-frequency information, improving the model's ability to segment key areas, but also enhances the accuracy and reliability of the segmentation results, allowing the model to focus on areas with higher semantic significance (such as lesions), thereby enhancing the model's segmentation capabilities in key areas.

[0047] The maximum pooling and average pooling operations involved in the noise filtering mechanism are used to extract the spatial features of low-frequency information. The noise filtering mechanism weights each spatial location of low-frequency information, emphasizing important areas and suppressing irrelevant areas.

[0048] S3, extracting and processing the output feature map using a multi-head perception visual state space module in the FDMUNet network model to obtain global information features of the output feature map at different scales;

[0049] Specifically, step S3 includes: using the multi-head perception visual state space module in the bottleneck layer of the FDMUNet network model to Perform extraction processing to extract the global information features of the output feature maps at different scales. The formula of the multi-head perception visual state space module is: ,in, is a 3x3 depth-wise separable convolution operation, For VSS module, is layer normalization, It is a multi-scale adaptive feature fusion module. are the extracted fine-grained features and coarse-grained features;

[0050] The multi-scale adaptive feature fusion module divides the output feature map into three branches of different scales for processing, and performs feature fusion on the outputs of the three branches. The formula of the multi-scale adaptive feature fusion module is: , , , , among which ConCat To splice by channel for feature fusion, is the output of the first branch, is the output of the second branch, is the output of the third branch, is 1x1 average adaptive pooling, is 3x3 average adaptive pooling, is 5x5 average adaptive pooling, yes The result of feature fusion;

[0051] After four consecutive multi-head perception visual state space modules, the output feature map Extraction process to obtain The feature map of the fourth bottleneck layer is represented as , , , .in, , , .

[0052] In this example, to effectively process the multi-scale features in skin disease images, a multi-head perception visual state space (MPVSS) module is proposed in the bottleneck layer. This module uses a multi-scale feature fusion mechanism to enhance the model's understanding of small lesions and complex structures. This module extracts global image information from different scales, effectively capturing the various scale features of skin diseases through a multi-scale adaptive feature fusion module. Finally, the visual state space module effectively models the global context of the image, thereby enhancing the model's perception of lesion areas of different scales during the segmentation process.

[0053] Specifically, the multi-head perceptual visual state space module is integrated into the bottleneck layer of the FDMUNet network model. The core of this module lies in its ability to extract global information features from the output feature maps at different scales, which is crucial for understanding the complex structures in medical images. For example, in skin disease images, lesions can vary in size, from tiny lesions to larger lesion areas. The multi-head perceptual visual state space module enables the model to simultaneously capture features at these different scales, thereby achieving a more comprehensive understanding of the image content. Simply put, the model enhances its understanding capabilities by extracting information from different scales. This process ensures feature stability through layer normalization (LN), models the visual state space using the VSS module, extracts local features through 3x3 depthwise separable convolutions, and finally fuses features from different scales through the multi-scale adaptive feature fusion module (MAF) to obtain a feature representation that incorporates global information. This module effectively captures both fine-grained (such as small lesions) and coarse-grained (such as overall anatomical structures) features, ensuring sufficient feature representation at different resolutions. This multi-scale adaptive feature fusion is particularly well-suited for complex scenes in medical imaging.

[0054] In addition, the multi-scale adaptive feature fusion module is an important component of the multi-head perception visual state space module. This module divides the input feature map into three branches of different scales for processing, and each branch is responsible for extracting feature information of a specific scale. The first branch expands the number of channels through a 1x1 convolution layer, and then uses a batch normalization layer and a ReLU activation layer for normalization. Finally, 1x1 adaptive average pooling is used to capture global information while reducing the amount of computation. The second branch processes the input features through 3x3 adaptive pooling, reduces the spatial dimension and extracts local context information, and then passes through batch normalization and ReLU activation function, and finally obtains the output features through 1x1 convolution. The third branch processes the feature map through a 5x5 adaptive pooling window to extract a larger range of context information, and then passes through batch normalization and ReLU activation function, and finally obtains the output features through 1x1 convolution. The outputs of the three branches are recorded as 、 and , they obtain the final output feature map through feature fusion, which can effectively extract features from different scales and enhance the network's multi-scale perception and information integration capabilities when processing complex medical images (such as skin disease images).

[0055] Through this multi-scale adaptive feature fusion mechanism, the model can effectively capture feature information from different scales, enhancing the network's multi-scale perception and information integration capabilities when processing complex medical images. For example, when processing skin disease images containing tiny lesions, the model can simultaneously focus on the detailed features of the lesion area and the macroscopic features of the overall anatomical structure, thereby performing more accurate segmentation. This multi-scale feature fusion mechanism not only improves the model's segmentation accuracy for lesions of different scales, but also ensures that the network has sufficient feature expression capabilities at different resolutions, which is very important for medical image segmentation tasks. In addition, the design of the multi-head perception visual state space module also takes computational efficiency into consideration. By combining operations such as depthwise separable convolution and adaptive pooling, the module can effectively extract global information features while maintaining efficient computation.

[0056] S4, use the focused context attention module in the FDMUNet network model to perform boundary enhancement on the feature map to obtain enhanced features;

[0057] Specifically, step S4 includes: in the skip connection process, according to the focus context attention module in the FDMUNet network model, the feature map 、 and Boundary enhancement is performed separately, and local area features are obtained by combining average pooling and 1x1 convolution;

[0058] Use two depth-wise separable large kernel convolutions to expand the features, and use 1x1 convolution and sigmoid activation function to perform convolution activation on the processed features to obtain enhanced features. The formula is: ,in, To enhance the features, is an 11x1 depth-wise separable convolution operation, It is a 1x11 depth-wise separable convolution operation;

[0059] Similarly, the features of the three layers after being enhanced by the focused context attention module are expressed as, , , ,in, , .

[0060] In this embodiment, in traditional UNets, skip connections simply concatenate the corresponding feature maps of the encoder to the feature maps of the decoder, lacking the ability to model contextual dependencies between distant pixels. In medical images, the distribution of diseases is often localized, but the contextual information of the lesion area (such as the entire organ structure) is crucial for segmentation. Traditional skip connections are often insufficient to capture this long-range contextual information. To overcome these shortcomings, this method introduces a contextual focused attention module (CFA) in the skip layer connection. While transmitting low-level features, the CFA module can also strengthen the network's attention to global information through the guidance of contextual information, thereby improving the model's performance in segmentation tasks. This module is a simple and effective attention mechanism scheme used to mine multi-granularity information around edges at each stage in the skip connection.

[0061] The focused contextual attention module is integrated into the skip connections of the FDMUNet network model. Skip connections are a key component of the U-shaped network architecture, directly passing low-level features from the encoder to the decoder to preserve image details. This method introduces the focused contextual attention module, which, guided by contextual information, strengthens the network's focus on global information, improving the model's performance in segmentation tasks.

[0062] Specifically, in this embodiment, during the boundary enhancement process, local region features are first acquired using a combination of average pooling and 1x1 convolution. Average pooling downsamples the feature map to extract local information, while 1x1 convolution adjusts the number of channels in the feature map to make it more suitable for subsequent processing. This combination not only effectively extracts local region features, but also reduces computational effort and improves model efficiency.

[0063] Subsequently, two depthwise separable large kernel convolutions are used to expand the features. Depthwise separable convolution is an efficient convolution operation that decomposes the standard convolution into two steps: depthwise convolution and pointwise convolution, significantly reducing the amount of computation. In this method, 11x1 depthwise separable convolution and 1x11 depthwise separable convolution are used. These two operations expand the feature map horizontally and vertically, respectively, thereby expanding the model's receptive field and enabling the model to better capture contextual information in the feature map.

[0064] Finally, the processed features are activated using a 1x1 convolution and a sigmoid activation function to generate enhanced features. The 1x1 convolution adjusts the number of channels in the feature map to match the subsequent network structure, while the sigmoid activation function maps each element in the feature map to a value between 0 and 1, resulting in a probability map representing the probability that each pixel belongs to the lesion area. This allows the model to more accurately identify the boundaries of the lesion area, thereby improving segmentation accuracy.

[0065] By focusing on the boundary enhancement processing of the contextual attention module, this method can effectively enhance the boundary information in the feature map, enabling the model to more accurately identify the edges of the lesion area. This improvement not only improves the accuracy of segmentation, but also enhances the model's adaptability to complex medical images. For example, when processing skin disease images containing subtle lesions, the model can more clearly identify the boundaries of the lesion area, thereby performing more accurate segmentation. In addition, by adopting depthwise separable convolution and sigmoid activation functions, this module can effectively extract and enhance boundary features while maintaining efficient computation, further improving the model's performance.

[0066] S5, use the decoder to restore the global information features and pre-process them, and fuse the processed global information features with the enhanced features to obtain the segmentation result. Repeat this operation until the preset number of times is reached to generate a high-resolution feature map.

[0067] Specifically, step S5 includes: using the decoder to decode the global information output by the multi-head perception visual state space module Enhanced features obtained with the focused context attention module Fusion is performed by channel and bilinear interpolation upsampling is performed to obtain , to restore the controlled size of the image;

[0068] The convolution operation is used to perform feature dimension reduction and batch normalization on the processed global information features, and the feature information is retained through the ReLU activation function;

[0069] The enhanced features obtained by the jump connection and the retained global information features are added and fused to obtain the features restored by the decoder, which is expressed as follows: ,in, is the upsampling operation, To obtain the features restored by the decoder;

[0070] After three consecutive feature fusions and upsampling, the final high-resolution feature map G is obtained. , where the features recovered by the decoder are , ,in, , .

[0071] In this embodiment, the decoder first performs bilinear interpolation upsampling on the global information features. Bilinear interpolation upsampling is a commonly used image magnification technique that generates new pixel values ​​by calculating the weighted average of adjacent pixels, thereby restoring the spatial resolution of the image. In this embodiment, this operation is used to restore the low-resolution global information features processed by the encoder to a resolution close to that of the original image for fusion with the enhanced features. This process not only effectively restores the spatial structure of the image but also provides a foundation for subsequent feature fusion. Subsequently, convolution operations are used to perform feature dimensionality reduction and batch normalization on the processed global information features. Feature dimensionality reduction reduces the number of channels in the feature map through convolution, thereby reducing computational complexity and extracting a more compact feature representation. Batch normalization is a commonly used regularization technique that standardizes the activation values ​​of each feature map, making network training more stable and accelerating convergence. Through convolution and batch normalization, the representation of the global information features can be further optimized, making them more suitable for fusion with the enhanced features.

[0072] To preserve feature information, the processed global features undergo a nonlinear transformation using the ReLU activation function. The ReLU activation function sets negative values ​​in the feature map to zero while retaining positive values, thereby enhancing feature sparsity and improving the model's ability to distinguish features. This operation not only effectively preserves important information in the global features but also provides a clearer feature representation for subsequent feature fusion. Finally, the enhanced features are additively fused with the retained global features to produce the final segmentation result. Additive fusion is a simple yet effective feature fusion method. It combines the boundary information in the enhanced features with the background information in the global features by pixel-by-pixel summation of the two feature maps. This process not only fully utilizes the boundary details in the enhanced features but also incorporates contextual information in the global features, resulting in more accurate segmentation results. After four consecutive upsampling operations, the model ultimately generates a high-resolution feature map. This process ensures that the model can integrate global background information and local details, resulting in more accurate segmentation results. Through continuous fusion and learning, the decoder effectively recovers the spatial structure of the image and performs complex medical image segmentation tasks.

[0073] Through this step, the method can effectively fuse global information features with enhanced features to ensure the accuracy and completeness of the segmentation results. The bilinear interpolation upsampling process can restore the spatial resolution of the image, the convolution operation and batch normalization process can optimize the feature expression, the ReLU activation function can retain important feature information, and the additive fusion process can make full use of the boundary details in the enhanced features and the contextual information in the global information features. The combination of these steps not only improves the accuracy of segmentation, but also enhances the model's adaptability to complex medical images. For example, when processing skin disease images containing tiny lesions, the model can more clearly identify the boundaries of the lesion area and accurately segment it in combination with global background information. This fusion mechanism of global and local information significantly improves the model's ability to identify lesion areas, thereby providing a new technical solution for the field of medical image segmentation and has important clinical application value.

[0074] Preferably, before calling the pre-trained FDMUNet network model to perform feature extraction processing on the pre-processed medical image, the method further includes:

[0075] Acquiring training data, performing data preprocessing on the training data, and obtaining a medical segmentation task dataset, wherein the medical segmentation task dataset includes training set images and training segmentation true value images corresponding to the training set images;

[0076] The data preprocessing includes: size scaling, random cropping, horizontal mirroring, and random cropping;

[0077] The FDMUNet network model is trained according to the medical segmentation task dataset to obtain a trained FDMUNet network model.

[0078] Specifically, in this embodiment, medical image data for training is first obtained. This data typically comes from medical imaging libraries in hospitals or research institutions and includes various types of medical images, such as CT scans and MRI images. These images and their corresponding segmentation ground truth images constitute the medical segmentation task dataset. The training set images are used to train the model, while the training ground truth images serve as supervisory signals to guide the model in learning the correct segmentation method.

[0079] Before model training, preprocessing the training data is a crucial step. This includes scaling the images to ensure uniform size for model processing; performing random cropping to increase data diversity, allowing the model to learn different parts of the image; performing horizontal mirroring to further enrich the dataset and improve the model's generalization capabilities; and finally, performing random cropping again to enhance the model's adaptability to different image sizes and orientations. These preprocessing steps not only improve the model's robustness but also effectively prevent overfitting, enabling the model to maintain good performance on new, unseen images.

[0080] After data preprocessing, the FDMUNet network model is trained using the preprocessed medical segmentation task dataset. During the training process, the model gradually adjusts its own parameters by learning the mapping relationship between the training set images and the corresponding training segmentation ground truth images to minimize the difference between the predicted segmentation results and the true segmentation results. This training process usually involves a large number of iterative optimizations, as well as continuous evaluation and adjustment of model performance. Through this carefully designed training strategy, the FDMUNet network model is able to learn rich feature representations, thereby performing well in subsequent feature extraction and segmentation tasks. For example, when processing medical images containing complex lesions, the pre-trained model can more accurately identify the boundaries of the lesion area while combining global background information for precise segmentation. This fusion mechanism of global and local information significantly improves the model's ability to identify lesion areas, thereby providing a new technical solution for the field of medical image segmentation with important clinical application value.

[0081] In summary, this medical image segmentation method based on spatial perception and frequency domain information aims to address key challenges in existing medical image segmentation techniques, such as loss of edge details, omission of lesion regions, and insufficient extraction of multi-scale lesion features. This method significantly improves the accuracy and robustness of medical image segmentation and is applicable to a variety of medical image segmentation tasks.

[0082] Specifically, during the data preprocessing stage, medical images are resized, randomly cropped, and horizontally mirrored to enhance the model's generalization and adaptability to different image sizes. These preprocessing steps not only improve the model's robustness but also effectively prevent overfitting, enabling the model to maintain good performance even when faced with new, unseen images. During the feature extraction stage, an improved contextual feature extraction module focusing on low-frequency information (FLICEB) is introduced. This module leverages the multi-scale analysis capabilities of frequency domain transforms to decompose the input image into low-frequency and high-frequency components. The low-frequency component contains global information about the image, while the high-frequency component contains detailed information. Through a learnable noise filtering mechanism, the module can automatically identify and suppress noise or irrelevant background information in the low-frequency portion, enabling the model to more accurately focus on the lesion area. This innovative design significantly enhances the model's ability to capture both global information and detailed features, improving segmentation accuracy.

[0083] Furthermore, a multi-head perceptual visual state-space module (MPVSS) is designed at the bottleneck layer. This module extracts global image features from different scales through a multi-scale adaptive feature fusion mechanism. Combining the local perception advantages of CNNs with the efficient global modeling capabilities of state-space models, this module effectively captures multi-scale features, from subtle lesions to overall anatomical structures. This multi-scale feature fusion mechanism not only improves the model's segmentation accuracy for lesions of varying scales but also ensures sufficient feature representation at all resolutions, particularly for small lesions and complex structures. Within the skip connections, a focused contextual attention module (CFA) is introduced to enhance the boundaries of feature maps. Local features are obtained through a combination of average pooling and 1x1 convolutions. Depthwise separable large kernel convolutions are then used to expand the model's receptive field. Finally, enhanced features are obtained through 1x1 convolutions and sigmoid activation. This module design not only strengthens the model's focus on global information but also significantly enhances its ability to detect lesion edges, further improving segmentation accuracy.

[0084] Finally, the decoder performs bilinear interpolation upsampling on the global information features to restore the image's spatial resolution. This is then additively fused with the enhanced features to produce the final segmentation result. After continuous upsampling, the model ultimately generates a high-resolution feature map. This process not only combines global context with local details, but also ensures the accuracy and integrity of the segmentation results through feature fusion.

[0085] In summary, this method for medical image segmentation based on spatial perception and frequency domain information significantly improves the accuracy and robustness of medical image segmentation through the complementary fusion of frequency and spatial information, adaptive optimization of multi-scale features, and effective utilization of contextual information. This method is not only theoretically innovative but also demonstrates significant performance improvements in practical applications, providing a new technical solution for the field of medical image segmentation and possessing important clinical application value.

[0086] See also Figure 6 A second embodiment of the present invention provides a medical image segmentation device based on spatial perception and frequency domain information, comprising:

[0087] A preprocessing unit 101 is configured to acquire a medical image to be processed, perform data preprocessing on the medical image, and perform quality assessment on the preprocessed medical image, calculate the signal-to-noise ratio and peak signal-to-noise ratio of the image, assess the contrast and clarity of the image, detect artifacts and noise levels in the image, and if the signal-to-noise ratio is lower than a preset threshold, perform adaptive histogram equalization based on the image contrast and selectively apply gamma correction to locally enhance the detected artifact area;

[0088] The first extraction unit 102 is configured to call a pre-trained FDMUNet network model and perform feature extraction on the pre-processed medical image using a context feature extraction module focusing on low-frequency information to obtain an output feature map;

[0089] The second extraction unit 103 is used to extract the output feature map using the multi-head perception visual state space module in the FDMUNet network model to obtain global information features of the output feature map at different scales;

[0090] An enhancement unit 104 is configured to perform boundary enhancement processing on the feature map using a focused context attention module in the FDMUNet network model to obtain enhanced features;

[0091] The fusion unit 105 is used to use the decoder to perform recovery preprocessing on the global information features, and fuse the processed global information features with the enhanced features to obtain a segmentation result. This operation is repeated until a preset number of times is reached to generate a high-resolution feature map.

[0092] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A medical image segmentation method based on spatial perception and frequency domain information, characterized in that: include: Acquiring a medical image to be processed, performing data preprocessing on the medical image, and performing quality assessment on the preprocessed medical image to calculate the signal-to-noise ratio and peak signal-to-noise ratio of the image, assess the contrast and clarity of the image, detect artifacts and noise levels in the image, and if the signal-to-noise ratio is lower than a preset threshold, performing adaptive histogram equalization based on the image contrast, selectively applying gamma correction to locally enhance the detected artifact area; Call the pre-trained FDMUNet network model and use the context feature extraction module focusing on low-frequency information to perform feature extraction on the pre-processed medical image to obtain the output feature map; The output feature map is extracted and processed using the multi-head perception visual state space module in the FDMUNet network model to obtain global information features of the output feature map at different scales; Use the focused context attention module in the FDMUNet network model to perform boundary enhancement on the feature map to obtain enhanced features; Use the decoder to restore the global information features and pre-process them, and fuse the processed global information features with the enhanced features to obtain the segmentation result. Repeat this operation until the preset number of times is reached to generate a high-resolution feature map. Call the pre-trained FDMUNet network model to perform feature extraction on the pre-processed medical image to obtain the output feature map, specifically: For the input preprocessed medical feature map , H, W, C represent the height, width and number of channels of the image respectively. The context feature extraction module focusing on low-frequency information of the FDMUNet network model is used to perform feature decomposition processing on the medical feature map X at different preset frequency scales to capture different frequency features in the medical image. The formula is: , , T , ,in, is the frequency domain transform operation, is the low-frequency component of the medical image, Both are high-frequency components of medical images. is the output feature map obtained after the noise filtering mechanism, is the sigmoid function, is the maximum pooling operation, is the average pooling operation, Indicates connection by channel number, is a 1x1 convolution operation, is the batch normalization operation, is the ReLU activation function, is the characteristic information after high-frequency information and low-frequency information are connected through channels, is the extracted feature map; After three consecutive feature extractions, we get The output feature map of the three-times downsampling feature extraction is expressed as ,in, , , ; ， ， 。 2. The medical image segmentation method based on spatial perception and frequency domain information according to claim 1, characterized in that: The data preprocessing includes: size scaling, random cropping, and horizontal mirroring.

3. The medical image segmentation method based on spatial perception and frequency domain information according to claim 1, characterized in that: Perform quality assessment on pre-processed medical images, calculate the signal-to-noise ratio and peak signal-to-noise ratio of the image, evaluate the contrast and clarity of the image, detect artifacts and noise levels in the image, and if the signal-to-noise ratio is lower than the preset threshold, perform adaptive histogram equalization based on the image contrast and selectively apply gamma correction to perform local enhancement on the detected artifact areas. Specifically: Calculate the signal-to-noise ratio (SNR) and peak signal-to-noise ratio (PSNR) of the image to quantitatively evaluate image quality. The SNR measures image quality by calculating the ratio of the average power of the image signal to the average power of the noise, while the PSNR evaluates image fidelity by calculating the logarithmic relationship between the maximum possible pixel value and the mean square error. The root mean square contrast method is used to evaluate the global contrast of the image, and the second-order derivative of the image is calculated using the Laplacian operator to evaluate the clarity; To address the artifacts and noise in images, a noise detection mechanism based on high-frequency component analysis is designed, combining directional gradient histogram features and Radon transform to identify existing striping artifacts and motion artifacts. When the signal-to-noise ratio is detected to be lower than a preset threshold, contrast-limited adaptive histogram equalization is used for enhancement, where the cropping limit parameters are dynamically adjusted according to the signal-to-noise ratio value; For areas with insufficient contrast, adaptive gamma correction based on image statistical characteristics is introduced to automatically adjust the gamma value by analyzing the average brightness of the image; For the detected artifact area, local enhancement processing based on guided filtering is adopted, and Poisson fusion is used to ensure a natural transition between the processed area and the surrounding images.

4. The medical image segmentation method based on spatial perception and frequency domain information according to claim 3, characterized in that: The multi-head perception visual state space module in the FDMUNet network model is used to extract and process the output feature map to obtain the global information features of the output feature map at different scales, specifically: The output feature map is processed by the multi-head perception visual state space module in the bottleneck layer of the FDMUNet network model. Perform extraction processing to extract the global information features of the output feature maps at different scales. The formula of the multi-head perception visual state space module is: ,in, is a 3x3 depth-wise separable convolution operation, For VSS module, is layer normalization, It is a multi-scale adaptive feature fusion module. are the extracted fine-grained features and coarse-grained features; The multi-scale adaptive feature fusion module divides the output feature map into three branches of different scales for processing, and performs feature fusion on the outputs of the three branches. The formula of the multi-scale adaptive feature fusion module is: , , , , among which ConCat To splice by channel for feature fusion, is the output of the first branch, is the output of the second branch, is the output of the third branch, is 1x1 average adaptive pooling, is 3x3 average adaptive pooling, is 5x5 average adaptive pooling, yes The result of feature fusion; After four consecutive multi-head perception visual state space modules, the output feature map Extraction process to obtain The feature map of the fourth bottleneck layer is represented as , , , ,in, , , .

5. The medical image segmentation method based on spatial perception and frequency domain information according to claim 1, characterized in that: The focused context attention module in the FDMUNet network model is used to perform boundary enhancement on the feature map to obtain enhanced features, specifically: During the jump connection process, the feature map is focused on the context attention module in the FDMUNet network model. 、 and Boundary enhancement is performed separately, and local area features are obtained by combining average pooling and 1x1 convolution; Use two depth-wise separable large kernel convolutions to expand the features, and use 1x1 convolution and sigmoid activation function to perform convolution activation on the processed features to obtain enhanced features. The formula is: ,in, To enhance the features, is an 11x1 depth-wise separable convolution operation, It is a 1x11 depth-wise separable convolution operation; Similarly, the features of the three layers after being enhanced by the focused context attention module are expressed as, , , ,in, , .

6. The medical image segmentation method based on spatial perception and frequency domain information according to claim 4, characterized in that: Use the decoder to restore the global information features and pre-process them, and fuse the processed global information features with the enhanced features to obtain the segmentation result. Repeat this operation until the preset number of times is reached to generate a high-resolution feature map, specifically: Use the decoder to process the global information output by the multi-head perception visual state space module Enhanced features obtained with the focused context attention module Fusion is performed by channel and bilinear interpolation upsampling is performed to obtain , to restore the controlled size of the image; The convolution operation is used to perform feature dimension reduction and batch normalization on the processed global information features, and the feature information is retained through the ReLU activation function; The enhanced features obtained by the jump connection and the retained global information features are added and fused to obtain the features restored by the decoder. The formula is: ,in, is the upsampling operation, To obtain the features restored by the decoder; After three consecutive feature fusions and upsampling, the final high-resolution feature map G is obtained. , where the features recovered by the decoder are , ,in, , .

7. The medical image segmentation method based on spatial perception and frequency domain information according to claim 1, characterized in that: Before calling the pre-trained FDMUNet network model to perform feature extraction on the pre-processed medical image, it also includes: Acquiring training data, performing data preprocessing on the training data, and obtaining a medical segmentation task dataset, wherein the medical segmentation task dataset includes training set images and training segmentation true value images corresponding to the training set images; The data preprocessing includes: size scaling, random cropping, and horizontal mirroring; The FDMUNet network model is trained according to the medical segmentation task dataset to obtain a trained FDMUNet network model.

8. A medical image segmentation device based on spatial perception and frequency domain information, characterized in that: A method for implementing the medical image segmentation method based on spatial perception and frequency domain information according to any one of claims 1 to 7, comprising: a preprocessing unit, configured to acquire a medical image to be processed, perform data preprocessing on the medical image, and perform quality assessment on the preprocessed medical image, calculate the signal-to-noise ratio and peak signal-to-noise ratio of the image, assess the contrast and clarity of the image, detect artifacts and noise levels in the image, and if the signal-to-noise ratio is lower than a preset threshold, perform adaptive histogram equalization based on the image contrast, selectively apply gamma correction to perform local enhancement processing on the detected artifact area; The first extraction unit is used to call the pre-trained FDMUNet network model and use the context feature extraction module focusing on low-frequency information to perform feature extraction on the pre-processed medical image to obtain an output feature map; A second extraction unit is used to extract the output feature map using a multi-head perception visual state space module in the FDMUNet network model to obtain global information features of the output feature map at different scales; The enhancement unit is used to perform boundary enhancement processing on the feature map using the focused context attention module in the FDMUNet network model to obtain enhanced features; The fusion unit is used to use the decoder to restore the global information features and fuse the processed global information features with the enhanced features to obtain the segmentation result. This operation is repeated until a preset number of times is reached to generate a high-resolution feature map.

Images