Medical image segmentation method based on Laplacian pyramid and dynamic Transform

Abstract

The invention discloses a medical image segmentation method based on a Laplacian pyramid and a dynamic Transform. The method comprises the following steps: S1, forming a preprocessed medical image; s2, decomposing the preprocessed medical image into a plurality of image sub-layers with different scales; s3, generating a corresponding multi-scale feature representation; s4, inputting the multi-scale feature representation into a dynamic Transform module, and generating a fused global feature representation; and S5, decoding the fused global feature representation to generate a preliminary segmentation image, applying a post-processing algorithm to the preliminary segmentation image, and executing morphological operation, edge repair and noise removal to obtain a medical image segmentation result with clear boundary and regional coherence. According to the method, a medical image segmentation result with clear boundary, coherent region and accurate semantics is generated, the practicability of the model in clinical application is structurally improved, and the method has significant advantages in processing images with fuzzy tumor boundary and dense and overlapped tissues.

Description

Technical Field

[0001] The present invention relates to the field of medical technology, and in particular to a medical image segmentation method based on Laplacian pyramid and dynamic Transformer. Background Art

[0002] With the development of medical artificial intelligence and computer vision technology, medical image segmentation has gradually become a core link in key tasks such as auxiliary diagnosis, preoperative planning and lesion monitoring. The goal of medical image segmentation is to accurately separate different tissue structures, lesion areas and functional areas in various medical images such as CT, MRI and ultrasound, thereby providing accurate spatial positioning and contour description for downstream analysis and decision-making. At present, the mainstream methods of medical image segmentation mostly use structures based on deep convolutional neural networks, including U-Net and ResNet, to achieve semantic segmentation of images through end-to-end training. However, due to the multi-scale lesion morphology, variable tissue structure and noise interference characteristics of medical images, traditional methods still have obvious shortcomings in processing image details and collaborative modeling of global information.

[0003] Specifically, existing convolution-based image segmentation models mostly rely on fixed receptive fields for local feature extraction, which makes it difficult to capture global semantic relationships while retaining detailed information. When segmenting small lesions, areas with blurred edges or complex tissue structures, problems such as unclear boundaries, inaccurate positioning, and missed lesions often occur. Some studies have introduced multi-scale input or feature pyramid structures to alleviate this problem, but lack an adaptive mechanism in the feature fusion stage and cannot dynamically adjust the contribution ratio of features at each scale according to the image content. In addition, in recent years, the Transformer structure has been introduced into medical image segmentation to enhance the global modeling capability of the model, but existing Transformer modules mostly use static parameter mapping, which makes it difficult to achieve dynamic feature adjustment based on the heterogeneity of lesion areas in the image, thereby limiting the adaptability and segmentation accuracy of the model in complex image scenarios.

[0004] Therefore, there is an urgent need for a segmentation method that can integrate multi-scale detail information, has a dynamic adaptive mechanism, and adapts to the structural complexity of medical images, so as to improve the accuracy, robustness and clinical practicality of the segmentation model. Summary of the Invention

[0005] One purpose of the present invention is to propose a medical image segmentation method based on Laplacian pyramid and dynamic Transformer. The present invention generates medical image segmentation results with clear boundaries, coherent regions, and accurate semantics. The structure improves the practicality of the model in clinical applications and has significant advantages in processing images with blurred tumor boundaries and densely overlapping tissues.

[0006] A medical image segmentation method based on a Laplacian pyramid and a dynamic Transformer according to an embodiment of the present invention includes the following steps:

[0007] S1. collecting an input medical image and performing preprocessing on the medical image to form a preprocessed medical image;

[0008] S2. Perform multi-scale decomposition on the pre-processed medical image using a Laplacian pyramid, decomposing the pre-processed medical image into multiple image sub-layers of different scales;

[0009] S3. Extract features from each scale image sub-layer separately, use a convolutional neural network to encode local features of each scale image sub-layer, and generate corresponding multi-scale feature representations;

[0010] S4. Input the multi-scale feature representation into the dynamic Transformer module, dynamically weight the features of each scale through the self-attention mechanism to generate a fused global feature representation;

[0011] S5. Decode the fused global feature representation and restore the fused global feature representation mapping to the same spatial resolution as the original medical image using upsampling, deconvolution, and skip connection techniques to generate a preliminary segmentation map. Apply post-processing algorithms to the preliminary segmentation map to perform morphological operations, edge repair, and noise removal to obtain a medical image segmentation result with clear boundaries and regional coherence.

[0012] Optionally, the medical image is a CT image, an MRI image or an ultrasound image.

[0013] Optionally, the S1 includes the following steps:

[0014] S11. Obtain input medical image data I raw , Medical Image Data I raw It is the original medical image matrix of a two-dimensional or three-dimensional structure, containing multiple grayscale pixel value distributions, which are used to characterize the difference in medical image intensity between the lesion area and normal tissue:

[0015] I raw ={i x,y ∣i x,y ∈[0,I max ],x∈[1,H],y∈[1,W]};

[0016] Among them, i x,y Represents the grayscale value of the pixel at position (x, y), H and W represent the height and width of the medical image respectively, I max The maximum grayscale value supported by the medical imaging system;

[0017] S12. Using Gaussian filtering operation to process medical image data I raw The grayscale value of each pixel in the image is weighted averaged with its neighboring pixels, and the filtering weight is calculated based on the two-dimensional Gaussian distribution function to obtain a noise-suppressed medical image.

[0018] S13. Normalizing the grayscale values of all pixels in the noise-suppressed medical image by the difference between the minimum grayscale value and the maximum grayscale value, and converting the image into a normalized medical image within the interval [0, 1];

[0019] S14. Using the linear grayscale transformation method, the grayscale value of each pixel in the normalized medical image is scaled and shifted by a constant to enhance the overall brightness and darkness contrast of the medical image, and the contrast-enhanced medical image I is obtained. pre , contrast-enhanced medical image I pre Used to improve the boundary visibility between the lesion area and background tissue in medical images.

[0020] Optionally, the S2 includes the following steps:

[0021] S21. Contrast-enhanced medical image I pre As the initial input medical image, a multi-scale Gaussian medical image sequence {G0, G1, ..., G N}, where G0 = I pre , for each scale Gaussian pyramid image G k The next level of multi-scale Gaussian medical image G is generated by smoothing and downsampling the guided Gaussian convolution method of integrating the local edge gradient response of the medical image. k+1 :

[0022]

[0023] Among them, ω(u,v) is the normalized two-dimensional Gaussian kernel, r represents the radius of the Gaussian convolution kernel, u and υ represent the offset of the Gaussian convolution kernel in the horizontal and vertical directions respectively, G k is the Gaussian pyramid image of the kth layer, G k+1 (x, y) is the pixel value of the Gaussian pyramid image at the spatial position (x, y) in the k+1th layer, λ is the edge enhancement factor, E k (x,y) represents the edge gradient response intensity at the pixel position (x,y) in the current scale k medical image:

[0024]

[0025] in, and It is the first-order gradient of the medical image in the horizontal and vertical directions, which is used to highlight the grayscale change interface between the lesion edge and normal tissue;

[0026] S22. For each scale Gaussian pyramid image G k Instead of upsampling medical images Up(G k+1 ) performs a significant guided difference operation to obtain the high-frequency Laplacian medical image sublayer L k The saliency guided difference operation process is achieved by fusing the medical image saliency response map S k Reweighting:

[0027] L k (x,y)=G k (x,y)-S k (x,y)·Up(G k+1 )(x,y);

[0028] Significant response map S k (x,y)∈[0,1] represents the importance of the region at position (x,y) to the overall segmentation target. The saliency response is calculated by the lesion morphology, texture abnormality, and structural asymmetry features in the medical image. It is used to enhance the layer response of the lesion area and suppress the influence of the background area.

[0029] S23. Repeat steps S21 and S22 to construct a complete multi-scale Laplacian pyramid medical image set

[0030]

[0031] Among them, L N-1 is the N-1th high-frequency Laplacian medical image sublayer, G N is the lowest level Gaussian medical image.

[0032] Optionally, S3 includes the following steps:

[0033] S31. Multi-scale Laplacian Pyramid Medical Image Collection Each scale medical image sublayer L k and the bottom global medical image G N Perform feature encoding separately and use scale-sharing convolutional encoding module F conv Extract the local feature response of each medical image sub-layer and obtain the initial encoding feature map F at each scale k :

[0034]

[0035] Among them, F k(x,y) represents the feature response value of the initial encoding feature map at position (x,y) at scale k, represents the kth Laplacian medical image sublayer or the bottom global medical image of the input medical image channel i, C in is the number of input channels, represents the weight parameter of the convolution kernel of scale k at position (u, v) under channel i, b k is the bias term, σ represents the nonlinear activation function, and the convolution kernel size is (2r+1)×(2r+1);

[0036] S32. The initial encoding feature map F at each scale k The corresponding medical image saliency response map S k Perform pixel-by-pixel weighted fusion to generate a saliency enhanced feature map Saliency Enhancement Feature Map The enhancement value of each pixel in the initial encoding feature map F is the pixel k The original value and its corresponding position in the medical image saliency response map S k The sum of the products of the response values in , plus the result of scaling the original value itself, is used to enhance the response intensity of the lesion area in the feature map:

[0037]

[0038] Among them, γ is the significance enhancement coefficient, which is used to control the enhancement degree of the characteristic response of the lesion area, S k (x,y)∈[0,1] represents the significant response value at the pixel position (x,y) at scale k, which is used to highlight the local features of lesion contours, edge abnormalities and texture discontinuities in medical images;

[0039] S33. Saliency Enhanced Feature Maps at All Scales Perform dimension unification and structural adjustment, and adjust the feature maps of different scales to the preset unified resolution (H′, W′) and unified channel number C′ through channel transformation and spatial alignment operations to form a standardized multi-scale feature representation set F multi :

[0040]

[0041] Among them, φ(·) represents the composite transformation operation of upsampling, channel mapping and feature normalization, H k With W k is the spatial dimension of the original scale k, C k is the number of channels, and (H′, W′, C′) is the dimension and number of channels of the uniformly aligned target space.

[0042] Optionally, the S4 includes the following steps:

[0043] S41. Input the standardized multi-scale feature representation set into the dynamic Transformer module, and enhance the saliency feature map at each scale The input feature sequence flattened into a two-dimensional sequence representation Where H′·W′ is the sequence length and C′ is the channel dimension of each position;

[0044] S42. The dynamic parameter generation module driven by image content is used to generate the input feature sequence X at each scale. k To perform encoding processing, first input feature sequence X k Perform global average pooling operation to obtain the statistical representation vector at scale k The statistical representation vector μ k Input into three multi-layer perception networks with shared structure but independent parameters to generate the query weight matrix of Transformer at this scale Bond weight matrix and the value weight matrix

[0045] S43. Based on query weight matrix Bond weight matrix and the value weight matrix For the input feature sequence X k Perform matrix multiplication to obtain the query matrix Q at scale k k , key matrix K k Sum matrix V k ;

[0046] S44. Calculate the attention weight matrix within the scale Use scaled dot product attention mechanism for attention allocation;

[0047] S45. Introduce content-aware gating mechanism to dynamically adjust attention distribution at pixel level, according to the input medical image saliency response map S k , construct pixel attention gating factor G k , and embed the pixel attention gating factor into the attention weighted output to obtain the dynamic attention output sequence Z k :

[0048] Z k =(G k ⊙A k )·V k ;

[0049] where ⊙ represents the element-wise multiplication with row broadcasting, G k(i) = 1 + γ·S k (i) used to enhance the dynamic response strength of pixels in the lesion area in the attention output;

[0050] S46. Dynamic attention output sequence Z at all scales k Perform scale-level fusion and introduce global adaptive fusion weight α k The global adaptive fusion weight is generated by the statistical characteristics of the attention output at each scale, and the global adaptive fusion weight is dynamically adjusted according to the response intensity of the lesion area at each scale:

[0051]

[0052] Among them, Z k represents the dynamic attention output sequence obtained after processing by the dynamic Transformer module at scale k, Z j represents the dynamic attention output sequence obtained after processing by the dynamic Transformer module at scale j, mean(Z k ) represents the dynamic attention output sequence Z k The global average pooling result, that is, the average of the feature responses of all spatial positions, is used to extract the main semantic features expressed by scale k in the overall image, reflecting the overall response intensity of the lesion area at this scale. w is the weight vector in the adaptive fusion function, and b is the bias term in the adaptive fusion function.

[0053] S47. According to the global adaptive fusion weight α k Perform weighted summation on the dynamic outputs of all scales to obtain the fused global feature representation F global .

[0054] Optionally, the S5 includes the following steps:

[0055] S51. The fused global feature is represented by F global Input to the decoding module, the fused global feature representation F global The decoding module is a three-dimensional tensor of size H′×W′×d. The spatial resolution is gradually restored by upsampling and deconvolution operations layer by layer, and a sequence of decoding feature maps with increasing resolution is constructed until the spatial resolution H×W of the original medical image is restored. The number of channels and spatial size of each layer of decoding feature maps are adjusted in sequence according to the number of decoding layers, and finally a decoding output with the same spatial resolution as the original medical image is obtained.

[0056] S52. Introduce the skip connection mechanism in the decoding process to enhance the multi-scale saliency feature map The feature map with the same spatial resolution as the output resolution of the current decoding layer is spatially aligned through upsampling or convolution operations, and then spliced with the current decoding feature map in the channel dimension to obtain a fused feature map, which is used to enhance the ability to preserve fine-grained structures and boundary features during the decoding process and avoid the loss of lesion area information during the upsampling process;

[0057] S53. Perform a channel mapping operation on the last layer of decoded feature maps, using a convolution kernel whose number of channels is equal to the number of medical image segmentation categories to map the final decoded feature maps into a preliminary medical image segmentation map. The spatial size of the preliminary medical image segmentation map is H×W, and the number of output channels of each pixel is equal to the number of segmentation categories, indicating the probability distribution of the pixel belonging to each category. The probability distribution is processed by a Softmax normalization function so that the output of each pixel satisfies the category sum to 1.

[0058] S54. Performing post-processing operations on the preliminary medical image segmentation map, the post-processing operations including morphological operations, edge repair operations, and noise removal operations;

[0059] S55. Output the post-processed medical image segmentation result as the final medical image segmentation result, wherein the final medical image segmentation result is a two-dimensional label map of size H×W, in which each pixel corresponds to a unique medical image segmentation category.

[0060] Optionally, in the morphological operation, a closing operation and an opening operation are combined using a structuring element to perform dilation followed by erosion. The closing operation is used to fill small holes in the lesion area, and the opening operation is used to remove small isolated artifacts, thereby enhancing the connectivity and integrity of the segmented area.

[0061] In the edge restoration operation, Gaussian filtering and curvature smoothing methods are used to adjust the edge of the segmented area to make the segmented edge smoother and more natural in order to solve the jagged, discontinuous or overly sharp boundary problems existing in the preliminary medical image segmentation map.

[0062] In the noise removal operation, through regional connectivity analysis and shape constraint judgment, false detection areas and isolated pixels in the background are automatically identified and eliminated, thereby optimizing the accuracy and stability of the lesion area in the medical image segmentation results.

[0063] The beneficial effects of the present invention are:

[0064] (1) Based on the traditional Laplacian pyramid, the present invention introduces the edge response guidance and saliency response map control mechanism, uses guided Gaussian convolution to perform edge-sensitive Gaussian blur processing on the image, and performs pixel-level difference enhancement by fusing the saliency map when constructing the Laplacian pyramid layer. This not only improves the response intensity of the lesion area, but also suppresses the interference information of the background area. It can enhance the multi-scale high-frequency detail feature expression while maintaining the global structural semantics, and is suitable for the segmentation modeling of lesion areas in multi-modality, low-contrast medical images.

[0065] (2) The present invention combines the image content-driven mechanism with the Transformer structure. By statistically analyzing the global distribution of features at each scale, the query, key, and value weight matrices are dynamically generated, so that the attention calculation of the Transformer no longer relies on static parameters, but is adjusted in real time with the image content. Combined with the content-aware gating mechanism and the saliency response map, the attention output is dynamically adjusted pixel by pixel in the spatial dimension, effectively improving the model's sensitivity to lesion boundaries, morphological changes, and heterogeneous regions. The dynamic Transformer module significantly enhances the segmentation model's adaptability to fine-grained structures in complex medical images, making up for the deficiency of the traditional attention mechanism in modeling the joint dynamic relationship of space and scale.

[0066] (3) Based on the fused global feature representation, the present invention designs a step-by-step upsampling decoding network, combines the skip connection mechanism of the multi-scale feature map, supplements the detail information in the decoding process, effectively avoids the semantic ambiguity and edge information loss during feature upsampling, introduces a structured post-processing module in the output stage, uses morphological closing operations to enhance regional connectivity, uses edge repair filters to smooth the segmentation contours, and performs regional connectivity detection to eliminate isolated artifacts, ultimately generating medical image segmentation results with clear boundaries, regional coherence, and accurate semantics. The structure improves the practicability of the model in clinical applications, and has significant advantages in processing images with blurred tumor boundaries and densely overlapping tissues. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0068] Figure 1 This is a flowchart of a medical image segmentation method based on Laplacian pyramid and dynamic Transformer proposed by the present invention;

[0069] Figure 2 This is a schematic diagram of the segmentation results on a brain MRI image in Example 1 of a medical image segmentation method based on Laplacian pyramid and dynamic Transformer proposed in the present invention. DETAILED DESCRIPTION

[0070] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0071] refer to Figure 1 , a medical image segmentation method based on Laplacian pyramid and dynamic Transformer, comprising the following steps:

[0072] S1. collecting an input medical image and performing preprocessing on the medical image to form a preprocessed medical image;

[0073] S2. Perform multi-scale decomposition on the pre-processed medical image using a Laplacian pyramid, decomposing the pre-processed medical image into multiple image sub-layers of different scales;

[0074] S3. Extract features from each scale image sub-layer separately, use a convolutional neural network to encode local features of each scale image sub-layer, and generate corresponding multi-scale feature representations;

[0075] S4. Input the multi-scale feature representation into the dynamic Transformer module, dynamically weight the features of each scale through the self-attention mechanism to generate a fused global feature representation;

[0076] S5. Decode the fused global feature representation and restore the fused global feature representation mapping to the same spatial resolution as the original medical image using upsampling, deconvolution, and skip connection techniques to generate a preliminary segmentation map. Apply post-processing algorithms to the preliminary segmentation map to perform morphological operations, edge repair, and noise removal to obtain a medical image segmentation result with clear boundaries and regional coherence.

[0077] In this embodiment, the medical image is a CT image, an MRI image, or an ultrasound image.

[0078] In this embodiment, S1 includes the following steps:

[0079] S11. Obtain input medical image data I raw , Medical Image Data I raw It is the original medical image matrix of a two-dimensional or three-dimensional structure, containing multiple grayscale pixel value distributions, which are used to characterize the difference in medical image intensity between the lesion area and normal tissue:

[0080] I raw ={i x,y ∣i x,y ∈[0,I max ],x∈[1,H],y∈[1,W]};

[0081] Among them, i x,y Represents the grayscale value of the pixel at position (x, y), H and W represent the height and width of the medical image respectively, I max The maximum grayscale value supported by the medical imaging system;

[0082] S12. Using Gaussian filtering operation to process medical image data I raw The grayscale value of each pixel in the image is weighted averaged with its neighboring pixels, and the filtering weight is calculated based on the two-dimensional Gaussian distribution function to obtain a noise-suppressed medical image.

[0083] S13. Normalizing the grayscale values of all pixels in the noise-suppressed medical image by the difference between the minimum grayscale value and the maximum grayscale value, and converting the image into a normalized medical image within the interval [0, 1];

[0084] S14. Using the linear grayscale transformation method, the grayscale value of each pixel in the normalized medical image is scaled and shifted by a constant to enhance the overall brightness and darkness contrast of the medical image, and the contrast-enhanced medical image I is obtained. pre , contrast-enhanced medical image I pre Used to improve the boundary visibility between the lesion area and background tissue in medical images.

[0085] In this embodiment, S2 includes the following steps:

[0086] S21. Contrast-enhanced medical image I pre As the initial input medical image, a multi-scale Gaussian medical image sequence {G0, G1, ..., G N}, where G0 = I pre , for each scale Gaussian pyramid image G k The next level of multi-scale Gaussian medical image G is generated by smoothing and downsampling the guided Gaussian convolution method of integrating the local edge gradient response of the medical image. k+1 :

[0087]

[0088] Among them, ω(u,v) is the normalized two-dimensional Gaussian kernel, r represents the radius of the Gaussian convolution kernel, u and υ represent the offset of the Gaussian convolution kernel in the horizontal and vertical directions respectively, G k is the Gaussian pyramid image of the kth layer, G k+1 (x, y) is the pixel value of the Gaussian pyramid image at the spatial position (x, y) in the k+1th layer, λ is the edge enhancement factor, E k (x,y) represents the edge gradient response intensity at the pixel position (x,y) in the current scale k medical image:

[0089]

[0090] in, and It is the first-order gradient of the medical image in the horizontal and vertical directions, which is used to highlight the grayscale change interface between the lesion edge and normal tissue;

[0091] S22. For each scale Gaussian pyramid image G k Instead of upsampling medical images Up(G k+1 ) performs a significant guided difference operation to obtain the high-frequency Laplacian medical image sublayer L k The saliency guided difference operation process is achieved by fusing the medical image saliency response map S k Reweighting:

[0092] L k (x,y)=G k (x,y)-S k (x,y)·Up(G k+1 )(x,y);

[0093] Significant response map S k (x,y)∈[0,1] represents the importance of the region at position (x,y) to the overall segmentation target. The saliency response is calculated by the lesion morphology, texture abnormality, and structural asymmetry features in the medical image. It is used to enhance the layer response of the lesion area and suppress the influence of the background area.

[0094] S23. Repeat steps S21 and S22 to construct a complete multi-scale Laplacian pyramid medical image set

[0095]

[0096] Among them, L N-1 is the N-1th high-frequency Laplacian medical image sublayer, G N is the lowest level Gaussian medical image.

[0097] In this embodiment, S3 includes the following steps:

[0098] S31. Multi-scale Laplacian Pyramid Medical Image Collection Each scale medical image sublayer L k and the bottom global medical image G N Perform feature encoding separately and use scale-sharing convolutional encoding module F conv Extract the local feature response of each medical image sub-layer and obtain the initial encoding feature map F at each scale k :

[0099]

[0100] Among them, F k (x,y) represents the feature response value of the initial encoding feature map at position (x,y) at scale k, represents the kth Laplacian medical image sublayer or the bottom global medical image of the input medical image channel i, C in is the number of input channels, represents the weight parameter of the convolution kernel of scale k at position (u, v) under channel i, b k is the bias term, σ represents the nonlinear activation function, and the convolution kernel size is (2r+1)×(2r+1);

[0101] S32. The initial encoding feature map F at each scale k The corresponding medical image saliency response map S k Perform pixel-by-pixel weighted fusion to generate a saliency enhanced feature map Saliency Enhancement Feature Map The enhancement value of each pixel in the initial encoding feature map F is the pixel k The original value and its corresponding position in the medical image saliency response map S k The sum of the products of the response values in , plus the original values themselves scaled:

[0102]

[0103] Among them, γ is the significance enhancement coefficient, which is used to control the enhancement degree of the characteristic response of the lesion area, S k (x,y)∈[0,1] represents the significant response value at the pixel position (x,y) at scale k, which is used to highlight the local features of lesion contours, edge abnormalities and texture discontinuities in medical images;

[0104] S33. Saliency Enhanced Feature Maps at All Scales Perform dimension unification and structural adjustment, and adjust the feature maps of different scales to the preset unified resolution (H′, W′) and unified channel number C′ through channel transformation and spatial alignment operations to form a standardized multi-scale feature representation set F multi :

[0105]

[0106] Among them, φ(·) represents the composite transformation operation of upsampling, channel mapping and feature normalization, H k With W k is the spatial dimension of the original scale k, C kis the number of channels, and (H′, W′, C′) is the dimension and number of channels of the uniformly aligned target space.

[0107] In this embodiment, S4 includes the following steps:

[0108] S41. Input the standardized multi-scale feature representation set into the dynamic Transformer module, and enhance the saliency feature map at each scale The input feature sequence flattened into a two-dimensional sequence representation Where H′·W′ is the sequence length and C′ is the channel dimension of each position;

[0109] S42. The dynamic parameter generation module driven by image content is used to generate the input feature sequence X at each scale. k To perform encoding processing, first input feature sequence X k Perform global average pooling operation to obtain the statistical representation vector at scale k The statistical representation vector μ k Input into three multi-layer perception networks with shared structure but independent parameters to generate the query weight matrix of Transformer at this scale Bond weight matrix and the value weight matrix

[0110] S43. Based on query weight matrix Bond weight matrix and the value weight matrix For the input feature sequence X k Perform matrix multiplication to obtain the query matrix Q at scale k k , key matrix K k Sum matrix V k ;

[0111] S44. Calculate the attention weight matrix within the scale Use scaled dot product attention mechanism for attention allocation;

[0112] S45. Introduce content-aware gating mechanism to dynamically adjust attention distribution at pixel level, according to the input medical image saliency response map S k , construct pixel attention gating factor G k , and embed the pixel attention gating factor into the attention weighted output to obtain the dynamic attention output sequence Z k :

[0113] Z k =(G k ⊙A k )·V k ;

[0114] where ⊙ represents the element-wise multiplication with row broadcasting, G k (i) = 1 + γ·S k (i) used to enhance the dynamic response strength of pixels in the lesion area in the attention output;

[0115] S46. Dynamic attention output sequence Z at all scales k Perform scale-level fusion and introduce global adaptive fusion weight α k The global adaptive fusion weight is generated by the statistical characteristics of the attention output at each scale, and the global adaptive fusion weight is dynamically adjusted according to the response intensity of the lesion area at each scale:

[0116]

[0117] Among them, Z k represents the dynamic attention output sequence obtained after processing by the dynamic Transformer module at scale k, Z j represents the dynamic attention output sequence obtained after processing by the dynamic Transformer module at scale j, mean(Z k ) represents the dynamic attention output sequence Z k The global average pooling result, that is, the average of the feature responses of all spatial positions, is used to extract the main semantic features expressed by scale k in the overall image, reflecting the overall response intensity of the lesion area at this scale. w is the weight vector in the adaptive fusion function, and b is the bias term in the adaptive fusion function.

[0118] S47. According to the global adaptive fusion weight α k Perform weighted summation on the dynamic outputs of all scales to obtain the fused global feature representation F global .

[0119] In this embodiment, S5 includes the following steps:

[0120] S51. The fused global feature is represented by F global Input to the decoding module, the fused global feature representation F global The decoding module is a three-dimensional tensor of size H′×W′×d. The spatial resolution is gradually restored by upsampling and deconvolution operations layer by layer, and a sequence of decoding feature maps with increasing resolution is constructed until the spatial resolution H×W of the original medical image is restored. The number of channels and spatial size of each layer of decoding feature maps are adjusted in sequence according to the number of decoding layers, and finally a decoding output with the same spatial resolution as the original medical image is obtained.

[0121] S52. Introduce the skip connection mechanism in the decoding process to enhance the multi-scale saliency feature map The feature map with the same spatial resolution as the output resolution of the current decoding layer is spatially aligned through upsampling or convolution operations, and then spliced with the current decoding feature map in the channel dimension to obtain a fused feature map, which is used to enhance the ability to preserve fine-grained structures and boundary features during the decoding process and avoid the loss of lesion area information during the upsampling process;

[0122] S53. Perform a channel mapping operation on the last layer of decoded feature maps, using a convolution kernel whose number of channels is equal to the number of medical image segmentation categories to map the final decoded feature maps into a preliminary medical image segmentation map. The spatial size of the preliminary medical image segmentation map is H×W, and the number of output channels of each pixel is equal to the number of segmentation categories, indicating the probability distribution of the pixel belonging to each category. The probability distribution is processed by a Softmax normalization function so that the output of each pixel satisfies the category sum to 1.

[0123] S54. Performing post-processing operations on the preliminary medical image segmentation map, the post-processing operations including morphological operations, edge repair operations, and noise removal operations;

[0124] S55. Output the post-processed medical image segmentation result as the final medical image segmentation result. The final medical image segmentation result is a two-dimensional label map of size H×W, where each pixel corresponds to a unique medical image segmentation category.

[0125] In this embodiment, in the morphological operation, a combination of closing and opening operations is performed using a structuring element, performing the steps of dilation followed by erosion. The closing operation is used to fill small holes in the lesion area, and the opening operation is used to remove small isolated artifacts, thereby enhancing the connectivity and integrity of the segmented area.

[0126] In the edge restoration operation, Gaussian filtering and curvature smoothing methods are used to adjust the edge of the segmented area to address the jagged, discontinuous or overly sharp boundary problems in the preliminary medical image segmentation map, making the segmentation edge smoother and more natural.

[0127] During the noise removal operation, through regional connectivity analysis and shape constraint judgment, false detection areas and isolated pixels in the background are automatically identified and eliminated, optimizing the accuracy and stability of the lesion area in the medical image segmentation results.

[0128] Example 1:

[0129] In the example, a 56-year-old male patient was admitted to the neurosurgery department of a tertiary general hospital with complaints of recurrent headaches and blurred vision. His medical history showed that he had undergone meningioma resection 5 years ago. After this admission, the doctor arranged for him to undergo a head MRI to assess for tumor recurrence.

[0130] The Department of Imaging acquired the patient's complete multimodal MRI data, including T1, T1ce, T2, and FLAIR sequences, each with a resolution of 240×240×155, and uploaded them to the hospital's PACS system. Due to the blurred lesion boundaries and the presence of high-density vascular areas in the surrounding tissue, the initial segmentation results generated by the traditional U-Net model in the PACS system showed blurred outlines of the suspected lesion, making it difficult for doctors to clearly define the recurrence boundary and posing a risk of residual recurrence in the surgical area.

[0131] Subsequently, the hospital's intelligent image analysis platform automatically calls the method of the present invention to perform fine segmentation processing:

[0132] First, the system automatically acquires the input MRI image and performs image preprocessing, including noise suppression, grayscale normalization and contrast enhancement. Taking T1ce mode as an example, the system generates an image matrix I after preprocessing. pre , and it shows that the edge of the lesion in the enhanced image is significantly clearer than that in the original image.

[0133] The system activated a multi-scale decomposition module based on the Laplacian pyramid. For this case image, five image sublayers were constructed, each retaining information at a different spatial frequency. The bottom Gaussian layer preserved the general shape of the tumor, while the high-frequency layer clearly displayed the irregular enhancement at the edge of the suspected lesion.

[0134] Then, the system enters the feature encoding stage. Each scale layer is enhanced by the saliency mechanism and combined with the saliency response map S k The system automatically detected patchy signal enhancement points located at the junction of the right frontal and parietal lobes in the second-layer sub-image, with a significant response value of 0.89, while the average response of the background brain tissue was only 0.12.

[0135] The enhanced multi-scale features were fed into the dynamic Transformer module, which analyzed the feature distribution of the patient's image in real time, dynamically generated a weight matrix corresponding to each scale, and implemented an attention-weighted mechanism. After dynamic adjustment, the attention value of the suspected lesion area on the right side in the fused feature map reached the highest value of 0.93, the highest in the entire image. This mechanism effectively suppressed false responses in the background sulci and venous sinuses.

[0136] The system completes upsampling and decoding operations, injects edge details from the multi-scale feature map into the decoding network through skip connections, and generates a medical image segmentation map with a spatial resolution of 240×240. In the preliminary segmentation map, the lesion volume detected in the right frontal lobe is 5.6cm 3 The morphology is irregular nodular with clear boundaries and continuous signals.

[0137] The system automatically performs morphological processing and edge smoothing filtering, performs closing operations on the lesion area to eliminate stray small dots, and performs boundary curvature filtering to eliminate segmentation artifacts and remove isolated signal artifacts in the left occipital lobe.

[0138] Final segmentation result S final The system automatically overlays the segmentation map onto the original MRI image, creating a 3D visualization. The physician, after reviewing the results on a graphics workstation, confirmed that the system's segmentation results completely covered the clinically suspicious area, and that the lesion's edge contours were highly consistent with the enhanced MRI. The physician determined that the likelihood of tumor recurrence was extremely high, and recommended preoperative planning.

[0139] To verify the effectiveness of the proposed method compared with traditional methods, the hospital retrospectively selected imaging data of 40 patients with frontal lobe gliomas (including 20 low-grade gliomas and 20 high-grade gliomas) from the imaging database over the past 6 months for comparative experiments. The comparison models included traditional U-Net and the proposed method (LP-DT). Both were evaluated under the same training set and the same hardware (NVIDIA RTX3090):

[0140] Table 1 Comparison of segmentation performance indicators between traditional U-Net and the proposed method (LP-DT)

[0141] Segmentation performance indicators U-Net (traditional method) Method of the present invention (LP-DT) Dice coefficient (average) 0.841 0.893 HD95 boundary error (mm) 6.02 3.68 <![CDATA[Small lesion detection rate (<3 cm 3 )]]> 73.5％ 91.2％ Number of mis-segmented regions (average) 2.3 0.7 Reasoning time (per image) 0.94 seconds 1.31 seconds

[0142] In addition, in one patient sample of the frontal lobe tumor, the traditional U-Net missed a long and thin lesion (1.2 cm in volume) behind the tumor close to the brain pool. 3 ), and the method of the present invention accurately segments the area through the dynamic fusion of multi-scale saliency guidance mechanism and dynamic Transformer. Doctors said that if this area is missed, it will directly affect the surgical margin.

[0143] It can be seen from this Example 1 that the method proposed in the present invention can effectively adapt to complex lesion morphology, solve the problems of blurred boundaries and misidentification, and is suitable for medical image segmentation tasks in areas with complex structures and dense tissues such as brain gliomas, liver cancer, and pancreatic cancer, and has extremely high clinical practical value.

[0144] refer to Figure 2 As can be seen, the background in the figure is an actual MRI scan image of the patient's brain with a resolution of 240×240. The grayscale distribution shows the characteristics of normal brain tissue imaging. The highlighted area in the figure represents the lesion area automatically identified by the present invention after fusing the multi-scale saliency-guided Laplacian pyramid with the dynamic Transformer module.

[0145] As shown in the figure, the right frontal-parietal junction area shows a patchy structure with continuous boundaries and enhanced signal intensity. The system automatically detects and segments the lesion area. The contour of the area fits the actual tumor enhancement edge, and the segmentation boundary is highly consistent with the manual annotation results of medical experts. There are no redundant false detection areas or background mis-segmentation, which fully demonstrates the effectiveness of the present invention in lesion boundary detection, fine-grained segmentation and morphology preservation.

[0146] The method of the present invention uses a saliency map to guide the Laplacian pyramid decomposition to enhance the response of the lesion area. Combined with the dynamic Transformer mechanism driven by image content, it realizes adaptive scale feature fusion of the input image and finally obtains the medical image segmentation output as shown in the figure.

[0147] Based on the traditional Laplacian pyramid, the present invention introduces edge response guidance and saliency response map control mechanism, uses guided Gaussian convolution to perform edge-sensitive Gaussian blur processing on the image, and performs pixel-level difference enhancement by fusing the saliency map when constructing the Laplacian pyramid layer. It not only improves the response intensity of the lesion area, but also suppresses the interference information of the background area. It can enhance the multi-scale high-frequency detail feature expression while maintaining the global structural semantics, and is suitable for the segmentation modeling of lesion areas in polymorphic, low-contrast medical images.

[0148] The present invention combines the image content-driven mechanism with the Transformer structure. By statistically analyzing the global distribution of features at each scale, it dynamically generates query, key, and value weight matrices, so that the Transformer's attention calculation no longer relies on static parameters, but is adjusted in real time with the image content. Combined with the content-aware gating mechanism and the saliency response map, the attention output is dynamically adjusted pixel by pixel in the spatial dimension, effectively improving the model's sensitivity to lesion boundaries, morphological changes, and heterogeneous regions. The dynamic Transformer module significantly enhances the segmentation model's adaptability to fine-grained structures in complex medical images, making up for the deficiency of the traditional attention mechanism in modeling the joint dynamic relationship of space and scale.

[0149] Based on the fused global feature representation, the present invention designs a step-by-step upsampling decoding network, combines the skip connection mechanism of the multi-scale feature map, supplements detail information in the decoding process, effectively avoids semantic ambiguity and edge information loss during feature upsampling, introduces a structured post-processing module in the output stage, uses morphological closing operations to enhance regional connectivity, uses edge repair filters to smooth segmentation contours, and performs regional connectivity detection to eliminate isolated artifacts, ultimately generating medical image segmentation results with clear boundaries, regional coherence, and accurate semantics. The structure improves the practicability of the model in clinical applications, and has significant advantages in processing images with blurred tumor boundaries and densely overlapping tissues.

[0150] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A medical image segmentation method based on Laplacian pyramid and dynamic Transformer, characterized in that: The steps include: S1. collecting an input medical image and performing preprocessing on the medical image to form a preprocessed medical image; S2. Perform multi-scale decomposition on the pre-processed medical image using a Laplacian pyramid, decomposing the pre-processed medical image into multiple image sub-layers of different scales; S3. Extract features from each scale image sub-layer separately, use a convolutional neural network to encode local features of each scale image sub-layer, and generate corresponding multi-scale feature representations; S4. Input the multi-scale feature representation into the dynamic Transformer module, dynamically weight the features of each scale through the self-attention mechanism to generate a fused global feature representation; S5. Decode the fused global feature representation and restore the fused global feature representation mapping to the same spatial resolution as the original medical image using upsampling, deconvolution, and skip connection techniques to generate a preliminary segmentation map. Apply post-processing algorithms to the preliminary segmentation map to perform morphological operations, edge repair, and noise removal to obtain a medical image segmentation result with clear boundaries and regional coherence.

2. The medical image segmentation method based on Laplacian pyramid and dynamic Transformer according to claim 1, characterized in that: The medical image is a CT image, an MRI image or an ultrasound image.

3. The medical image segmentation method based on Laplacian pyramid and dynamic Transformer according to claim 1, characterized in that: Said S1 comprises the following steps: S11. Obtain input medical image data I raw , Medical Image Data I raw It is the original medical image matrix of a two-dimensional or three-dimensional structure, containing multiple grayscale pixel value distributions, which are used to characterize the difference in medical image intensity between the lesion area and normal tissue: I raw ={i x,y ∣i x,y ∈[0,I max ],x∈[1,H],y∈[1,W]}; Among them, i x,y Represents the grayscale value of the pixel at position (x, y), H and W represent the height and width of the medical image respectively, I max The maximum grayscale value supported by the medical imaging system; S12. Using Gaussian filtering operation to process medical image data I raw The grayscale value of each pixel in the image is weighted averaged with its neighboring pixels, and the filtering weight is calculated based on the two-dimensional Gaussian distribution function to obtain a noise-suppressed medical image. S13. Normalizing the grayscale values of all pixels in the noise-suppressed medical image by the difference between the minimum grayscale value and the maximum grayscale value, and converting the image into a normalized medical image within the interval [0, 1]; S14. Using the linear grayscale transformation method, the grayscale value of each pixel in the normalized medical image is scaled and shifted by a constant to enhance the overall brightness and darkness contrast of the medical image, and the contrast-enhanced medical image I is obtained. pre .

4. The medical image segmentation method based on Laplacian pyramid and dynamic Transformer according to claim 3, characterized in that: The S2 comprises the following steps: S21. Contrast-enhanced medical image I pre As the initial input medical image, a multi-scale Gaussian medical image sequence {G0, G1, ..., G N }, where G0 = I pre , for each scale Gaussian pyramid image G k The next level of multi-scale Gaussian medical image G is generated by smoothing and downsampling the guided Gaussian convolution method of integrating the local edge gradient response of the medical image. k+1 : Among them, ω(u,v) is the normalized two-dimensional Gaussian kernel, r represents the radius of the Gaussian convolution kernel, u and υ represent the offset of the Gaussian convolution kernel in the horizontal and vertical directions respectively, G k is the Gaussian pyramid image of the kth layer, G k+1 (x, y) is the pixel value of the Gaussian pyramid image at the spatial position (x, y) in the k+1th layer, λ is the edge enhancement factor, E k (x, y) represents the edge gradient response intensity at the pixel position (x, y) in the current scale k medical image; S22. For each scale Gaussian pyramid image G k Instead of upsampling medical images Up(G k+1 ) performs a significant guided difference operation to obtain the high-frequency Laplacian medical image sublayer L k The saliency guided difference operation process is achieved by fusing the medical image saliency response map S k Reweighting: L k (x,y)=G k (x,y)-S k (x,y)·Up(G k+1 )(x,y); Significant response map S k (x,y)∈[0,1] represents the importance of the region at position (x,y) to the overall segmentation target; S23. Repeat steps S21 and S22 to construct a complete multi-scale Laplacian pyramid medical image set Among them, L N-1 is the N-1th high-frequency Laplacian medical image sublayer, G N is the lowest level Gaussian medical image.

5. The medical image segmentation method based on Laplacian pyramid and dynamic Transformer according to claim 4, characterized in that: The S3 includes the following steps: S31. Multi-scale Laplacian Pyramid Medical Image Collection Each scale medical image sublayer L k and the bottom global medical image G N Perform feature encoding separately and use scale-sharing convolutional encoding module F conv Extract the local feature response of each medical image sub-layer and obtain the initial encoding feature map F at each scale k : Among them, F k (x,y) represents the feature response value of the initial encoding feature map at position (x,y) at scale k, represents the kth Laplacian medical image sublayer or the bottom global medical image of the input medical image channel i, C in is the number of input channels, represents the weight parameter of the convolution kernel of scale k at position (u, v) under channel i, b k is the bias term, σ represents the nonlinear activation function; S32. The initial encoding feature map F at each scale k The corresponding medical image saliency response map S k Perform pixel-by-pixel weighted fusion to generate a saliency enhanced feature map Saliency Enhancement Feature Map The enhancement value of each pixel in the initial encoding feature map F is the pixel k The original value and its corresponding position in the medical image saliency response map S k The sum of the products of the response values in , plus the original values themselves scaled; S33. Saliency Enhanced Feature Maps at All Scales Perform dimension unification and structural adjustment, and adjust the feature maps of different scales to the preset unified resolution (H) through channel transformation and spatial alignment operations. ′ ,W ′ ) and the uniform channel number C ′ , forming a standardized multi-scale feature representation set F multi : Among them, φ(·) represents the composite transformation operation of upsampling, channel mapping and feature normalization, H k With W k is the spatial dimension of the original scale k, C k is the number of channels, (H ′ ,W ′ ,C ′ ) is the target space dimension and number of channels for uniform alignment.

6. The medical image segmentation method based on Laplacian pyramid and dynamic Transformer according to claim 5, characterized in that: The S4 comprises the following steps: S41. Input the standardized multi-scale feature representation set into the dynamic Transformer module, and enhance the saliency feature map at each scale Flatten the input feature sequence into a two-dimensional sequence representation Among them H ′ W ′ is the sequence length, C ′ is the channel dimension for each position; S42. The dynamic parameter generation module driven by image content is used to generate the input feature sequence X at each scale. k Perform encoding processing on the input feature sequence X k Perform global average pooling operation to obtain the statistical representation vector at scale k The statistical representation vector μ k Input into three multi-layer perception networks with shared structure but independent parameters to generate the query weight matrix of Transformer at this scale Bond weight matrix and the value weight matrix S43. Based on query weight matrix Bond weight matrix and the value weight matrix For the input feature sequence X k Perform matrix multiplication to obtain the query matrix Q at scale k k , key matrix K k Sum matrix V k ; S44. Calculate the attention weight matrix within the scale Use scaled dot product attention mechanism for attention allocation; S45. Introduce content-aware gating mechanism to dynamically adjust attention distribution at pixel level, according to the input medical image saliency response map S k , construct pixel attention gating factor G k , and embed the pixel attention gating factor into the attention weighted output to obtain the dynamic attention output sequence Z k : With k =(G k ⊙A k )·V k ; where ⊙ represents the element-wise multiplication with row broadcasting, G k (i) = 1 + γ·S k (i) is used to enhance the dynamic response intensity of the pixels in the lesion area in the attention output, where γ is the weight coefficient; S46. Dynamic attention output sequence Z at all scales k Perform scale-level fusion and introduce global adaptive fusion weight α k The global adaptive fusion weight is generated by the statistical characteristics of the attention output at each scale, and the global adaptive fusion weight is dynamically adjusted according to the response intensity of the lesion area at each scale: Among them, Z k represents the dynamic attention output sequence obtained after processing by the dynamic Transformer module at scale k, Z j represents the dynamic attention output sequence obtained after processing by the dynamic Transformer module at scale j, mean(Z k ) represents the dynamic attention output sequence Z k The global average pooling result, that is, the average of the feature responses of all spatial positions, is used to extract the main semantic features expressed by scale k in the overall image. w is the weight vector in the adaptive fusion function, and b is the bias term in the adaptive fusion function. S47. According to the global adaptive fusion weight α k Perform weighted summation on the dynamic outputs of all scales to obtain the fused global feature representation F global .

7. The medical image segmentation method based on Laplacian pyramid and dynamic Transformer according to claim 6, characterized in that: The S5 comprises the following steps: S51. The fused global feature is represented by F global Input to the decoding module, the fused global feature representation F global The size is H ′ ×W ′ ×d three-dimensional tensor, the decoding module gradually restores the spatial resolution through layer-by-layer upsampling and deconvolution operations, constructing a sequence of decoding feature maps with increasing resolution layer by layer until the spatial resolution H×W is restored to the same as the original medical image. The number of channels and spatial size of each layer of decoding feature maps are adjusted in sequence according to the number of decoding layers to obtain a decoding output consistent with the spatial resolution of the original medical image; S52. Introduce the skip connection mechanism in the decoding process to enhance the multi-scale saliency feature map The feature map with the same spatial resolution as the output resolution of the current decoding layer is spatially aligned through upsampling or convolution operations, and then spliced with the current decoding feature map in the channel dimension to obtain a fused feature map; S53. Perform a channel mapping operation on the last layer of decoded feature maps, using a convolution kernel whose number of channels is equal to the number of medical image segmentation categories to map the final decoded feature maps into a preliminary medical image segmentation map. The spatial size of the preliminary medical image segmentation map is H×W, and the number of output channels of each pixel is equal to the number of segmentation categories, indicating the probability distribution of the pixel belonging to each category. The probability distribution is processed by a Softmax normalization function so that the output of each pixel satisfies the category sum to 1. S54. Performing post-processing operations on the preliminary medical image segmentation map, the post-processing operations including morphological operations, edge repair operations, and noise removal operations; S55. Output the post-processed medical image segmentation result as the final medical image segmentation result, wherein the final medical image segmentation result is a two-dimensional label map of size H×W, in which each pixel corresponds to a unique medical image segmentation category.

8. The medical image segmentation method based on Laplacian pyramid and dynamic Transformer according to claim 7, characterized in that: In the morphological operation, a closing operation and an opening operation are combined with a structural element to perform dilation and then erosion operations. The closing operation is used to fill small holes in the lesion area, and the opening operation is used to remove small isolated artifacts. In the edge repair operation, Gaussian filtering and curvature smoothing methods are used to adjust the edge of the segmented area to address the jagged, discontinuous or overly sharp boundary problems existing in the preliminary medical image segmentation map; In the noise removal operation, false detection areas and isolated pixels in the background are automatically identified and removed through regional connectivity analysis and shape constraint judgment.

Images