Medical image enhancement processing and tissue boundary intelligent identification method

Abstract

The invention relates to the field of medical image processing, and discloses a medical image enhancement processing and tissue boundary intelligent identification method, and the method comprises the steps: constructing a medical image segmentation model, and enabling a learnable filter module to comprise a Gabor filter group and an LoG filter group which are parallel; the multi-modal feature fusion module fuses the texture features output by the plurality of Gabor filters, the edge features output by the plurality of LoG filters and the original input image; the Swin Transform is used for dividing the fusion feature map into an image block sequence and extracting a multi-scale feature map from the image block sequence; the segmentation head is used for recovering the multi-scale feature map to the original resolution and outputting a segmentation mask; training the constructed medical image segmentation model; and performing image segmentation processing on a clinical medical image by using the trained medical image segmentation model, and outputting pixel-level segmentation masks capable of clearly displaying different tissue structures and tissue boundaries. The method can effectively improve the recognition precision of the tissue boundary in the medical image.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and more specifically, to a method for medical image enhancement processing and intelligent recognition of tissue boundaries. Background Art

[0002] Medical image semantic segmentation is crucial for precision medicine. For example, in radiotherapy and surgical planning, its accuracy directly affects the treatment effect of patients. Although traditional convolutional neural networks (CNNs) have been applied in this field, due to the locality of convolutional operations, it is difficult to effectively model global context information. In recent years, Vision Transformer (ViT) has shown great potential in medical image segmentation tasks due to its ability to capture long-range dependencies. However, when existing ViT models process small datasets without large-scale pre-training, their ability to model local spatial information is insufficient, making it difficult to accurately segment the boundaries of fine anatomical structures, such as the boundaries of small organs, small tumors, or blood vessels. Summary of the Invention

[0003] The objective of the present invention is to propose a method for medical image enhancement processing and intelligent recognition of tissue boundaries, so as to effectively improve the recognition accuracy of tissue boundaries in medical images.

[0004] To achieve the above objective, the present invention proposes a method for medical image enhancement processing and intelligent recognition of tissue boundaries, including: Constructing a medical image segmentation model, where the medical image segmentation model includes a learnable filter module, a multi-modal feature fusion module, a Swin Transformer backbone network, and a segmentation head connected in sequence; the learnable filter module includes a parallel Gabor filter bank and LoG filter bank, the Gabor filter bank is used to extract multi-scale and multi-directional texture features of different tissues in the input image, and the LoG filter bank is used to extract multi-scale edge features of different tissues in the input image; the multi-modal feature fusion module is used to fuse the texture features output by multiple Gabor filters, the edge features output by multiple LoG filters, and the original input image to generate a fused feature map with enhanced texture and edge features; the Swin Transformer is used to divide the fused feature map into a sequence of image patches and extract multi-scale feature maps from the sequence of image patches; the segmentation head is used to restore the multi-scale feature maps to the original resolution and output segmentation masks; Training the constructed medical image segmentation model with training data; the training data includes medical images and segmentation masks corresponding to them with pixel-by-pixel labeled tissue categories; Use the trained medical image segmentation model to perform image segmentation on clinical medical images, and output a pixel-level segmentation mask that can clearly show different tissue structures and tissue boundaries.

[0005] Optionally, the Gabor filter bank includes a plurality of parallel Gabor filters, each Gabor filter has independent trainable parameters, and the trainable parameters of the plurality of Gabor filters are different from each other; During the training process, a plurality of parallel learnable Gabor filters dynamically learn the multi-scale and multi-directional texture features of different tissues in the input image.

[0006] Optionally, the function of the Gabor filter is defined as: , Where: λ, θ , ψ, σ and γ are trainable parameters; λ is the sine wavelength, which is used to control the texture scale; θ is the orientation angle, which is used to set the texture detection direction; ψ is the phase offset, which is used to adjust the sine wave phase; σ is the standard deviation of the Gaussian envelope, which is used to control the receptive field size; γ is the spatial aspect ratio, which is used to control the ellipticity of the filter; x θ = x cos θ + y sin θ , y θ =− x sin θ + y cos θ , x , y is the pixel coordinate in the original image, x θ , y θ are the rotated x and y coordinates; Gabor( x θ , y θ ) represents the Gabor filter response value calculated in the rotated local coordinate system ( x θ , y θ ), which is used to quantify the texture intensity of the image at a specific direction θ and scale λ.

[0007] Optionally, the LoG filter bank includes a plurality of parallel LoG filters, each LoG filter having independent trainable parameters, and the trainable parameters of the plurality of LoG filters are different from each other; During the training process, the plurality of learnable LoG filters in parallel dynamically learn the multi-scale edge features of different tissues in the input image.

[0008] Optionally, the function of the LoG filter is defined as: , where: σ is a trainable parameter representing the Gaussian smoothing scale, which is used to control the sensitivity of the filter to edge detection at different scales; x , y are the pixel coordinates in the original image; LoG ( x , y ) represents the LoG filter response value calculated in the original coordinate system ( x , y ), which is used to quantify the edge intensity of the image at a specific scale σ .

[0009] Optionally, the Swin Transformer backbone network includes a convolutional embedding module and an encoder and a decoder with a symmetric structure; The convolutional embedding module is used to divide the fused feature map into image patches, convert each image patch into an embedding vector, and generate a feature map; The encoder is used to perform multi-stage downsampling and multi-scale global feature extraction on the feature map; The decoder is used to restore the spatial resolution by multi-stage upsampling the feature map, and fuse the multi-scale features of the encoder through skip connections to accurately locate the anatomical structure.

[0010] Optionally, the operations of the convolutional embedding module specifically include: Dividing the fused feature map into non-overlapping 4×4 image patches through a 4×4 convolutional layer with a stride of 4; Converting each image patch into an embedding vector through linear projection; Arranging the embedding vectors of all image patches into a 2D grid according to the spatial position of the original image to form a feature map.

[0011] Optionally, the operations of the encoder specifically include: The first stage: First, process the feature map through a plurality of Swin Transformer blocks, then perform an image patch merging operation, and halve the resolution and double the number of channels through a 2×2 convolutional layer with a stride of 2; Second stage: First, process the feature map output in the first stage through multiple Swin Transformer blocks, then perform an image patch merging operation, halve the resolution by a 2×2 convolution with a stride of 2, and double the number of channels; Third stage: First, process the feature map output in the second stage through multiple Swin Transformer blocks, then perform an image patch merging operation, halve the resolution by a 2×2 convolution with a stride of 2, and double the number of channels; At the end of the encoder, use multiple Swin Transformer blocks to further extract deep features from the feature map output in the third stage; Among them, the processing of each Swin Transformer block includes: dividing the feature map into windows including multiple image patches, calculating the multi-head self-attention within the window to capture local context, and enhancing the communication between windows by using shifted window multi-head self-attention; generating context-aware features through feature transformation.

[0012] Optionally, the operations of the decoder specifically include: Third stage: First, perform an upsampling operation on the feature map output at the end of the encoder to double the resolution, splice the feature map output in the third stage of the encoder with the upsampled feature map through a skip connection, and then process the spliced feature map through multiple Swin Transformer blocks; Second stage: First, perform an upsampling operation on the feature map output in the third stage of the decoder to double the resolution, splice the feature map output in the second stage of the encoder with the upsampled feature map through a skip connection, and then process the spliced feature map through multiple Swin Transformer blocks; First stage: First, perform an upsampling operation on the feature map output in the second stage of the decoder to double the resolution, splice the feature map output in the first stage of the encoder with the upsampled feature map through a skip connection, and then process the spliced feature map through multiple Swin Transformer blocks.

[0013] Optionally, the operations of the segmentation head include: Restore the feature map to the original size through upsampling; Map the number of channels to the number of classes through linear projection to generate a segmentation probability map; Use the Softmax activation function to calculate the class probability for each pixel point; Output a segmentation mask.

[0014] The beneficial effects of the present invention are as follows: The medical image segmentation model constructed in the present invention innovatively introduces learnable Gabor filters and LoG filters into the Transformer architecture. The two work together. The Gabor filter is responsible for extracting multi-scale and multi-directional texture information of different tissues in the image, while the LoG filter focuses on edge detection and boundary delineation and is responsible for extracting multi-scale edge features of different tissues in the image, enabling the model to perform excellently in capturing local features. The filters are responsible for finely depicting local textures and boundary details, and the Swin Transformer captures the global dependencies between image patches through the self-attention mechanism to generate feature representations containing global context information and understand the image structure from a global perspective. The two complement each other, enabling the model to take into account both the whole and the local when dealing with complex medical image segmentation tasks, thus accurately identifying the boundaries and structures of various tissue organs and making up for the deficiencies of traditional Transformer models in this regard. At the same time, the parameters of the learnable Gabor filters and LoG filters are dynamically adjusted during the model training process, enabling the model to automatically optimize the filter parameters according to the characteristics of different medical images to better adapt to various types of image data, greatly enhancing the generality and adaptability of the model to different medical image data and enhancing its practical value.

[0015] The system of the present invention has other characteristics and advantages that will be obvious from the accompanying drawings incorporated herein and the subsequent detailed description, or will be described in detail in the accompanying drawings incorporated herein and the subsequent detailed description, which together are used to explain the specific principles of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] By describing the exemplary embodiments of the present invention in more detail in conjunction with the accompanying drawings, the above and other objects, features, and advantages of the present invention will become more obvious. In the exemplary embodiments of the present invention, the same reference numerals generally represent the same components.

[0017] Figure 1 FIG. shows a schematic structural diagram of a medical image segmentation model in a medical image enhancement processing and tissue boundary intelligent recognition method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] The present invention will be described in more detail below with reference to the accompanying drawings. Although the preferred embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0019] As Figure 1As shown in the figure, this embodiment provides a method for medical image enhancement processing and intelligent tissue boundary recognition, including: S1: Construct a medical image segmentation model, which includes a learnable filter module, a multi-modal feature fusion module, a Swin Transformer backbone network, and a segmentation head connected in sequence; the learnable filter module includes a parallel Gabor filter bank and a LoG (Laplacian of Gaussian) filter bank, the Gabor filter bank is used to extract multi-scale and multi-directional texture features of different tissues in the input image, and the LoG filter bank is used to extract multi-scale edge features of different tissues in the input image; the multi-modal feature fusion module is used to fuse the texture features output by multiple Gabor filters, the edge features output by multiple LoG filters, and the original input image to generate a fused feature map with enhanced texture and edge features; the Swin Transformer is used to divide the fused feature map into a sequence of image patches (Patch) and extract multi-scale feature maps from the sequence of image patches (Patch); the segmentation head is used to restore the multi-scale feature map to the original resolution and output a segmentation mask; In this embodiment, the Gabor filter bank includes a plurality of parallel Gabor filters, each Gabor filter has independent trainable parameters, and the trainable parameters of the plurality of Gabor filters are different from each other; During the training process, a plurality of parallel learnable Gabor filters dynamically learn the multi-scale and multi-directional texture features of different tissues in the input image and dynamically optimize the parameters through backpropagation.

[0020] The function definition of the Gabor filter is: , where: λ, θ , ψ, σ and γ are trainable parameters; λ is the sine wavelength, which is used to control the texture scale; θ is the orientation angle, which is used to set the texture detection direction; ψ is the phase offset, which is used to adjust the sine wave phase; σ is the standard deviation of the Gaussian envelope, which is used to control the receptive field size; γ is the spatial aspect ratio, which is used to control the ellipticity of the filter; x θ = x cos θ + y sin θ ,y θ = - x sin θ + y cos θ , x , y are the pixel coordinates in the original image, x θ 、 y θ are the rotated x and y coordinates; Gabor( x θ , y θ ) represents the Gabor filter response value calculated in the rotated local coordinate system ( x θ , y θ ) and is used to quantify the texture intensity of the image at a specific direction θ and scale λ.

[0021] In this embodiment, the LoG filter bank includes a plurality of parallel LoG filters, each LoG filter has independent trainable parameters, and the trainable parameters of the plurality of LoG filters are different; During the training process, the plurality of learnable parallel LoG filters dynamically learn the multi-scale edge features of different tissues in the input image and dynamically optimize the parameters through backpropagation.

[0022] The function definition of the LoG filter is: , where: σ is a trainable parameter representing the Gaussian smoothing scale, which is used to control the sensitivity of the filter to edge detection at different scales; x , y are the pixel coordinates in the original image; LoG ( x , y ) represents the LoG filter response value calculated in the original coordinate system ( x , y ) and is used to quantify the edge intensity of the image at a specific scale σ .

[0023] In this embodiment, the Swin Transformer backbone network includes a convolutional embedding module and an encoder and a decoder with a symmetric structure; The convolutional embedding module is used to divide the fused feature map into image patches and convert each image patch into an embedding vector, and generate a feature map; The encoder is used to perform multi-stage downsampling and multi-scale global feature extraction on the feature map; The decoder is used to restore the spatial resolution by multi-stage upsampling the feature map, and fuse the multi-scale features of the encoder through skip connections to accurately locate the anatomical structure.

[0024] Among them, the operations of the convolutional embedding module specifically include: Divide the fused feature map into non-overlapping 4×4 image patches (Patches) through a 4×4 convolutional layer with a stride of 4; Convert each image patch (Patch) into an embedding vector through linear projection; Arrange the embedding vectors of all image patches (Patches) in a 2D grid according to the spatial position of the original image to form a feature map.

[0025] The operations of the encoder specifically include: The first stage: First, process the feature map through multiple Swin Transformer blocks (CST Blocks), then perform a patch merging operation, and halve the resolution and double the number of channels through a 2×2 convolutional layer with a stride of 2; The second stage: First, process the feature map output in the first stage through multiple Swin Transformer blocks (CST Blocks), then perform a patch merging operation, and halve the resolution and double the number of channels through a 2×2 convolutional layer with a stride of 2; The third stage: First, process the feature map output in the second stage through multiple Swin Transformer blocks (CST Blocks), then perform a patch merging operation, and halve the resolution and double the number of channels through a 2×2 convolutional layer with a stride of 2; At the end of the encoder, multiple Swin Transformer blocks (CST Blocks) are used to further extract deep features from the feature map output in the third stage; Among them, the processing of each Swin Transformer block (CST Block) includes: Divide the feature map into windows including multiple image patches, calculate the multi-head self-attention within the windows to capture local context, and use shifted window multi-head self-attention to enhance communication between windows; Generate context-aware features through feature transformation (Layer Norm, MLP, etc.).

[0026] Specifically, the processing of data by the Swin Transformer Block (CST Block) includes: (1) Window partitioning: The feature map is partitioned into non-overlapping windows of a fixed size. For example, the feature map is usually partitioned into multiple windows of size 7×7 or 14×14. Self-attention calculations are performed among the feature points within each window. Such window partitioning makes the calculation more efficient while retaining local spatial information; (2) Shifted window strategy (SW-MSA): In specific Transformer blocks, the shifted window method is adopted. Compared with the standard window partitioning, the shifted window first performs a shift operation on the feature map by half of the window size and then conducts window partitioning. This enables the feature points between adjacent windows to interact in different Transformer blocks, increasing the receptive field and helping to capture richer context information; (3) Multi-head self-attention mechanism within the window (W-MSA): Within each window, the multi-head self-attention mechanism is used to model the mutual relationships among the feature points. The feature vectors within each window are divided into multiple heads, and each head independently performs self-attention calculations. By extracting and aggregating features from different heads, the mutual dependencies in different feature subspaces can be captured, enhancing the feature expression ability of the model; The feature vector of each window undergoes a non-linear transformation through a multi-layer perceptron (MLP). The MLP usually consists of two fully connected layers with the GELU activation function in the middle, which is used to introduce non-linearity and enhance the expression ability and fitting ability of the model. In the Swin Transformer Block, residual connections are used to add the input features and the output features. Such residual connections help to alleviate the vanishing gradient problem in deep networks, promote the flow of information, improve the training efficiency and model performance. In each Swin Transformer Block, layer normalization is used to normalize the features. This helps to stabilize the training process of the model, accelerate convergence, and improve the model performance. Through operations such as W-MSA / SW-MSA and MLP mentioned above, the Swin Transformer Block gradually extracts and models the local and global features of the image, thereby achieving efficient feature extraction and pattern recognition of image data. This structure can effectively capture the long-range dependencies and rich context information in the image when dealing with medical image segmentation tasks, laying a foundation for the high-precision segmentation performance of the model.

[0027] The operations of the decoder specifically include: The third stage: First, perform upsampling on the feature map output at the end of the encoder to double the resolution. Concatenate the feature map output in the third stage of the encoder with the upsampled feature map through skip connections, and then process the concatenated feature map through multiple Swin Transformer blocks (CST Block); The second stage: First, perform upsampling on the feature map output in the third stage of the decoder to double the resolution. Concatenate the feature map output in the second stage of the encoder with the upsampled feature map through skip connections, and then process the concatenated feature map through multiple Swin Transformer blocks (CST Block); The first stage: First, perform upsampling on the feature map output in the second stage of the decoder to double the resolution. Concatenate the feature map output in the first stage of the encoder with the upsampled feature map through skip connections, and then process the concatenated feature map through multiple Swin Transformer blocks (CST Block).

[0028] The operations of the segmentation head include: Restore the feature map to the original size through upsampling; Map the number of channels to the number of classes through a 1×1 convolutional layer for linear projection to generate a segmentation probability map; Adopt a Softmax activation function to calculate the class probability for each pixel point; Output a segmentation mask.

[0029] In an example, the data processing flow of the medical image segmentation model is as follows: 1. Input layer Input data: Original medical images (such as CT / MRI), with a size of H (Height) × W (Width) × C (Number of Channels) (by default 224×224×3, a single-channel grayscale image); Preprocessing: Normalization: Adjust according to the modality (such as HU value truncation for CT and z-score normalization for MRI); Size alignment: Ensure that the image and label sizes are consistent (such as 224×224); Output: The normalized image tensor I ∈R 224×224×3 .

[0030] 2. Learnable Filters (including a parallel Gabor filter bank and a LoG filter bank) (1) Gabor Filtering Path (N parallel Gabor filters) Operations: 1) Dynamic Parameter Generation: The parameters (θ, σ, λ, ψ, γ) of each Gabor filter are learned through backpropagation, with initial values based on traditional Gabor settings; 2) Convolution Operation: The input image is convolved with multiple Gabor filters respectively to generate multi-scale and multi-directional texture features; 3) Output: N Gabor feature maps F Gabor ∈R 224×224×N .

[0031] (2) LoG Filtering Path (M LoG filters) Operations: 1) Dynamic σ Learning: The σ parameter of each LoG filter is optimized through training, with the initial value based on experience (e.g., σ = 1.0); 2) Convolution Operation: The input image is convolved with the LoG filter to extract edge and boundary features; Output: M LoG feature maps F LoG ∈R 224×224×M .

[0032] 3. Multimodal Feature Fusion (Concat) Input: Original image I ∈R 224×224×3 ; Gabor features F Gabor ∈R 224×224×N ; LoG features F LoG ∈R 224×224×M ; Operations: Channel Concatenation: Concatenate the three groups of features along the channel dimension to obtain F fused ∈R 224×224×(3+N+M) , for example, using 5 Gabor + 5 LoG filters, the number of channels after fusion is 13; Purpose: Combine the original image with radiomics features to enhance the ability to express local details.

[0033] 4. Convolutional Patch Embedding Input: fused features F fused ∈R 224×224×(3+N+M) ; Operation: (1) Convolutional block: Using a 4×4 convolutional kernel with a stride of 4, the image is divided into 56×56 non-overlapping patches, each patch being 4×4 in size. Each convolutional kernel uses the GELU activation function, and a normalization layer (LN, Layer Normalization) is added to the second convolutional kernel; (2) Channel mapping: The number of output channels is adjusted to a fixed dimension D embed = 96, obtaining the embedding vector E patch ∈R 56 ×56×96 ; Purpose: To retain spatial locality and provide structured input for the Transformer.

[0034] 5. Swin Transformer Encoder The encoder consists of 4 stages, gradually downsampling and extracting multi-scale global features: (1) Stage 1 Input: E patch ∈R 56×56×96 ; Operation: Swin Transformer Block (CST Block) ×2: Window partitioning: The 56×56 feature is divided into 7×7 windows (each window is an 8×8 patch); Window multi-head self-attention (W-MSA): Calculate the multi-head self-attention within the window to capture local context; Shifted window (SW-MSA): Shift the window position in the next layer to promote cross-window information interaction; Output: Remain 56×56×96; Patch Merging: Halve the resolution and double the number of channels through a 2×2 convolution (stride 2); For example, input 56×56×96 → output 28×28×192.

[0035] (2) Stage 2 - 3 Input: The output of the previous stage; Operation: Swin Transformer Block ×2: Same as in Stage 1, the window size is reduced (e.g., the window in Stage 2 is 4×4); Patch Merging: Halve the resolution and double the number of channels through a 2×2 convolution (stride 2); Output: Stage 2: 14×14×384; Stage 3: 7×7×768.

[0036] 6. Swin Transformer Decoder The decoder restores details through upsampling and skip connections: (1) Stage 3 Input: The output of the encoder's Stage 3, 7×7×768; Operations: Up sampling: Upsample by a factor of 2 using a transposed convolution (kernel 2×2, stride 2); Output 14×14×384; Skip connection: Concatenate with the output of the encoder's Stage 3, 14×14×384, and compress the number of channels to 384; Swin Transformer Block ×2: Process the fused features; Output: 14×14×384.

[0037] (2) Stage 3 → Stage 2 Input: The output of the decoder's Stage 3, 14×14×384; Operations: Upsample to 28×28×192, concatenate with the features of the encoder's Stage 2, and process as above; Output: 28×28×192; (3) Stage 2 → Stage 1 Input: The output of the decoder's Stage 2, 28×28×192; Operations: Upsample to 56×56×96, concatenate with the features of the encoder's Stage 1, and process as above; Output: 56×56×96.

[0038] 7. Output Layer (Segmentation Head) Input: The final output of the decoder, 56×56×96; Operations: Final upsampling (Patch expanding): Restore the decoder output feature map from 56×56×96 to the original size of 224×224×96 through bilinear interpolation or transposed convolution; Convolutional prediction: 1×1 convolution: Map the number of channels to the number of classes K , generating a segmentation probability map P ∈R 224×224×K ; Softmax activation: Calculate the class probability for each pixel; Output: Segmentation mask Y ∈R 224×224×K , and take argmax to obtain the final class label.

[0039] S2: Use the training data to train the constructed medical image segmentation model; the training data includes medical images and the corresponding segmentation masks that label tissue classes pixel by pixel; Specifically, existing datasets can be used to train the model, such as the Synapse multi-organ segmentation dataset, the ACDC cardiac segmentation dataset, etc. The model starts training from randomly initialized weights, for example, with a training period of 300 epochs. The AdamW optimizer is used to optimize the model. This optimizer combines the adaptive learning rate adjustment of Adam and the weight decay technique, which helps prevent the model from overfitting. During the training process, different learning rates can be set for different datasets. Different numbers of filters can be used according to the datasets. For example, for the ACDC dataset, 5 Gabor filters and 5 LoG filters can be trained; while for the Synapse dataset, 2 Gabor filters and 5 LoG filters are used. The parameters of these filters are learnable during the training process and can be dynamically adjusted according to the characteristics of the data to optimize the effect of feature extraction. Through training, the model can learn the feature representations of tissues and organs in medical images, including local texture, edges, boundary information, and global context relationships, so as to achieve accurate segmentation of medical images and accurately identify the boundaries of different tissues and organs. During the training process, by monitoring the performance of the model on the validation set, such as evaluation metrics like the Dice similarity coefficient (DSC) and the 95th percentile Hausdorff distance (HD95), the performance of the model is evaluated, and the model parameters and training strategies are adjusted as needed to optimize the segmentation effect of the model.

[0040] S3: Use the trained medical image segmentation model to perform image segmentation on clinical medical images, and output a pixel-level segmentation mask that can clearly show different tissue structures and tissue boundaries.

[0041] In summary, this method combines learnable Gabor and LoG filters with the Transformer architecture, effectively addressing the problem of insufficient local feature capture in traditional Vision Transformers for medical image segmentation. By introducing the Gabor filter to extract multi-scale and multi-directional texture information, and the LoG filter to enhance edge detection and boundary delineation accuracy, the model can capture rich local detail features in medical images. At the same time, leveraging the powerful global context modeling ability of the Transformer, it can integrate and understand the global information of the image. This organic combination of local and global feature extraction enables the model to accurately identify the boundaries and structures of various tissues and organs when dealing with complex medical image segmentation tasks. Even in small dataset scenarios, with its lightweight design and dynamically adjustable filter parameters, it can achieve high-precision segmentation results with low computational resource consumption, demonstrating strong adaptability and practicality, providing an efficient, accurate, and applicable solution for various segmentation tasks in the field of medical image segmentation.

[0042] The embodiments of the present invention have been described above. The above description is exemplary and not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations are obvious to those of ordinary skill in the art in the technical field without departing from the scope and spirit of the described embodiments.

Claims

1. A method for medical image enhancement processing and intelligent recognition of tissue boundaries, characterized in that, Including: Construct a medical image segmentation model, which includes a learnable filter module, a multi-modal feature fusion module, a Swin Transformer backbone network, and a segmentation head connected in sequence; the learnable filter module includes a parallel Gabor filter bank and a LoG filter bank. The Gabor filter bank is used to extract multi-scale and multi-directional texture features of different tissues in the input image, and the LoG filter bank is used to extract multi-scale edge features of different tissues in the input image; the multi-modal feature fusion module is used to fuse the texture features output by multiple Gabor filters, the edge features output by multiple LoG filters, and the original input image to generate a fused feature map with enhanced texture and edge features; the Swin Transformer is used to divide the fused feature map into a sequence of image patches and extract multi-scale feature maps from the sequence of image patches; the segmentation head is used to restore the multi-scale feature map to the original resolution and output a segmentation mask; Use the training data to train the constructed medical image segmentation model; the training data includes medical images and segmentation masks corresponding to them with pixel-level labeled tissue categories; Use the trained medical image segmentation model to perform image segmentation processing on clinical medical images, and output a pixel-level segmentation mask that can clearly show different tissue structures and tissue boundaries.

2. The method according to claim 1, wherein The Gabor filter bank includes multiple parallel Gabor filters, each Gabor filter has independent trainable parameters, and the trainable parameters of multiple Gabor filters are different from each other; During the training process, multiple parallel learnable Gabor filters dynamically learn the multi-scale and multi-directional texture features of different tissues in the input image, and dynamically optimize the parameters through backpropagation.

3. The method according to claim 2, wherein The function definition of the Gabor filter is: ， Wherein: λ, θ , ψ, σ and γ are trainable parameters; λ is the sine wavelength, used to control the texture scale; θ is the orientation angle, used to set the texture detection direction; ψ is the phase offset, used to adjust the sine wave phase; σ is the standard deviation of the Gaussian envelope, used to control the receptive field size; γ is the spatial aspect ratio, used to control the ellipticity of the filter; x θ = x cos θ + y sin θ , y θ =− x sin θ + y cos θ , x , y are the pixel coordinates in the original image, x θ 、 y θ are the rotated x and y coordinates; Gabor( x θ , y θ ) represents the Gabor filter response value calculated in the rotated local coordinate system ( x θ , y θ ), used to quantify the texture intensity of the image at a specific direction θ and scale λ.

4. The method according to claim 1, wherein The LoG filter bank includes multiple parallel LoG filters, each LoG filter has independent trainable parameters, and the trainable parameters of multiple LoG filters are different from each other; During the training process, multiple parallel learnable LoG filters dynamically learn the multi-scale edge features of different tissues in the input image, and dynamically optimize the parameters through backpropagation.

5. The method according to claim 4, characterized in that, The function definition of the LoG filter is: ， Wherein: σ is a trainable parameter representing the Gaussian smoothing scale, which is used to control the sensitivity of the filter to edge detection at different scales; x , y is the pixel coordinate in the original image; LoG ( x , y ) represents the LoG filter response value calculated in the original coordinate system ( x , y ), which is used to quantify the edge intensity of the image at a specific scale σ .

6. The method according to claim 1, wherein The Swin Transformer backbone network includes a convolutional embedding module and symmetric encoder and decoder structures; The convolutional embedding module is used to divide the fused feature map into image patches and convert each image patch into an embedding vector, and generate a feature map; The encoder is used to perform multi-stage downsampling and multi-scale global feature extraction on the feature map; The decoder is used to restore the spatial resolution by multi-stage upsampling the feature map, and fuse the multi-scale features of the encoder through skip connections to accurately locate anatomical structures.

7. The method according to claim 6, wherein The operations of the convolutional embedding module specifically include: Divide the fused feature map into non-overlapping 4×4 image patches through a 4×4 convolutional layer with a stride of 4; Convert each image patch into an embedding vector through linear projection; Arrange the embedding vectors of all image patches into a 2D grid according to the spatial positions of the original image to form a feature map.

8. The method according to claim 6, wherein The operations of the encoder specifically include: The first stage: First, process the feature map through multiple Swin Transformer blocks, then perform an image patch merging operation, and halve the resolution and double the number of channels through a 2×2 convolution with a stride of 2; The second stage: First, process the feature map output by the first stage through multiple Swin Transformer blocks, then perform an image patch merging operation, and halve the resolution and double the number of channels through a 2×2 convolution with a stride of 2; The third stage: First, process the feature map output by the second stage through multiple Swin Transformer blocks, then perform an image patch merging operation, and halve the resolution and double the number of channels through a 2×2 convolution with a stride of 2; At the end of the encoder, use multiple Swin Transformer blocks to further extract deep features from the feature map output by the third stage; Among them, the processing of each Swin Transformer block includes: dividing the feature map into windows including multiple image patches, calculating the multi-head self-attention within the window to capture local context, and using shifted window multi-head self-attention to enhance communication between windows; generating context-aware features through feature transformation.

9. The method according to claim 8, wherein The operations of the decoder specifically include: The third stage: First, perform an upsampling operation on the feature map output at the end of the encoder to double the resolution, splice the feature map output by the third stage of the encoder with the upsampled feature map through a skip connection, and then process the spliced feature map through multiple Swin Transformer blocks; The second stage: First, perform an upsampling operation on the feature map output by the third stage of the decoder to double the resolution, splice the feature map output by the second stage of the encoder with the upsampled feature map through a skip connection, and then process the spliced feature map through multiple Swin Transformer blocks; The first stage: First, perform an upsampling operation on the feature map output by the second stage of the decoder to double the resolution, splice the feature map output by the first stage of the encoder with the upsampled feature map through a skip connection, and then process the spliced feature map through multiple Swin Transformer blocks.

10. The method according to claim 9, wherein The operations of the segmentation head include: Restore the feature map to the original size through upsampling; Map the number of channels to the number of classes through linear projection to generate a segmentation probability map; Use the Softmax activation function to calculate the class probability for each pixel point; Output the segmentation mask.

Images