A parameter frozen fuzzy attention method for 3D abdominal organ segmentation
By employing parameter freezing and fuzzy attention methods, the problems of high training resource consumption and blurred boundaries of large models were solved, achieving efficient and accurate 3D abdominal organ segmentation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANTONG UNIV
- Filing Date
- 2026-02-12
- Publication Date
- 2026-06-09
AI Technical Summary
Existing large-scale medical image segmentation models consume significant training resources and struggle to effectively handle blurred transition zones at organ boundaries in 3D abdominal images, leading to misjudgments in segmentation results.
An image encoder based on the Vision Transformer architecture is constructed using a parameter freezing strategy. It combines a fuzzy spatial attention module and a bidirectional cross-attention mechanism to enhance deep semantic features through a fuzzy membership matrix and utilizes geometric cue information for segmentation.
It reduces training resource consumption, improves segmentation accuracy, reduces edge jaggedness and oversmoothing, and achieves efficient and accurate 3D abdominal organ segmentation.
Smart Images
Figure CN122176296A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision, medical image processing and deep learning, and in particular to a parameter-freezing fuzzy attention method for 3D abdominal organ segmentation. Background Technology
[0002] Accurate abdominal organ segmentation is one of the core tasks of modern medical image analysis, directly related to key clinical processes such as computer-aided diagnosis. In clinical practice, 3D abdominal imaging is crucial. Computed tomography (CT) imaging, with its high spatial resolution and fast imaging speed, has become the preferred imaging modality for diagnosing abdominal diseases. For a long time, radiologists relied on manual layer-by-layer delineation to obtain organ masking information, which was not only time-consuming and labor-intensive but also highly dependent on the physician's clinical experience. With the advancement of artificial intelligence technology, deep learning-based segmentation algorithms have become the main method for solving this problem. Early research mainly focused on fully convolutional neural networks. Especially with Encoder-decoder architectures, represented by convolutional neural networks and their derivatives, effectively fuse deep semantic features with shallow detail features through symmetrical structures and skip connections, achieving excellent performance in many standardized medical image segmentation tasks. Its inherent translation invariance and locality allow it to extract local texture features well, but also limit its further development. The local receptive field of the convolutional kernel means that the network struggles to establish dependencies between pixels over long distances. (This is particularly relevant when dealing with the abdomen.) In such scenarios, due to the significant differences in the shapes of various organs and the extremely low contrast between soft tissues, relying solely on local features is insufficient. It is often difficult to capture the implicit positional constraints between organs, which cannot meet the requirements of high-precision clinical applications.
[0003] To overcome the shortcomings of convolutional neural networks in global context modeling capabilities, It introduced the field of computer vision. It abandons local convolution operations and instead segments the image into a series of... The sequence is analyzed, and a multi-head self-attention mechanism is used to calculate the correlation between each element in the sequence and all other elements. This mechanism gives the model a global receptive field, thus enabling a more effective understanding of the relative positions of abdominal organs. With the development of technology, [further details are needed]. The basic large-scale model, represented by [model name], has demonstrated zero-shot generalization ability. Through transfer learning and fine-tuning on large-scale medical datasets, [achievements / results] have emerged. Large-scale segmentation models specifically designed for medical scenarios. These models leverage general parameters obtained through pre-training on a large amount of data, and Its powerful feature representation capabilities demonstrate superior performance compared to traditional methods when handling abdominal organ segmentation tasks with complex anatomical structures. Their robustness is remarkable. They can accurately identify organs and infer ambiguous boundaries using contextual information, making them a recognized mainstream technology and research hotspot in the field of medical image processing.
[0004] Although based on While large-scale medical segmentation models have achieved significant breakthroughs in theoretical performance, translating them into practical clinical tools still faces challenges. Firstly, there is the exorbitant computational resource consumption. Existing large-scale medical segmentation models typically contain hundreds of millions of parameters, significantly increasing the computational demands for training and inference. Traditional full-scale fine-tuning strategies require updating all weight parameters in the network, but most primary hospitals and research institutions have limited hardware resources. Therefore, it is necessary to explore parameter freezing, an efficient parameter fine-tuning strategy, to reduce memory usage and training time while preserving as much of the pre-training knowledge of the large model as possible. The inherent uncertainty of medical images themselves presents another challenge. (Abdominal...) Significant partial volume effects exist in the images, causing the boundary between organs and background to often not be a clear binary outline, but rather a blurred transition zone with gradual grayscale changes. Most existing deep learning models employ hard attention mechanisms based on binary logic or standard... The classification layer forcibly classifies these blurred pixels as either foreground or background. This processing logic ignores the blurred nature of boundary regions, easily leading to misclassifications at crucial lesion boundaries and affecting subsequent diagnostic accuracy. Therefore, developing a method that can achieve lightweight training through parameter freezing strategies while effectively modeling and enhancing the blurred features of edges using fuzzy set theory has become a driving force. The key to truly implementing assisted diagnosis lies in this. Summary of the Invention
[0005] The purpose of this invention is to provide a parameter-freezing fuzzy attention method for 3D abdominal organ segmentation, aiming to solve the technical problem of high training resource consumption of existing large-scale medical image segmentation models, and to overcome the limitations of 3D abdominal... The technical challenge of blurred boundaries in images.
[0006] The inventive concept of this invention is as follows: This invention provides a parameter-freezing fuzzy attention method for 3D abdominal organ segmentation, comprising the following steps: S1: Obtaining 3D abdominal organ medical image data and preprocessing the data; S2: Constructing an image encoder based on the Vision Transformer architecture, and preprocessing general weight parameters on a large-scale natural image dataset; S3: Inputting the preprocessed 3D abdominal organ image data into the image encoder, executing a parameter-freezing strategy, and setting the gradient calculation attributes of all weight parameters and bias parameters. for S4: The obtained deep semantic features are fed into the fuzzy spatial attention module. A fuzzy mapping relationship is established through convolution to obtain a fuzzy membership matrix, which is then used to enhance the original deep semantic features. S5: Geometric cue information of the target region is obtained. A cue encoder generates a geometric cue embedding. The enhanced image features and the geometric cue embedding are input together into a mask decoder, outputting a segmentation mask for the slice. S6: Following the spatial index order of the original medical image, the continuously generated two-dimensional binary segmentation masks are stacked and reconstructed to output the final 3D abdominal organ segmentation volume data. This invention reduces training memory usage through parameter freezing and effectively solves the segmentation problem of blurred abdominal organ edges using fuzzy logic.
[0007] To achieve the above-mentioned objectives, the present invention employs the following technical solution: A parameter-freezing fuzzy attention method for 3D abdominal organ segmentation, comprising the following steps:
[0008] S1: Obtain 3D abdominal organ medical image data and preprocess the data;
[0009] S2: Building based on An image encoder architecture that loads generic weight parameters obtained by preprocessing on a large-scale natural image dataset;
[0010] S3: Input the preprocessed 3D abdominal organ image data into the image encoder, execute the parameter freezing strategy, and forcibly set the gradient calculation attributes of all weight parameters and bias parameters. for The original high-dimensional deep semantic features of the image are extracted using only forward propagation;
[0011] S4: The extracted deep semantic features are fed into the fuzzy spatial attention module. Fuzzy mapping relationship is established through convolution to obtain the fuzzy membership matrix. This matrix is then used to enhance the original deep semantic features, suppress background noise interference, and strengthen edge uncertainty features.
[0012] S5: Obtain geometric cue information for the target region and generate geometric cue embedding through the cue encoder; input the blurred and enhanced image features and the geometric cue embedding into the mask decoder, and output a two-dimensional binary segmentation mask for each slice after bidirectional cross-attention fusion, upsampling and prediction processing;
[0013] S6: Following the spatial index order of the original medical image, the continuously generated two-dimensional binary segmentation masks are stacked and reconstructed to output the final 3D abdominal organ segmentation volume data, thus completing the segmentation task of the target region in the medical image.
[0014] As a further optimization scheme of the parameter freezing fuzzy attention method for 3D abdominal organ segmentation provided by the present invention, step S1 specifically includes the following steps:
[0015] Step S1.1: Read the 3D abdomen Image data, data format is ;
[0016] Step S1.2: Place the 3D abdomen Image data is segmented into a two-dimensional slice sequence;
[0017] Step S1.3: Adjust the size of the obtained two-dimensional slice sequence to suit subsequent operations. The model input dimensions are then normalized.
[0018] As a further optimization scheme of the parameter freezing fuzzy attention method for 3D abdominal organ segmentation provided by the present invention, step S2 specifically includes the following steps:
[0019] Step S2.1: Construct with The backbone network is based on this;
[0020] Step S2.2: Use the weight parameters pre-trained on a large-scale dataset as the basis for subsequent segmentation tasks;
[0021] As a further optimization scheme of the parameter freezing fuzzy attention method for 3D abdominal organ segmentation provided by the present invention, step S3 specifically includes the following steps:
[0022] Step 3.1: During the training phase, set the gradient calculation properties of the weight and bias parameters of the image encoder. for ;
[0023] Step S3.2: In backpropagation, the encoder parameters are not updated; only forward computation is used to extract the original high-dimensional deep semantic features. To shorten the model convergence time;
[0024] As a further optimization scheme of the parameter freezing fuzzy attention method for 3D abdominal organ segmentation provided by the present invention, step S4 specifically includes the following steps:
[0025] S4.1: Input the deep semantic features of the 3D abdominal organs output from the image encoder into a kernel with a size of [missing value]. In spatial attention convolutional layers, the channel dimensions are represented by linear projection. Compress to Thus, the initial attention map is obtained. The calculation formula is as follows:
[0026]
[0027] in, Indicates the convolution kernel weights, Indicates the bias term. This represents the convolution operation;
[0028] S4.2: Construct a Gaussian blur layer, initialize the convolution kernel weights using the two-dimensional Gaussian distribution formula, and prepare to perform smoothing filtering on the initial attention map to simulate the boundary transition characteristics of the fuzzy set. The calculation formula for the Gaussian weights is as follows:
[0029]
[0030] in, Is the convolution kernel in coordinates The weight value at that location, It is the standard deviation of the Gaussian distribution. It is the coordinate offset relative to the center of the convolution kernel;
[0031] S4.3: Normalize the generated Gaussian weight matrix so that the sum of the elements of the entire matrix is 1. The normalized convolutional kernel is used to apply the initial attention map. Perform discrete convolution operations to obtain a smoothed attention map. The calculation formula is as follows:
[0032]
[0033] in, Image pixel coordinates, The relative coordinates within the convolution kernel. The radius of the convolution kernel;
[0034] S4.4: Utilizing The activation function will smooth the attention map. Each pixel value is non-linearly mapped to Intervals, generating fuzzy membership matrices The calculation formula is as follows:
[0035]
[0036] in, Representing coordinates The fuzzy membership degree at the location is close to the value. The region represents the blurred boundary region in the image;
[0037] S4.5: The generated fuzzy membership matrix is used to weight the original semantic features extracted in step S3, and residual connections are introduced. The calculation formula for feature enhancement is as follows:
[0038]
[0039] in, This represents the feature image after blur enhancement. Represents the original semantic features. Represents the element-wise Hadamard product. A learnable scaling factor;
[0040] S4.6: Through the above calculations, the feature response of the background noise region with low membership is suppressed, and the feature expression of the 3D abdominal organ target region with high membership and the blurred boundary region is enhanced, thereby outputting a feature map with a high signal-to-noise ratio.
[0041] As a further optimization scheme of the parameter freezing fuzzy attention method for 3D abdominal organ segmentation provided by the present invention, step S5 specifically includes the following steps:
[0042] S5.1: The cue encoder receives bounding box coordinates for a 3D abdominal organ, maps the coordinates of its top-left and bottom-right corners to position embedding vectors, and superimposes a learnable box cue type embedding. Generate sparse hint embeddings The calculation formula is as follows:
[0043]
[0044] in, This represents a position encoding function used to map coordinate values to a high-dimensional feature space;
[0045] S5.2: The mask decoder utilizes a bidirectional cross-attention mechanism to process the blurred and enhanced image features. Embedded with prompts The core operation of this mechanism utilizes scaled dot product attention, calculated using the following formula:
[0046]
[0047] in, They represent queries respectively. ,key Sum matrix, , which is the scaling factor for the feature dimension;
[0048] S5.3: Construct a cue-to-image attention flow to embed the cue. for Image features for and Extract image features relevant to the cues; simultaneously construct an image-to-cue attention flow, using image features as... The prompt is embedded as and Update the prompt features to perceive the geometric information of the image, and realize the deep interactive fusion of multimodal features;
[0049] S5.4: The interactively fused features are upsampled through a deconvolution layer to restore the feature map resolution to the original slice size, and the predicted values are output using a linear projection layer. picture The calculation formula is as follows:
[0050]
[0051] in, For the fused feature tensor, For upsampling convolution kernel weights, For transpose convolution operation, For bias terms;
[0052] S5.5: Utilizing Activation function will The graph is converted into a probabilistic graph, and a threshold is set to generate the final binary segmentation mask. The calculation formula is as follows:
[0053]
[0054] Subsequently, following the spatial index order of the original images, the binary segmentation masks of all slices were stacked and reconstructed into 3D abdominal organ segmentation volume data.
[0055] Meanwhile, the present invention proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed, it implements the steps of the method described in the present invention.
[0056] Furthermore, the present invention proposes a computer-readable storage medium having a computer program stored thereon, the computer program being configured to implement the steps of the method described in the present invention when invoked by a processor.
[0057] Finally, the present invention provides a computer program product comprising a computer program / instructions that, when executed by a processor, implement the steps of the method described in the present invention.
[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0059] 1) This invention introduces a parameter freezing strategy to lock a large number of parameters during the training process. The backbone network parameters are updated only for gradients in the lightweight fuzzy module and decoder. This method reduces memory usage and computational cost, and shortens the model convergence time.
[0060] 2) To address the common problem of blurred boundaries in 3D abdominal organs, this invention designs a fuzzy spatial attention module. By introducing Gaussian fuzzy logic to generate a membership matrix, the model can identify and enhance the features of uncertain edge regions, effectively avoiding the jagged edges and oversmoothing caused by traditional segmentation, and significantly improving the segmentation precision.
[0061] 3) Through a bidirectional cross-attention mechanism, this invention achieves the fusion of visual features and geometric cue information. This mechanism not only utilizes… The extracted global deep semantics, combined with human-provided localization information, enable the model to quickly and accurately locate and segment specific 3D abdominal organs, which has high clinical value. Attached Figure Description
[0062] The technical solution and its beneficial effects of the present invention will become apparent from the following detailed description of specific embodiments in conjunction with the accompanying drawings.
[0063] Figure 1 This is a flowchart of a parameter-freezing fuzzy attention method for 3D abdominal organ segmentation according to the present invention.
[0064] Figure 2 This is an image encoder parameter freeze diagram for a parameter-freezing fuzzy attention method for 3D abdominal organ segmentation according to the present invention.
[0065] Figure 3 This is a fuzzy spatial attention structure diagram of a parameter-freezing fuzzy attention method for 3D abdominal organ segmentation according to the present invention. Detailed Implementation
[0066] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. Of course, the examples described with reference to the accompanying drawings are only for explaining the invention and should not be construed as limiting it.
[0067] Example 1: See Figures 1 to 3 This embodiment provides a parameter-freezing fuzzy attention method for 3D abdominal organ segmentation. This embodiment processes a set of standard 3D abdominal... Taking scanned data as an example, the process from data input to final output of 3D segmented volume data includes the following steps:
[0068] S1: Obtain 3D abdominal organ medical image data and preprocess the data. First, read... 3D abdomen in format Image data. For abdominal soft tissue, window width and level were set to remove irrelevant bone and air noise. Subsequently, the 3D data was divided into a series of two-dimensional slice sequences along the axis, and the resolution was uniformly adjusted to adapt to the size of the model input. Finally, normalization processing was performed.
[0069] S2: Building based on An image encoder architecture that loads general weight parameters preprocessed from large-scale natural image datasets. Constructing a standard... As the backbone network for feature extraction, the model is loaded with features to leverage general visual feature representation capabilities. Pre-trained weight parameters on large-scale datasets serve as the basis for subsequent segmentation tasks.
[0070] S3: Input the preprocessed 3D abdominal organ image data into the image encoder, and execute a parameter freezing strategy, extracting the original high-dimensional deep semantic features of the image only using forward propagation. During the training phase, the gradient calculation properties of all weight and bias parameters in the image encoder are forcibly locked, preventing them from being updated during backpropagation. The input two-dimensional slices are mapped to high-dimensional deep semantic features solely through forward computation. This strategy significantly reduces video memory usage.
[0071] S4: The extracted deep semantic features are fed into the fuzzy spatial attention module. Fuzzy mapping relationships are established through convolution to obtain the fuzzy membership matrix, and the original deep semantic features are enhanced based on this matrix.
[0072] Gaussian fuzzy logic is used to simulate biological vision's perception of fuzzy boundaries. First, an initial attention map is generated through convolution, then smoothed using Gaussian smoothing. A fuzzy membership matrix is generated by activation. This matrix is then used to weight the original features, suppress background noise, and enhance the feature representation of the blurred edges of abdominal organs, resulting in enhanced features. .
[0073] S5: Obtain geometric cue information for the target region and generate a geometric cue embedding through a cue encoder; input the blurred and enhanced image features and the geometric cue embedding into a mask decoder, perform bidirectional cross-attention fusion, upsampling, and prediction processing, and output a two-dimensional binary segmentation mask for each slice. Generate bounding box cue based on the approximate location of the target organ. After encoding the cue into a vector, it interacts bidirectionally with the image features output in step S4 in the decoder. The fused features are upsampled to restore resolution, and a predicted probability map is output, which is then binarized to obtain the current segmentation mask.
[0074] S6: Following the spatial index order of the original medical image, the continuously generated two-dimensional binary segmentation masks are stacked and reconstructed to output the final 3D abdominal organ segmentation volume data. After completing the above reasoning for all slices, based on the spatial location information of the original data, the generated two-dimensional mask sequence is stacked and restored to 3D volume data, completing the overall segmentation of the target organ.
[0075] Specifically, the implementation process of step S4 above is as follows:
[0076] Step S4.1: Extract the deep semantic features of the 3D abdominal organs from the image encoder. Input to core size is In spatial attention convolutional layers, the channel dimensions are represented by linear projection. Compress to Thus, the initial attention map is obtained. The calculation formula is as follows:
[0077]
[0078] in, Indicates the convolution kernel weights, Indicates the bias term. This represents the convolution operation.
[0079] Step S4.2: Construct a Gaussian blur layer, initialize the convolution kernel weights using the two-dimensional Gaussian distribution formula, and prepare to perform smoothing filtering on the initial attention map to simulate the boundary transition characteristics of the fuzzy set. The calculation formula for the Gaussian weights is as follows:
[0080]
[0081] in, Is the convolution kernel in coordinates The weight value at that location, It is the standard deviation of the Gaussian distribution. It is the coordinate offset relative to the center of the convolution kernel.
[0082] Step S4.3: Normalize the generated Gaussian weight matrix so that the sum of the elements of the entire matrix is 0. The normalized convolutional kernel is used to apply the initial attention map. Perform discrete convolution operations to obtain a smoothed attention map. The calculation formula is as follows:
[0083]
[0084] in, Image pixel coordinates, The relative coordinates within the convolution kernel. is the radius of the convolution kernel.
[0085] Step S4.4: Utilize The activation function will smooth the attention map. Each pixel value is non-linearly mapped to the interval (0, 1) to generate a fuzzy membership matrix. The calculation formula is as follows: Specifically, the specific steps of step S4 are as follows:
[0086]
[0087] in, Representing coordinates The fuzzy membership degree at a given location, with values close to 0.5 representing fuzzy boundary regions in the image.
[0088] Step S4.5: The generated fuzzy membership matrix is used to weight the original deep semantic features extracted in step S3, and residual connections are introduced. The calculation formula for feature enhancement is as follows:
[0089]
[0090] in, This represents the feature image after blur enhancement. Represents the original deep semantic features. This represents the element-wise Hadamard product. This is a learnable scaling factor.
[0091] Step S4.6: Through the above calculation, the feature response of the background noise region with low membership is suppressed, while the feature expression of the 3D abdominal organ target region with high membership and the blurred boundary region is enhanced, thereby outputting a feature map with high signal-to-noise ratio.
[0092] Specifically, the implementation process of step S5 above is as follows:
[0093] Step S5.1: The cue encoder receives the bounding box coordinates for the 3D abdominal organ, maps the coordinates of its top-left and bottom-right corners to position embedding vectors, and superimposes the learnable box cue type embedding. Generate sparse hint embeddings The calculation formula is as follows:
[0094]
[0095] in, This represents a position encoding function used to map coordinate values to a high-dimensional feature space.
[0096] Step S5.2: The mask decoder uses a bidirectional cross-attention mechanism to process the blurred and enhanced image features. (i.e., the above) ) and hint embedding The core operation of this mechanism utilizes scaled dot product attention, calculated using the following formula:
[0097]
[0098] in, These represent the query, key, and value matrices, respectively. is the scaling factor for the feature dimension.
[0099] Step S5.3: Construct a cue-to-image attention flow to embed the cue. for Image features for and Extract image features relevant to the cues; simultaneously construct an image-to-cue attention flow, using image features as... The prompt is embedded as and This update prompts the perception of geometric information in the image by features, enabling deep interactive fusion of multimodal features.
[0100] Step S5.4: Upsample the interactively fused features through a deconvolution layer to restore the feature map resolution to the original slice size, and output the predicted Logits map using a linear projection layer. The calculation formula is as follows:
[0101]
[0102] in, For the fused feature tensor, For upsampling convolution kernel weights, For transpose convolution operation, This is a bias term.
[0103] Step S5.5: Utilize Activation function will The graph is converted into a probabilistic graph, and a threshold is set to generate the final binary segmentation mask. The calculation formula is as follows:
[0104]
[0105] Subsequently, following the spatial index order of the original images, the binary segmentation masks of all slices were stacked and reconstructed into 3D abdominal organ segmentation volume data.
[0106] Example 2: This example proposes an electronic system, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method steps of the present invention.
[0107] Example 3: This example proposes a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the steps of the method described in this invention, which will not be repeated here.
[0108] Example 4: This example proposes a computer program product, including a computer program / instructions. When the computer program / instructions are executed by a processor, they implement the steps of the method described in this invention, which will not be repeated here.
[0109] It should be noted that the processing flow of embodiments 2-4 corresponds to the specific steps of the method provided in embodiment 1 of the present invention, and has the corresponding functional modules and beneficial effects of the method. Technical details not described in detail in this embodiment can be found in the method provided in embodiment 1 of the present invention.
[0110] The program code used to implement the methods of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.
[0111] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A parameter-freezing fuzzy attention method for 3D abdominal organ segmentation, characterized in that, The process includes the following steps: Step S1: Obtain 3D abdominal organ medical image data and preprocess the data; Step S2: Build based on An image encoder architecture that loads generic weight parameters obtained by preprocessing on a large-scale natural image dataset; Step S3: Input the preprocessed 3D abdominal organ image data into the image encoder, execute the parameter freezing strategy, and set the gradient calculation properties of all weight parameters and bias parameters. for High-dimensional deep semantic features of the image are obtained through forward propagation; Step S4: The obtained deep semantic features are fed into the fuzzy spatial attention module. Fuzzy mapping relationship is established through convolution to obtain the fuzzy membership matrix. The fuzzy membership matrix is then used to enhance the original deep semantic features to suppress background noise interference and strengthen edge uncertainty features. Step S5: Obtain geometric cue information of the target region, and obtain the geometric cue embedding through the cue encoder; The blurred and enhanced image features and geometric cues are embedded and input into the mask decoder. After bidirectional cross-attention fusion, upsampling and prediction processing, a two-dimensional binary segmentation mask for each slice is output. Step S6: Stack the continuously generated two-dimensional binary segmentation masks according to the spatial index order of the original medical image, and output the final 3D abdominal organ segmentation volume data to complete the segmentation task of the target region in the medical image.
2. The parameter-freezing fuzzy attention method for 3D abdominal organ segmentation according to claim 1, characterized in that, Step S1 includes the following steps: Step S1.1: Read the 3D abdomen Image data, data format is ; Step S1.2: Place the 3D abdomen Image data is segmented into a two-dimensional slice sequence; Step S1.3: Adjust the size of the obtained two-dimensional slice sequence to suit subsequent operations. The model input dimensions are then normalized.
3. The parameter-freezing fuzzy attention method for 3D abdominal organ segmentation according to claim 1, characterized in that, Step S2 includes the following steps: Step S2.1: Construct with The backbone network is based on this; Step S2.2: Use the weight parameters pre-trained on a large-scale dataset as the basis for subsequent segmentation tasks.
4. The parameter-freezing fuzzy attention method for 3D abdominal organ segmentation according to claim 1, characterized in that, Step S3 includes the following steps: Step 3.1: During the training phase, set the gradient calculation properties of the weight and bias parameters of the image encoder. for ; Step S3.2: In backpropagation, the encoder parameters are not updated; only forward computation is used to extract the original high-dimensional deep semantic features. This is to shorten the model convergence time.
5. The parameter-freezing fuzzy attention method for 3D abdominal organ segmentation according to claim 1, characterized in that, Step S4 includes the following steps: Step S4.1: Extract the deep semantic features of the 3D abdominal organs from the image encoder. Input to core size is In spatial attention convolutional layers, the channel dimensions are represented by linear projection. Compress to The initial attention map is obtained. The calculation formula is as follows: ； in, Indicates the convolution kernel weights, Indicates the bias term. This represents the convolution operation; Step S4.2: Construct a Gaussian blur layer, initialize the convolution kernel weights using the two-dimensional Gaussian distribution formula, and perform smoothing filtering on the initial attention map to simulate the boundary transition characteristics of the fuzzy set. The formula for calculating the Gaussian weights is as follows: ； in, Is the convolution kernel in The weight value at that location, It is the standard deviation of the Gaussian distribution. It is the coordinate offset relative to the convolution and the center; Step S4.3: Normalize the Gaussian weight matrix so that the sum of the elements of the entire matrix is 0. Using normalized convolution kernels For the initial attention map Perform discrete convolution operations to obtain a smoothed attention map. ， The calculation formula is as follows: ； in, Image pixel coordinates, The relative coordinates within the convolution kernel. The radius of the convolution kernel; Step S4.4: Utilize The activation function will smooth the attention map. Each pixel value is non-linearly mapped to Intervals, generating fuzzy membership matrices The calculation formula is as follows: ； in, Representing coordinates The fuzzy membership degree at the location is close to the value. The region represents the blurred boundary region in the image; Step S4.5: The original semantic features extracted in step S3 are weighted using the generated fuzzy membership matrix, and residual connections are introduced. The calculation formula for feature enhancement is as follows: ； in, This represents the feature image after blur enhancement. Represents the original semantic features. This represents the element-wise Hadamard product. A learnable scaling factor; Step S4.6: Through the above calculations, the feature response of the low membership background noise region is suppressed, and the feature expression of the high membership 3D abdominal organ target region and the blurred boundary region is enhanced, thereby outputting a feature map with a high signal-to-noise ratio.
6. The parameter-freezing fuzzy attention method for 3D abdominal organ segmentation according to claim 1, characterized in that, Step S5 includes the following steps: Step S5.1: Prompt the encoder to receive the bounding box coordinates of the 3D abdominal organs. The coordinates of its top-left and bottom-right corners are mapped to position embedding vectors, and then a learnable box hint type embedding is superimposed on them. Generate sparse hint embeddings The calculation formula is as follows: ； in, This represents a position encoding function used to map coordinate values to a high-dimensional feature space; Step S5.2: The mask decoder uses a bidirectional cross-attention mechanism to process the blurred and enhanced image features. Embedded with prompts The core operation of this mechanism utilizes scaled dot product attention, calculated using the following formula: ； in, They represent queries respectively. ,key Sum matrix, , which is the scaling factor for the feature dimension; Step S5.3: Embed with prompts for Image features for and Extract image features relevant to the cues; and construct an image-to-cue attention flow, using image features as the basis for... The prompt is embedded as and Update the prompt features to perceive the geometric information of the image, and realize the interactive fusion of multimodal features; Step S5.4: Upsample the interactively fused features through a deconvolution layer to restore the feature map resolution to the original slice size, and then use a linear projection layer to output the predicted value. picture The calculation formula is as follows: ； in, For the fused feature tensor, For upsampling convolution kernel weights, For transpose convolution operation, For bias terms; Step S5.5: Utilize Activation function will The graph is converted into a probabilistic graph, and a threshold is set to generate the final binary segmentation mask. The calculation formula is as follows: ； Following the spatial index order of the original images, the binary segmentation masks of all slices are stacked and reconstructed into 3D abdominal organ segmentation volume data.
7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed, it implements the steps of the method as described in any one of claims 1 to 6.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is configured to implement the steps of the method according to any one of claims 1 to 6 when invoked by a processor.
9. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 6.