An intravascular ultrasound image segmentation method based on fuzzy fusion and contrast distillation

Abstract

The application discloses an intravascular ultrasound image segmentation method based on fuzzy fusion and contrast distillation, and comprises the following steps: preprocessing an acquired intravascular ultrasound image, performing multi-scale feature extraction on the preprocessed image, and generating a pixel-level uncertainty map corresponding to each scale feature; constructing a fusion uncertainty map generation module to realize adaptive weighted fusion of cross-scale pixel-level uncertainty maps and obtain a fused uncertainty map; constructing a fused feature generation module to generate fused features and fused disturbance features; constructing a multi-objective optimization module to obtain an optimal segmentation model; and segmenting the collected intravascular ultrasound image to obtain a final segmentation result. The application can improve the segmentation accuracy and robustness of the intravascular ultrasound image, solve the problems of inaccurate boundary positioning and incorrect area judgment caused by noise interference, structure blur and low-efficiency multi-scale feature fusion in the intravascular ultrasound image segmentation, and realize high-precision analysis of complex pathological characteristics.

Description

A method for intravascular ultrasound image segmentation based on fuzzy fusion and contrastive distillation Technical Field

[0001] This invention belongs to the fields of medical image processing and computer vision, and specifically relates to a method for intravascular ultrasound image segmentation based on fuzzy fusion and contrastive distillation. By fusing the uncertainty information of multi-scale features and constructing a self-supervised contrastive distillation mechanism, high-precision identification and segmentation of vascular structures can be achieved. Background Technology

[0002] Intravascular ultrasound (IVUS), a core technology for the diagnosis and treatment of coronary artery disease, acquires real-time cross-sectional images of the coronary arteries through an intravascular catheter. This clearly displays the vessel wall structure and plaque composition, providing crucial guidance for interventional treatment. Precise segmentation of the vessel lumen and plaque boundaries is fundamental for calculating important clinical parameters such as stenosis rate and assessing plaque stability. However, the speckle noise and artifacts inherent in IVUS images, along with the high complexity of vascular structures (such as bifurcation lesions and calcified plaques), pose significant challenges to traditional segmentation methods.

[0003] Early IVUS segmentation research was primarily based on traditional image processing methods. Thresholding segmentation methods distinguish vascular structures using gray-level differences, but are sensitive to noise and cannot handle heterogeneous plaques; region growing methods rely on seed point selection and are susceptible to oversegmentation due to blurred boundaries; edge detection-based algorithms (such as Canny and active contour models) can locate parts of the vessel wall, but struggle to handle the texture similarity between high-density calcified areas and thrombus formations. Traditional methods generally suffer from frequent manual parameter adjustments, poor model generalization, and segmentation accuracy limited by operator experience. With breakthroughs in deep learning technology, convolutional neural networks (CNNs) have gradually become the mainstream solution for IVUS segmentation. U-Net and its variants achieve multi-scale feature extraction through an encoder-decoder structure, achieving superior performance compared to traditional algorithms on certain standardized datasets. However, significant limitations remain in practical applications: the low signal-to-noise ratio of IVUS images makes it easy for models to confuse normal vascular tissue with noise artifacts, especially with insufficient sensitivity to thin-layer fibrous plaques and micro-dissections; existing models lack quantitative assessment of the reliability of segmentation results, are prone to misclassification in areas with blurred boundaries, and cannot provide uncertainty hints; the fusion strategy of multi-level features fails to effectively utilize spatial context association and ignores the complementary representation of vascular morphology by features at different scales; supervised learning models that rely on a large amount of labeled data are difficult to adapt to the differences in data distribution in clinical scenarios (such as different ultrasound probe frequencies and individualized vascular morphology of patients), resulting in a significant decrease in cross-device segmentation performance.

[0004] In recent years, uncertainty modeling and self-supervised learning have provided new approaches for medical image segmentation. Uncertainty estimation methods identify regions with low confidence through probability prediction or Monte Carlo sampling, but existing techniques typically use them as post-processing rather than embedding them into the feature learning process, failing to achieve feature-level bias correction. Contrastive learning enhances feature discriminativity by constructing positive and negative sample pairs, but in IVUS scenarios, it is necessary to address the problem of negative sample confusion caused by vascular structural similarity. Furthermore, existing distillation methods mainly focus on model compression and have not explored their role in unifying modal features of ultrasound images. Therefore, there is an urgent need to construct a novel IVUS segmentation framework that integrates uncertainty and contrast mechanisms to overcome the performance bottlenecks of current techniques under complex pathological conditions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method for intravascular ultrasound image segmentation based on fuzzy fusion and contrastive distillation. This method can improve the segmentation accuracy and robustness of intravascular ultrasound images. By fusing the uncertainty information of multi-scale features and constructing a self-supervised contrastive distillation mechanism, it solves the problems of inaccurate boundary positioning and misjudgment of regions caused by noise interference, structural fuzziness, and inefficient multi-scale feature fusion in intravascular ultrasound image segmentation, thereby achieving high-precision analysis of complex pathological features.

[0006] The present invention adopts the following technical solution.

[0007] A method for intravascular ultrasound image segmentation based on fuzzy fusion and contrast distillation includes:

[0008] Step S1: Preprocess the acquired intravascular ultrasound image. Based on the Res2Net-50 network, extract multi-scale features from the preprocessed image and generate pixel-level uncertainty maps corresponding to each scale feature based on deep evidence learning.

[0009] Step S2: Construct a fusion uncertainty map generation module. This module introduces learnable scale weight parameters and uses spatial attention to achieve adaptive weighted fusion of cross-scale pixel-level uncertainty maps to obtain a fusion uncertainty map, which is used to calculate the pixel classification loss value.

[0010] Step S3: Construct a fusion feature generation module. This module generates corresponding perturbation features based on the multi-scale features of the preprocessed image, and modulates the multi-scale features and their corresponding perturbation features using a pixel-level uncertainty map to generate fusion features and fusion perturbation features, which are used to calculate the contrastive distillation loss value.

[0011] Step S4: Construct a multi-objective optimization module by calculating the original segmentation loss value. Pixel classification loss value And compare distillation loss values The weighted sum is used to obtain the total loss value. The total loss value is then used to perform joint optimization training on the segmentation model composed of the Res2Net-50 network, the uncertainty graph generation module, and the feature generation module to obtain the optimal segmentation model.

[0012] Step S5: Using the optimal segmentation model, the acquired intravascular ultrasound image is segmented to obtain the final segmentation result.

[0013] In one embodiment of the present invention, step S1 includes:

[0014] Step S11: The acquired intravascular ultrasound image is scaled and linearly normalized to meet the input size constraints of the segmentation model.

[0015] Step S12: Based on the Res2Net-50 network, extract L features from different stages of the preprocessed image as its multi-scale features F. l Where l = 1, 2, ..., L, for each scale feature, the evidence generation layer generates a corresponding pixel-level uncertainty map.

[0016] In one embodiment of the present invention, in step S12, the pixel-level uncertainty map M corresponding to each scale feature is generated using the following formula. l :

[0017]

[0018] Where K represents the number of segmentation categories, It is the Dirichlet concentration parameter of each pixel in the l-th scale feature corresponding to the k-th category. It is the l-th evidence value corresponding to the k-th category.

[0019] In one embodiment of the present invention, step S2 includes:

[0020] Step S21: Adaptive weighted fusion of the pixel-level uncertainty map is performed using learnable scale weight parameters to obtain the initial fusion uncertainty map;

[0021] Step S22: Based on the spatial attention map, the initial fusion uncertainty map is dynamically calibrated in terms of spatial dimensions to obtain the final fusion uncertainty map.

[0022] In one embodiment of the present invention, in step S21, the pixel-level uncertainty map is adaptively weighted and fused using the following formula to obtain an initial fused uncertainty map M'. fused :

[0023]

[0024] Among them, w' l M is the l-th learnable scale weight parameter initialized randomly. l Let w' be the pixel-level uncertainty map corresponding to the l-th scale feature, where l = 1, 2, ..., L. s Let s be the s-th learnable scale weight parameter that is randomly initialized, where s = 1, 2, ..., L.

[0025] In one embodiment of the present invention, in step S22, the final fusion uncertainty map M is obtained using the following formula. f :

[0026] M f =M′ fused ⊙A,

[0027] Among them, M' fused This is the initial fusion uncertainty map, where ⊙ represents pixel-wise multiplication and A represents the spatial attention map:

[0028] A=σ(Conv(ReLU(Conv(Concat(M 1 ,…,M l ,…,M L ))))), where σ is the Sigmoid function, Conv is the convolution operation, ReLU is the activation function, Concat is the concatenation operation by scale, and M l Let l be the pixel-level uncertainty map corresponding to the l-th scale feature, where l = 1, 2, ..., L.

[0029] In one embodiment of the present invention, step S3 includes:

[0030] Step S31, for the multi-scale features F of the preprocessed image l Using the perturbation information, corresponding perturbation features are generated.

[0031] Step S32, using the pixel-level uncertainty map to analyze the multi-scale feature F l and their corresponding disturbance characteristics The modulated features and their corresponding perturbation features are processed by adaptive pooling and convolution mapping, and then concatenated along the channel dimension to obtain fused features and fused perturbation features.

[0032] In one embodiment of the present invention, step S32 includes the following steps:

[0033] Step S321: Using the pixel-level uncertainty map, the multi-scale feature F is... l and their corresponding disturbance characteristics Modulation was performed separately to obtain the modulation multi-scale features G. l and their corresponding modulation perturbation characteristics

[0034] G l =F l ⊙(1+M l ),

[0035]

[0036] Among them, M l This is the pixel-level uncertainty map corresponding to the l-th scale feature;

[0037] Step S322, using the scale-characteristic still d l (·) For the modulation multi-scale feature G l and their corresponding modulation perturbation characteristics Adaptive pooling and convolutional mapping are performed to obtain the distilled multi-scale features z. l and its corresponding distillation perturbation characteristics

[0038]

[0039] Where Conv is the convolution operation, ReLU is the activation function, and GAP represents global average pooling;

[0040] Step S323: Process the distillation multi-scale features z using a fully connected network. l and its corresponding distillation perturbation characteristics A nonlinear mapping is performed to obtain the fused feature y and the fused perturbation feature.

[0041] y = W2·GELU(W1·Concat(z) 1 ,…,z l ,…,z L )+b1)+b2,

[0042]

[0043] Where W1 and W2 are the weights of the fully connected layer, b1 and b2 are the biases of the fully connected layer, GELU is the activation function, and Concat is the concatenation operation by scale.

[0044] In one embodiment of the present invention, in step S4, the total loss value The following formula is used to calculate:

[0045]

[0046] In one embodiment of the present invention, in step S4, the model parameters of the segmentation model are iteratively optimized based on the obtained total loss value to obtain the optimal segmentation model:

[0047]

[0048] Where θ' is the optimized model parameter, θ is the model parameter of the previous cycle, and ρ is the learning rate.

[0049] The beneficial effects of this invention are that, compared with the prior art, it can improve the segmentation accuracy and robustness of intravascular ultrasound images. By fusing the uncertainty information of multi-scale features and constructing a self-supervised comparative distillation mechanism, it solves the problems of inaccurate boundary positioning and misjudgment of regions caused by noise interference, structural ambiguity and inefficient multi-scale feature fusion in intravascular ultrasound image segmentation, and achieves high-precision analysis of complex pathological features. Attached Figure Description

[0050] Figure 1 is a flowchart of an intravascular ultrasound image segmentation method based on fuzzy fusion and contrastive distillation according to an embodiment of the present invention;

[0051] Figure 2 is a label image corresponding to the original intravascular ultrasound image for training and an example image of the original intravascular ultrasound image according to an embodiment of the present invention.

[0052] Figure 3 is a schematic diagram of the segmentation results of an intravascular ultrasound test image according to an embodiment of the present invention. Detailed Implementation

[0053] The present application will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention, and should not be construed as limiting the scope of protection of the present application.

[0054] Figure 1 is a flowchart of an intravascular ultrasound image segmentation method based on fuzzy fusion and contrast distillation according to an embodiment of the present invention. As shown in Figure 1, the intravascular ultrasound image segmentation method based on fuzzy fusion and contrast distillation includes the following steps:

[0055] Step S1: Preprocess the acquired intravascular ultrasound image. Based on the Res2Net-50 network, extract multi-scale features from the preprocessed image and generate pixel-level uncertainty maps corresponding to each scale feature based on deep evidence learning.

[0056] In one embodiment of the present invention, step S1 includes:

[0057] Step S11: The acquired intravascular ultrasound image is scaled and linearly normalized to meet the input size constraints of the segmentation model.

[0058] In one embodiment of the present invention, the dataset contains 1800 intravascular ultrasound images, and an example of the original intravascular ultrasound image is shown in Figure 2. The dataset is randomly divided into a training set and a test set according to a preset ratio (e.g., 8:2). Each image is labeled with two categories (intima and media), that is, the intravascular ultrasound image is labeled as the intima part and the media part. Figure 2 shows the label image corresponding to the example of the original intravascular ultrasound image, and all image data are adjusted to a preset resolution size, such as 256×256.

[0059] Step S12: Based on the Res2Net-50 network, extract L features from different stages of the preprocessed image as its multi-scale features F. l Where l = 1, 2, ..., L, in one embodiment of the present invention, L = 4, and for each scale feature, the evidence generation layer generates a corresponding pixel-level uncertainty map.

[0060] In one embodiment of the present invention, in step S12, the obtained multi-scale features are connected to the evidence generation layer, and the channels of the multi-scale features are mapped to K dimensions through independent 1×1 dilated convolutions (where K is the number of segmentation categories; in the above embodiment, for the label image shown in Figure 2, the number of label categories is 2, and since there is also a background category in the image, the number of segmentation categories K = 3). The Softplus activation function is used to constrain the evidence value e. l ∈R + Let l = 1, 2, ..., L. Based on the evidence values ​​corresponding to each scale feature, a pixel-level uncertainty map M corresponding to each scale feature is generated using the following uncertainty quantification formula. l :

[0061]

[0062] in, It is the Dirichlet concentration parameter of each pixel in the l-th scale feature corresponding to the k-th category. It is the l-th evidence value corresponding to the k-th category.

[0063] Step S2: Construct a fusion uncertainty map generation module. This module introduces learnable scale weight parameters and uses spatial attention to achieve adaptive weighted fusion of cross-scale pixel-level uncertainty maps to obtain a fusion uncertainty map, which is used to calculate the pixel classification loss value.

[0064] Step S2 includes:

[0065] Step S21: Adaptive weighted fusion of the pixel-level uncertainty map is performed using learnable scale weight parameters to obtain the initial fusion uncertainty map;

[0066] In one embodiment of the present invention, step S21 uses the following formula to perform adaptive weighted fusion of the pixel-level uncertainty map to obtain an initial fusion uncertainty map M'. fused :

[0067]

[0068] Among them, w' l M is the l-th learnable scale weight parameter initialized randomly. l Let w' be the pixel-level uncertainty map corresponding to the l-th scale feature, where l = 1, 2, ..., L. s Let s be the s-th learnable scale weight parameter that is randomly initialized, where s = 1, 2, ..., L.

[0069] Step S22: Based on the spatial attention map, the initial fusion uncertainty map is dynamically calibrated in terms of spatial dimensions to obtain the final fusion uncertainty map.

[0070] In one embodiment of the present invention, step S22 uses the following formula to obtain the final fusion uncertainty map M. f :

[0071] M f =M′ fused ⊙A,

[0072] Where ⊙ represents pixel-wise multiplication, used to achieve dynamic calibration of spatial dimensions, and A represents the spatial attention map:

[0073] A=σ(Conv(ReLU(Conv(Concat(M 1 ,…,M l ,…,M L ))))),

[0074] Where σ is the Sigmoid function, Conv is the convolution operation, ReLU is the activation function, and Concat is the concatenation operation by scale.

[0075] Step S3: Construct a fusion feature generation module. This module generates corresponding perturbation features based on the multi-scale features of the preprocessed image, and modulates the multi-scale features and their corresponding perturbation features using a pixel-level uncertainty map to generate fusion features and fusion perturbation features, which are used to calculate the contrastive distillation loss value.

[0076] Step S3 includes:

[0077] Step S31, for the multi-scale features F of the preprocessed image l Using the perturbation information, corresponding perturbation features are generated.

[0078] The perturbation information can be, for example, random noise or perturbation information generated by operations such as Dropout or masking. In one embodiment of the present invention, the perturbation information generated by the Dropout perturbation operation is used for the multi-scale feature F. l Generate perturbation features corresponding to the perturbation. The perturbation ratio, or the discard ratio, is 0.2.

[0079] Step S32, using the pixel-level uncertainty map to analyze the multi-scale feature F l and their corresponding disturbance characteristics The modulated features and their corresponding perturbation features are processed by adaptive pooling and convolution mapping, and then concatenated along the channel dimension to obtain fused features and fused perturbation features.

[0080] Step S32 further includes the following steps:

[0081] Step S321: Using the pixel-level uncertainty map, the multi-scale feature F is... l and their corresponding disturbance characteristics Modulation was performed separately to obtain the modulation multi-scale features G. l and their corresponding modulation perturbation characteristics

[0082] G l =F l ⊙(1+M l ),

[0083]

[0084] Step S322, using the scale-characteristic still d l (·) For the modulation multi-scale feature G l and their corresponding modulation perturbation characteristics Adaptive pooling and convolutional mapping are performed to obtain the distilled multi-scale features z. l and its corresponding distillation perturbation characteristics

[0085] z l =d l (G l =Conv(ReLU(Conv(GAP(G l )))),

[0086]

[0087] Where Conv is the convolution operation, ReLU is the activation function, and GAP represents global average pooling.

[0088] Step S323: Process the distillation multi-scale features z using a fully connected network. l and its corresponding distillation perturbation characteristics A nonlinear mapping is performed to obtain the fused feature y and the fused perturbation feature.

[0089] y = W2·GELU(W1·Concat(z) 1 ,…,z l ,…,z L )+b1)+b2,

[0090]

[0091] Where W1 and W2 are the weights of the fully connected layer, b1 and b2 are the biases of the fully connected layer, and GELU is the activation function.

[0092] Step S4: Construct a multi-objective optimization module by calculating the original segmentation loss value. Pixel classification loss value And compare distillation loss values The weighted sum is used to obtain the total loss value. The total loss value is then used to perform joint optimization training on the segmentation model composed of the Res2Net-50 network, the uncertainty graph generation module, and the feature generation module to obtain the optimal segmentation model.

[0093] Wherein, the original segmentation loss value It can be calculated using the following formula:

[0094]

[0095] in, and These are the ground truth label and predicted probability of the c-th category corresponding to the i-th scale feature, respectively, where K is the number of segmentation categories, and ∈ is a very small value to ensure computational stability.

[0096] Wherein, the pixel classification loss value The following can be calculated based on the fusion uncertainty graph:

[0097]

[0098] Where K is the number of segmentation categories, p represents the number of pixels, Ω is the set of pixels, γ is a hyperparameter used to control the intensity of uncertainty enhancement, and y c (p) represents the actual pixel label value. To predict pixel label values.

[0099] In one embodiment of the present invention, the hyperparameter γ that controls the uncertainty enhancement intensity can be set to 0.5.

[0100] Among them, the comparative distillation loss value The following can be calculated based on the fusion features and fusion perturbation features:

[0101]

[0102] P={(i,j)|S ij >U(S ij ),i≠j},

[0103] Where P represents the set of positive sample pairs, This represents the similarity between the fused features of the i-th sample and the fused features of the j-th sample. This represents the similarity between the fused perturbation features of the i-th sample and the fused perturbation features of the j-th sample. The similarity between the fused perturbation features of the i-th sample and the fused perturbation features of the k-th sample is represented by y and First, normalization is performed using the L2 norm, where T represents the transpose operation, U(·) represents finding the median of the matrix triangle, B is the batch size, and τ is the temperature coefficient that controls the steepness of the probability distribution.

[0104] In one embodiment of the present invention, B = 8 and τ = 0.07.

[0105] Wherein, the total loss value It can be calculated using the following formula:

[0106]

[0107] Based on the obtained total loss value, the model parameters of the segmentation model are iteratively optimized to obtain the optimal segmentation model:

[0108]

[0109] Where θ' is the optimized model parameter, θ is the model parameter of the previous cycle, and ρ is the learning rate.

[0110] In one embodiment of the present invention, the Adam optimizer can be used as the training optimizer, the initial learning rate can be set to 1e-4, a linear decay strategy is adopted, the batch size is set to 8, and the training cycle is 100 times.

[0111] It is worth noting that the training method and learning rate optimization strategy of the above segmentation model are non-restrictive and preferred options. Those skilled in the art can select the training method and parameter configuration scheme of the segmentation model based on various indicators such as recognition accuracy and efficiency.

[0112] Step S5: Using the optimal segmentation model, the acquired intravascular ultrasound image is segmented to obtain the final segmentation result.

[0113] In one embodiment of the present invention, the optimal segmentation model obtained through training was used to segment 360 intravascular ultrasound images in the test set. An example of the segmentation results is shown in Figure 3. The results show that the segmentation model of this embodiment can accurately identify and segment the intima and media structures in intravascular ultrasound images. Simultaneously, fusing the uncertainty map can effectively indicate low-confidence regions, thereby improving the segmentation accuracy of the model in boundary regions and complex lesion sites.

[0114] The beneficial effects of this invention are that, compared with the prior art, it can improve the segmentation accuracy and robustness of intravascular ultrasound images. By fusing the uncertainty information of multi-scale features and constructing a self-supervised comparative distillation mechanism, it solves the problems of inaccurate boundary positioning and misjudgment of regions caused by noise interference, structural ambiguity and inefficient multi-scale feature fusion in intravascular ultrasound image segmentation, and achieves high-precision analysis of complex pathological features.

[0115] The applicant of this invention has provided a detailed description of the embodiments of the invention in conjunction with the accompanying drawings. However, those skilled in the art should understand that the above embodiments are merely preferred embodiments of the invention. The detailed description is only intended to help readers better understand the spirit of the invention and is not intended to limit the scope of protection of the invention. On the contrary, any improvements or modifications made based on the inventive spirit of the invention should fall within the scope of protection of the invention.

Claims

1. A method for segmenting intravascular ultrasound images based on fuzzy fusion and contrastive distillation, characterized in that, The method includes the following steps: Step S1, preprocessing the acquired intravascular ultrasound image, extracting multi-scale features from the preprocessed image based on the Res2Net-50 network, and generating pixel-level uncertainty maps corresponding to features at each scale based on deep evidence learning; Step S2, constructing a fusion uncertainty map generation module, which introduces learnable scale weight parameters and achieves adaptive weighted fusion of cross-scale pixel-level uncertainty maps based on spatial attention to obtain a fusion uncertainty map, which is used to calculate the pixel classification loss value. Step S3: Construct a fusion feature generation module. This module generates corresponding perturbation features based on the multi-scale features of the preprocessed image, and modulates the multi-scale features and their corresponding perturbation features using a pixel-level uncertainty map to generate fusion features and fusion perturbation features, which are used to calculate the contrastive distillation loss value. Step S4: Construct a multi-objective optimization module by calculating the original segmentation loss value. Pixel classification loss value And compare distillation loss values The weighted sum is used to obtain the total loss value. The total loss value is then used to perform joint optimization training on the segmentation model composed of the Res2Net-50 network, the uncertainty graph generation module, and the feature generation module to obtain the optimal segmentation model. In step S5, the acquired intravascular ultrasound image is segmented using the optimal segmentation model to obtain the final segmentation result.

2. The method according to claim 1, characterized in that, Step S1 includes: Step S11, scaling and linearly normalizing the acquired intravascular ultrasound image to meet the input size constraints of the segmentation model; Step S12, extracting L different stage features of the preprocessed image as its multi-scale features F based on the Res2Net-50 network. l Where l = 1, 2, ..., L, for each scale feature, the evidence generation layer generates a corresponding pixel-level uncertainty map.

3. The method according to claim 2, characterized in that, In step S12, the pixel-level uncertainty map M corresponding to each scale feature is generated using the following formula. l : Where K represents the number of segmentation categories, It is the Dirichlet concentration parameter of each pixel in the l-th scale feature corresponding to the k-th category. It is the l-th evidence value corresponding to the k-th category.

4. The method according to claim 1, characterized in that, Step S2 includes: Step S21, using learnable scale weight parameters to adaptively weight and fuse the pixel-level uncertainty map to obtain an initial fused uncertainty map; Step S22, based on the spatial attention map, dynamically calibrating the spatial dimension of the initial fused uncertainty map to obtain the final fused uncertainty map.

5. The method according to claim 4, characterized in that, In step S21, the pixel-level uncertainty map is adaptively weighted and fused using the following formula to obtain the initial fused uncertainty map M′. fused : Among them, w′ l M is the l-th learnable scale weight parameter initialized randomly. l Let w' be the pixel-level uncertainty map corresponding to the l-th scale feature, where l = 1, 2, ..., L. s Let s be the s-th learnable scale weight parameter that is randomly initialized, where s = 1, 2, ..., L.

6. The method according to claim 4, characterized in that, In step S22, the final fusion uncertainty map M is obtained using the following formula. f M f =M fused ⊙A, where M′ fused For the initial fusion uncertainty map, ⊙ represents pixel-wise multiplication, and A represents the spatial attention map: A = σ(Conv(ReLU(Conv(Concat(M))) 1 ,…,M l ,…,M L ))))), where σ is the Sigmoid function, Conv is the convolution operation, ReLU is the activation function, Concat is the concatenation operation by scale, and M l Let l be the pixel-level uncertainty map corresponding to the l-th scale feature, where l = 1, 2, ..., L.

7. The method according to claim 1, characterized in that, Step S3 includes: Step S31, for the multi-scale features F of the preprocessed image l Using the perturbation information, corresponding perturbation features are generated. l = 1, 2, ..., L; Step S32, use the pixel-level uncertainty map to analyze the multi-scale feature F. l and their corresponding disturbance characteristics The modulated features and their corresponding perturbation features are processed by adaptive pooling and convolution mapping, and then concatenated along the channel dimension to obtain fused features and fused perturbation features.

8. The method according to claim 7, characterized in that, Step S32 includes the following steps: Step S321, using the pixel-level uncertainty map, the multi-scale feature F l and their corresponding disturbance characteristics Modulation was performed separately to obtain the modulation multi-scale features G. l and their corresponding modulation perturbation characteristics :G l =F l ⊙(1+M l ), Among them, M l The pixel-level uncertainty map corresponding to the l-th scale feature; step S322, using the scale feature distiller d l (·) For the modulation multi-scale feature G l and their corresponding modulation perturbation characteristics Adaptive pooling and convolutional mapping are performed to obtain the distilled multi-scale features z. l and its corresponding distillation perturbation characteristics :z l =d l (G l )=Conv(ReLU(Conv(GAP(G l )))), Where Conv represents convolution operation, ReLU represents activation function, and GAP represents global average pooling; step S323, the distilled multi-scale features z are processed through a fully connected network. l and its corresponding distillation perturbation characteristics A nonlinear mapping is performed to obtain the fused feature y and the fused perturbation feature. :y=W2·GELU(W1·Concat(z 1 ,…,from l ,…,from L )+b1)+b2, Where W1 and W2 are the weights of the fully connected layer, b1 and b2 are the biases of the fully connected layer, GELU is the activation function, and Concat is the concatenation operation by scale.

9. The method according to claim 1, characterized in that, In step S4, the total loss value The following formula is used to calculate:

10. The method according to claim 1, characterized in that, In step S4, the model parameters of the segmentation model are iteratively optimized based on the obtained total loss value to obtain the optimal segmentation model: Where θ′ is the optimized model parameter, θ is the model parameter of the previous cycle, and ρ is the learning rate.

Images