Intravascular ultrasound image segmentation method based on fuzzy fusion and contrast distillation

Abstract

The invention discloses an intravascular ultrasound image segmentation method based on fuzzy fusion and contrast distillation, and the method comprises the steps: carrying out the preprocessing of an obtained intravascular ultrasound image, carrying out the multi-scale feature extraction of the preprocessed image, and generating a pixel-level uncertainty map corresponding to each scale feature; constructing a fusion uncertainty map generation module, and realizing self-adaptive weighted fusion of the cross-scale pixel-level uncertainty map to obtain a fusion uncertainty map; constructing a fusion feature generation module, and generating fusion features and fusion disturbance features; constructing a multi-objective optimization module to obtain an optimal segmentation model; and segmenting the acquired intravascular ultrasound image to obtain a final segmentation result. According to the method, the segmentation precision and robustness of the intravascular ultrasonic image can be improved, the problems of inaccurate boundary positioning and region misjudgment caused by noise interference, fuzzy structure and low-efficiency multi-scale feature fusion in intravascular ultrasonic image segmentation are solved, and high-precision analysis of complex pathological features is realized.

Description

Technical Field

[0001] The present invention belongs to the fields of medical image processing and computer vision, and particularly relates to an intravascular ultrasound image segmentation method 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 are achieved. Background Art

[0002] Intravascular ultrasound (IVUS), as a core technology for the diagnosis and treatment of coronary heart disease, can obtain real-time cross-sectional images of coronary arteries through an intracavitary catheter, clearly showing the vascular wall structure and plaque components, and providing key guidance for interventional treatment. Precise segmentation of the vascular lumen and plaque boundaries is the basis for calculating important clinical parameters such as the vascular stenosis rate and evaluating plaque stability. However, the unique speckle noise, artifact interference in IVUS images, and the high complexity of vascular structures (such as bifurcation lesions and calcified plaques) pose severe challenges to traditional segmentation methods.

[0003] Early IVUS segmentation research mainly relied on traditional image processing methods. The threshold segmentation method uses gray-scale differences to distinguish vascular structures, but is sensitive to noise and cannot handle heterogeneous plaques; the region growing method depends on seed point selection and is easily affected by fuzzy boundaries, resulting in over-segmentation; algorithms based on edge detection (such as Canny and active contour models) can locate some vascular walls, but it is difficult to handle the texture similarity in high-density calcified regions and thrombus formation. Traditional methods generally have problems such as frequent manual parameter adjustment, poor model generalization, and the segmentation accuracy being limited by the operator's experience. With the breakthrough of 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 and obtain better results than traditional algorithms on some standardized datasets. However, there are still significant limitations in practical applications: the low signal-to-noise ratio of IVUS images causes the model to easily confuse normal vascular tissue with noise artifacts, especially with insufficient sensitivity to thin fibrous plaques and micro-dissections; existing models lack quantitative evaluation of the credibility of segmentation results, are prone to misclassification in regions with fuzzy boundaries, and cannot provide uncertainty cues; the fusion strategy of multi-level features fails to effectively utilize spatial context associations and ignores the complementary representation of different scale features on vascular morphology; the supervised learning mode that relies on a large amount of labeled data is difficult to adapt to the data distribution differences in clinical scenarios (such as different ultrasound probe frequencies and patient-specific vascular morphologies), and the cross-device segmentation performance significantly decreases.

[0004] In recent years, uncertainty modeling and self-supervised learning have provided new ideas for medical image segmentation. Uncertainty estimation methods identify regions with low confidence through probability prediction or Monte Carlo sampling. However, existing techniques usually use it as post-processing rather than embedding it into the feature learning process, making it impossible to correct biases at the feature level. Contrastive learning enhances feature discriminability by constructing positive and negative sample pairs. However, in the IVUS scenario, it is necessary to solve the problem of negative sample confusion caused by the similarity of vascular structures. In addition, existing distillation methods mainly focus on model compression and do not explore their role in unifying the features of ultrasound image modalities. Based on this, there is an urgent need to construct a new IVUS segmentation framework that integrates uncertainty and contrast mechanisms to break through the performance bottleneck of current technologies under complex pathological conditions. Summary of the Invention

[0005] To address the deficiencies in the prior art, the object of the present invention is to provide an intravascular ultrasound image segmentation method 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 localization and regional misjudgment caused by noise interference, structural blurring, and inefficient multi-scale feature fusion in intravascular ultrasound image segmentation, and achieves high-precision analysis of complex pathological features.

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

[0007] An intravascular ultrasound image segmentation method based on fuzzy fusion and contrastive distillation, comprising:

[0008] Step S1: Preprocess the acquired intravascular ultrasound image, perform multi-scale feature extraction on the preprocessed image based on the Res2Net-50 network, and generate a pixel-level uncertainty map corresponding to each scale feature based on deep evidence learning;

[0009] Step S2: Construct a fused uncertainty map generation module. The fused uncertainty map generation module adaptively weights and fuses cross-scale pixel-level uncertainty maps based on spatial attention by introducing learnable scale weight parameters to obtain a fused uncertainty map for calculating the pixel classification loss value

[0010] Step S3: Construct a fused feature generation module. The fused feature generation module generates corresponding perturbed features based on the multi-scale features of the preprocessed image, and modulates the multi-scale features and their corresponding perturbed features using the pixel-level uncertainty map to generate fused features and fused perturbed features for calculating 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 the contrastive distillation loss value The weighted sum of them is used to obtain the total loss value, and the total loss value is used to perform joint optimization training on the segmentation model composed of the Res2Net-50 network, the fusion uncertainty map generation module and the fusion feature generation module to obtain the optimal segmentation model;

[0012] Step S5, using the optimal segmentation model, segment the collected intravascular ultrasound image to obtain the final segmentation result.

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

[0014] Step S11, perform image scaling and linear normalization on the obtained intravascular ultrasound image to make it meet the input size constraint of the segmentation model;

[0015] Step S12, based on the Res2Net-50 network, extract the features of L different stages of the preprocessed image as its multi-scale features F l , where l = 1, 2,..., L. For each scale feature, use the evidence generation layer to generate its corresponding pixel-level uncertainty map.

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

[0017]

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

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

[0020] Step S21, use the learnable scale weight parameter to perform adaptive weighted fusion on the pixel-level uncertainty map to obtain the initial fusion uncertainty map;

[0021] Step S22, based on the spatial attention map, perform dynamic calibration on the initial fusion uncertainty map in the spatial dimension to obtain the final fusion uncertainty map.

[0022] In an embodiment of the present invention, in the step S21, the following formula is used to perform adaptive weighted fusion on the pixel-level uncertainty map to obtain the initial fusion uncertainty map M' fused :

[0023]

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

[0025] In an embodiment of the present invention, in the step S22, the final fused uncertainty map M is obtained by the following formula f :

[0026] M f = M′ fused ⊙ A,

[0027] where M' fused is the initial fused uncertainty map, ⊙ represents element-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 according to the number of scales, and M l is the pixel-level uncertainty map corresponding to the l-th scale feature, where l = 1, 2, …, L.

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

[0030] Step S31, for the multi-scale features F l of the preprocessed image, generate the corresponding perturbed features by means of perturbation information

[0031] Step S32, use the pixel-level uncertainty map to modulate the multi-scale features F l and their corresponding perturbed features respectively, perform adaptive pooling and convolution mapping processing on the modulated multi-scale features and their corresponding perturbed features, and concatenate them along the channel dimension to obtain the fused features and the fused perturbed features.

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

[0033] Step S321, use the pixel-level uncertainty map to modulate the multi-scale features F land its corresponding perturbation features Modulate them respectively to obtain the modulated multi-scale feature G l and its corresponding modulated perturbation feature

[0034] G l = F l ⊙ (1 + M l ),

[0035]

[0036] where M l is the pixel-level uncertainty map corresponding to the l-th scale feature;

[0037] Step S322, use the scale feature distiller d l (·) to perform adaptive pooling and convolutional mapping processing on the modulated multi-scale feature G l and its corresponding modulated perturbation feature to obtain the distilled multi-scale feature z l and its corresponding distilled perturbation feature

[0038]

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

[0040] Step S323, perform non-linear mapping on the distilled multi-scale feature z l and its corresponding distilled perturbation feature 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 according to the number of scales.

[0044] In an embodiment of the present invention, in the step S4, the total loss value is calculated using the following formula:

[0045]

[0046] In one embodiment of the present invention, in step S4, based on the obtained total loss value, the model parameters of the segmentation model are iteratively optimized to obtain an 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 effect of the present invention is 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 contrastive distillation mechanism, it solves the problems of inaccurate boundary localization and regional misjudgment caused by noise interference, structural blurring, and inefficient multi-scale feature fusion in intravascular ultrasound image segmentation, and realizes high-precision analysis of complex pathological features. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

[0052] Figure 3 is a schematic diagram of the segmentation result of the intravascular ultrasound test image according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] The following further describes the present application 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 used to limit the protection scope of the present application.

[0054] Figure 1 is a flowchart of a method for segmenting intravascular ultrasound images based on fuzzy fusion and contrastive distillation according to an embodiment of the present invention, as Figure 1 shown, the method for segmenting intravascular ultrasound images based on fuzzy fusion and contrastive distillation includes the following steps:

[0055] Step S1, preprocess the obtained intravascular ultrasound image, perform multi-scale feature extraction on the preprocessed image based on the Res2Net-50 network, and generate a pixel-level uncertainty map corresponding to each scale feature based on deep evidence learning;

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

[0057] Step S11: Scale and linearly normalize the acquired intravascular ultrasound images to meet the input size constraints of the segmentation model;

[0058] In an embodiment of the present invention, the dataset contains a total of 1800 intravascular ultrasound image data. An example of the original intravascular ultrasound image is as Figure 2 shown. The dataset is randomly divided into a training set and a test set according to a preset ratio (such as 8:2). The segmentation label for each image is 2 classes (intima and media), that is, the intravascular ultrasound image is labeled as the intima part and the media part. Figure 2 The label image corresponding to the example of the original intravascular ultrasound image is shown, 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 the features of L different stages of the preprocessed image as its multi-scale features F l , where l = 1, 2, …, L. In an embodiment of the present invention, L = 4. For each scale feature, an evidence generation layer is used to generate its corresponding pixel-level uncertainty map.

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

[0061]

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

[0063] Step S2: Construct a fused uncertainty map generation module. The fused uncertainty map generation module adaptively weights and fuses the cross-scale pixel-level uncertainty maps based on spatial attention by introducing learnable scale weight parameters to obtain a fused uncertainty map for calculating the pixel classification loss value.

[0064] Among them, the step S2 includes:

[0065] Step S21, using the learnable scale weight parameter to perform adaptive weighted fusion on the pixel-level uncertainty map to obtain an initial fusion uncertainty map;

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

[0067]

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

[0069] Step S22, based on the spatial attention map, perform dynamic calibration on the initial fusion uncertainty map in the spatial dimension to obtain the final fusion uncertainty map.

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

[0071] M f = M′ fused ⊙A,

[0072] where ⊙ represents element-wise multiplication for realizing dynamic calibration in the spatial dimension, 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 according to the number of scales.

[0075] Step S3, construct a fusion feature generation module, which generates corresponding perturbation features based on the multi-scale features of the preprocessed image, and uses the pixel-level uncertainty map to modulate the multi-scale features and their corresponding perturbation features to generate fusion features and fusion perturbation features for calculating the contrastive distillation loss value

[0076] Among them, the step S3 includes:

[0077] Step S31: For the multi-scale feature F of the preprocessed image l , generate a corresponding perturbed feature by means of perturbation information

[0078] Among them, the perturbation information can be, for example, perturbation information such as random noise, or perturbation information generated by perturbation operations such as Dropout or masks. In an embodiment of the present invention, the perturbation information generated by the Dropout perturbation operation is used for the multi-scale feature F l to perform perturbation to generate a corresponding perturbed feature Among them, the perturbation ratio, that is, the dropout ratio, is 0.2.

[0079] Step S32: Use the pixel-level uncertainty map to respectively modulate the multi-scale feature F l and its corresponding perturbed feature After the modulated multi-scale feature and its corresponding perturbed feature are processed by adaptive pooling and convolutional mapping, and concatenated along the channel dimension, a fused feature and a fused perturbed feature are obtained.

[0080] Among them, the step S32 further includes the following steps:

[0081] Step S321: Use the pixel-level uncertainty map to respectively modulate the multi-scale feature F l and its corresponding perturbed feature to obtain a modulated multi-scale feature G l and its corresponding modulated perturbed feature

[0082] G l = F l ⊙ (1 + M l )

[0083]

[0084] Step S322: Use the scale feature distiller d l (·) to perform adaptive pooling and convolutional mapping processing on the modulated multi-scale feature G l and its corresponding modulated perturbed feature to obtain a distilled multi-scale feature z l and its corresponding distilled perturbed feature

[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, perform a non - linear mapping on the distilled multi - scale feature z l and its corresponding distilled perturbation feature to obtain a fused feature y and a 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, and obtain the total loss value by calculating the weighted sum of the original segmentation loss value pixel classification loss value and the contrastive distillation loss value Perform joint optimization training on the segmentation model composed of the Res2Net - 50 network, the fused uncertainty map generation module, and the fused feature generation module using the total loss value to obtain the optimal segmentation model;

[0093] where the original segmentation loss value can be calculated according to the following formula:

[0094]

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

[0096] where the pixel classification loss value can be calculated based on the fused uncertainty map:

[0097]

[0098] Among them, K is the number of segmentation categories, p represents a pixel point, Ω is the set of pixel points, γ is a hyperparameter used to control the intensity of uncertainty enhancement, and y c (p) is the true pixel label value, is the predicted pixel label value.

[0099] In an embodiment of the present invention, the hyperparameter γ for controlling the intensity of uncertainty enhancement can be set to 0.5.

[0100] Among them, the contrastive distillation loss value can be calculated based on the fused feature and the fused perturbation feature:

[0101]

[0102] P = {(i, j)|S ij >U(S ij ), i ≠ j},

[0103] Among them, P represents the set of positive sample pairs, represents the similarity between the fused feature of the i-th sample and the fused feature of the j-th sample, represents the similarity between the fused perturbation feature of the i-th sample and the fused perturbation feature of the j-th sample, represents the similarity between the fused perturbation feature of the i-th sample and the fused perturbation feature of the k-th sample, y and are both normalized using the L2 norm first, T represents the transpose operation, U(·) represents finding the matrix triangular median, B is the batch size, and τ is the temperature coefficient for controlling the steepness of the probability distribution.

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

[0105] Among them, the total loss value can be calculated using the following formula:

[0106]

[0107] Iteratively optimize the model parameters of the segmentation model based on the obtained total loss value to obtain the optimal segmentation model:

[0108]

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

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

[0111] It should be noted that the training method of the above segmentation model and the parameter configuration of the learning rate optimization strategy are a non-restrictive and optimal choice. Those skilled in the art can select the training method of the segmentation model and the parameter configuration scheme according to various indicators such as recognition accuracy and efficiency.

[0112] Step S5: Use the optimal segmentation model to segment the collected intravascular ultrasound image to obtain the final segmentation result.

[0113] In an embodiment of the present invention, the trained optimal segmentation model is used to segment 360 intravascular ultrasound images in the test set, and an example of the segmentation result is as Figure 3 shown. The results show that the segmentation model of this embodiment can accurately identify and segment the intima and media structures in the intravascular ultrasound image. At the same time, the fused uncertainty map can effectively indicate the low-confidence regions, thereby improving the segmentation accuracy of the segmentation model in the boundary region and complex lesion sites.

[0114] The beneficial effect of the present invention is 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 contrastive distillation mechanism, it solves the problems of inaccurate boundary localization and regional misjudgment caused by noise interference, structural blurring and inefficient multi-scale feature fusion in intravascular ultrasound image segmentation, and realizes high-precision analysis of complex pathological features.

[0115] The applicant of the present invention has made a detailed description and illustration of the embodiments of the present invention in combination with the accompanying drawings of the specification. However, those skilled in the art should understand that the above embodiments are only the preferred implementation schemes of the present invention, and the detailed description is only to help readers better understand the spirit of the present invention, rather than a limitation on the protection scope of the present invention. On the contrary, any improvement or modification made based on the spirit of the present invention should fall within the protection scope of the present invention.

Claims

1. An intravascular ultrasound image segmentation method based on fuzzy fusion and contrastive distillation, characterized in that The method includes the following steps: Step S1: Preprocess the acquired intravascular ultrasound image, extract multi-scale features from the preprocessed image based on the Res2Net-50 network, and generate a pixel-level uncertainty map corresponding to each scale feature based on deep evidence learning; Step S2, construct a fused uncertainty map generation module. The fused uncertainty map generation module realizes the adaptive weighted fusion of cross-scale pixel-level uncertainty maps based on spatial attention by introducing learnable scale weight parameters, and obtains a fused uncertainty map for calculating a pixel classification loss value. Step S3, construct a fusion feature generation module. The fusion feature generation module generates corresponding perturbation features based on the multi-scale features of the preprocessed image, and uses the pixel-level uncertainty map to modulate the multi-scale features and their corresponding perturbation features 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, and calculate the original segmentation loss value pixel classification loss value and the weighted sum of the contrastive distillation loss value to obtain the total loss value, and use the total loss value to perform joint optimization training on the segmentation model composed of the Res2Net-50 network, the fusion uncertainty map generation module, and the fusion feature generation module to obtain the optimal segmentation model; Step S5: Use the optimal segmentation model to segment the acquired intravascular ultrasound image to obtain the final segmentation result.

2. The method according to claim 1, wherein The said Step S1 includes: Step S11: Scale and linearly normalize the acquired intravascular ultrasound image to meet the input size constraint of the segmentation model; Step S12: Based on the Res2Net-50 network, extract the features of L different stages of the preprocessed image as its multi-scale feature F l , where l = 1, 2, …, L. For each scale feature, use the evidence generation layer to generate its corresponding pixel-level uncertainty map.

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

4. The method according to claim 1, wherein The said Step S2 includes: Step S21: Use 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, perform dynamic calibration on the initial fused uncertainty map in the spatial dimension to obtain the final fused uncertainty map.

5. The method according to claim 4, wherein In the step S21, the pixel-level uncertainty map is adaptively weighted and fused by the following formula to obtain an initial fused uncertainty map M′ fused : where, w′ l is the l-th learnable scale weight parameter initialized randomly, M l is the pixel-level uncertainty map corresponding to the l-th scale feature, l = 1, 2, …, L, and w′ s is the s-th learnable scale weight parameter initialized randomly, s = 1, 2, …, L.

6. The method according to claim 4, characterized in that, In the step S22, the final fused uncertainty map M is obtained by the following formula f :[[]]END]] M f = M fused ⊙ A, Among them, M' fused is 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 according to the number of scales, and M l is the pixel-level uncertainty map corresponding to the feature of the l-th scale, where l = 1, 2, …, L.

7. The method according to claim 1, characterized in that, The said Step S3 includes: Step S31: For the multi-scale feature F of the preprocessed image l , generate the corresponding perturbation feature by means of perturbation information l = 1, 2, …, L; Step S32, using the pixel-level uncertainty map to modulate the multi-scale feature F l and its corresponding perturbed feature respectively, subjecting the modulated multi-scale feature and its corresponding perturbed feature to adaptive pooling and convolutional mapping processing, and concatenating them along the channel dimension to obtain a fused feature and a fused perturbed feature.

8. The method according to claim 7, wherein The said Step S32 includes the following steps: Step S321, using the pixel-level uncertainty map, modulate the multi-scale feature F l and its corresponding perturbed feature respectively to obtain the modulated multi-scale feature G l and its corresponding modulated perturbed feature : G l = F l ⊙ (1 + M l ) Among them, M l is the pixel-level uncertainty map corresponding to the l-th scale feature; Step S322, using the scale feature distiller d l (·) for the modulated multi-scale feature G l and its corresponding modulated perturbation feature perform adaptive pooling and convolutional mapping processing to obtain the distilled multi-scale feature z l and its corresponding distilled perturbation feature : z l = d l (G l ) = Conv(ReLU(Conv(GAP(G l )))) Where, Conv is the convolution operation, ReLU is the activation function, and GAP represents global average pooling; Step S323, perform a non-linear mapping on the distilled multi-scale feature z l and its corresponding distilled perturbation feature to obtain a fused feature y and a fused perturbation feature : y = W2·GELU(W1·Concat(z 1 ,…,z l ,…,z 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 according to the number of scales.

9. The method according to claim 1, wherein In the step S4, the total loss value is calculated by the following formula:

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

Images