A medical image segmentation method and system based on guidance information and multi-dimensional attention mechanism

Abstract

The application discloses a medical image segmentation method and system based on guide information and a multi-dimensional attention mechanism, which comprises the following steps: collecting original dermoscope images for preprocessing; constructing a segmentation model, including a double-path image encoder, a guide information encoder and a mask decoder; the double-path image encoder is used for extracting local detail features and global context semantic information in the image; the guide information encoder is used for converting the coarse segmentation mask predicted by the previous round of network into guide feature information; the mask decoder fuses image features and guide feature information, gradually restores and refines the coarse-grained feature map, and finally outputs an accurate lesion segmentation mask; a loss function is constructed, and the segmentation model is trained by using the preprocessed data; and the original dermoscope image to be segmented is input into the trained segmentation model, and a lesion area segmentation mask of the image is output. The application can improve the segmentation precision and the model generalization ability, and enhance the processing capacity for multi-scale lesions.

Description

Technical Field

[0001] This invention belongs to the field of medical image segmentation, and in particular relates to a medical image segmentation method and system based on guidance information and multidimensional attention mechanism. Background Technology

[0002] Skin cancer is one of the most common types of cancer worldwide, with melanoma receiving particular attention due to its highly aggressive nature and high mortality rate. In recent years, the incidence and mortality rates of melanoma have been on the rise, especially in areas with strong ultraviolet radiation, where the incidence remains consistently high. Current diagnostic procedures primarily rely on dermatologists manually interpreting dermoscopic images, a process that is not only time-consuming and labor-intensive but also demands a high level of expertise and experience from physicians. With the continuous increase in the number of skin disease cases, traditional manual diagnostic methods face serious challenges in terms of efficiency and accuracy, failing to meet the clinical requirements for timeliness and consistency.

[0003] Furthermore, during image segmentation, subjective human factors often lead to differing interpretations of the same image among different observers, affecting the consistency and reliability of diagnostic results. Therefore, there is an urgent need for an efficient, accurate, and standardized automatic image segmentation method to alleviate the workload of medical personnel and improve diagnostic quality and efficiency.

[0004] Chinese patent document CN116912270A discloses a method and system for skin disease image segmentation based on an improved Unet network; Chinese patent document CN119672041A discloses a method for skin lesion segmentation based on error localization and adaptive optimization.

[0005] Both of the above schemes use a typical U-shaped segmentation network as their basic framework. In the encoder stage, both schemes combine the feature extraction capabilities of convolutional neural networks (CNNs) and Transformers, progressively extracting deep semantic features of the image through multi-layer downsampling operations. In the decoder stage, skip connections are used to fuse the feature maps of corresponding layers of the encoder into the decoding path, while spatial resolution is gradually restored through layer-by-layer upsampling to achieve accurate segmentation of skin lesion areas.

[0006] However, existing methods for segmenting skin lesion images still have shortcomings in terms of accuracy, robustness, and adaptability to complex lesion morphologies, specifically in the following aspects:

[0007] 1. Limited segmentation accuracy: Although existing methods have combined CNN and Transformer structures, they are still prone to omissions or misclassifications when dealing with skin lesions with blurred boundaries, similar colors, or complex shapes, which affects the accuracy of the final diagnosis.

[0008] 2. Insufficient utilization of contextual information: Although the traditional U-shaped structure has a certain global receptive field, it is still insufficient for unified modeling of large-scale semantic associations and fine-grained features, and cannot fully capture the contextual semantic information of the skin lesion area.

[0009] 3. Weak adaptability to multi-scale lesions: Skin lesions vary significantly in size, shape and distribution. Existing segmentation models have poor generalization ability when faced with multi-scale targets, making it difficult to balance fine segmentation of small targets with complete recognition of large targets. Summary of the Invention

[0010] This invention provides a medical image segmentation method and system based on guidance information and multidimensional attention mechanism, aiming to improve segmentation accuracy and model generalization ability, enhance the ability to process lesions at multiple scales, and reduce computational overhead and improve inference efficiency while ensuring performance, thereby better serving the auxiliary diagnosis and treatment decision support of skin diseases.

[0011] A medical image segmentation method based on guidance information and multidimensional attention mechanism, characterized by comprising:

[0012] (1) Collect raw dermoscopy images and perform preprocessing;

[0013] (2) Construct a segmentation model, which includes a dual-channel image encoder, a guiding information encoder, and a mask decoder;

[0014] The dual-channel image encoder integrates a deep convolutional neural network and a Transformer structure to extract local detail features and global contextual semantic information from dermoscopy images. The guidance information encoder converts the coarse segmentation mask predicted by the previous network into guidance feature information, guiding the model to focus on potential lesion areas. The mask decoder is responsible for fusing the image features output by the dual-channel image encoder with the guidance feature information output by the guidance information encoder, gradually restoring and refining the coarse-grained feature map, and finally outputting an accurate lesion segmentation mask.

[0015] (3) Construct a loss function and train the segmentation model using the preprocessed data;

[0016] (4) Input the original dermoscopy image to be segmented into the trained segmentation model and output the lesion area segmentation mask of the image.

[0017] In step (1), the preprocessing includes:

[0018] The image was downsampled to 512×512 pixels, denoised, and then the pixel values ​​were normalized. Data augmentation was performed by rotating, flipping, color changing, and adding adversarial noise.

[0019] In step (2), the dual-channel image encoder includes two independent branches. The first branch is composed of multiple cascaded deep convolutional modules (CNNs) for extracting local detail features in the dermoscopy image. The second branch is composed of multiple cascaded Transformer modules for capturing global contextual information in the dermoscopy image.

[0020] After fusing the local detail features output from the two branches with the global context information, the feature is further input into the triple attention module to obtain image features.

[0021] The triple attention module models the fusion features from three dimensions: channel attention, spatial horizontal attention, and spatial vertical attention.

[0022] In each dimension, firstly, max pooling and average pooling operations are performed on the input fusion features to obtain two sets of feature representations; then, the two sets of feature representations are concatenated and fed into a shared convolution operation to extract attention maps; finally, normalized attention weights are generated through the Sigmoid activation function and then element-wise weighted and fused with the original input fusion features to complete the attention enhancement for the corresponding dimension.

[0023] In step (2), the operation of the guidance information encoder is as follows:

[0024] First, the coarse segmentation mask output from the mask decoder in the previous round of the network is input into the dense embedding representation module. This module consists of two concatenated convolutional sub-modules. Each convolutional sub-module includes the following steps: the input features are first spatially downsampled through a convolutional layer; then they are processed sequentially through layer normalization and the GELU activation function to enhance the feature representation capability; after the two convolutional sub-modules, a 1×1 convolutional layer is used to map the final features to the preset embedding dimension to generate the final guiding feature information.

[0025] In step (2), the mask decoder includes multiple cascaded decoding subunits. Each decoding subunit includes a parallel Transformer module and a triple attention module, and the outputs of the two modules are fused.

[0026] The input to the mask decoder is the image features obtained by the dual-channel image encoder and the guiding feature information obtained by the guiding information encoder. After the image features and guiding feature information are concatenated, they are input into the first decoding subunit to obtain the corresponding fused features, and then input into the next decoding subunit. The fused features output by the last decoding subunit are entered into the upsampling module for spatial resolution restoration and output a fine segmentation mask.

[0027] The upsampling module includes two sequentially connected transposed convolutional layers. The first layer uses a transposed convolution with a kernel size of 2×2 and a stride of 2 to achieve initial upsampling, and then completes feature normalization and nonlinear transformation through layer normalization and the GELU activation function. The second layer also uses a transposed convolution with a kernel size of 2×2 and a stride of 2 for further upsampling, and is then input into the multilayer perceptron module after GELU activation to output a fine segmentation mask with the same size as the original dermoscopy image.

[0028] In step (3), the loss function constructed is:

[0029]

[0030] In the formula, λ1, λ2, and λ3 represent weighting coefficients. For Dice's loss; For binary cross-entropy loss; Uncertainty-weighted boundary loss.

[0031] The process of constructing the uncertainty-weighted boundary loss is as follows:

[0032] The uncertainty graph is defined as the prediction variance of the segmentation model at each pixel location:

[0033] U(x) = Var(p1(x),p2(x),…,p n (x));

[0034] In the formula, x represents the position of a pixel in the image; p i (x) represents the prediction probability of the pixel in the i-th forward propagation of the segmentation model, which is the probability that the pixel belongs to the foreground; n represents the number of inferences performed on the model, i.e., the n predictions generated using different perturbations (e.g., Dropout, data augmentation, or model ensemble); Var(.) represents the variance of the n predictions at pixel position x, which is used to measure the uncertainty of the prediction result at that position; U(x) is the uncertainty map generated in the end, which is used to assign different weights to the boundary regions in the loss function.

[0035] Let B(x) be the boundary mask extracted from the true mask by the Sobel or Morphological method, then the uncertainty-weighted boundary loss is defined as:

[0036]

[0037] In the formula, B(x) represents the indicator function for whether pixel x is in the boundary region, with a value of 1 for the boundary region and 0 for the non-boundary region; g(x) is the absolute value of the true label prediction error of pixel x, and |p(x)-g(x)| measures the difference between the prediction and the true label.

[0038] A medical image segmentation system based on guidance information and multidimensional attention mechanism includes a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they implement the above-mentioned medical image segmentation method.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. This invention constructs a hybrid decoder architecture that integrates guidance information: by embedding previous prediction results into the guidance information encoder, the focus on the target region is enhanced. This achieves coarse-fine two-stage segmentation coupling, effectively improving the model convergence speed and localization accuracy.

[0041] 2. This invention introduces a triple attention mechanism decoder that combines Transformer and CNN: It adopts a two-way Transformer cross-attention mechanism and a triple attention module (channel-space-scale) for joint modeling, which strengthens the expression of structural hierarchical features and improves the anatomical rationality and boundary sensitivity of segmentation.

[0042] 3. The present invention designs a loss function with boundary sensitivity and uncertainty perception capabilities: an uncertainty-weighted boundary loss is introduced and jointly optimized with Dice+BCE, which focuses on penalizing prediction errors in areas with blurred boundaries and uncertainty, effectively enhancing the stability and robustness of the model in structurally complex regions. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the segmentation model in an embodiment of the present invention.

[0044] Figure 2 This is a schematic diagram of the triple attention module in an embodiment of the present invention.

[0045] Figure 3 This is a schematic diagram of the upsampling module in an embodiment of the present invention. Detailed Implementation

[0046] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not constitute any limitation thereof.

[0047] A medical image segmentation method based on guidance information and multidimensional attention mechanism is proposed. This method utilizes deep learning technology to automatically segment lesion areas in dermoscopy images, thereby reducing the workload of medical staff and improving the efficiency and accuracy of subsequent diagnoses.

[0048] The method of this invention first preprocesses the original dermoscopy image. The preprocessing includes: downsampling the image to a standard size (512×512 pixels); performing noise reduction using Gaussian filtering, median filtering, and other methods; normalizing the image pixel values; and performing data augmentation through rotation, flipping, color transformation, and adding adversarial noise to improve the model's generalization ability.

[0049] The preprocessed image is input into the segmentation model constructed in this invention for lesion region segmentation. The structure of the segmentation model is as follows: Figure 1 As shown, it consists of three parts: a dual-channel image encoder, a guiding information encoder, and a mask decoder. Each module has a clear division of labor and works together to achieve robust and high-precision segmentation results.

[0050] The dual-channel image encoder integrates a deep convolutional neural network and a Transformer structure, enabling it to fully extract contextual semantic information and fine-grained lesion structural features from dermoscopic images. The guidance information encoder models coarse mask guidance information and transforms it into sparse and dense feature representations, thus providing a regional attention mechanism for segmentation tasks. The mask decoder is responsible for fusing image features and guidance feature information, progressively restoring and refining the coarse-grained feature map, and ultimately outputting an accurate lesion segmentation mask.

[0051] The dual-path image encoder comprises two independent branches. The first branch consists of multiple deep convolutional modules used to extract local detail features from dermoscopy images. The second branch consists of multiple cascaded Transformer modules capable of capturing global contextual information in the image. This dual-branch structure enables the collaborative extraction of local and global features, effectively improving the comprehensiveness and discriminative power of feature representation.

[0052] The deep convolutional network (CNN) module employs a residual network structure, achieving efficient feature extraction through stacked convolutional layers, batch normalization layers, and non-linear activation functions (such as ReLU or GELU). The Transformer module utilizes a scaled dot product attention mechanism with relative positional bias to construct a multi-head self-attention structure, effectively capturing long-distance dependencies between different locations in the image. In this structure, the introduction of relative positional encoding enhances the model's perception of spatial location information, thereby improving the accuracy and expressive power of global feature modeling. The Transformer module further combines residual connections and layer normalization to ensure stability during training and enhance the network's non-linear expressive power.

[0053] The first branch, composed of deep convolutional modules, and the second branch, composed of Transformer modules, perform feature fusion at the bottom of the encoder to effectively integrate local and global feature information. The fused features are further input into the Triplet Attention module to enhance the expressive power of key region features.

[0054] Triple attention module, such as Figure 2 As shown, the fusion features are modeled from three dimensions: channel attention, spatial horizontal attention, and spatial vertical attention. This fully extracts the importance information of different dimensions in the image, thereby improving the recognition and segmentation accuracy of skin lesion areas. In each dimension, the module first performs max pooling and average pooling operations on the input features to obtain two sets of feature representations. Then, the pooled features are concatenated along the corresponding direction and fed into a shared convolution operation to extract the attention map. Finally, normalized attention weights are generated through the Sigmoid activation function and fused with the original input features element-wise to complete the attention enhancement of the corresponding dimension.

[0055] The guided information encoder proposed in this invention aims to fully utilize the coarse segmentation results obtained in the previous round of network prediction, using them as dynamic guided information input to guide the model to focus more efficiently on potential lesion areas, thereby accelerating model convergence and improving the segmentation accuracy at the current stage. Figure 1 As shown, the overall processing flow of the guide information encoder is as follows:

[0056] First, the coarse segmentation mask output from the mask decoder in the previous round is input into the dense embedding representation module. This module consists of two cascaded convolutional sub-modules. Each convolutional sub-module includes the following steps: the input features are first spatially downsampled through a convolutional layer; then processed sequentially through layer normalization and the GELU activation function to enhance feature representation; after the two convolutional sub-modules, a 1×1 convolutional layer maps the final features to a preset embedding dimension, generating the final guiding feature information. Through this structural design, the guiding information encoder achieves an effective conversion from the coarse segmentation mask to the semantic embedding representation, which not only significantly enhances the model's ability to perceive the target region but also provides context-dependent prior knowledge support for the entire decoding process.

[0057] The mask decoder proposed in this invention adopts a fusion architecture that combines the advantages of Transformer and Convolutional Neural Network (CNN), aiming to recover high-resolution volume segmentation masks from image feature embedding and guiding information embedding.

[0058] The input to the mask decoder includes: (1) image feature embeddings extracted by the dual-path image encoder, and (2) dense guided embedding representations generated by the guided information encoder. To fully integrate global context and local structural information, the decoder integrates a Transformer-based self-attention mechanism and a CNN-based triple attention mechanism into a unified framework. This design not only effectively models long-range dependencies but also enhances the model's ability to discriminate complex anatomical boundaries and fine-grained regions, thereby improving the representational power and segmentation accuracy during the decoding stage.

[0059] After feature enhancement is complete, the fused features are fed into the upsampling module for spatial resolution restoration. The upsampling module is as follows: Figure 3 As shown, it includes two sequentially connected transposed convolutional layers: the first layer uses a transposed convolution with a kernel size of 2×2 and a stride of 2 to achieve initial upsampling, and then completes feature normalization and nonlinear transformation through layer normalization and GELU activation function; the second layer also uses a transposed convolution with a kernel size of 2×2 and a stride of 2 for further upsampling, and is then input into a multilayer perceptron (MLP) module after GELU activation to output a fine segmentation mask with the same size as the original image.

[0060] This decoder structure fully combines the global modeling capabilities of Transformer with the efficient local enhancement features of CNN. With the guidance mechanism and triple attention module, it effectively improves the model's sensitivity to lesion areas and overall segmentation performance.

[0061] To improve the robustness and accuracy of the model in volumetric medical image segmentation tasks, this invention employs a boundary-uncertainty joint guided loss function to supervise and optimize the difference between the predicted results and the true labels. This loss function consists of three parts: a basic Dice loss to handle foreground / background imbalance; a binary cross-entropy to optimize pixel classification; and an uncertainty-weighted boundary loss, which is the innovative core of the loss function.

[0062]

[0063] Traditional boundary loss only considers edge regions, but in medical images, boundaries are often accompanied by uncertainty (such as blurred transitions between lesions and normal tissue). This invention introduces a weighting mechanism based on uncertainty maps at the boundary locations, making the network pay more attention to blurred boundary regions.

[0064] The uncertainty graph is defined as the prediction variance of the model at each pixel location.

[0065] U(x) = Var(p1(x),p2(x),…,p n (x));

[0066] Let B(x) be the boundary mask extracted from the true mask by the Sobel or Morphological method, then the boundary loss is defined as:

[0067]

[0068] This loss encourages the model to focus on learning in high-uncertainty boundary regions, thereby improving the prediction accuracy of lesion contours.

[0069] In practical applications, the workflow of the guided information encoder is as follows:

[0070] Initial stage (round 1): The input image is predicted using the basic segmentation model to generate a preliminary coarse segmentation mask; since no guiding information is available at this time, the input of the guiding information encoder is empty.

[0071] Subsequent training phase (round n): Use the model from the previous round (round n-1) to infer the same image and obtain a coarse segmentation mask; then sample the coarse segmentation mask to generate a guiding mask;

[0072] The guiding mask serves as the guiding information input to the guiding information encoder. It is input into the model along with the current round of image features, guiding the model to focus on key regions, thereby optimizing the segmentation quality, especially for boundary and complex regions.

[0073] This invention innovatively introduces a guided information encoder, using the coarse segmentation mask generated in the previous stage as the guided input, and constructs a dense guided embedding through convolution, layer normalization, and GELU activation. This embedding is jointly modeled with image features, significantly enhancing the model's ability to focus on lesion regions and providing effective constraints in cases of early small lesions or blurred boundaries.

[0074] This invention designs a hybrid decoder that integrates Transformer and CNN, which integrates a Transformer module and a triple attention module design. The triple attention module combines channel attention, spatial attention and scale attention modules to fully explore the salient features of multidimensional information, effectively enhance the response of lesion areas, and improve the precision and robustness of segmentation results.

[0075] To overcome the low accuracy of traditional Dice / BCE in boundary and uncertain regions, this invention proposes a joint loss function, where the weights of each component can be flexibly adjusted according to the actual task. This loss function is particularly suitable for medical image segmentation scenarios with blurred boundaries and class imbalance, significantly improving the model's performance in difficult-to-segment regions.

[0076] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A medical image segmentation method based on guiding information and multidimensional attention mechanism, characterized in that, include: (1) Collect raw dermoscopy images and perform preprocessing; (2) Construct a segmentation model, which includes a dual-channel image encoder, a guiding information encoder, and a mask decoder; The dual-channel image encoder integrates a deep convolutional neural network and a Transformer structure to extract local detail features and global contextual semantic information from dermoscopy images. The guidance information encoder converts the coarse segmentation mask predicted by the previous network into guidance feature information, guiding the model to focus on potential lesion areas. The mask decoder is responsible for fusing the image features output by the dual-channel image encoder with the guidance feature information output by the guidance information encoder, gradually restoring and refining the coarse-grained feature map, and finally outputting an accurate lesion segmentation mask. (3) Construct a loss function and train the segmentation model using the preprocessed data; (4) Input the original dermoscopy image to be segmented into the trained segmentation model and output the lesion area segmentation mask of the image.

2. The medical image segmentation method based on guidance information and multidimensional attention mechanism according to claim 1, characterized in that, In step (1), the preprocessing includes: The image was downsampled to 512×512 pixels, denoised, and then the pixel values ​​were normalized. Data augmentation was performed by rotating, flipping, color changing, and adding adversarial noise.

3. The medical image segmentation method based on guidance information and multidimensional attention mechanism according to claim 1, characterized in that, In step (2), the dual-channel image encoder includes two independent branches. The first branch is composed of multiple cascaded deep convolutional modules (CNNs) for extracting local detail features in the dermoscopy image. The second branch is composed of multiple cascaded Transformer modules for capturing global contextual information in the dermoscopy image. After fusing the local detail features output from the two branches with the global context information, the feature is further input into the triple attention module to obtain image features.

4. The medical image segmentation method based on guidance information and multidimensional attention mechanism according to claim 3, characterized in that, The triple attention module models the fusion features from three dimensions: channel attention, spatial horizontal attention, and spatial vertical attention. In each dimension, firstly, max pooling and average pooling operations are performed on the input fusion features to obtain two sets of feature representations; then, the two sets of feature representations are concatenated and fed into a shared convolution operation to extract attention maps; finally, normalized attention weights are generated through the Sigmoid activation function and then element-wise weighted and fused with the original input fusion features to complete the attention enhancement for the corresponding dimension.

5. The medical image segmentation method based on guidance information and multidimensional attention mechanism according to claim 1, characterized in that, In step (2), the operation of the guidance information encoder is as follows: First, the coarse segmentation mask output from the mask decoder in the previous round of the network is input into the dense embedding representation module. This module consists of two concatenated convolutional sub-modules. Each convolutional sub-module includes the following steps: the input features are first spatially downsampled through a convolutional layer; then they are processed sequentially through layer normalization and the GELU activation function to enhance the feature representation capability; after the two convolutional sub-modules, a 1×1 convolutional layer is used to map the final features to the preset embedding dimension to generate the final guiding feature information.

6. The medical image segmentation method based on guidance information and multidimensional attention mechanism according to claim 1, characterized in that, In step (2), the mask decoder includes multiple cascaded decoding subunits. Each decoding subunit includes a parallel Transformer module and a triple attention module, and the outputs of the two modules are fused. The input to the mask decoder is the image features obtained by the dual-channel image encoder and the guiding feature information obtained by the guiding information encoder. After the image features and guiding feature information are concatenated, they are input into the first decoding subunit to obtain the corresponding fused features, and then input into the next decoding subunit. The fused features output by the last decoding subunit are entered into the upsampling module for spatial resolution restoration and output a fine segmentation mask.

7. The medical image segmentation method based on guiding information and multidimensional attention mechanism according to claim 6, characterized in that, The upsampling module includes two sequentially connected transposed convolutional layers; the first layer uses a transposed convolution with a kernel size of 2×2 and a stride of 2 to achieve initial upsampling, and then completes feature normalization and nonlinear transformation through layer normalization and the GELU activation function; The second layer also uses a transposed convolution with a kernel size of 2×2 and a stride of 2 for further upsampling, and is then activated by GELU before being input into a multilayer perceptron module to output a fine segmentation mask with the same size as the original dermoscopy image.

8. The medical image segmentation method based on guidance information and multidimensional attention mechanism according to claim 1, characterized in that, In step (3), the loss function constructed is: In the formula, λ1, λ2, and λ3 represent weighting coefficients. For Dice's loss; For binary cross-entropy loss; Uncertainty-weighted boundary loss.

9. The medical image segmentation method based on guidance information and multidimensional attention mechanism according to claim 8, characterized in that, The process of constructing the uncertainty-weighted boundary loss is as follows: The uncertainty graph is defined as the prediction variance of the segmentation model at each pixel location: U(x)=Var(p1(x),p2(x),…,p n (x)); In the formula, x represents the position of a pixel in the image; p i (x) represents the prediction probability of pixel x by the segmentation model in the i-th forward propagation, which is the probability that the pixel belongs to the foreground; n represents the number of inferences performed on the model, i.e., n predictions generated using different perturbations; Var(.) represents the variance of the n predictions at pixel position x, which is used to measure the uncertainty of the prediction result at that position; U(x) is the uncertainty map generated in the end, which is used to assign different weights to the boundary regions in the loss function; Let B(x) be the boundary mask extracted from the true mask by the Sobel or Morphological method, then the uncertainty-weighted boundary loss is defined as: In the formula, B(x) represents the indicator function for whether pixel x is in the boundary region, with a value of 1 for the boundary region and 0 for the non-boundary region; g(x) is the absolute value of the true label prediction error of pixel x, and |p(x)-g(x)| measures the difference between the prediction and the true label.

10. A medical image segmentation system based on guiding information and multidimensional attention mechanism, characterized in that, The device includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the medical image segmentation method according to any one of claims 1-9.

Images