Medical image segmentation method and system based on guiding information and multi-dimensional attention mechanism

Abstract

The invention discloses a medical image segmentation method and system based on guidance information and a multi-dimensional attention mechanism. The method comprises the following steps: collecting an original dermatoscope image for preprocessing; constructing a segmentation model, wherein the segmentation model comprises a double-path image encoder, a guide information encoder and a mask decoder; the two-way 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 a coarse segmentation mask predicted by the last round of network into guide feature information; the mask decoder fuses the image feature information and the guide feature information, gradually restores and refines the coarse-grained feature map, and finally outputs an accurate lesion segmentation mask; constructing a loss function, and training the segmentation model by using the preprocessed data; and inputting a to-be-segmented original dermatoscope image into the trained segmentation model, and outputting a lesion region segmentation mask of the image. According to the method, the segmentation precision and the model generalization ability can be improved, and the multi-scale lesion processing ability is enhanced.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of medical image segmentation, and particularly relates to a medical image segmentation method and system based on guide information and a multi-dimensional attention mechanism. BACKGROUND

[0002] Skin cancer is one of the highest incidence types of cancer worldwide, among which melanoma is widely concerned due to its high invasiveness and high mortality rate. In recent years, the incidence and mortality of melanoma are on the rise, especially in areas with strong ultraviolet radiation, the incidence rate remains at a high level. The existing diagnosis process mainly relies on the manual interpretation of dermatoscope images by dermatologists, which not only consumes time and effort, but also requires high professional knowledge and experience of doctors. With the continuous growth of skin disease cases, traditional manual diagnosis method faces severe challenges in efficiency and accuracy, and it is difficult to meet the requirements of timeliness and consistency in clinic.

[0003] In addition, during image segmentation, due to the interference of human subjective factors, there are often differences in the interpretation of the same image between different observers, which affects the consistency and reliability of the diagnosis results. Therefore, there is an urgent need for an efficient, accurate and standardized automatic image segmentation method to alleviate the work burden of medical personnel and improve the diagnosis quality and efficiency.

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

[0005] The above two schemes both use a typical U-shaped structure segmentation network as the basic framework. Among them, the feature extraction capabilities of convolutional neural network (CNN) and Transformer structure are combined in the encoder stage, and the deep semantic features of the image are gradually extracted through multi-layer downsampling operations; in the decoder stage, the feature maps of the corresponding layers of the encoder are fused into the decoding path through the way of skip connection, and the spatial resolution is gradually restored through layer-by-layer upsampling, so as to realize the accurate segmentation of the skin lesion area.

[0006] However, the existing skin lesion image segmentation method still has deficiencies in accuracy, robustness and adaptability to complex lesion morphology, which are manifested in the following aspects:

[0007] 1. Limited segmentation accuracy: Although existing methods have combined CNN and Transformer structure, when facing boundary fuzzy, color similar or complex lesion area, it is still easy to miss or misclassify, which affects the accuracy of the final diagnosis.

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

[0009] 3. Weak adaptability to multi-scale lesions: Skin lesions have significant differences in size, shape and distribution. Existing segmentation models have poor generalization ability when facing multi-scale targets, making it difficult to balance the fine segmentation of small targets and the complete identification of large targets. SUMMARY

[0010] The present application provides a medical image segmentation method and system based on guide information and multi-dimensional attention mechanism, aiming to improve segmentation accuracy and model generalization ability, enhance processing capability for multi-scale lesions, reduce computational overhead while ensuring performance, and improve inference efficiency, so as to better serve the auxiliary diagnosis and treatment decision support of skin diseases.

[0011] A medical image segmentation method based on guide information and multi-dimensional attention mechanism, characterized by comprising:

[0012] (1) Collecting original dermoscopy images for preprocessing;

[0013] (2) Constructing a segmentation model, which includes a dual-path image encoder, a guide information encoder and a mask decoder;

[0014] Wherein, the dual-path image encoder combines deep convolutional neural network and Transformer structure, which is used to extract local detail features and global context semantic information in dermoscopy images; the guide information encoder is used to convert the coarse segmentation mask predicted by the previous network into guide feature information, and the guide model focuses on the potential lesion area; the mask decoder is responsible for fusing the image features output by the dual-path image encoder and the guide feature information output by the guide information encoder, and gradually restoring and refining the coarse-grained feature map, and finally outputting the accurate lesion segmentation mask;

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

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

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

[0018] Downsample the image to 512x512 pixels, perform denoising, and then normalize the image pixel values; and perform data augmentation by rotating, flipping, color transformation, and adding adversarial noise.

[0019] In step (2), the dual-path image encoder comprises two independent branches, wherein the first branch is composed of a plurality of deep convolution modules CNN cascaded, for extracting local detail features in the dermoscopy image; the second branch is composed of a plurality of Transformer modules cascaded, for capturing global context information in the dermoscopy image.

[0020] After the local detail features and the global context information output by the two branches are fused, the fused features are further input into a triple attention module to obtain image features.

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

[0022] In each dimension, first, the input fused features are respectively subjected to maximum pooling and average pooling operations to obtain two groups of feature representations; then the two groups of feature representations are spliced and sent to a shared convolution operation to extract attention mapping; finally, normalized attention weights are generated through a Sigmoid activation function, and the original input fused features are element-wise weighted and fused to complete attention enhancement in the corresponding dimension.

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

[0024] First, the coarse segmentation mask output by the mask decoder in the last round of network is input into a dense embedding representation module; the module is composed of two serial convolution submodules, each convolution submodule includes the following operation steps: the input features are first subjected to spatial down-sampling through a convolution layer; then they are sequentially processed by layer normalization and GELU activation function to enhance the feature expression ability; after the two convolution submodules, a 1x1 convolution layer is used to map the final feature to the preset embedding dimension to generate the final guidance feature information.

[0025] In step (2), the mask decoder comprises a plurality of cascaded decoding subunits, each decoding subunit comprising a Transformer module and a triple attention module in parallel, and the outputs of the two modules are fused;

[0026] The input of the mask decoder is the image features obtained by the dual-path image encoder and the guidance feature information obtained by the guidance information encoder; after the image features and the guidance feature information are spliced, the fused features are obtained after inputting into the first decoding subunit, and then inputting into the next decoding subunit; the fused features output by the last decoding subunit are input into the up-sampling module for spatial resolution recovery, and the fine segmentation mask is output.

[0027] The upsampling module includes two sequentially connected transpose convolution layers; the first layer adopts a transpose convolution with a kernel size of 2*2 and a step of 2 to realize preliminary upsampling, and then completes feature normalization and nonlinear transformation through layer normalization and a GELU activation function; the second layer also adopts a transpose convolution with a kernel size of 2*2 and a step of 2 to further upsample, and after GELU activation, inputs a multilayer perception module to output a fine segmentation mask consistent with the size of the original dermoscope image.

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

[0029]

[0030] In the formula, λ1, λ2 and λ3 represent weight coefficients, is a Dice loss; is a binary cross-entropy loss; is an uncertainty weighted boundary loss.

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

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

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

[0034] In the formula, x represents a pixel position 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 of the model, that is, n predictions generated using different perturbations (such as Dropout, data augmentation or model integration); Var(.) represents the variance calculated on the n prediction values at the pixel position x, which is used to measure the uncertainty of the prediction result at the position; U(x) is the final generated uncertainty map, which is used to assign different weights to the boundary region in the loss function.

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

[0036]

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

[0038] A medical image segmentation system based on guidance information and multi-dimensional attention mechanism, comprising a memory and one or more processors, the memory having stored executable code, the one or more processors executing the executable code to implement the above medical image segmentation method.

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

[0040] 1、The present application constructs a hybrid decoder architecture fused with guidance information: the previous prediction results are embedded through a guidance information encoder to strengthen the attention to the target area. The coupling of coarse-fine two-stage segmentation is realized, effectively improving the model convergence speed and positioning accuracy.

[0041] 2、The present application introduces a triple attention mechanism decoder combining Transformer and CNN: Two-way Transformer cross attention mechanism and triple attention module (channel-space-scale) are jointly modeled to strengthen the structural hierarchical feature expression and improve the anatomical rationality and boundary sensitivity of segmentation.

[0042] 3、The present application designs a loss function with boundary sensitivity and uncertainty perception ability: an uncertainty weighted boundary loss is introduced, combined with Dice+BCE optimization, focusing on punishing the prediction error of the fuzzy boundary and uncertain area, effectively enhancing the stability and robustness of the model in the complex structure area. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The structure diagram of the segmentation model in the embodiment of the present application.

[0044] Figure 2 The structure diagram of the triple attention module in the embodiment of the present application.

[0045] Figure 3 The structure diagram of the up-sampling module in the embodiment of the present application. DETAILED DESCRIPTION

[0046] The present application will be further described in detail below in conjunction with the drawings and embodiments, it should be pointed out that the following described embodiments are intended to facilitate the understanding of the present application, and do not have any limiting effect on it.

[0047] A medical image segmentation method based on guidance information and multi-dimensional attention mechanism, which realizes automatic segmentation of lesion area in dermatoscope image by using deep learning technology, so as to reduce the work burden of medical personnel and improve the subsequent diagnosis efficiency and accuracy.

[0048] The method of the present application first preprocesses the original dermoscopy image. The preprocessing includes: downsampling the image to a standard size (512x512 pixels); denoising processing by using Gaussian filtering, median filtering and the like; normalizing the image pixel values; and data enhancement by rotation, flipping, color transformation, adding adversarial noise and the like to improve the generalization ability of the model.

[0049] The preprocessed image is input into the segmentation model constructed by the present application for lesion region segmentation. The structure of the segmentation model is as shown in Figure 1 The segmentation model is composed of three parts: a dual-path image encoder, a guide information encoder and a mask decoder. Each module has clear division of labor and cooperative work to achieve robust and high-precision segmentation effect.

[0050] The dual-path image encoder fuses deep convolutional neural network and Transformer structure, which can fully extract the context semantic information and fine-grained lesion structure features in the dermoscopy image. The guide information encoder is used for modeling rough mask guide information and converting it into sparse and dense feature representation, thereby providing a regional attention mechanism for the segmentation task. The mask decoder is responsible for fusing image features and guide feature information, gradually restoring and refining the coarse-grained feature map, and finally outputting an accurate lesion segmentation mask.

[0051] The dual-path image encoder includes two independent branches. The first branch is composed of multiple deep convolution modules for extracting local detail features in the dermoscopy image. The second branch is composed of multiple Transformer modules cascaded, which can capture global context information in the image. This dual-branch structure realizes the collaborative extraction of local and global features, effectively improving the comprehensiveness and discriminability of feature expression.

[0052] The deep convolution module (CNN module) adopts a residual network structure, which realizes efficient feature extraction by stacking multiple convolution layers, batch normalization layers (Batch Normalization) and nonlinear activation functions (such as ReLU or GELU); the Transformer module adopts a scaled dot-attention mechanism with relative position bias, constructs a multi-head self-attention structure, and can effectively capture the long-distance dependency relationship between different positions in the image. In this structure, the introduction of relative position encoding can enhance the model's perception ability of spatial position information, thereby improving the accuracy and expression ability of global feature modeling. The Transformer module further combines residual connection and layer normalization to ensure the stability during training and enhance the non-linear expression ability of the network.

[0053] The first branch composed of a deep convolutional module and the second branch composed of a Transformer module are fused at the bottom of the encoder to effectively integrate local and global feature information.

[0054] The triplet attention module is shown in Figure 2 The triplet attention module models the fused features from three dimensions of channel attention, spatial horizontal attention and spatial vertical attention, which can fully exploit the importance information in different dimensions of the image, thereby improving the recognition and segmentation accuracy of the skin lesion area. 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 attention mapping; finally, the normalized attention weights are generated through a Sigmoid activation function, and the original input features are element-wise weighted and fused to complete the attention enhancement in the corresponding dimension.

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

[0056] First, the coarse segmentation mask output by the mask decoder in the previous round of network is input into the dense embedding representation module; the module is composed of two serial convolutional sub-modules, each of which includes the following operation steps: the input feature is first spatially down-sampled through a convolutional layer; then it is processed through layer normalization and GELU activation function in turn to enhance the feature expression ability; after the two convolutional sub-modules, a 1x1 convolutional layer is used to map the final feature to the preset embedding dimension to generate the final guide feature information. Through the above structural design, the guide information encoder realizes the effective conversion from the coarse segmentation mask to the semantic embedding representation, not only significantly enhancing the model's perception ability of the target area, but also providing prior knowledge support with context dependency for the entire decoding process.

[0057] The mask decoder proposed in the present application adopts a fusion architecture combining the advantages of Transformer and convolutional neural network (CNN), aiming to recover a high-resolution volumetric segmentation mask from image feature embedding and guide information embedding.

[0058] The input of the mask decoder includes: (1) image feature embedding extracted by a dual-channel image encoder, and (2) a dense guide embedding representation generated by a guide information encoder. In order to fully integrate global context and local structure 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 distinguish complex anatomical boundaries and fine-grained regions, thereby enhancing the representation ability and segmentation accuracy in the decoding stage.

[0059] After feature enhancement is completed, the fused features enter the upsampling module for spatial resolution recovery. The upsampling module, as shown in Figure 3 The upsampling module includes two sequentially connected transpose convolution layers: the first layer uses a transpose convolution with a kernel size of 2x2 and a step size of 2 to perform preliminary upsampling, and then performs feature normalization and nonlinear transformation through layer normalization and a GELU activation function; the second layer also uses a transpose convolution with a kernel size of 2x2 and a step size of 2 for further upsampling, and after GELU activation, inputs a multi-layer perceptron (MLP) module to output a fine segmentation mask consistent with the original image size.

[0060] This decoder structure fully combines the global modeling capability of Transformer and the efficient local enhancement characteristics of CNN, and cooperates with the guide mechanism and triple attention module, effectively improving the sensitivity of the model to the lesion area and the overall segmentation performance.

[0061] To improve the robustness and accuracy of the model in the volumetric medical image segmentation task, the present application uses a boundary-uncertainty joint guide loss function to supervise and optimize the difference between the prediction result and the true label. The loss is composed of three parts: a basic Dice loss for handling foreground / background imbalance; a binary cross-entropy for optimizing pixel classification; and an uncertainty-weighted boundary loss, which is the core part of the loss function.

[0062]

[0063] Traditional boundary loss only considers edge regions, while in medical images, boundaries are often accompanied by uncertainty (such as the transition ambiguity between lesions and normal tissue). The present application introduces a weighting mechanism based on the uncertainty map at the boundary position, making the network pay more attention to the ambiguous boundary region.

[0064] The uncertainty map is defined as the prediction variance of the model at each pixel position

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

[0066] Let B(x) be the boundary mask extracted from the real mask by 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 the lesion contour.

[0069] In the actual application process, the workflow of the guide information encoder is as follows:

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

[0071] Subsequent training stage (round n): use the model of the last round (round n-1) to infer the same image to obtain a coarse segmentation mask; the guide mask is generated by upsampling on the coarse segmentation mask.

[0072] The guide mask is input into the guide information encoder as guide information, and the image features of the current round are input into the model together, so that the model can focus on the key areas, thereby optimizing the segmentation quality, especially the boundary and complex areas.

[0073] The application innovatively introduces a guide information encoder, which takes the coarse segmentation mask generated in the last stage as the guide input, and constructs a dense guide embedding through convolution, layer normalization and GELU activation. The embedding is jointly modeled with the image features, which significantly enhances the model's attention to the lesion area and provides effective constraints in the case of early small lesions or fuzzy boundaries.

[0074] The application designs a hybrid decoder that combines Transformer and CNN, which integrates a Transformer module and a triple attention module. The triple attention module combines channel attention, spatial attention and scale attention modules to fully exploit the salient features of multi-dimensional information and effectively enhance the response of the lesion area, improving the fineness and robustness of the segmentation result.

[0075] To overcome the problem of low accuracy of traditional Dice / BCE in boundary and uncertain areas, the application proposes a joint loss function, in which the weights of each term can be flexibly adjusted according to the actual task. This loss function is particularly suitable for medical image segmentation scenarios with fuzzy boundaries and multi-class imbalance, significantly improving the performance of the model in difficult areas.

[0076] The above embodiments describe the technical solutions and advantages of the present application in detail. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the present application. Any modification, supplement and equivalent replacement made within the principle range of the present application shall be included in the protection range of the present application.

Claims

1. A medical image segmentation method based on guidance information and multi-dimensional attention mechanism, characterized in that: include: (1) Collect original dermatoscopic images for preprocessing; (2) constructing a segmentation model, wherein the segmentation model includes a dual-path image encoder, a guidance information encoder, and a mask decoder; The dual-path image encoder combines a deep convolutional neural network with a Transformer architecture to extract local detail features and global contextual semantic information from dermatoscope 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-path image encoder with the guidance feature information output by the guidance information encoder, gradually restoring and refining the coarse-grained feature map, and ultimately outputting an accurate lesion segmentation mask. (3) Construct a loss function and use the preprocessed data to train the segmentation model; (4) The original dermatoscopic image to be segmented is input into the trained segmentation model, and the segmentation mask of the lesion area of ​​the image is output.

2. The medical image segmentation method based on guidance information and multi-dimensional attention mechanism according to claim 1, characterized in that: In step (1), the pretreatment includes: The image is downsampled to 512×512 pixels, and the image pixel values ​​are normalized after denoising. Data enhancement is performed by rotating, flipping, color conversion, and adding adversarial noise.

3. The medical image segmentation method based on guidance information and multi-dimensional attention mechanism according to claim 1, characterized in that: In step (2), the dual-path image encoder includes two independent branches, wherein the first branch is composed of a cascade of multiple deep convolutional modules (CNN) for extracting local detail features in the dermoscopic image; and the second branch is composed of a cascade of multiple Transformer modules for capturing global context information in the dermoscopic image. After the local detail features output by the two branches are fused with the global context information, they are further input into the triple attention module to obtain image features.

4. The medical image segmentation method based on guidance information and multi-dimensional 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, the maximum pooling and average pooling operations are first performed on the input fusion features to obtain two sets of feature representations; then the two sets of feature representations are spliced ​​and sent to a shared convolution operation to extract the attention map; finally, the normalized attention weight is generated through the Sigmoid activation function, and is element-wise weighted fused with the original input fusion features to complete the attention enhancement of the corresponding dimension.

5. The medical image segmentation method based on guidance information and multi-dimensional attention mechanism according to claim 1, characterized in that: In step (2), the working process of the guidance information encoder is as follows: First, the coarse segmentation mask output by the mask decoder in the previous round of the network is input into the dense embedding representation module; this module consists of two convolutional sub-modules in series, each of which contains the following steps: the input features are first spatially downsampled through a convolutional layer; then they are processed in sequence through layer normalization and GELU activation function to enhance the feature expression 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 guided feature information.

6. The medical image segmentation method based on guidance information and multi-dimensional 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 feature fused; The input of the mask decoder is the image features obtained by the dual-path image encoder and the guidance feature information obtained by the guidance information encoder; after the image features and the guidance feature information are spliced, they are input into the first decoding subunit to obtain the corresponding fusion features, and then input into the next decoding subunit; the fusion features output by the last decoding subunit enter the upsampling module to restore the spatial resolution and output a fine segmentation mask.

7. The medical image segmentation method based on guidance information and multi-dimensional attention mechanism according to claim 6, characterized in that: The upsampling module consists of 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 preliminary upsampling, followed by 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 after being activated by GELU, it is input into the multi-layer perceptron module to output a fine segmentation mask consistent with the size of the original dermatoscope image.

8. The medical image segmentation method based on guidance information and multi-dimensional attention mechanism according to claim 1, characterized in that: In step (3), the loss function constructed is: Where λ1, λ2 and λ3 represent weight coefficients, for Dice loss; is the binary cross entropy loss; is the uncertainty-weighted margin loss.

9. The medical image segmentation method based on guidance information and multi-dimensional attention mechanism according to claim 8, characterized in that: The construction process of uncertainty weighted boundary loss is: Define the uncertainty map as the prediction variance of the segmentation model at each pixel location: U(x)=Var(p1(x),p2(x),…,p n (x)); Where x represents a pixel position in the image; p i (x) represents the predicted 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 times the model is inferred, that is, n predictions are generated using different perturbations; Var(.) represents the variance calculated for the n predicted values ​​at pixel position x, which is used to measure the uncertainty of the prediction result at that position; U(x) is the final uncertainty map generated, which is used to assign different weights to the boundary area in the loss function; Let B(x) be the boundary mask extracted by Sobel or Morphological method in the real mask, then the uncertainty weighted boundary loss is defined as: Where B(x) is the indicator function of whether pixel x is in the boundary area, the boundary area value is 1, and the non-boundary area is 0; 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 guidance information and multi-dimensional attention mechanism, characterized in that: The medical image segmentation method comprises a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, the medical image segmentation method is implemented as described in any one of claims 1 to 9.

Images