Medical image segmentation method based on multi-scale Transformer
By setting up a multi-scale Transformer module and a skip connection encoding/decoding structure at the bottom layer of the encoding structure, the problem of insufficient multi-scale information capture in multimodal cardiac image segmentation is solved, and more efficient and accurate multimodal cardiac image segmentation is achieved.
Patent Information
- Application Number
- CN202210795156.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-07
- Publication Date
- 2026-03-03
- Estimated Expiration
- 2042-07-07
AI Technical Summary
Existing Transformer-based medical image segmentation networks struggle to capture multi-scale information in multimodal cardiac image segmentation tasks, resulting in poor segmentation accuracy, especially when the heart structure is complex and the boundaries are blurred.
A multi-scale Transformer module and a skip connection encoding/decoding structure are adopted. By setting a multi-scale Transformer module at the bottom layer of the encoding structure to extract multi-scale features, and adding skip connections between corresponding layers of the encoding/decoding structure, the size of local feature maps is restored, and global and local feature maps are fused to improve the feature extraction capability of the generator.
It achieves more efficient and accurate multimodal cardiac image segmentation, improves image resolution and segmentation accuracy, and adapts to multi-scale changes in multi-organ segmentation tasks.
Smart Images
Figure CN115170584B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image processing technology, and in particular to a medical image segmentation method based on multi-scale Transformer. Background Technology
[0002] The heart is one of the most vital organs in the human body. With rapid societal development, unhealthy lifestyles have led to a continuous increase in the incidence of cardiovascular diseases. According to statistics from the World Health Organization (WHO), it is estimated that nearly 23.6 million people will die from cardiovascular diseases by 2030, primarily from heart disease and stroke, making cardiovascular disease one of the leading causes of death globally. Cardiovascular diseases are characterized by high incidence and high mortality rates. Because cardiovascular diseases pose a serious threat to human health, their diagnosis and treatment are extremely important. Precise segmentation of cardiac images enables physicians to perform non-invasive qualitative and quantitative assessments of cardiac anatomy and function, playing a crucial role in disease prevention, diagnosis, and treatment.
[0003] Cardiac imaging often presents multiple modalities in actual diagnostic processes, and images with different physical properties can play a complementary role in diagnosis. Therefore, effectively adapting a trained network model to another modality of data is of great significance for medical clinical research. However, the significant differences in the distribution of medical image data across different modalities lead to the problem of domain shift.
[0004] To address the aforementioned issues, current mainstream deep learning methods can be broadly categorized into supervised and unsupervised approaches. While supervised models demonstrate excellent performance on labeled datasets, the limited cost of labeling medical images results in a severe shortage of labeled data. Consequently, training these models on unlabeled datasets can lead to a significant performance degradation, hindering their applicability in medical image segmentation tasks. Unsupervised models, on the other hand, require no labeled data for training, saving manpower and exhibiting strong robustness. However, due to the complex structure of the heart, its blurred boundaries, and the presence of adhesions, unsupervised segmentation models generally perform poorly in multimodal cardiac image segmentation tasks.
[0005] Existing Transformer-based image segmentation networks expand the receptive field by extracting long-range dependencies in the data, thereby improving the matching accuracy of Regions of Interest (ROIs). However, because the Transformer divides the input image into fixed-size blocks, it loses useful information and struggles to capture multi-scale information in multi-organ segmentation tasks, resulting in poor performance in multimodal cardiac image segmentation tasks where the sub-regions to be segmented vary in size.
[0006] Therefore, how to provide a medical image segmentation method that can effectively improve segmentation accuracy is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0007] In view of this, this invention proposes a medical image segmentation method based on multi-scale Transformer. This method, by setting a multi-scale Transformer module at the bottom layer of the encoding structure, transforms the source domain image x, after downsampling encoding processing by the encoding structure, into a multi-scale medical image segmentation method. A The input is fed into the multi-scale Transformer module to extract multi-scale features, outputting a global feature map. This global feature map is then passed to the decoding structure for upsampling and decoding to restore the global feature map size. A spatial pyramid module is also constructed to assist the multi-scale Transformer module in extracting multi-scale features. Simultaneously, the generator G... A Skip connections are added between corresponding layers of the encoding and decoding structures. Local feature maps from the corresponding layers of the encoding structure are passed to the corresponding layers of the decoding structure, and upsampling decoding is performed to restore the size of the local feature maps and preserve the low-level features of the network. A similar target domain image x is obtained by fusing global and local feature maps. A→B This method effectively integrates global and local features through multi-scale Transformer modules and skip connections, modeling both global and local features of cardiac images and improving the generator G. A Its feature extraction capabilities enable the generator to reconstruct images with higher resolution, thereby achieving more efficient and accurate multimodal cardiac image segmentation.
[0008] To achieve the above objectives, the present invention adopts the following technical solution:
[0009] Medical image segmentation methods based on multi-scale Transformer include:
[0010] Step (1): Obtain the multimodal cardiac dataset, perform data preprocessing on the multimodal cardiac dataset, and obtain the training set.
[0011] Step (2): Input the training set into a segmentation network model containing multi-scale Transformer modules and skip connections for iterative training to construct the overall objective function model L with optimal weight parameters.
[0012] Step (3): Input the image to be tested into the overall objective function model L with optimal weight parameters to obtain the segmentation result.
[0013] Optionally, in step (1), data preprocessing includes: cropping, rotating, and normalizing.
[0014] Optionally, in step (2), the segmentation network model is specifically as follows:
[0015] Using the training set as the source domain image x A Input to generator G A The transformation yields an image similar to the target domain, x. A→B Discriminator D B With generator G A Competition, by identifying losses Optimize and identify similar target domain images x A→B and target domain image x B .
[0016] Image x, similar to the target domain A→B and target domain image x B Features are extracted from the input to the shared encoder E, and these features are simultaneously input to the decoder U and the classifier C:
[0017] The shared encoder E and decoder U constitute the inverse generator G. B Output the reconstructed source domain image x A→B→A and similar source domain image x B→A Discriminator D A With the inverse generator G B Competition, by identifying losses Optimize and identify the source domain image x A and similar source domain image x B→A By determining the loss Optimize and identify the reconstructed source domain image x. A→B→A and similar source domain image x B→A and using the reverse generator G B Applying cycle consistency loss L cyc .
[0018] The shared encoder E and classifier C form a segmentation network, utilizing the segmentation loss L. seg Optimize the segmentation network and output the segmentation prediction result P:x A→B and P:x B Discriminator D P By identifying losses Distinguish between P:x A→B and P:x B .
[0019] Optional, generator G A The encoding structure is built from convolutional layers, and the decoding structure is built from deconvolutional layers.
[0020] Optionally, in step (2), the multi-scale Transformer module is at the bottom layer of the encoding structure; the generator G A Add skip connections between corresponding layers of the encoding / decoding structure.
[0021] Optional, multi-scale Transformer modules are as follows:
[0022]
[0023]
[0024] in, For the input of the Multi-Head Self-Attention Module (MHSA), The output features of the Multi-Head Self-Attention Module (MHSA); The input features are fed into the Forward Propagation Network (FFD) and transformed by the Seq2Img sequence transformation layer, outputting the output features of the multi-scale Transformer module.
[0025] Optionally, it also includes building a spatial pyramid module to help the multi-scale Transformer module extract multi-scale features. The spatial pyramid module is as follows:
[0026]
[0027] f ASPP =Concat([y1,…,y 24 |d∈(1,8,16,24)])
[0028]
[0029] Where f and y represent the input and output feature maps of the spatial pyramid module, respectively; S represents the number of convolution kernels, which is equivalent to the depth of the feature map; i represents the pixel position where the dilated convolution is performed; d∈[1,8,16,24] represents the dilation rate; w[s] represents the feature parameters of the s-th layer; f ASPP This represents the characteristics of a spatial pyramid pooling layer; This represents the feature sequence of the image after being transformed by the activation layers GeLU and Img2Seq.
[0030] Optionally, determine the loss. as follows:
[0031]
[0032] Among them, G A D represents a generator; B Indicates the discriminator; x B Represents the target domain image; x A Represents the source domain image.
[0033] Determine the loss as follows:
[0034]
[0035] Among them, G B D represents a generator; A Indicates the discriminator; x A Represents the source domain image; x B Represents the target domain image.
[0036] Determine the loss as follows:
[0037]
[0038] Among them, D A Indicates discriminator; E indicates shared encoder; x B Represents the target domain image; x A→B This represents a target domain image.
[0039] Cyclic Consistent Loss L cyc ,as follows:
[0040]
[0041] Among them, G A D represents a generator; A Indicates discriminator; E indicates shared encoder; x A Represents the source domain image; x B Represents the target domain image; x A→B This represents a target domain image.
[0042] Segmentation loss L seg ,as follows:
[0043] L seg (E,C)=H(y A ,C(E(x A→B )))+Dice(y A ,C(E(x A→B )))
[0044] Where E represents the shared encoder; C represents the classifier; y A H represents the source domain label image; H(·) represents the cross-entropy loss; Dice(·) represents the overlap loss.
[0045] Determine the loss as follows:
[0046]
[0047] Where E represents the shared encoder; C represents the classifier; D P Indicates the discriminator; x B Represents the target domain image; x A→BThis represents a target domain image.
[0048] Optionally, in step (2), the iterative training specifically involves: training the segmentation network model using an end-to-end training method, optimizing the overall objective function model L driven by the Adam optimizer, setting the initial learning rate to 2e-4, and the overall objective function model L as follows:
[0049]
[0050] in, It is x A→B and x B The discriminant loss between them; It is x A and x B→A The discriminant loss between them; It is x B →A and x A→B→A Discriminant loss between; L cyc This represents the cycle-consistent loss, used to determine x. A and x A→B→A ; and Indicates y A and P:x A→B The segmentation loss between them; and It is P:x A→B and P:x B The discriminant loss between them; is a coefficient.
[0051] Optionally, after step (3), the process further includes: evaluating the segmentation results by applying the overlap coefficient and the average surface distance.
[0052] The overlap coefficient is defined as follows:
[0053]
[0054] Where A represents the prediction region; B represents the Ground Truth region.
[0055] The average surface distance is defined as follows:
[0056]
[0057] Where A and B represent the 3D prediction result and the actual value, respectively; S(·) represents the voxel set in the 3D surface; This represents the shortest distance from any voxel a to b.
[0058] As can be seen from the above technical solution, compared with the prior art, this invention proposes a medical image segmentation method based on multi-scale Transformer. This method, by setting a multi-scale Transformer module at the bottom layer of the encoding structure, transforms the source domain image x, after downsampling encoding processing by the encoding structure, into a multi-scale medical image segmentation method. A The input is fed into the multi-scale Transformer module to extract multi-scale features, outputting a global feature map. This global feature map is then passed to the decoding structure for upsampling and decoding to restore the global feature map size. A spatial pyramid module is also constructed to assist the multi-scale Transformer module in extracting multi-scale features. Simultaneously, the generator G... A Skip connections are added between corresponding layers of the encoding and decoding structures. Local feature maps from the corresponding layers of the encoding structure are passed to the corresponding layers of the decoding structure, and upsampling decoding is performed to restore the size of the local feature maps and preserve the low-level features of the network. A similar target domain image x is obtained by fusing global and local feature maps. A→B This method effectively integrates global and local features through a multi-scale Transformer module and skip connections, models the global and local features of cardiac images, improves the feature extraction capability of the generator GA, and makes the image reconstructed by the generator have higher resolution, thereby achieving more efficient and accurate multimodal cardiac image segmentation. Attached Figure Description
[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0060] Figure 1 This is a schematic diagram of the process of the present invention.
[0061] Figure 2 This is a schematic diagram of the model of the present invention.
[0062] Figure 3 The generator G of this invention A Schematic diagram.
[0063] Figure 4 This is a schematic diagram of the ASPP module structure of the present invention. Detailed Implementation
[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0065] This invention discloses a medical image segmentation method based on multi-scale Transformer, including:
[0066] Step (1): Obtain the multimodal heart dataset and perform data preprocessing on the multimodal heart dataset, including cropping, rotation, and normalization. Specifically, the 3D training images are cropped to the heart region and sliced into coronal image slices of shape 256×256. Then, rotation is performed to increase the amount of training data and improve the generalization ability of the model. The rotation angles are set to 60°, 90°, 120°, 150°, 210°, 240°, and 300°. Finally, normalization is performed to transform the pixel values of the data to the range of 0-1 to obtain the training set.
[0067] Step (2): Input the training set into a segmentation network model containing multi-scale Transformer modules and skip connections for iterative training to construct the overall objective function model L with optimal weight parameters.
[0068] like Figure 2 As shown, the segmentation network model is specifically as follows:
[0069] Using the training set as the source domain image x A Input to generator G A The transformation yields an image similar to the target domain, x. A→B Discriminator D B With generator G A Competition, by identifying losses Optimize and identify similar target domain images x A→B and target domain image x B .
[0070] Image x, similar to the target domain A→B and target domain image x B Features are extracted from the input to the shared encoder E, and these features are simultaneously input to the decoder U and the classifier C:
[0071] The shared encoder E and decoder U constitute the inverse generator G. B Output the reconstructed source domain image x A→B→A and similar source domain image x B→A Discriminator D AWith the inverse generator G B Competition, by identifying losses Optimize and identify the source domain image x A and similar source domain image x B→A By determining the loss Optimize and identify the reconstructed source domain image x. A→B→A and similar source domain image x B→A and using the reverse generator G B Applying cycle consistency loss L cyc .
[0072] The shared encoder E and classifier C form a segmentation network, utilizing the segmentation loss L. seg Optimize the segmentation network and output the segmentation prediction result P:x A→B and P:x B Discriminator D P By identifying losses Distinguish between P:x A→B and P:x B .
[0073] like Figure 3 As shown, generator G A The encoding structure is built from convolutional layers, and the decoding structure is built from deconvolutional layers.
[0074] The multi-scale Transformer module is located at the bottom layer of the encoding structure. Its input is the source domain image x, which has undergone downsampling encoding processing by the encoding structure. A Multi-scale features are extracted, and a global feature map is output. This global feature map is then passed to the decoding structure for upsampling and decoding to restore the size of the global feature map. Simultaneously, in the generator G... A Skip connections are added between corresponding layers of the encoding and decoding structures. Local feature maps from the corresponding layers of the encoding structure are passed to the corresponding layers of the decoding structure, and upsampling decoding is performed to restore the size of the local feature maps and preserve the low-level features of the network. By fusing global and local feature maps, a similar target domain image xA is obtained. →B .
[0075] The multi-scale Transformer module is as follows:
[0076]
[0077]
[0078] in, For the input of the Multi-Head Self-Attention Module (MHSA), The output features of the Multi-Head Self-Attention Module (MHSA); The input features are fed into the Forward Propagation Network (FFD) and transformed by the Seq2Img sequence transformation layer, outputting the output features of the multi-scale Transformer module.
[0079] It also includes building a spatial pyramid module to help the multi-scale Transformer module extract multi-scale features. The spatial pyramid module is as follows:
[0080]
[0081] f ASPP =Concat([y1,…,y 24 |d∈(1,8,16,24)])
[0082]
[0083] Where f and y represent the input and output feature maps of the spatial pyramid module, respectively; S represents the number of convolution kernels, which is equivalent to the depth of the feature map; i represents the pixel position where the dilated convolution is performed; d∈[1,8,16,24] represents the dilation rate; w[s] represents the feature parameters of the s-th layer; f ASPP This represents the characteristics of a spatial pyramid pooling layer; This represents the feature sequence of the image after being transformed by the activation layers GeLU and Img2Seq.
[0084] Determine the loss as follows:
[0085]
[0086] Among them, G A D represents a generator; B Indicates the discriminator; x B Represents the target domain image; x A Represents the source domain image.
[0087] Determine the loss as follows:
[0088]
[0089] Among them, G B D represents a generator; A Indicates the discriminator; x A Represents the source domain image; x B Represents the target domain image.
[0090] Determine the loss as follows:
[0091]
[0092] Among them, D A Indicates discriminator; E indicates shared encoder; x B Represents the target domain image; x A→B This represents a target domain image.
[0093] Cyclic Consistent Loss L cyc ,as follows:
[0094]
[0095] Among them, G A D represents a generator; A Indicates discriminator; E indicates shared encoder; x A Represents the source domain image; x B Represents the target domain image; x A→B This represents a target domain image.
[0096] Segmentation loss L seg ,as follows:
[0097] L seg (E,C)=H(y A ,C(E(x A→B )))+Dice(y A ,C(E(x A→B )))
[0098] Where E represents the shared encoder; C represents the classifier; y A H represents the source domain label image; H(·) represents the cross-entropy loss; Dice(·) represents the overlap loss.
[0099] Determine the loss as follows:
[0100]
[0101] Where E represents the shared encoder; C represents the classifier; D P Indicates the discriminator; x B Represents the target domain image; x A→B This represents a target domain image.
[0102] The iterative training process involves training the segmentation network model using an end-to-end training method, optimizing the overall objective function model L driven by the Adam optimizer, and performing 20,000 iterations of training. The initial learning rate is set to 2e-4, and the overall objective function model L is as follows:
[0103]
[0104] in, It is x A→Band x B The discriminant loss between them; It is x A and x B→A The discriminant loss between them; It is x B →A and x A→B→A Discriminant loss between; L cyc This represents the cycle-consistent loss, used to determine x. A and x A→B→A ; and Indicates y A and P:x A→B The segmentation loss between them; and It is P:x A→B and P:x B The discriminant loss between them; The coefficients are set to {0.1, 10, 1.0, 0.1, 0.1, 0.01, 0.1} respectively.
[0105] Step (3): Input the image to be tested into the overall objective function model L with the optimal weight parameters to obtain the segmentation result. Then, use the overlap coefficient and the average surface distance to evaluate the segmentation result.
[0106] The overlap coefficient is defined as follows:
[0107]
[0108] Where A represents the predicted region; B represents the Ground Truth region; the closer the overlap ratio DSC is to 1, the more accurate the image segmentation result.
[0109] The average surface distance is defined as follows:
[0110]
[0111] Where A and B represent the 3D prediction result and the actual value, respectively; S(·) represents the voxel set in the 3D surface; It represents the shortest distance from any voxel a to b; the lower the ASD, the higher the segmentation accuracy.
[0112] This invention discloses a medical image segmentation method based on multi-scale Transformer. This method incorporates a multi-scale Transformer module at the bottom layer of the encoding structure to segment the source domain image x after downsampling encoding. AThe input is fed into the multi-scale Transformer module to extract multi-scale features, outputting a global feature map. This global feature map is then passed to the decoding structure for upsampling and decoding to restore the global feature map size. A spatial pyramid module is also constructed to assist the multi-scale Transformer module in extracting multi-scale features. Simultaneously, the generator G... A Skip connections are added between corresponding layers of the encoding and decoding structures. Local feature maps from the corresponding layers of the encoding structure are passed to the corresponding layers of the decoding structure, and upsampling decoding is performed to restore the size of the local feature maps and preserve the low-level features of the network. A similar target domain image x is obtained by fusing global and local feature maps. A→B This method effectively integrates global and local features through multi-scale Transformer modules and skip connections, modeling both global and local features of cardiac images and improving the generator G. A Its feature extraction capabilities enable the generator to reconstruct images with higher resolution, thereby achieving more efficient and accurate multimodal cardiac image segmentation.
[0113] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0114] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A medical image segmentation method based on multi-scale Transformer, characterized in that, include: Step (1): Obtain the multimodal cardiac dataset, perform data preprocessing on the multimodal cardiac dataset to obtain the training set; Step (2): Input the training set into a segmentation network model containing multi-scale Transformer modules and skip connections for iterative training to construct an overall objective function model L with optimal weight parameters; Step (3): Input the image to be tested into the overall objective function model L with the optimal weight parameters to obtain the segmentation result; In step (2), the segmentation network model is specifically as follows: The training set is used as the source domain image. Input to generator The conversion yields an image similar to the target domain. Discriminator With generator Competition, by identifying losses Optimize and identify similar target domain images. and target domain image ; The similar target domain image and the target domain image Input to shared encoder Features are extracted and simultaneously input into the decoder. and classifier middle: The shared encoder With the decoder Constructing a reverse generator Output the reconstructed source domain image and similar source domain images ; Discriminator With the reverse generator Competition, by identifying losses Optimize and identify the source domain image. and similar source domain images By determining the loss Optimization is performed to identify the reconstructed source domain image. and similar source domain images and using a reverse generator Applying cycle consistency loss ; The shared encoder With the classifier Construct a segmentation network and utilize segmentation loss Optimize the segmentation network and output segmentation prediction results. and Discriminator By identifying losses distinguish and ; The discriminant loss ,as follows: in, Indicates a generator; Indicates the discriminator; Represents the target domain image; Represents the source domain image; The discriminant loss ,as follows: in, Indicates a generator; Indicates the discriminator; Represents the source domain image; Represents the target domain image; The discriminant loss ,as follows: = in, Indicates the discriminator; Indicates a shared encoder; Represents the target domain image; Represents images similar to the target domain; The cycle-consistent loss ,as follows: in, Indicates a generator; Indicates the discriminator; Indicates a shared encoder; Represents the source domain image; Represents the target domain image; Represents images similar to the target domain; The segmentation loss ,as follows: in, Indicates a shared encoder; Represents a classifier; Represents the source domain labeled image; H( ) represents cross-entropy loss; Dice( () indicates the loss of overlap. The discriminant loss ,as follows: in, Indicates a shared encoder; Represents a classifier; Indicates the discriminator; Represents the target domain image; This represents a target domain image.
2. The medical image segmentation method based on multi-scale Transformer according to claim 1, characterized in that, In step (1), the data preprocessing includes: cropping, rotation, and normalization.
3. The medical image segmentation method based on multi-scale Transformer according to claim 1, characterized in that, The generator The encoding structure is constructed from convolutional layers, and the decoding structure is constructed from deconvolutional layers.
4. The medical image segmentation method based on multi-scale Transformer according to claim 3, characterized in that, In step (2), the multi-scale Transformer module is at the bottom layer of the encoding structure; the generator Add skip connections between corresponding layers of the encoding / decoding structure.
5. The medical image segmentation method based on multi-scale Transformer according to claim 1, characterized in that, The multi-scale Transformer module is as follows: in, For multi-head self-attention modules Input, For the multi-head self-attention module The output characteristics; Input to the forward propagation network And through the sequence conversion layer After transformation, the output features of the multi-scale Transformer module are output. .
6. The medical image segmentation method based on multi-scale Transformer according to claim 1, characterized in that, It also includes building a spatial pyramid module to help the multi-scale Transformer module extract multi-scale features. The spatial pyramid module is as follows: in, and These represent the input and output feature maps of the spatial pyramid module, respectively. The number of convolution kernels is equal to the depth of the feature map. Indicates the pixel position where the dilated convolution is performed; Indicates the expansion rate; Indicates the first Layer characteristic parameters; This represents the characteristics of a spatial pyramid pooling layer; This indicates that the image has passed through the activation layer. and The transformed feature sequence.
7. The medical image segmentation method based on multi-scale Transformer according to claim 1, characterized in that, In step (2), the iterative training specifically involves: training the segmentation network model using an end-to-end training method, optimizing the overall objective function model L designed using the Adam optimizer, with an initial learning rate set to 2e-4. The overall objective function model L is as follows: in, yes and The discriminant loss between them; yes and The discriminant loss between them; yes and The discriminant loss between them; Represents the cycle-consistent loss, used to determine... and ; and express and The segmentation loss between them; and yes and Discriminant loss between; , , , , , , } represents the coefficient.
8. The medical image segmentation method based on multi-scale Transformer according to claim 1, characterized in that, After step (3), the method further includes: evaluating the segmentation results by applying the overlap coefficient and the average surface distance; The overlap coefficient is defined as follows: in, Represents the predicted region; Represents the Ground Truth region; The average surface distance is defined as follows: in and These represent the 3D prediction result and the actual value, respectively. Represents a set of voxels in a 3D surface; Represents arbitrary voxels arrive The shortest distance.