A medical image continuous passive domain adaptation segmentation system
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2024-07-02
- Publication Date
- 2026-06-26
AI Technical Summary
Existing continuous passive domain adaptation methods have limited domain adaptation capabilities and suffer from catastrophic forgetting problems, making them unable to effectively adapt to changes in the target domain.
A continuous passive domain-adaptive segmentation system for medical images is constructed. By jointly training medical image segmentation models in the first and second data domains, combining shape-aware label generation units and data-domain-aware image generation units, the influence of anatomical structures on the segmentation results is adjusted using causal chain theory, and a loss function is designed to balance domain adaptation and knowledge preservation capabilities.
It significantly improves the model's domain adaptability, solves the catastrophic forgetting problem, achieves stable segmentation performance on the target domain, and is suitable for clinical applications.
Smart Images
Figure CN118710654B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical image processing technology, specifically relating to a continuous passive domain adaptive segmentation system for medical images. Background Technology
[0002] The advent of deep neural networks has revolutionized the landscape of medical image segmentation. However, this tremendous success is built on the assumption that training and testing data come from the same distribution. But in real-world clinical scenarios, due to different acquisition protocols or imaging methods, there is significant domain shift between the training (source domain) and testing (target domain) datasets. This distributional discrepancy typically degrades model performance in the target domain.
[0003] In medical image processing, we often need to transfer models from a source domain to a target domain. This transfer process is called domain adaptation (DA). The data distributions of the source and target domains may differ, and the goal of domain adaptation is to enable the model to achieve good performance in the target domain. In traditional domain adaptation tasks, we need data samples from both the source and target domains. However, in practical applications, we may not have access to source domain data. In such cases, we need to use a special domain adaptation method called source-free domain adaptation (SFDA). Its goal is to transfer the model from the source domain to the target domain using only target domain data. Unlike standard source-free domain adaptation, which assumes a specific target domain, continuous source-free domain adaptation (Continual SFDA) considers the adaptation problem of constantly changing target data. In recent years, continuous source-free domain adaptation has become popular due to the need to protect the data privacy of the source domain.
[0004] However, the existing continuous passive domain adaptation methods still have limited domain adaptation capabilities, and the models suffer from catastrophic forgetting of previous target domains. Therefore, it is essential to propose a new method to solve these problems. Summary of the Invention
[0005] The purpose of this invention is to address the problems of limited domain adaptation capability and catastrophic forgetting of previous target domains in existing continuous passive domain adaptation methods, and to propose a continuous passive domain adaptation segmentation system for medical images.
[0006] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:
[0007] A continuous passive domain adaptive segmentation system for medical images, the segmentation system comprising an MRI image acquisition unit, an MRI image processing unit, a first data domain medical image segmentation model, and a second data domain medical image segmentation model, wherein the first data domain medical image segmentation model includes a Unet segmentation network, a shape-aware label generation unit, and a data domain-aware image generation unit; the second data domain medical image segmentation model employs the Unet segmentation network;
[0008] The MRI image acquisition unit is used to acquire MRI images of T target domains and one source domain;
[0009] The MRI image processing unit is used to slice the MRI images of the source domain and each target domain to obtain slice images of the source domain and each target domain, respectively.
[0010] The first data domain medical image segmentation model and the second data domain medical image segmentation model are jointly trained based on slice images from the source domain and each target domain;
[0011] The trained second-data-domain medical image segmentation model is used to segment the MRI image to be segmented.
[0012] Furthermore, the joint training process of the first data domain medical image segmentation model and the second data domain medical image segmentation model is as follows:
[0013] Step 1: Use image slices from the source domain to pre-train the medical image segmentation model in the second data domain until the loss function converges, then stop the pre-training to obtain the pre-trained medical image segmentation model in the second data domain.
[0014] Then, the parameters of the Unet segmentation network in the first data domain medical image segmentation model are updated to the parameters of the second data domain medical image segmentation model pre-trained on the source domain image slices, and the parameters of the shape-aware label generation unit and the data domain-aware image generation unit are calculated based on the parameters of the Unet segmentation network in the first data domain medical image segmentation model.
[0015] Step 2: Initialize the training phase t=1;
[0016] Step 3: Freeze all parameters of the first data domain medical image segmentation model, and train the second data domain medical image segmentation model using the image slice of the t-th target domain; the specific training process is as follows:
[0017] The image slice of the t-th target domain is used as the input of the Unet segmentation network in the first data domain medical image segmentation model and the second data domain medical image segmentation model, respectively. The output of the Unet segmentation network in the first data domain medical image segmentation model and the image slice of the t-th target domain are used as the input of the shape-aware label generation unit in the first data domain medical image segmentation model. The domain adaptation loss is calculated based on the output of the shape-aware label generation unit in the first data domain medical image segmentation model and the output of the second data domain medical image segmentation model.
[0018] The image slice of the t-th target domain is then used as the input of the data domain perceptual image generation unit in the first data domain medical image segmentation model. The output of the data domain perceptual image generation unit is then used as the input of the Unet segmentation network in the first data domain medical image segmentation model and the Unet segmentation network in the second data domain medical image segmentation model, respectively. The knowledge retention loss is then calculated based on the output of the Unet segmentation network in the first data domain medical image segmentation model and the Unet segmentation network in the second data domain medical image segmentation model.
[0019] The training in stage t is stopped when the sum of the domain adaptation loss and the knowledge retention loss converges, and the second data domain medical image segmentation model trained in stage t is obtained.
[0020] Step 4: Update all parameters of the first data domain medical image segmentation model based on the parameters of the second data domain medical image segmentation model trained in stage t.
[0021] Step 5: Determine if t < T is satisfied;
[0022] If t < T, then let t = t + 1 and return to step three;
[0023] If t = T, then the second data domain medical image segmentation model trained in the last training stage will be used as the final second data domain medical image segmentation model.
[0024] Furthermore, before being input into the model, the image slices first need to undergo a cropping operation to obtain image slices of size n×n.
[0025] Furthermore, the shape-aware label generation unit includes a gradient category activation module and a guided inverse convolution module; the shape-aware label generation unit is used to generate pseudo-labels;
[0026] The working process of the shape-aware label generation unit is as follows:
[0027] For an image slice x, the output of the Unet segmentation network after image slice x is a segmentation probability map y. After multiplying x and y, the multiplication result is sent to the gradient category activation module. The gradient category activation module outputs the heat map of the segmented region h = grad_CAM(x*y, f).
[0028] x is fed into the Guided Backward Convolution module, and the Guided Backward Convolution module outputs the segmented region texture map t' = GuideBP(x, f);
[0029] Then the pseudo-label y_dom = y*h+t'.
[0030] Furthermore, the process of generating the heatmap of the segmented region is as follows:
[0031] Calculate the gradient of the final output of the gradient class activation module relative to the feature map of each channel of the output of the last convolutional layer in the gradient class activation module, and then calculate the weight of each channel of the feature map of the output of the last convolutional layer based on the gradient.
[0032] The weights are then summed with the feature maps of each channel to obtain a heatmap of the segmented region.
[0033] Furthermore, the data domain-aware image generation unit employs a gradient category activation module. This unit is used to generate pseudo-images, and the pseudo-image generation process is as follows:
[0034] For an image slice x, x is fed into the data domain sensing image generation unit to obtain a global heatmap g = grad_CAM(x, f). Then, the pseudo-image with old domain information is x*g.
[0035] Furthermore, in the pre-training process of step one, the loss function used is... seg for:
[0036]
[0037] Wherein, TP represents the number of true positive pixels in the predicted output of the second data domain medical image segmentation model, FP represents the number of false positive pixels in the predicted output of the second data domain medical image segmentation model, and FN represents the number of false negative pixels in the predicted output of the second data domain medical image segmentation model.
[0038] Furthermore, when the number of training stages t≥1, the loss function used is:
[0039] loss = loss da +λloss rk
[0040] Where, loss daIt is the domain adaptation loss of the second data domain medical image segmentation model. rk λ is the knowledge retention loss of the medical image segmentation model in the second data domain, where λ is the weight coefficient.
[0041] Furthermore, the domain adaptation loss da for:
[0042]
[0043] Where ζ represents the total number of image slices x in the t-th target domain of the input. t Let ζ be the set of all pixels in ζ, and |ζ| be the number of pixels contained in the set ζ. x represents all image slices of the t-th target domain from the input. t The i-th pixel in express The output of the second data domain medical image segmentation model express After passing through the Unet segmentation network in the first data domain medical image segmentation model, grad_CAM(·) represents the output of the gradient class activation module, and GuideBP(·) represents the output of the guided inverse convolution module;
[0044] The knowledge retention loss rk for:
[0045]
[0046] in, This refers to the pseudo-image generated by the data domain-aware image generation module. This represents the predicted output of the pseudo-image after passing through the Unet segmentation network in the first data domain medical image segmentation model. This represents the predicted output of the pseudo-image after passing through the Unet segmentation network in the second data domain medical image segmentation model.
[0047] Furthermore, the method for segmenting the MRI image to be segmented is as follows:
[0048] The MRI image to be segmented is sliced to obtain the sliced image. The obtained image is then input into the trained second data domain medical image segmentation model to obtain the segmentation result.
[0049] The beneficial effects of this invention are:
[0050] This invention constructs a causal chain between the image domain D, anatomical structure S, image x, and segmentation result y. In the causal chain of medical image segmentation, anatomical structure S and image domain D jointly determine image x, while anatomical structure S and image x jointly influence segmentation result y. In the domain adaptation sub-chain, adaptation to the new target domain is achieved by adjusting the influence of anatomical structure S on segmentation result y. A shape-aware label generation unit is designed to improve the model's domain adaptation capability. In the knowledge retention sub-chain, the influence of image domain D on image x and the influence of image x on segmentation result y are processed, thereby retaining the memory of old knowledge while learning new knowledge. A data domain-aware image generation unit is designed to improve the model's ability to resist forgetting.
[0051] The causal reasoning-based continuous passive domain adaptation model proposed in this invention is a novel continuous passive domain adaptation framework with high knowledge extensibility. It can significantly improve domain adaptation ability and solve the catastrophic forgetting problem of existing SFDA methods, and can be more widely used in clinical applications. Attached Figure Description
[0052] Figure 1 This is a causal chain diagram for the causal reasoning aspect of this invention;
[0053] Figure 2 This is a training flowchart for a medical image continuous passive domain adaptive segmentation system according to the present invention. Detailed Implementation
[0054] Specific Implementation Method 1: The medical image continuous passive domain adaptive segmentation system described in this implementation method includes an MRI image acquisition unit, an MRI image processing unit, a first data domain medical image segmentation model, and a second data domain medical image segmentation model. The first data domain medical image segmentation model includes a Unet segmentation network, a shape-aware label generation unit, and a data domain-aware image generation unit. The second data domain medical image segmentation model uses the Unet segmentation network.
[0055] The MRI image acquisition unit is used to acquire MRI images of T target domains and one source domain;
[0056] The MRI image processing unit is used to slice the MRI images of the source domain and each target domain to obtain slice images of the source domain and each target domain, respectively.
[0057] The first data domain medical image segmentation model and the second data domain medical image segmentation model are jointly trained based on slice images from the source domain and each target domain;
[0058] The trained second-data-domain medical image segmentation model is used to segment the MRI image to be segmented.
[0059] The Unet segmentation network consists of an encoder and a decoder. In the encoder, the encoder layer receives an input image of size 1×192×192 and, after a series of convolutional operations, outputs an image of size 64×192×192. The down1 unit includes a max-pooling layer and two convolutional layers to halve the size of the feature map; that is, the down1 unit outputs a feature map of size 128×96×96. In the down2 unit of the encoder, similar to the down1 unit, the feature map size is further halved, outputting a feature map of size 256×48×48. In the down3 unit, the feature map size is halved again, outputting a feature map of size 512×24×24. Finally, in the down4 unit of the encoder, the feature map size is halved again, outputting a feature map of size 1024×12×12.
[0060] In the decoder section, unit up1 contains an upsampling module that restores the feature map size to 512×24×24. Unit up2 is similar to unit up1, restoring the feature map size to 256×48×48. Unit up3 is similar to unit up1, restoring the feature map size to 128×96×96. Finally, in unit up4 of the decoder, the feature map size is restored again to 64×192×192.
[0061] The final output layer maps the output of the decoder to the category space, with an output size of 1×192×192, where 1 represents the number of categories.
[0062] Specific Implementation Method Two: Combining Figure 2 This embodiment is described below. The difference between this embodiment and specific embodiment one is that the joint training process of the first data domain medical image segmentation model and the second data domain medical image segmentation model is as follows:
[0063] Step 1: Use image slices from the source domain to pre-train the medical image segmentation model in the second data domain until the loss function converges, then stop the pre-training to obtain the pre-trained medical image segmentation model in the second data domain.
[0064] Then, the parameters of the Unet segmentation network in the first data domain medical image segmentation model are updated to the parameters of the second data domain medical image segmentation model pre-trained on the source domain image slices. The parameters of the shape-aware label generation unit and the data domain-aware image generation unit are calculated based on the parameters of the Unet segmentation network in the first data domain medical image segmentation model (the existing parameter calculation method is used here).
[0065] Step 2: Initialize the training phase t=1;
[0066] Step 3: Freeze all parameters of the first data domain medical image segmentation model, and train the second data domain medical image segmentation model using the image slice of the t-th target domain; the specific training process is as follows:
[0067] The image slice of the t-th target domain is used as the input of the Unet segmentation network in the first data domain medical image segmentation model and the second data domain medical image segmentation model, respectively. The output of the Unet segmentation network in the first data domain medical image segmentation model and the image slice of the t-th target domain are used as the input of the shape-aware label generation unit in the first data domain medical image segmentation model. The domain adaptation loss is calculated based on the output of the shape-aware label generation unit in the first data domain medical image segmentation model and the output of the second data domain medical image segmentation model.
[0068] The image slice of the t-th target domain is then used as the input of the data domain perceptual image generation unit in the first data domain medical image segmentation model. The output of the data domain perceptual image generation unit is then used as the input of the Unet segmentation network in the first data domain medical image segmentation model and the Unet segmentation network in the second data domain medical image segmentation model, respectively. The knowledge retention loss is then calculated based on the output of the Unet segmentation network in the first data domain medical image segmentation model and the Unet segmentation network in the second data domain medical image segmentation model.
[0069] The training in stage t is stopped when the sum of the domain adaptation loss and the knowledge retention loss converges, and the second data domain medical image segmentation model trained in stage t is obtained.
[0070] Step 4: Update all parameters of the first data domain medical image segmentation model based on the parameters of the second data domain medical image segmentation model trained in stage t.
[0071] Step 5: Determine if t < T is satisfied;
[0072] If t < T, then let t = t + 1 and return to step three;
[0073] If t = T, then the second data domain medical image segmentation model trained in the last training stage will be used as the final second data domain medical image segmentation model.
[0074] The other steps and parameters are the same as in Specific Implementation Method 1.
[0075] This invention has the following characteristics:
[0076] 1. Causal intervention mechanisms: such as Figure 1As shown, a causal chain was constructed between the image domain D, anatomical structure S, image x, and segmentation result y using causal intervention theory. By adjusting the influence of anatomical structure S on segmentation result y, adaptation to the new target domain was achieved, and a shape-aware label generation unit was designed to improve the model's domain adaptability. The influence of image domain D on image x and the influence of image x on segmentation result y were addressed, thus retaining old knowledge while learning new knowledge. A data domain-aware image generation unit was designed to improve the model's resistance to forgetting.
[0077] 2. Dual-module design: A shape-aware label generation unit and a data-domain-aware image generation unit were designed. The shape-aware label generation unit generates pseudo-labels for domain adaptation using Grad_CAM and GuideBP, improving the model's segmentation performance in the target domain; the data-domain-aware image generation unit generates pseudo-images with old domain information using Grad_CAM, providing the segmentation model with training data to retain old knowledge and enhancing the model's resistance to forgetting.
[0078] Grad_CAM and GuideBP are both tools for interpreting model behavior, but they employ different approaches to achieve this goal. Grad_CAM primarily focuses on the model's overall region of interest in the input image, while GuideBP places greater emphasis on the sensitivity to gradient changes at the individual pixel level.
[0079] 3. Passive Domain Adaptation Framework: This invention enables the model to migrate from the source domain to the target domain using only target domain data, and can continuously adapt to the constantly changing target data, thus solving the problem of source domain data privacy protection.
[0080] 4. Loss function design: During training, the domain adaptation loss and knowledge retention loss are combined. By adjusting the weight coefficients, the domain adaptation ability and knowledge retention ability of the model are balanced to ensure the stability of the model's performance on both new and old domain data.
[0081] 5. Practical Application: This invention has broad application prospects in clinical applications and can significantly improve the robustness and stability of medical image segmentation models under different data distributions.
[0082] Specific Implementation Method 3: This implementation method differs from Specific Implementation Method 1 or 2 in that the image slices need to undergo a cropping operation before being input into the model to obtain image slices of size n×n.
[0083] Other steps and parameters are the same as in specific implementation method one or two.
[0084] In this embodiment, the value of n can be 192, but is not limited to 192. This can avoid interference from inconsistent image sizes during training and improve the segmentation accuracy of the final model.
[0085] Specific implementation method four: Combination Figure 2 This embodiment is described below. The difference between this embodiment and one of the specific embodiments one to three is that the shape-aware label generation unit includes a gradient category activation module Grad_CAM and a guided inverse convolution module GuideBP; the shape-aware label generation unit is used to generate pseudo-labels.
[0086] The working process of the shape-aware label generation unit is as follows:
[0087] For an image slice x, the output of the Unet segmentation network after image slice x is a segmentation probability map y. After multiplying x and y, the multiplication result is sent to the gradient category activation module. The gradient category activation module outputs the heat map of the segmented region h = grad_CAM(x*y, f).
[0088] x is fed into the Guided Backward Convolution module, and the Guided Backward Convolution module outputs the segmented region texture map t' = GuideBP(x, f);
[0089] Then the pseudo-label y_dom used for domain adaptation is y*h+t'.
[0090] The other steps and parameters are the same as those in one of the specific implementation methods one to three.
[0091] This invention utilizes pseudo-labels y_dom to represent anatomical structures S. During the backpropagation process, GuideBP only retains the gradients of positive activations, which can generate more refined texture activation maps and thus better explain the behavior of the model.
[0092] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that the process of generating the heat map of the segmented region is as follows:
[0093] Calculate the gradient of the final output of the gradient class activation module relative to the feature map of each channel of the output of the last convolutional layer in the gradient class activation module, and then calculate the weight of each channel of the feature map of the output of the last convolutional layer based on the gradient.
[0094] The weights are then summed with the feature maps of each channel to obtain a heatmap of the segmented region.
[0095] The other steps and parameters are the same as those in one of the specific implementation methods one to four.
[0096] The gradient category activation module (grad_CAM) is used to determine which parts of the image are important for the model's predictions and to generate a heatmap to visualize the important regions.
[0097] Specific Implementation Method Six: Combination Figure 2This embodiment is described below. The difference between this embodiment and one of the specific embodiments one to five is that the data domain-aware image generation unit uses a gradient category activation module (Grad_CAM). The data domain-aware image generation unit is used to generate pseudo-images, and the pseudo-image generation process is as follows:
[0098] For an image slice x, x is fed into the data domain sensing image generation unit to obtain a global heatmap g = grad_CAM(x, f). Then, the pseudo-image with old domain information is x*g.
[0099] The other steps and parameters are the same as those in one of the specific implementation methods one to five.
[0100] The image domain D is represented by pseudo-images, and the pseudo-images generated in this embodiment are provided to the segmentation model to retain training data of old knowledge and solve the problem of catastrophic forgetting.
[0101] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that, in the pre-training process of Step One, the loss function used is... seg for:
[0102]
[0103] Wherein, TP represents the number of true positive pixels in the predicted output of the second data domain medical image segmentation model, FP represents the number of false positive pixels in the predicted output of the second data domain medical image segmentation model, and FN represents the number of false negative pixels in the predicted output of the second data domain medical image segmentation model.
[0104] The other steps and parameters are the same as those in one of the specific implementation methods one to six.
[0105] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One through Seven in that, when the number of training stages t≥1, the loss function used is:
[0106] loss = loss da +λloss rk
[0107] Where, loss da It is the domain adaptation loss of the second data domain medical image segmentation model. rk λ is the knowledge retention loss of the medical image segmentation model in the second data domain, where λ is the weight coefficient.
[0108] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.
[0109] In this embodiment, the weight coefficient λ is set to 1. By setting an appropriate weight coefficient, the segmentation accuracy of the model can be improved.
[0110] Specific Implementation Method Nine: Combining Figure 2 This embodiment is described below. The difference between this embodiment and specific embodiments one through eight is that the domain adaptation loss... da for:
[0111]
[0112] Where ζ represents the total number of image slices x in the t-th target domain of the input. t Let ζ be the set of all pixels in ζ, and |ζ| be the number of pixels contained in the set ζ. x represents all image slices of the t-th target domain from the input. t The i-th pixel in express The output of the second data domain medical image segmentation model express After passing through the Unet segmentation network in the first data domain medical image segmentation model, grad_CAM(·) represents the output of the gradient class activation module, and GuideBP(·) represents the output of the guided inverse convolution module;
[0113] The knowledge retention loss rk for:
[0114]
[0115] in, This refers to the pseudo-image generated by the data domain-aware image generation module. This represents the predicted output of the pseudo-image after passing through the Unet segmentation network in the first data domain medical image segmentation model. This represents the predicted output of the pseudo-image after passing through the Unet segmentation network in the second data domain medical image segmentation model.
[0116] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.
[0117] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods One to Nine in that the method for segmenting the MRI image to be segmented is as follows:
[0118] The MRI image to be segmented is sliced to obtain the sliced image. The obtained image is then input into the trained second data domain medical image segmentation model to obtain the segmentation result.
[0119] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.
[0120] The segmentation process in this embodiment is equivalent to:
[0121] y = f θ (x)
[0122] Where x represents the input medical image slice to be segmented, and y represents the output of the trained second data domain medical image segmentation model.
[0123] When using the segmentation model trained by this invention to segment MRI images, a source domain image is not required.
[0124] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A medical image continuous passive domain adaptive segmentation system, characterized in that, The segmentation system includes an MRI image acquisition unit, an MRI image processing unit, a first data domain medical image segmentation model, and a second data domain medical image segmentation model. The first data domain medical image segmentation model includes a Unet segmentation network, a shape-aware label generation unit, and a data domain-aware image generation unit. The second data domain medical image segmentation model uses the Unet segmentation network. The MRI image acquisition unit is used to acquire MRI images of T target domains and one source domain; The MRI image processing unit is used to slice the MRI images of the source domain and each target domain to obtain slice images of the source domain and each target domain, respectively. The first data domain medical image segmentation model and the second data domain medical image segmentation model are jointly trained based on slice images from the source domain and each target domain; The joint training process of the first data domain medical image segmentation model and the second data domain medical image segmentation model is as follows: Step 1: Use image slices from the source domain to pre-train the medical image segmentation model in the second data domain until the loss function converges, then stop the pre-training to obtain the pre-trained medical image segmentation model in the second data domain. Then, the parameters of the Unet segmentation network in the first data domain medical image segmentation model are updated to the parameters of the second data domain medical image segmentation model pre-trained on the source domain image slices, and the parameters of the shape-aware label generation unit and the data domain-aware image generation unit are calculated based on the parameters of the Unet segmentation network in the first data domain medical image segmentation model. Step 2: Initialize the training phase t=1; Step 3: Freeze all parameters of the first data domain medical image segmentation model, and train the second data domain medical image segmentation model using the image slice of the t-th target domain; the specific training process is as follows: The image slice of the t-th target domain is used as the input of the Unet segmentation network in the first data domain medical image segmentation model and the second data domain medical image segmentation model, respectively. The output of the Unet segmentation network in the first data domain medical image segmentation model and the image slice of the t-th target domain are used as the input of the shape-aware label generation unit in the first data domain medical image segmentation model. The domain adaptation loss is calculated based on the output of the shape-aware label generation unit in the first data domain medical image segmentation model and the output of the second data domain medical image segmentation model. The image slice of the t-th target domain is then used as the input of the data domain perceptual image generation unit in the first data domain medical image segmentation model. The output of the data domain perceptual image generation unit is then used as the input of the Unet segmentation network in the first data domain medical image segmentation model and the Unet segmentation network in the second data domain medical image segmentation model, respectively. The knowledge retention loss is then calculated based on the output of the Unet segmentation network in the first data domain medical image segmentation model and the Unet segmentation network in the second data domain medical image segmentation model. The training in stage t is stopped when the sum of the domain adaptation loss and the knowledge retention loss converges, and the second data domain medical image segmentation model trained in stage t is obtained. Step 4: Update all parameters of the first data domain medical image segmentation model based on the parameters of the second data domain medical image segmentation model trained in stage t. Step 5: Determine if t < T is satisfied; If t < T, then let t = t + 1 and return to step three; If t=T, then the second data domain medical image segmentation model trained in the last training stage will be used as the final second data domain medical image segmentation model. The trained second-data-domain medical image segmentation model is used to segment the MRI image to be segmented.
2. The medical image continuous passive domain adaptive segmentation system according to claim 1, characterized in that, Before being input into the model, the image slices first need to undergo a cropping operation to obtain a size of [size missing]. Image slices.
3. The medical image continuous passive domain adaptive segmentation system according to claim 1, characterized in that, The shape-aware label generation unit includes a gradient category activation module and a guided inverse convolution module; the shape-aware label generation unit is used to generate pseudo-labels. The working process of the shape-aware label generation unit is as follows: For an image slice x, the output of the Unet segmentation network after image slice x is a segmentation probability map y. After multiplying x and y, the multiplication result is sent to the gradient category activation module. The gradient category activation module outputs the heat map of the segmented region h=grad_CAM(x * y, f). x is fed into the Guided Backward Convolution module, and the Guided Backward Convolution module outputs the segmented region texture map t'=GuideBP(x, f); Then the pseudo-label y_dom = y * h + t'.
4. A medical image continuous passive domain adaptive segmentation system according to claim 3, characterized in that, The process of generating the heatmap of the segmented region is as follows: calculate the gradient of the final output of the gradient category activation module relative to the feature map of each channel output of the last convolutional layer in the gradient category activation module, and then calculate the weight of each channel feature map output of the last convolutional layer based on the gradient. The weights are then summed with the feature maps of each channel to obtain a heatmap of the segmented region.
5. A medical image continuous passive domain adaptive segmentation system according to claim 4, characterized in that, The data domain-aware image generation unit employs a gradient category activation module. This unit is used to generate pseudo-images, and the pseudo-image generation process is as follows: For an image slice x, x is fed into the data domain sensing image generation unit to obtain a global heatmap g=grad_CAM(x,f), and the pseudo-image with old domain information is x*g.
6. A medical image continuous passive domain adaptive segmentation system according to claim 5, characterized in that, The loss function used in the pre-training process of step one. for: Wherein, TP represents the number of true positive pixels in the predicted output of the second data domain medical image segmentation model, FP represents the number of false positive pixels in the predicted output of the second data domain medical image segmentation model, and FN represents the number of false negative pixels in the predicted output of the second data domain medical image segmentation model.
7. A medical image continuous passive domain adaptive segmentation system according to claim 6, characterized in that, When the number of training stages t≥1, the loss function used is... for: in, It is the domain adaptation loss of the medical image segmentation model in the second data domain. It is the knowledge preservation loss of the medical image segmentation model in the second data domain. These are the weighting coefficients.
8. A medical image continuous passive domain adaptive segmentation system according to claim 7, characterized in that, The domain adaptation loss for: in, This represents all image slices of the t-th target domain from the input. The set of all pixels in the set. It is a set The number of pixels contained in This represents all image slices of the t-th target domain from the input. The i-th pixel in express The output of the second data domain medical image segmentation model express The output after passing through the Unet segmentation network in the first data domain medical image segmentation model. The output of the gradient category activation module, GuideBP This indicates the output of the guided inverse convolution module; The loss of knowledge retention for: in, This refers to the pseudo-image generated by the data domain-aware image generation module. This represents the predicted output of the pseudo-image after passing through the Unet segmentation network in the first data domain medical image segmentation model. This represents the predicted output of the pseudo-image after passing through the Unet segmentation network in the second data domain medical image segmentation model.
9. A medical image continuous passive domain adaptive segmentation system according to claim 8, characterized in that, The method for segmenting the MRI image to be segmented is as follows: The MRI image to be segmented is sliced to obtain the sliced image. The obtained image is then input into the trained second data domain medical image segmentation model to obtain the segmentation result.