Shape-guided unsupervised domain adaptation method and system for medical image segmentation

Abstract

The application discloses a shape-guided unsupervised domain adaptation medical image segmentation method and system, which constructs a framework in an unsupervised domain adaptation manner, can greatly relieve the serious performance degradation, and can learn the structure information of the medical image in an unsupervised condition, is helpful to more accurately segment different tissue structures and organs, and improves the segmentation precision. Meanwhile, the training mode of learning the shape information is more easily trained than the generative adversarial network. The shape and structure information of the organ is learned in a manner of completing a label mask, so that the shape learning model can complete the organ shape of each person, instead of only learning the average template of the organ in a data set. In this way, the domain distance between different modalities and even different organs can be greatly reduced, so that the performance is not sharply reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and medical image processing, and more particularly, to a shape-guided unsupervised domain adaptation medical image segmentation method and system. BACKGROUND

[0002] One of the main problems in the clinic is that the preferred way and scanning scheme adopted by different hospitals can vary greatly. Different CT machines and schemes will lead to different scanning intervals, slice thicknesses, and variations in organ intensity and texture, which results in significant domain differences between medical data sets. Traditional supervised learning methods are often limited in practical applications and are difficult to achieve satisfactory performance in unlabeled domains. That is, a model trained on a specific source domain in a certain hospital will usually reduce performance if directly applied to data obtained from other hospitals without fine-tuning for the target data. In addition, medical image segmentation requires large amounts of annotated data to achieve good performance, which is available in the labeled source domain but not in the unknown target domain. In addition, there are difficulties in labeled samples because professional medical annotation usually requires a lot of time and professional knowledge.

[0003] In previous unsupervised domain adaptation medical image segmentation, the main focus was on solving texture differences to narrow the domain gap. However, due to the complexity of texture, it is very challenging to obtain a general model that performs well on target data sets. Previous shape learning methods either constrain the shape by constraining the segmentation boundary, or use variational autoencoders to learn the shape of the organ, or use implicit space representation to learn the shape of the organ. However, these methods are often only suitable for small domain distances, but when faced with cross-modal situations, the performance will decrease significantly. In addition, existing shape modeling methods often only learn the average shape of a specific organ within a certain data set, and have poor generalizability. When faced with a large domain distance, the source domain segmentation model is often severely affected by noise, resulting in a significant degradation in segmentation result quality.

[0004] The prior art provides an unsupervised domain adaptation image segmentation method and system, which includes training a segmentation model and a generative model based on images with segmentation labels in the source domain data; inputting images without segmentation labels in the target domain data into the segmentation model to obtain corresponding segmentation prediction labels; inputting the target domain data and the corresponding segmentation prediction labels into the generative model to obtain matching reference images and segmentation reference labels; taking segmentation prediction labels with an overlap rate with the segmentation reference labels that meets a preset value as quasi-segmentation labels, and training the segmentation model again based on the quasi-segmentation labels and the corresponding target domain data and re-determining the quasi-segmentation labels until the segmentation accuracy of the segmentation model meets a preset condition to end the training. SUMMARY

[0005] The present application provides a shape-guided unsupervised domain adaptation medical image segmentation method and system to overcome the above-mentioned defects of excessive domain distance, severe noise and poor generalization caused by few organ types of the learner.

[0006] To solve the above technical problems, the technical scheme of the present application is as follows:

[0007] The present application provides a shape-guided unsupervised domain adaptation medical image segmentation method, which comprises:

[0008] Obtaining source domain image data and target domain image data;

[0009] Preprocessing the source domain image and the target domain image to obtain the preprocessed source domain image and the target domain image;

[0010] Training the source segmentation model and the shape guide model constructed according to the preprocessed source domain image to obtain the trained shape guide model, the trained source segmentation model and the network parameters thereof;

[0011] Training the pseudo-label segmentation model constructed based on the preprocessed target domain image data and the trained shape guide model to obtain the trained pseudo-label segmentation model and the network parameters thereof; wherein the network parameters of the trained source segmentation model are used as the initial network parameters of the constructed pseudo-label segmentation model;

[0012] Training the target segmentation model constructed based on the preprocessed target domain image data, the trained shape guide model and the trained pseudo-label segmentation model to obtain the trained target segmentation model; wherein the network parameters of the trained pseudo-label segmentation model are used as the initial network parameters of the constructed target segmentation model;

[0013] Inputting the image to be segmented into the trained target segmentation model to obtain the target image segmentation result.

[0014] Preferably, the specific method for preprocessing is:

[0015] Adjusting the source domain image and the target domain image to the same layer thickness, performing random rotation and random translation operation, and then performing random intensity scaling to obtain the preprocessed source domain image and the target domain image.

[0016] Preferably, the specific method for obtaining the trained source segmentation model is:

[0017] Constructing a segmentation model, wherein the segmentation model comprises an encoder and a decoder connected in sequence;

[0018] The preprocessed source domain image is extracted by using an encoder to obtain a source domain image feature;

[0019] The source domain image feature is decoded by a decoder to obtain a source domain segmentation result;

[0020] A first Dice loss function is set, the first Dice loss function value is calculated according to the source domain segmentation result and a real label, and the parameter adjustment of the constructed segmentation model is performed;

[0021] When the first Dice loss function value reaches a preset value or reaches a preset training number of times, a trained segmentation model is obtained.

[0022] Preferably, the first Dice loss function is:

[0023]

[0024] Wherein, is the source domain segmentation result, y s is the real label.

[0025] Preferably, the specific method for obtaining the trained shape guide model is:

[0026] The segmentation label in the preprocessed source domain image is masked to obtain a masked label image;

[0027] The masked label image is input into the constructed shape guide model to obtain a reconstructed segmentation label image;

[0028] A reconstruction loss function is set, the reconstruction loss function value is calculated according to the reconstructed segmentation label image and the segmentation label image, and the parameter adjustment of the constructed shape guide model is performed;

[0029] When the reconstruction loss function value reaches a preset value, a trained shape guide model is obtained.

[0030] Preferably, the specific method for obtaining the trained pseudo label segmentation model and its network parameters is:

[0031] The preprocessed target domain image data is input into the constructed pseudo label segmentation model to obtain a pseudo label;

[0032] The pseudo label is input into the trained shape guide model to obtain a corresponding reconstructed pseudo label;

[0033] A second Dice loss function is set, the second Dice loss function value is calculated according to the reconstructed pseudo label and the pseudo label, and when the second Dice loss function value reaches a preset value, a trained pseudo label segmentation model and its network parameters are obtained.

[0034] Preferably, the second Dice loss function is:

[0035]

[0036] wherein, is a pseudo label; is a reconstructed pseudo label.

[0037] Preferably, the specific method for obtaining the trained target segmentation model is:

[0038] constructing a target segmentation model, taking the network parameters of the trained pseudo label segmentation model as the initial network parameters of the target segmentation model;

[0039] inputting the preprocessed target domain image data into the trained pseudo label segmentation model and the constructed target segmentation model respectively to obtain a pseudo label and a target segmentation label; inputting the target segmentation label into the trained shape guide model to obtain a target reconstructed label;

[0040] setting a total loss function, calculating the total loss function value according to the pseudo label, the target segmentation label and the target reconstructed label, and obtaining the trained target segmentation model when the total loss function value reaches a preset value.

[0041] Preferably, the total loss function includes a reconstruction loss function and a pseudo loss function, and the total loss function is:

[0042]

[0043] wherein, Λ pseudo is a hyperparameter, is a pseudo loss function; is a reconstruction loss function;

[0044] the reconstruction loss function is:

[0045]

[0046] wherein, is a target segmentation label; is a target reconstructed label;

[0047] the pseudo loss function pseudo:

[0048]

[0049] wherein, is a pseudo label.

[0050] The application also provides an unsupervised domain adaptation medical image segmentation system based on shape guidance, which is used to realize the above method, and the system comprises:

[0051] a data acquisition module, which acquires source domain image data and target domain image data;

[0052] a preprocessing module, which pre-processes the source domain image and the target domain image to obtain pre-processed source domain image and target domain image;

[0053] a first model training module, which trains a source segmentation model and a shape guide model constructed according to the pre-processed source domain image to obtain a trained shape guide model, a trained source segmentation model and network parameters thereof;

[0054] a second model training module, which trains a pseudo-label segmentation model constructed based on the pre-processed target domain image data and the trained shape guide model to obtain a trained pseudo-label segmentation model and network parameters thereof; wherein the network parameters of the trained source segmentation model are used as initial network parameters of the constructed pseudo-label segmentation model;

[0055] a target model training module, which trains a target segmentation model constructed based on the pre-processed target domain image data, the trained shape guide model and the trained pseudo-label segmentation model to obtain a trained target segmentation model; wherein the network parameters of the trained pseudo-label segmentation model are used as initial network parameters of the constructed target segmentation model;

[0056] a segmentation module, which inputs an image to be segmented into the trained target segmentation model to obtain a target image segmentation result.

[0057] Compared with the prior art, the technical scheme of the present application has the following beneficial effects:

[0058] The present application can learn the structural information of medical images in an unsupervised manner by constructing an unsupervised domain adaptation segmentation model based on shape guidance, which helps to more accurately segment different tissue structures and organs and improve segmentation accuracy. By learning the shape and structure information of organs through label mask completion, the shape guide model can complete the shape of each person's organ, rather than just learning the average template of organs in a data set, which can greatly alleviate the serious performance degradation. This method can greatly reduce the situation that the domain distance between different modalities or even different organs is too large, resulting in a sharp decline in performance. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 a flowchart of the shape guide-based unsupervised domain adaptation medical image segmentation method described in embodiment 1;

[0060] Figure 2 a structural schematic diagram of the first model training module, the second model training module and the target model training module in embodiment 2;

[0061] Figure 3 a structural schematic diagram of the source segmentation model in embodiment 2;

[0062] Figure 4 a structural schematic diagram of the second model training module in embodiment 2;

[0063] Figure 5 a structural schematic diagram of the target model training module in embodiment 2;

[0064] Figure 6 a structural schematic diagram of the shape guide based unsupervised domain adaptation medical image segmentation system in embodiment 3. DETAILED DESCRIPTION

[0065] The accompanying drawings are only used for illustrative purposes and should not be construed as limiting the patent;

[0066] In order to better illustrate the embodiments, some components in the drawings may be omitted, enlarged or reduced, and do not represent the actual size of the product;

[0067] It is understandable that some well-known structures and their descriptions in the drawings may be omitted for those skilled in the art.

[0068] The technical solutions of the present application will be further described below in combination with the drawings and embodiments.

[0069] Embodiment 1

[0070] The present embodiment provides a shape guide based unsupervised domain adaptation medical image segmentation method, as shown in Figure 1 The method comprises the following steps:

[0071] Obtain source domain image data and target domain image data;

[0072] Preprocess the source domain image and the target domain image to obtain the preprocessed source domain image and the target domain image;

[0073] According to the preprocessed source domain image, the source segmentation model and the shape guide model are trained to obtain the trained shape guide model, the trained source segmentation model and the network parameters thereof;

[0074] Based on the preprocessed target domain image data and the trained shape guide model, the pseudo label segmentation model is trained to obtain the trained pseudo label segmentation model and the network parameters thereof; wherein the network parameters of the trained source segmentation model are used as the initial network parameters of the constructed pseudo label segmentation model;

[0075] Based on the pre-processed target domain image data, the trained shape guide model and the trained pseudo-label segmentation model, a target segmentation model is trained to obtain a trained target segmentation model; wherein the network parameters of the trained pseudo-label segmentation model are used as the initial network parameters of the constructed target segmentation model;

[0076] The to-be-segmented image is input into the trained target segmentation model to obtain a target image segmentation result.

[0077] Embodiment 2

[0078] The embodiment provides a shape guide-based unsupervised domain adaptation medical image segmentation method, and the method comprises the following steps:

[0079] Obtain source domain image data and target domain image data;

[0080] We take the open-source database NIH as our source domain dataset, wherein the training set has 51 samples, the verification set has 10 samples, the test set has 21 samples, and the target domain dataset can be any. In the present application, we take the open-source MSD dataset as our target domain, wherein the training set has 180 samples, the verification set has 30 samples, and the test set has 72 samples.

[0081] The source domain image data contains labeled data and is represented as wherein, represents an image, represents an image of the corresponding ground truth, and N is the number of samples in the dataset. The target domain image data does not contain segmentation labels and is represented as wherein represents an image, and M is the number of samples in the dataset.

[0082] Pre-process the source domain image and the target domain image to obtain pre-processed source domain image and target domain image;

[0083] The specific method for pre-processing is as follows:

[0084] The source domain image and the target domain image are adjusted to the same layer thickness, and then random rotation and random translation operations are performed, and then random intensity scaling is performed to obtain the pre-processed source domain image and the target domain image.

[0085] Because the layer thickness of medical images collected by different devices is inconsistent, it undoubtedly increases the difficulty of model training, so we process the layer thickness of all medical images to be the same, and secondly, because the number of CT images is small, in order to make full use of limited CT images, we further randomly rotate the CT images within 20 degrees and randomly translate within 5 voxels, and apply random intensity scaling between 0.85-1.15 to increase the data quantity and diversity of limited CT images. At the same time, the data is processed as follows: 1. All CT images and corresponding labels are uniformly processed to (1x1x1) mm3. 2. Convert all image intensities to the range of -200 to 400, and then further normalize them to -1 to 1.

[0086] As shown in Figure 2 , the source segmentation model and the shape guide model are trained according to the preprocessed source domain image pairs, and the trained shape guide model, the trained source segmentation model and their network parameters are obtained;

[0087] As shown in Figure 3 , the specific method for obtaining the trained source segmentation model is:

[0088] The segmentation model is constructed, and the segmentation model comprises an encoder and a decoder connected in sequence;

[0089] The preprocessed source domain image is extracted by using the encoder, and the source domain image feature is obtained;

[0090] The source domain image feature is decoded by the decoder to obtain the source domain segmentation result;

[0091] The first Dice loss function is set, the first Dice loss function value is calculated according to the source domain segmentation result and the real label, and the parameter adjustment of the constructed segmentation model is performed;

[0092] When the first Dice loss function value reaches the preset value or the training iteration number reaches 100 epochs, the trained source segmentation model is obtained.

[0093] The first Dice loss function is:

[0094]

[0095] Wherein, is the source domain segmentation result, y s is the real label.

[0096] The specific method for obtaining the trained shape guide model is:

[0097] The segmentation label in the preprocessed source domain image is masked to obtain a masked label image;

[0098] input the image of the occluded label into the constructed shape guide model, to obtain a reconstructed segmentation label image;

[0099] A reconstruction loss function is set, and the reconstruction loss function value is calculated according to the reconstructed segmentation label image and the segmentation label image, and the parameter adjustment of the constructed shape guide model is performed;

[0100] When the reconstruction loss function value reaches a preset value, a trained shape guide model is obtained.

[0101] Specifically, we take the label mask of the source domain as input, and then sample the image blocks containing organ pixels that contain local shape features of the organ. Formally, first, input the organ real mask G∈R H×W×D×C Divide into regular non-overlapping 3D image blocks Where C is the input channel, (H, W, D) is the resolution of the organ label mask, P is the image block size, N p =H·W·D / P 3 represents the number of image blocks. Then randomly extract image blocks containing organ pixels for high-occlusion mask, and the remaining image blocks are sent into the Transformer block as the encoder block as visible image blocks. Finally, the learned mask token from the encoder and the occluded image block representation are input into the decoder as the input to obtain the reconstructed mask Gr. The loss function is the mean square error between the pixel-level input label and the pixel-level reconstructed label:

[0102]

[0103] Where, Ω is the set of all image blocks, G is the input label, and Gr is the reconstructed label.

[0104] As Figure 4 shown, based on the preprocessed target domain image data and the trained shape guide model, the constructed pseudo label segmentation model is trained to obtain a trained pseudo label segmentation model and network parameters thereof; wherein the network parameters of the trained source segmentation model are used as initial network parameters of the constructed pseudo label segmentation model;

[0105] The specific method for obtaining the trained pseudo label segmentation model and the network parameters thereof is:

[0106] The preprocessed target domain image data is input into the constructed pseudo label segmentation model to obtain a pseudo label;

[0107] The pseudo label is input into the trained shape guide model to obtain a corresponding reconstructed pseudo label;

[0108] The second Dice loss function is set, and the second Dice loss function value is calculated according to the reconstructed pseudo label and the pseudo label; when the second Dice loss function value reaches a preset value, the trained pseudo label segmentation model and network parameters thereof are obtained.

[0109] The input image I of the target domain data set t is taken as input of the segmentation network S R , the parameters of the segmentation network S R are copied from the source segmentation network S S , and the output of the segmentation network S R is obtained, denoted as Then, is taken as input and input into the shape guiding model to obtain the reconstructed pseudo label We use the Dice loss as the loss function between the output of the segmentation network and the output of the shape learning model , and the second Dice loss function is:

[0110]

[0111] Wherein, is the pseudo label; is the reconstructed pseudo label.

[0112] As shown in Figure 5 , based on the preprocessed target domain image data, the trained shape guiding model and the trained pseudo label segmentation model, the constructed target segmentation model is trained to obtain a trained target segmentation model; wherein the network parameters of the trained pseudo label segmentation model are taken as the initial network parameters of the constructed target segmentation model.

[0113] The specific method for obtaining the trained target segmentation model is:

[0114] The target segmentation model is constructed, and the network parameters of the trained pseudo label segmentation model are taken as the initial network parameters of the target segmentation model.

[0115] The preprocessed target domain image data is input into the trained pseudo label segmentation model and the constructed target segmentation model respectively, to obtain a pseudo label and a target segmentation label; the target segmentation label is input into the trained shape guiding model to obtain a target reconstructed label.

[0116] The total loss function is set, and the total loss function value is calculated according to the pseudo label, the target segmentation label and the target reconstructed label; when the total loss function value reaches a preset value, the trained target segmentation model is obtained.

[0117] In order to obtain the target segmentation network S T, the parameters of the pseudo model are copied to the target model as initialization parameters. The input image I t of the target domain dataset is input into the segmentation network S T and the pseudo label model S R respectively, and the segmentation result obtained by the pseudo label model is The segmentation result obtained by the target segmentation network S T is denoted as The loss is calculated between the result of the pseudo label model and the result of the target segmentation network Then, the result is input into the shape guiding model as , and the result after reconstruction of the shape guiding model is obtained The loss is calculated between the result of the target segmentation network and the reconstruction result of the shape guiding model The two losses jointly optimize the target network; the total loss function includes a reconstruction loss function and a pseudo loss function, and the total loss function is:

[0118]

[0119] Wherein, Λ pseudo is a hyperparameter, is a pseudo loss function; is a reconstruction loss function;

[0120] We add the shape guiding model to the optimization of the target segmentation model, and reduce the influence of noise on the segmentation result through the shape reconstruction loss , so as to obtain the target model S T with an explicit shape model, and the reconstruction loss function is:

[0121]

[0122] Wherein, is a target segmentation label; is a target reconstruction label;

[0123] However, the loss of is only related to the shape guiding model, which may cause the target segmentation model to collapse. Therefore, another loss function is needed to prevent the segmentation network S T from collapsing. Because the pseudo label segmentation network S R can provide more accurate pseudo labels and prevent the target segmentation model from generating the same prediction output with small reconstruction loss. Secondly, the combined effect of the pseudo label and the reconstructed image enables the target model to more accurately learn the shape information of the organ. Therefore, the pseudo label predicted by the pseudo label segmentation network S R can be used as a constraint to guide the reconstruction loss. The Dice loss between the predictions from the source and target segmentation networks is defined as the pseudo loss, and the pseudo loss function is:

[0124]

[0125] wherein, is a pseudo label.

[0126] inputting the image to be segmented into the trained target segmentation model to obtain a target image segmentation result.

[0127] Embodiment 3

[0128] The embodiment also provides an unsupervised domain adaptation medical image segmentation system based on shape guidance, which is used to implement the method of embodiments 1 or 2, as shown in Figure 6 The system comprises:

[0129] a data acquisition module, which acquires source domain image data and target domain image data;

[0130] a preprocessing module, which pre-processes the source domain image and the target domain image to obtain pre-processed source domain image and target domain image;

[0131] a first model training module, which trains a source segmentation model and a shape guidance model constructed according to the pre-processed source domain image, to obtain a trained shape guidance model, a trained source segmentation model and network parameters thereof;

[0132] a second model training module, which trains a pseudo label segmentation model constructed based on the pre-processed target domain image data and the trained shape guidance model, to obtain a trained pseudo label segmentation model and network parameters thereof; wherein the network parameters of the trained source segmentation model are used as initial network parameters of the constructed pseudo label segmentation model;

[0133] a target model training module, which trains a target segmentation model constructed based on the pre-processed target domain image data, the trained shape guidance model and the trained pseudo label segmentation model, to obtain a trained target segmentation model; wherein the network parameters of the trained pseudo label segmentation model are used as initial network parameters of the constructed target segmentation model;

[0134] a segmentation module, which inputs the image to be segmented into the trained target segmentation model to obtain a target image segmentation result.

[0135] The same or similar reference signs correspond to the same or similar components;

[0136] The terms describing the positional relationship in the drawings are only used for exemplary illustration, and cannot be understood as a limitation on the patent;

[0137] Obviously, the above embodiments of the present application are merely exemplary but not intended to limit the embodiments of the present application. Based on the above description, any other variations or changes can be made by those skilled in the art without departing from the spirit and principles of the present application. It is not necessary to list all the embodiments here. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present application shall fall within the scope of the claims of the present application.

Claims

1. A shape-guided unsupervised domain adaptation medical image segmentation method, characterized in that, The method comprises: obtaining source domain image data and target domain image data; preprocessing the source domain image and the target domain image to obtain preprocessed source domain image and target domain image; training the constructed source segmentation model and shape guide model according to the preprocessed source domain image, to obtain a trained shape guide model, a trained source segmentation model and network parameters thereof; training the constructed pseudo-label segmentation model based on the preprocessed target domain image data and the trained shape guide model, to obtain a trained pseudo-label segmentation model and network parameters thereof; wherein the network parameters of the trained source segmentation model are used as initial network parameters of the constructed pseudo-label segmentation model; training the constructed target segmentation model based on the preprocessed target domain image data, the trained shape guide model and the trained pseudo-label segmentation model, to obtain a trained target segmentation model; wherein the network parameters of the trained pseudo-label segmentation model are used as initial network parameters of the constructed target segmentation model; inputting the image to be segmented into the trained target segmentation model to obtain a target image segmentation result; wherein the method for obtaining the trained shape guide model comprises: masking the segmentation label in the preprocessed source domain image to obtain a masked label image; inputting the masked label image into the constructed shape guide model to obtain a reconstructed segmentation label image; setting a reconstruction loss function, calculating the reconstruction loss function value according to the reconstructed segmentation label image and the segmentation label image, and adjusting the parameters of the constructed shape guide model; when the reconstruction loss function value reaches a preset value, the trained shape guide model is obtained.

2. The shape-guide-based unsupervised domain adaptation medical image segmentation method according to claim 1, characterized in that, The specific method for preprocessing comprises: adjusting the source domain image and the target domain image to the same layer thickness, performing random rotation and random translation operations, and then performing random intensity scaling to obtain the preprocessed source domain image and the target domain image. 3.The shape-guided unsupervised domain adaptation medical image segmentation method of claim 1, wherein, The specific method for obtaining the trained source segmentation model comprises: constructing a segmentation model, wherein the segmentation model comprises an encoder and a decoder connected in sequence; extracting the preprocessed source domain image using the encoder to obtain a source domain image feature; decoding the source domain image feature through the decoder to obtain a source domain segmentation result; setting a first Dice loss function, calculating the first Dice loss function value according to the source domain segmentation result and the true label, and adjusting the parameters of the constructed segmentation model; when the first Dice loss function value reaches a preset value or reaches a preset training number of times, the trained source segmentation model is obtained.

4. The shape-guided unsupervised domain adaptation medical image segmentation method according to claim 3, characterized in that, The first Dice loss function is: wherein, is the source domain segmentation result, y s is the true label.

5. The shape-guide-based unsupervised domain adaptation medical image segmentation method according to claim 1, characterized in that, The specific method for obtaining the trained pseudo-label segmentation model and network parameters thereof comprises: inputting the preprocessed target domain image data into the constructed pseudo-label segmentation model to obtain a pseudo-label; inputting the pseudo-label into the trained shape guide model to obtain a corresponding reconstructed pseudo-label; setting a second Dice loss function, calculating the second Dice loss function value according to the reconstructed pseudo-label and the pseudo-label, and obtaining the trained pseudo-label segmentation model and network parameters thereof when the second Dice loss function value reaches a preset value.

6. The shape-guided unsupervised domain adaptation medical image segmentation method according to claim 5, characterized in that, The second Dice loss function is: wherein is a pseudo label; is a reconstructed pseudo label.

7. The shape-guide-based unsupervised domain adaptation medical image segmentation method according to claim 1, characterized in that, The specific method for obtaining the trained target segmentation model is: constructing a target segmentation model, taking the network parameters of the trained pseudo-label segmentation model as initial network parameters of the target segmentation model; inputting the preprocessed target domain image data into the trained pseudo-label segmentation model and the constructed target segmentation model respectively to obtain pseudo-labels and target segmentation labels; inputting the target segmentation labels into the trained shape guide model to obtain target reconstruction labels; setting a total loss function, calculating a total loss function value according to the pseudo-labels, the target segmentation labels and the target reconstruction labels, and obtaining the trained target segmentation model when the total loss function value reaches a preset value.

8. The shape-guided unsupervised domain adaptation medical image segmentation method according to claim 7, characterized in that, The total loss function includes a reconstruction loss function and a pseudo loss function, and the total loss function is: wherein, Λ pseudo is a hyperparameter, is a pseudo loss function; is a reconstruction loss function; The reconstruction loss function is: wherein, is a target segmentation label; is a target reconstruction label; The pseudo loss function is: wherein is a pseudo label.

9. A shape-guided unsupervised domain adaptation medical image segmentation system for implementing the method of any one of claims 1-8, characterized in that, The system includes: a data acquisition module that acquires source domain image data and target domain image data; a preprocessing module that pre-processes the source domain image and the target domain image to obtain preprocessed source domain image and target domain image; a first model training module that trains the constructed source segmentation model and shape guide model according to the preprocessed source domain image, and obtains the trained shape guide model, the trained source segmentation model and its network parameters; a second model training module that trains the constructed pseudo-label segmentation model based on the preprocessed target domain image data and the trained shape guide model, and obtains the trained pseudo-label segmentation model and its network parameters; wherein the network parameters of the trained source segmentation model are taken as initial network parameters of the constructed pseudo-label segmentation model; a target model training module that trains the constructed target segmentation model based on the preprocessed target domain image data, the trained shape guide model and the trained pseudo-label segmentation model, and obtains the trained target segmentation model; wherein the network parameters of the trained pseudo-label segmentation model are taken as initial network parameters of the constructed target segmentation model; a segmentation module that inputs an image to be segmented into the trained target segmentation model to obtain a target image segmentation result.

Images