A cross-modality medical image segmentation method, device, equipment and medium

Abstract

The application is suitable for the technical field of cross-modal medical image segmentation, and provides a cross-modal medical image segmentation method, device, equipment and medium, which comprises the following steps: obtaining source domain images, source domain labels and target domain images from source domain and target domain datasets; inputting the source domain images into an image generator to obtain pseudo target domain images; inputting the three-domain images into a segmentation model to obtain pseudo labels and segmentation probability maps of corresponding images; dividing the probability maps into stable regions and unstable regions through an entropy dynamic threshold strategy; obtaining global mean anchor points and local intra-class semantic anchor points from the source domain probability maps through a global mean strategy and a farthest point sampling strategy; obtaining positive and negative samples of corresponding domain images from the stable regions and the unstable regions; obtaining a segmentation model by taking the anchor points as targets to be pulled towards the positive samples and pushed away from the negative samples during training; and segmenting target domain images through the model.

Description

TECHNICAL FIELD

[0001] The application is suitable for the field of cross-modality medical image segmentation, and provides a cross-modality medical image segmentation method and device. BACKGROUND

[0002] Medical image segmentation, as a key supporting technology for clinical diagnosis and treatment planning, its core task is to accurately identify and outline the boundary profile of anatomical structure or pathological region. Although the current mainstream full-supervised convolutional neural network method shows high accuracy on standard data sets, its training process is heavily dependent on large-scale pixel-level labeled data. In the field of medical imaging, the acquisition of such labeling not only requires the deep participation of professional personnel such as radiologists, but also requires a lot of time for fine labeling, resulting in high data preparation cost and difficulty in meeting the actual clinical needs. More importantly, when the model is migrated from the source domain (such as MRI images collected by a specific device) to the target domain (such as CT images generated by different devices), due to the differences in imaging protocols, device parameters or patient groups, the model often encounters significant domain shift problem, causing the segmentation performance to drop sharply, which seriously hinders the promotion of the technology in real medical scenarios. In view of the domain shift challenge, unsupervised domain adaptation technology has become a research hotspot, aiming to use labeled source domain knowledge to guide the learning of unlabeled target domain. In the existing scheme, the method based on image conversion (such as CycleGAN and its medical special variant) reduces the pixel-level difference between the source domain and the target domain through style transfer. However, such methods often fail to maintain the semantic integrity of key anatomical structures during the conversion process, and are prone to introduce non-physiological artifacts or blur tissue boundary details, resulting in generated images that are visually similar but semantically distorted, ultimately compromising the reliability of the segmentation results. Another type of domain adaptation strategy based on contrastive learning attempts to learn domain-invariant features through positive and negative sample alignment in the feature space, but its effectiveness is highly dependent on the quality of the target domain pseudo-labels. In the initial adaptation stage, the model lacks understanding of the target domain features, and the generated pseudo-labels generally contain a large amount of noise and misjudgments. These unreliable supervision signals not only fail to effectively guide feature alignment, but also exacerbate the model's reliance on incorrect features, further weakening the cross-domain generalization ability. The inherent defects of existing technologies in semantic structure preservation and pseudo-label noise suppression pose serious challenges to the robustness and accuracy of cross-modality medical image segmentation. SUMMARY

[0003] The purpose of the present application is to provide a cross-modality medical image segmentation method, device, equipment and computer readable storage medium, which has enhanced model adaptation capability to the target domain, effectively suppresses pseudo-label noise, and improves the robustness and accuracy of cross-modality medical image segmentation.

[0004] The present application provides a cross-modality medical image segmentation method based on three-domain contrastive alignment guided by double anchor points, and the technical solution is as follows:

[0005] A cross-modal medical image segmentation method, comprising the following steps:

[0006] S1 obtains source domain images and source domain labels from a source domain dataset, obtains target domain images from a target domain dataset, inputs the source domain images into the image generator, and obtains pseudo target domain images;

[0007] S2 inputs the source domain images, the target domain images and the pseudo target domain images into a segmentation model to obtain source domain image segmentation probability maps, source domain pseudo labels, target domain image segmentation probability maps, target domain pseudo labels and pseudo target domain image segmentation probability maps;

[0008] S3 divides the image segmentation probability maps of the corresponding domains into stable regions and unstable regions based on the image segmentation probability maps, the source domain labels and the source domain pseudo labels through a dynamic threshold strategy of entropy;

[0009] S4 obtains global mean anchor points based on the stable regions of the source domain images through a global mean strategy, and obtains local intra-class semantic anchor points based on the source domain segmentation maps and the source domain labels through a farthest point sampling strategy;

[0010] S5 obtains positive and negative samples of the corresponding domain images from the stable regions and the unstable regions based on the source domain labels and the image segmentation probability maps; when training the segmentation model, different anchor points are selected for each domain based on the positive and negative samples and different domain segmentation stabilities through a multi-domain collaborative pixel-level contrast learning mechanism, and the parameters of the segmentation model are optimized by combining a contrast loss and a dynamic anchor point consistency loss to pull the anchor points to the positive samples and push the negative samples away as the target, thereby enhancing the adaptation ability of the model to the target domain and obtaining a trained segmentation model;

[0011] S6 segments the target domain images according to the segmentation model to obtain segmented target domain images.

[0012] Preferably, the segmentation model is an image segmentation operation strategy formed after pre-training using the source domain images, the source domain labels and the target domain images.

[0013] Preferably, the dynamic threshold strategy of entropy in step S3 specifically divides the image segmentation probability maps of the corresponding domains into stable regions and unstable regions as follows:

[0014] ;

[0015] ;

[0016] ;

[0017] ;

[0018] ;

[0019] ;

[0020] ;

[0021] wherein, is logical and operation, is source domain, is pseudo target domain, is target domain, is stable region of organ, is stable region of organ, is unstable region of organ, is unstable region of organ, is segmentation probability map, is pixel-level feature at index on segmentation probability map, is ground truth at on source domain label, is ground truth at on target domain pseudo label, is pixel-level predicted entropy map at , is dynamic threshold based on predicted entropy map , is normalized predicted probability at predicted as organ , target domain pseudo label is obtained by segmentation probability map of target domain image, , , is constant to ensure numerical stability.

[0022] Preferably, the global mean strategy of step S4 is:

[0023] ;

[0024] wherein, is global mean anchor point of organ , is stable region of organ in source image, is number of pixel-level features in stable region set, is a pixel-level feature vector in stable region set, represents global mean anchor point.

[0025] Preferably, the farthest point sampling strategy in step S4 is:

[0026] ;

[0027] ;

[0028] wherein, is an organ is the first local intra-class semantic anchor point, is the farthest point sampling function, i.e. input feature set and sampling number , output cosine distance farthest feature, is the pixel feature set in the class prediction result consistent with the source domain label, is the true value at on the source domain pseudo label.

[0029] Preferably, the step S5 of obtaining positive and negative samples of the corresponding domain image from the stable area and the unstable area based on the source domain label and the image segmentation probability map comprises the following steps:

[0030] define the source domain image segmentation probability map all pixel features belonging to the organ form a source domain positive sample set , the formula is:

[0031] ;

[0032] define the source domain image segmentation probability map all pixel features not belonging to the organ form a source domain negative sample set , the formula is:

[0033] ;

[0034] define the stable area of the pseudo target domain the lowest similarity of pixel features with the global mean anchor point form a pseudo target domain stable positive sample set , the formula is:

[0035] ;

[0036] wherein, is the cosine similarity, ; indicates the lowest , i.e. after sorting the similarity from small to large, the first pixel features are selected;

[0037] define the stable area of the non-organ of the pseudo target domain middle and global mean anchor points highest similarity The pixel features in the middle form the pseudo target domain stable negative sample set , the formula is:

[0038] ;

[0039] Among them, The highest , that is, after sorting the similarity from large to small, select the first Pixel features;

[0040] Define the unstable area of the pseudo target domain All pixel features belonging to the organ form the pseudo target domain unstable positive sample set , the formula is:

[0041] ;

[0042] Define the unstable area of the pseudo target domain All pixel features not belonging to the organ form the pseudo target domain unstable negative sample set , the formula is:

[0043] ;

[0044] Define the stable area of the target domain All pixel features belonging to the organ form the target domain positive sample set , the formula is:

[0045] ;

[0046] Define the stable area of the target domain All pixel features not belonging to the organ form the target domain negative sample set , the formula is:

[0047] .

[0048] Preferably, the loss function in step S5 during training :

[0049] + ;

[0050] Among them, is the source domain contrast learning loss, is the pseudo target domain stable area contrast learning loss, Contrast learning loss for unstable regions of pseudo-target domains For target domain contrastive learning loss, For dynamic anchor point consistency loss, The source domain cross-entropy loss is calculated based on the source domain labels. For the target domain cross-entropy loss, pseudo-labels are calculated based on entropy-based confidence constraints. This is the corresponding balance coefficient;

[0051] ;

[0052] ;

[0053] in, () represents the similarity calculation formula, which is used to calculate the feature vector. and anchor points The degree of similarity, The total number of organs that need to be segmented in the image, To control the temperature hyperparameter of similarity distribution smoothness, This represents a positive sample in the set of positive samples in the source domain. This represents a negative sample in the set of negative samples in the source domain;

[0054] ;

[0055] ;

[0056] in, This represents the set of positive and negative samples used when calculating the contrastive loss. Indicates organs The One positive or negative sample;

[0057] To reduce the computational overhead caused by a large number of pixels during training, positive and negative samples are randomly sampled from the positive and negative sample set. This represents the number of positive and negative sample pairs used for contrastive learning.

[0058] ;

[0059] ;

[0060] in, For local intra-class semantic anchor points, For all organs in the unstable region of the pseudo-target domain The average features of pixel features;

[0061] ;

[0062] ;

[0063] wherein, is the average feature of all pixel features belonging to the organ in the target domain stable region;

[0064] ;

[0065] ;

[0066] wherein, is the pixel feature in the target domain stable region, is the true value in the pseudo label corresponding to the pixel feature, is the organ currently to be considered, is all candidate organs, used for normalization to ensure the validity of the probability distribution.

[0067] A cross-modal medical image segmentation device, comprising:

[0068] a sample acquisition module, configured to acquire source domain samples and target domain samples from a source domain dataset and a target domain dataset, the source domain samples comprising source domain images and source domain labels, and the target domain samples comprising target domain images;

[0069] a training data acquisition module, configured to use a pre-trained image generator as a synthesis model, input the source domain images into the synthesis model, and acquire pseudo target domain images. Meanwhile, input the source domain images, the target domain images and the pseudo target domain images into a segmentation model, and acquire source domain image segmentation probability maps, source domain pseudo labels, target domain image segmentation probability maps, target domain pseudo labels and pseudo target domain image segmentation probability maps;

[0070] a stable region division module, configured to respectively divide the source domain image segmentation probability maps, the target domain image segmentation probability maps and the pseudo target domain image segmentation probability maps into stable regions and unstable regions through an entropy dynamic threshold strategy;

[0071] a double-anchor point acquisition module, configured to acquire a global mean anchor point based on the stable regions of the source domain images through a global mean strategy, and acquire a local intra-class semantic anchor point based on a source domain segmentation map and a source domain label through a farthest point sampling strategy;

[0072] a three-domain positive and negative sample acquisition module, configured to acquire source domain positive and negative samples, target domain positive and negative samples and pseudo target domain positive and negative samples from the stable regions and the unstable regions based on the source domain labels, the source domain image segmentation probability maps, the target domain image segmentation probability maps and the pseudo target domain image segmentation probability maps;

[0073] The training module is configured to, during training, select different anchor points for each domain based on different domain segmentation stability through a multi-domain collaborative pixel-level contrast learning mechanism based on the positive and negative samples, and combine a contrast loss and a dynamic anchor consistency loss to pull the anchor points towards the positive samples and push the negative samples away as a target, optimize parameters of a segmentation model, enhance adaptability of the model to a target domain, and obtain the segmentation model.

[0074] The segmentation module is configured to segment a target domain image according to the segmentation model and obtain a segmented target domain image.

[0075] A cross-modality medical image segmentation device includes a processor, a memory, and a computer program stored in the memory and executable by the processor to implement the cross-modality medical image segmentation method described above.

[0076] A computer-readable storage medium includes a stored computer program, wherein the computer-readable storage medium controls a device in which the computer-readable storage medium is located to perform the cross-modality medical image segmentation method described above when the computer program is running.

[0077] By adopting the above scheme, the cross-modality medical image segmentation method, device, equipment, and computer-readable storage medium provided by the present application have the following advantages and beneficial effects: the cross-modality medical image segmentation method, device, equipment, and computer-readable storage medium provided by the present application dynamically divide stable and unstable regions through a double-anchor guiding mechanism and optimize a segmentation model in combination with a three-domain contrast alignment strategy, effectively solve the domain shift problem and pseudo-label noise interference in cross-modality medical image segmentation, have enhanced adaptability of the model to a target domain, effectively suppress pseudo-label noise, and improve robustness and accuracy of cross-modality medical image segmentation. BRIEF DESCRIPTION OF DRAWINGS

[0078] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed to be used in the specific embodiments of the present application. It should be understood that the following drawings only show some specific embodiments of the present application, and therefore should not be regarded as a limitation on the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0079] Figure 1 is a flowchart of a cross-modality medical image segmentation method.

[0080] Figure 2 is a framework diagram of a cross-modality medical image segmentation method. DETAILED DESCRIPTION

[0081] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort should fall within the protection scope of the present application. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort should fall within the protection scope of the present application.

[0082] In the description of the present application, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the purpose of facilitating the description of the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0083] In addition, the terms "first", "second" are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise explicitly specified and limited.

[0084] In the present application, unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connection", "fixing" and the like should be understood broadly, for example, it can be fixed connection, or detachable connection, or integrally connected; it can be mechanical connection, or electrical connection; it can be directly connected, or indirectly connected through an intermediate medium, or the internal communication of two elements. For a person of ordinary skill in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0085] The preferred embodiments of the present application are described in detail below with reference to the accompanying drawings, so that the advantages and features of the present application can be more easily understood by those skilled in the art, and the protection scope of the present application is more clearly and definitely defined. Embodiment one, please refer to Figure 1 and Figure 2The first embodiment of the present application provides a cross-modal medical image segmentation method based on three-domain contrast alignment of double-anchor point guidance, which comprises steps S1 to S8.

[0086] S1, source domain images and source domain labels are obtained from a source domain dataset, and target domain images are obtained from a target domain dataset.

[0087] In this embodiment, the source domain images are magnetic resonance imaging (MRI) images. The target domain images are computed tomography (CT) images. In other embodiments, the source domain images and the target domain images can be obtained in other ways, and the present application does not make specific limitations on this.

[0088] The present application first defines a labeled source domain dataset and an unlabeled target domain Our goal is to train a segmentation model to transfer the knowledge of to . Wherein, is the source domain image, is the source domain label, is the number of source domain samples, is the target domain image, is the number of target domain samples.

[0089] Figure 2 The pre-training overall framework for unsupervised domain adaptation medical image segmentation of the present application is shown. The entire framework is composed of a segmentation model, a double-anchor point acquisition module and a three-domain positive and negative sample acquisition module, and the segmentation model is composed of an encoder and a decoder.

[0090] S2, pre-training the image generator based on the source domain image. The source domain image is input into the image generator to obtain a pseudo target domain image.

[0091] The image is input into the segmentation model for segmentation to obtain a segmentation probability map, and the double-anchor point acquisition module and the three-domain positive and negative sample acquisition module acquire anchor points and positive and negative samples according to the segmentation probability map. The specific operation is as steps S3 to S6.

[0092] S3, input the source domain image, the target domain image and the pseudo target domain image into the segmentation model to obtain the source domain image segmentation probability map, the source domain pseudo label, the target domain image segmentation probability map, the target domain pseudo label and the pseudo target domain image segmentation probability map.

[0093] S4, by the dynamic threshold strategy of entropy, based on the image segmentation probability map, the source domain label and the source domain pseudo label, the image segmentation probability map of the corresponding domain is divided into stable area and unstable area.

[0094] On the basis of the above-mentioned embodiments, in an optional embodiment of the present application, the dynamic threshold strategy of entropy is:

[0095] ;

[0096] ;

[0097] ;

[0098] ;

[0099] ;

[0100] ;

[0101] ;

[0102] wherein, is a source domain, is a pseudo target domain, is a target domain, is a stable region of an organ, is a non-stable region of an organ, is a segmentation probability map, is a pixel-level feature at an index on the segmentation probability map, is a true value at on a source domain label, is a true value at on a target domain pseudo label, is a pixel-level prediction entropy map at , is a dynamic threshold based on a percentile of the prediction entropy map, is a prediction probability at after normalization, which is obtained by segmenting a target domain image through the target domain pseudo label, , , , are constants for ensuring numerical stability.

[0103] ​​​​The application proposes a new strategy to divide the pixel features on the segmentation probability map into stable set and unstable set, specifically, due to the existence of uncertain area with high prediction noise, directly aligning all target domain pixels based on pseudo label may be misleading. In this case, as the training proceeds, the learning of the model will be misled to cause performance degradation, in order to enhance the robustness of feature alignment. In order to solve this problem, we divide the pixel features into stable area and unstable area based on the adaptive region division mechanism of dynamic entropy, and use different strategies to mine positive and negative samples for different regions.

[0104] S5, by global mean strategy, based on the stable area of the source domain image, the global mean anchor point is obtained, by the farthest point sampling strategy, based on the source domain segmentation map and the source domain label, the local intra-class semantic anchor point is obtained.

[0105] Based on the above embodiment, in an optional embodiment of the application, the dynamic threshold strategy of entropy is that the global mean strategy is:

[0106] ;

[0107] Among them, is the global mean anchor point of the organ , is the stable area of the organ in the source image, is the number of pixel-level features in the stable area set.

[0108] In an optional embodiment, the farthest point sampling strategy is:

[0109] ;

[0110] ;

[0111] Among them, is the local intra-class semantic anchor point of the organ , is the farthest point sampling function, that is, the input feature set and the sampling number , output the cosine distance farthest feature, is the pixel feature set whose prediction result is consistent with the source domain label in the class , is the true value at on the source domain pseudo label.

[0112] ​The application proposes a new bilingual semantic anchor construction mechanism to realize fine-grained cross-domain alignment and extract stable and comprehensive organ pixel feature representation from a supervised rich source domain. Specifically, the global mean anchor point can realize robust and consistent cross-domain feature alignment, but the global mean anchor point cannot capture the internal structure and morphological changes of the organ (such as organ sub-regions or shape differences). These anchors usually only represent the core features, resulting in significant differences between the edge features and the center. During training, the model is forced to pull the edge features to the center, hindering the effective learning of edge details. To solve this problem, the application introduces a local intra-class semantic anchor mechanism to capture multiple representative features within each organ and realize fine-grained structure alignment. In addition, these anchors are dynamically updated at each cycle, and as the model optimizes, the expression ability and domain adaptability of the anchors are continuously improved.

[0113] S6, based on the source domain label and the image segmentation probability map, obtaining positive and negative samples of the corresponding domain image from the stable area and the unstable area.

[0114] In an optional implementation, step S6 specifically includes steps S61 to S64.

[0115] S61, defining the source domain image segmentation probability map All pixel features belonging to the organ in the source domain image segmentation probability map form a source domain positive sample set , and the formula is:

[0116] ;

[0117] Defining the source domain image segmentation probability map All pixel features not belonging to the organ in the source domain image segmentation probability map form a source domain negative sample set , and the formula is:

[0118] .

[0119] S62, defining the stable area of the pseudo target domain The pixel features with the lowest similarity to the global mean anchor point in the stable area of the pseudo target domain form a stable positive sample set of the pseudo target domain , and the formula is:

[0120] ;

[0121] wherein, is the cosine similarity, Defining the stable area of the non-organ of the pseudo target domain The pixel features with the highest similarity to the global mean anchor point in the stable area of the pseudo target domain form a stable negative sample set of the pseudo target domain ​The pixel features in the organ , and the formula is:

[0122] .

[0123] S63, define the unstable area of the pseudo target domain The pixel features in the organ form a pseudo target domain unstable positive sample set , and the formula is:

[0124] ;

[0125] Define the unstable area of the pseudo target domain The pixel features in the organ form a pseudo target domain unstable negative sample set , and the formula is:

[0126] .

[0127] S64, define the stable area of the target domain The pixel features in the organ form a target domain positive sample set , and the formula is:

[0128] ;

[0129] Define the stable area of the target domain The pixel features in the organ form a target domain negative sample set , and the formula is:

[0130] .

[0131] The present application constructs a plurality of domain positive and negative samples for comparative learning, realizes alignment through shared source domain derived anchors, enhances feature consistency and discriminability, and introduces dynamic anchor consistency loss to provide semantic constraints for the target domain, so that the model can still learn more discriminative and universal representations under significant domain shift.

[0132] S7, during training, through a multi-domain cooperative pixel-level contrast learning mechanism, based on the positive and negative samples, based on different domain segmentation stability, select different anchor points for each domain, combine contrast loss and dynamic anchor consistency loss, and optimize the parameters of the segmentation model to enhance the adaptability of the model to the target domain, and obtain the segmentation model.

[0133] On the basis of the above embodiment, in an optional embodiment of the present application, the loss function is:

[0134] + ;

[0135] wherein, is a source domain contrastive learning loss, is a pseudo target domain stable region contrastive learning loss, is a pseudo target domain unstable region contrastive learning loss, is a target domain contrastive learning loss, is a dynamic anchor consistency loss, is a source domain cross-entropy loss, calculated based on source domain labels, is a target domain cross-entropy loss, calculated based on entropy-based confidence constraint pseudo labels, is a balance coefficient.

[0136] ;

[0137] ;

[0138] wherein, is the total number of organs to be segmented in the image, is a temperature hyperparameter to control the smoothness of the similarity distribution.

[0139] ;

[0140] ;

[0141] wherein, to reduce the computational overhead brought by a large number of pixels during training, positive and negative samples are randomly sampled within the positive and negative sample sets, is the number of positive and negative sample pairs used for contrastive learning.

[0142] ;

[0143] ;

[0144] wherein, is a local intra-class semantic anchor set, is the average feature of all pixel features belonging to the organ in the pseudo target domain unstable region.

[0145] ;

[0146] ;

[0147] wherein, is the average feature of all pixel features belonging to the organ in the target domain stable region.

[0148] ;

[0149] ;

[0150] wherein, is a pixel feature in the target domain stable area, is a true value in the pseudo label corresponding to the pixel feature, is the organ currently considered (i.e. the organ for which the membership probability of a certain pixel feature is calculated), is all candidate organs, used for normalization to ensure that the probability distribution is valid.

[0151] The application performs contrast learning from three domains to enhance the domain adaptation capability of the segmentation model. Specifically, the source domain contrast learning takes the global mean anchor as a stable reference, based on the high-quality real label of the source domain, takes the pixel features belonging to the same organ as positive samples, and takes the pixel features of other organs as negative samples, and constructs a pixel feature space with strong discriminability through contrast learning, laying a foundation for subsequent cross-domain adaptation. The pseudo target domain serves as a bridge for the style transfer of the source domain to the target domain, and distinguishes stable and unstable areas according to the prediction entropy: the stable area continues to take the global mean anchor as a reference, and selects the pixel features of the same organ with the lowest similarity to the global mean anchor as positive samples and the pixel features of the different organs with the highest similarity as negative samples through a difficult sample mining strategy, to strengthen the alignment effect; the unstable area uses a local intra-class semantic anchor instead, guiding the model to learn the intra-class structural details, thereby effectively transferring the semantic structure knowledge of the source domain to the target domain. The target domain is the core of unsupervised adaptation, and only focuses on the high-confidence stable pixel features predicted by the model, selects the local intra-class semantic anchor with the highest similarity to these pixel features as a reference, takes the stable pixel features of the same organ as positive samples and the stable pixel features of different organs as negative samples, and through contrast learning, promotes the alignment of the target domain features and the diversified structure representation captured in the source domain, finally realizes reliable cross-domain adaptation. Finally, to improve the alignment accuracy and reduce the influence of noisy pseudo labels in early training, the application proposes a dynamic anchor consistency loss, which regards the intra-class semantic anchor of each class organ as multiple sub-anchors, constrains the pseudo label and the intra-class semantic anchor through each other, uses a curriculum learning strategy to simultaneously improve the accuracy of the anchor and the pseudo label, and then improves the model.

[0152] S8, segmenting the target domain image according to the segmentation model to obtain a segmented target domain image.

[0153] As Figure 2As shown, the application is a three-domain contrast alignment cross-modal medical image segmentation method based on double anchor point guidance. First, the source domain image is directly input into the image generation model to synthesize pseudo target domain images. Second, the three-domain image is directly input into the segmentation model and is divided by the dynamic threshold strategy of entropy. Third, the global mean anchor point is obtained from the stable region of the source domain image by the global mean strategy, and the local intra-class semantic anchor point is obtained from the source domain segmentation map and the source domain label by the farthest point sampling strategy. Finally, the positive and negative samples are obtained from the segmentation probability map of the three-domain image, and the anchor points are pulled towards the positive samples and pushed away from the negative samples by contrast learning. This method focuses on the core representation and structural diversity of each organ class, realizes fine-grained semantic alignment, and introduces dynamic anchor point consistency loss to improve the robustness and adaptability of the model.

[0154] Embodiment two, the application provides a three-domain contrast alignment cross-modal medical image segmentation device based on double anchor point guidance, characterized by comprising a sample acquisition module, a training data acquisition module, a stable region division module, a double anchor point acquisition module, a three-domain positive and negative sample acquisition module, a training module and a segmentation module.

[0155] The sample acquisition module is used to obtain source domain samples and target domain samples from the source domain dataset and the target domain dataset. The source domain samples include source domain images and source domain labels, and the target domain samples include target domain images.

[0156] The training data acquisition module is used to input the source domain image into the synthesis model with the pre-trained image generator as the synthesis model, and obtain the pseudo target domain image. At the same time, the source domain image, the target domain image and the pseudo target domain image are input into the segmentation model to obtain the source domain image segmentation probability map, the source domain pseudo label, the target domain image segmentation probability map, the target domain pseudo label and the pseudo target domain image segmentation probability map.

[0157] The stable region division module is used to divide the source domain image segmentation probability map, the target domain image segmentation probability map and the pseudo target domain image segmentation probability map into stable regions and unstable regions by the dynamic threshold strategy of entropy.

[0158] The double anchor point acquisition module is used to obtain the global mean anchor point based on the stable region of the source domain image by the global mean strategy, and obtain the local intra-class semantic anchor point based on the source domain segmentation map and the source domain label by the farthest point sampling strategy.

[0159] The three-domain positive and negative sample acquisition module is used to obtain the source domain positive and negative samples, the target domain positive and negative samples and the pseudo target domain positive and negative samples from the stable region and the unstable region based on the source domain label, the source domain image segmentation probability map, the target domain image segmentation probability map and the pseudo target domain image segmentation probability map.

[0160] The training module is configured to, during training, select different anchor points for each domain based on different domain segmentation stability through a multi-domain collaborative pixel-level contrast learning mechanism based on the positive and negative samples, combine a contrast loss and a dynamic anchor consistency loss, and optimize parameters of a segmentation model to enhance adaptability of the model to a target domain and obtain the segmentation model.

[0161] The segmentation module is configured to segment a target domain image according to the segmentation model and obtain a segmented target domain image.

[0162] In the third embodiment, the present application provides a three-domain contrast alignment cross-modality medical image segmentation device based on double-anchor point guidance, which includes a processor, a memory, and a computer program stored in the memory. The computer program can be executed by the processor to implement the three-domain contrast alignment cross-modality medical image segmentation method based on double-anchor point guidance as described in any one of the first embodiment.

[0163] In the fourth embodiment, the present application provides a computer readable storage medium, which includes a stored computer program. When the computer program is running, the computer readable storage medium controls a device where the computer readable storage medium is located to execute the three-domain contrast alignment cross-modality medical image segmentation method based on double-anchor point guidance as described in any one of the first embodiment.

[0164] It can be understood that the classification device can be an electronic device with computing performance such as a portable notebook computer, a desktop computer, a server, a smart phone, or a tablet computer.

Claims

1. A cross-modality medical image segmentation method, characterized by, The method comprises the following steps: S1 obtains source domain images and source domain labels from a source domain data set, obtains target domain images from a target domain data set, inputs the source domain images into the image generator, and obtains pseudo target domain images; S2 inputs the source domain images, the target domain images, and the pseudo target domain images into a segmentation model to obtain source domain image segmentation probability maps, source domain pseudo labels, target domain image segmentation probability maps, target domain pseudo labels, and pseudo target domain image segmentation probability maps; S3 divides the image segmentation probability maps of the corresponding domains into stable regions and unstable regions based on the image segmentation probability maps, the source domain labels, and the source domain pseudo labels by using a dynamic threshold strategy of entropy, and the dynamic threshold strategy of entropy is specifically as follows: ; ; ; ; ; ; ; in, For logical AND operations, For source domain, For pseudo-target domain, For the target domain, For the source domain tag The truth value at the location, For the target domain pseudo tag The truth value at the location, for Pixel-level prediction entropy map at the location Dynamic threshold based on prediction entropy map Percentile dynamic determination for Predicted as an organ The process The normalized predicted probability and the target domain pseudo-label are obtained by analyzing the target domain image segmentation probability map. Calculated , Constants used to ensure numerical stability; S4 obtains global mean anchor points based on the stable regions of the source domain images by using a global mean strategy, and obtains local intra-class semantic anchor points based on the source domain segmentation maps and the source domain labels by using a farthest point sampling strategy; S5 obtains positive and negative samples of the images of the corresponding domains from the stable regions and the unstable regions based on the source domain labels and the image segmentation probability maps; when training the segmentation model, different anchor points are selected for each domain based on the positive and negative samples and different domain segmentation stabilities by using a multi-domain collaborative pixel-level contrast learning mechanism, and the parameters of the segmentation model are optimized by combining a contrast loss and a dynamic anchor point consistency loss to achieve the target of pulling the anchor points to the positive samples and pushing the negative samples away, thereby enhancing the adaptation capability of the model to the target domain and obtaining a trained segmentation model; S6 segments target domain images according to the segmentation model to obtain segmented target domain images.

2. The cross-modality medical image segmentation method of claim 1, wherein, The segmentation model is an image segmentation operation strategy formed after pre-training of the source domain images, the source domain labels, and the target domain images.

3. The cross-modality medical image segmentation method of claim 2, wherein, The global mean strategy of step S4 is as follows: ; wherein, is the global mean anchor point of the organ, is the global mean anchor point of the organ, is the stable region of the organ in the source domain, is the stable region of the organ in the source domain, is the number of pixel-level features in the stable region set in the source domain, is a pixel-level feature vector in the stable region set, denotes the global mean anchor point.

4. The cross-modality medical image segmentation method of claim 3, wherein, The farthest point sampling strategy in step S4 is as follows: ; ； in, For organs The Local class semantic anchors, The farthest point sampling function, i.e., the input feature set and the number of samples. Output The feature that is furthest from the cosine distance, For organs The set of pixel features whose prediction results are consistent with the source domain labels For source domain pseudo tags The truth value at that point.

5. The cross-modality medical image segmentation method of claim 4, wherein, The positive and negative samples of the images of the corresponding domains are obtained from the stable regions and the unstable regions based on the source domain labels and the image segmentation probability maps in step S5, and the specific steps are as follows: Defining source domain image segmentation probability maps All pixels belonging to organs in the source domain form the source domain positive sample set , which is given by ； Defining source domain image segmentation probability maps All pixels not belonging to organs Form source domain negative sample set with the formula: ； Stable regions defining pseudo-target domains Mid and global mean anchor points Lowest similarity Pixel feature forming pseudo-target domain stable positive sample set The formula is: ； wherein, is the cosine similarity, ; represents the lowest , i.e. the first pixel features are selected in the order of the similarity from small to large. Non-organ defining pseudo-target domain Stable region Global mean anchor point Most similar Pixel features form pseudo-target domain stable negative sample set The formula is: ； wherein, represents the highest , i.e. after sorting the similarity from large to small, the first pixel features are selected; Defining unstable regions of pseudo-target domains All pixels belonging to organs Forming unstable positive sample sets of pseudo-target domains The formula is: ; Defining unstable regions of pseudo-target domains All pixels not belonging to organs Forming unstable negative sample set of pseudo-target domains from pixel features , which is ； Defining stable regions of a target domain All pixels belonging to organs Forming a target domain positive sample set from pixel features , which is ; Defining stable regions of a target domain All pixels not belonging to organs Forming a target domain negative sample set from pixel features , which is: 。 6. The cross-modality medical image segmentation method of claim 5, wherein, The loss function in the step S5 of training is: + ; wherein, is a source domain contrastive learning loss, is a pseudo target domain stable region contrastive learning loss, is a pseudo target domain unstable region contrastive learning loss, is a target domain contrastive learning loss, is a dynamic anchor consistency loss, is a source domain cross-entropy loss, computed based on source domain labels, is a target domain cross-entropy loss, computed based on entropy-constrained pseudo labels, is a balancing coefficient corresponding to the pseudo target domain stable region contrastive learning loss, is a balancing coefficient corresponding to the pseudo target domain unstable region contrastive learning loss, is a balancing coefficient corresponding to the target domain contrastive learning loss, is a balancing coefficient corresponding to the dynamic anchor consistency loss, is a balancing coefficient corresponding to the source domain cross-entropy loss, is a balancing coefficient corresponding to the target domain cross-entropy loss; ; ； wherein, is a similarity calculation formula, which is used to calculate the similarity between the feature vector and the anchor point , is the total number of organs to be segmented in the image, is a temperature hyperparameter for controlling the smoothness of the similarity distribution, represents a positive sample in the source domain positive sample set, represents a negative sample in the source domain negative sample set; ; ; wherein, represents a set of positive and negative samples when calculating the contrast loss, represents an organ of the first positive or negative sample; To reduce the computational overhead of a large number of pixels during training, positive and negative samples are randomly sampled within the positive and negative sample sets, the number of positive and negative sample pairs for contrastive learning; ; ； wherein, is a local intra-class semantic anchor set, is an average feature of all pixel features belonging to organs in the pseudo target domain unstable region. ； ； wherein, is the average feature of all pixel features belonging to the organ of interest in the target domain stable region; ； ; where, is the pixel feature in the target domain stable region, is the true value in the pseudo label corresponding to the pixel feature, is the organ currently under consideration, is all candidate organs, used for normalization to ensure that the probability distribution is valid.

7. A cross-modality medical image segmentation apparatus characterized by comprising: Contains: A sample acquisition module is configured to acquire source domain samples and target domain samples from a source domain data set and a target domain data set, wherein the source domain samples include source domain images and source domain labels, and the target domain samples include target domain images; A training data acquisition module is configured to use an image generator that has been pre-trained as a synthetic model, input the source domain images into the synthetic model to obtain pseudo target domain images, and input the source domain images, the target domain images, and the pseudo target domain images into a segmentation model to obtain source domain image segmentation probability maps, source domain pseudo labels, target domain image segmentation probability maps, target domain pseudo labels, and pseudo target domain image segmentation probability maps; A stable region division module is configured to divide the source domain image segmentation probability maps, the target domain image segmentation probability maps, and the pseudo target domain image segmentation probability maps into stable regions and unstable regions respectively by using a dynamic threshold strategy of entropy, and the dynamic threshold strategy of entropy is specifically as follows: ; ; ; ; ; ; ; in, For logical AND operations, For source domain, For pseudo-target domain, For the target domain For the source domain tag The truth value at the location, For the target domain pseudo tag truth value at, for Pixel-level prediction entropy map at the location Dynamic threshold based on prediction entropy map Percentile dynamic determination for Predicted as an organ The process Normalized prediction probabilities and target domain pseudo-labels are obtained by analyzing the target domain image segmentation probability map. Calculated , Constants used to ensure numerical stability; The dual-anchor obtaining module is configured to obtain a global mean anchor point based on a stable region of the source domain image by a global mean strategy, and obtain a local intra-class semantic anchor point based on a source domain segmentation map and a source domain label by a farthest point sampling strategy; The three-domain positive and negative sample obtaining module is configured to obtain source domain positive and negative samples, target domain positive and negative samples, and pseudo target domain positive and negative samples from the stable region and the unstable region based on a source domain label, a source domain image segmentation probability map, a target domain image segmentation probability map, and a pseudo target domain image segmentation probability map; The training module is configured to, during training, select different anchor points for each domain based on different domain segmentation stabilities based on the positive and negative samples by a multi-domain collaborative pixel-level contrast learning mechanism, combine a contrast loss and a dynamic anchor point consistency loss, and optimize parameters of a segmentation model to enhance adaptability of the model to the target domain, so as to obtain the segmentation model. The segmentation module is configured to segment a target domain image according to the segmentation model to obtain a segmented target domain image.

8. A cross-modality medical image segmentation device comprising a processor, a memory, and a computer program stored in the memory, the computer program being executable by the processor to implement the cross-modality medical image segmentation method of any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium comprises a stored computer program, wherein the computer readable storage medium controls the device in which the computer readable storage medium is located to perform the cross-modality medical image segmentation method of any one of claims 1 to 6 when the computer program is running.

Images