Fetal magnetic resonance image brain region segmentation method and system, computer device and medium

Abstract

This invention provides a method, system, computer equipment, and medium for fetal magnetic resonance imaging (MRI) brain region segmentation, belonging to the field of automated segmentation technology for fetal brain MRI. The method includes: first, processing multi-planar acquired 2D low-resolution fetal MRI images to generate 3D high-resolution fetal brain images; second, performing selective inter-layer annotation on the 3D images and generating complete segmentation labels using a contour interpolation algorithm; and finally, using a dual-independent initialization segmentation network for mutual supervision learning, including steps such as data augmentation consistency constraints, cross-pseudo-label supervision, and feature comparison learning. This invention, employing the aforementioned fetal MRI brain region segmentation method, system, computer equipment, and medium, requires only a very small number of labels to train a high-performance segmenter, significantly reducing annotation costs and improving segmentation efficiency. It is suitable for practical clinical fetal brain MRI research, providing efficient technical support for fetal brain development assessment.

Description

Technical Field

[0001] This invention relates to the field of automated segmentation technology for fetal brain magnetic resonance imaging (MRI), and in particular to methods, systems, computer equipment, and media for segmenting brain regions in fetal MRI images. Background Technology

[0002] Fetal magnetic resonance imaging (MRI) is a core tool for assessing fetal brain development, especially when ultrasound examinations cannot provide sufficient structural detail. This technology can identify congenital brain developmental abnormalities at an early stage, such as cortical malformations and ventricular dilatation, providing crucial decision-making support for clinical intervention planning and prognostic assessment.

[0003] To achieve accurate quantitative analysis, precise segmentation of brain tissue (including gray matter, white matter, and cerebrospinal fluid) is a prerequisite, which is also the foundation for constructing fetal brain atlases and conducting developmental research. However, current manual segmentation relies on radiologists to annotate layer by layer, which is not only time-consuming and laborious, but may also affect the consistency of results due to subjective factors.

[0004] Deep learning-based automated segmentation methods hold immense potential for accurate analysis of fetal brain MRI. However, most current mainstream deep learning-based automated methods rely on supervised learning models, requiring substantial amounts of labeled data. In the field of fetal brain MRI, acquiring labeled data presents numerous challenges, severely hindering the widespread adoption and application of such automated segmentation methods in real-world clinical data. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, computer equipment, and medium for segmenting fetal brain regions in magnetic resonance imaging (MRI) images. When used in actual clinical fetal brain MRI research, only a very small number of labels need to be labeled according to the process to train a fetal brain region segmenter with excellent segmentation performance, thus avoiding the heavy dependence of existing supervised learning networks on labels when applied to actual clinical datasets.

[0006] To achieve the above objectives, the present invention provides a method for brain region segmentation in fetal magnetic resonance imaging, comprising the following steps:

[0007] Step S1: Process the 2D low-resolution fetal magnetic resonance images acquired from multiple angles to generate 3D high-resolution fetal brain images.

[0008] Step S2: Selectively perform interlayer annotation on the 3D high-resolution fetal brain image generated in step S1, and generate complete segmentation labels through contour interpolation algorithm;

[0009] Step S3: Based on the 3D high-resolution fetal brain image from step S1 and the segmentation labels from step S2, a deformer segmentation network with dual independent initialization that can utilize long-range attention mechanisms is used for mutual learning.

[0010] Preferably, in step S1, the 2D low-resolution fetal magnetic resonance images acquired from multiple perspectives are processed to generate 3D high-resolution fetal brain images, including the following steps:

[0011] Resampling was performed on 2D low-resolution fetal magnetic resonance images acquired from multiple angles to reduce resolution;

[0012] A localization convolutional neural network containing three convolutional layers and one pooling layer is used to perform preliminary brain localization on 2D low-resolution fetal MRI images and generate segmentation masks.

[0013] Perform 3D morphological closing and opening operations sequentially on the segmentation mask;

[0014] The maximum connected component is selected and a 3D bounding cube is fitted. The 3D bounding cube is then resampled to the original resolution to obtain the final localization result.

[0015] For the localization results, a segmentation convolutional neural network with multi-scale Dice loss function was used to extract the fetal brain. The segmentation convolutional neural network consists of a three-layer convolutional encoder and a three-layer convolutional decoder.

[0016] ;

[0017] in, Represents the multi-scale total loss function. Represents the Dice loss function. S Indicates different scaling scales; express s Predicted labels based on scale express s The true label of scale Indicates the original predicted label, Indicates the original, authentic label;

[0018] Bias field correction was performed on the acquired 2D low-resolution fetal magnetic resonance images;

[0019] The corrected image is registered to the selected orientation using symmetrical block matching.

[0020] Based on the registration results, Nadaraya-Watson kernel regression was used to estimate the initial 3D high-resolution fetal brain image:

[0021] ;

[0022] in, Indicates the coordinates of the point to be predicted. This represents the coordinates and values ​​of known data points. Represents the kernel function; Indicates the point to be predicted The kernel regression estimate at that location, Labels representing known data points, This indicates the number of known data points;

[0023] The final high-resolution 3D fetal brain image was obtained by iteratively solving an outlier-robust optimization problem.

[0024] ;

[0025] in, This represents the first 2D low-resolution fetal magnetic resonance image acquired in reality. k Layer data, This represents the forward operator of the data acquisition process. This represents the 3D high-resolution fetal brain image we want to obtain. Represents the regularization parameter. This represents the difference function.

[0026] Preferably, in step S2, selective interlayer annotation is performed on the 3D high-resolution fetal brain image generated in step S1, and complete segmentation labels are generated using a contour interpolation algorithm, including the following steps:

[0027] Step S21: Based on 3D high-resolution fetal brain images, select key layers according to the anatomical features of brain regions;

[0028] Step S22: Extract the contour boundaries of two adjacent labeled key layers to obtain the original key layer contours;

[0029] Step S23: Based on the extracted contour boundaries, calculate the symbolic distance maps for the two layers respectively;

[0030] Step S24: For the intermediate layer between two adjacent labeled key layers, merge the two symbol distance maps obtained in step S23 according to the inter-layer position ratio;

[0031] Step S25: Extract the zero-distance boundary from the fused symbolic distance map as the intermediate layer contour;

[0032] Step S26: Combine all the generated intermediate layer contours with the original key layer contours, perform a closing operation on the combined complete contour sequence, and output the final segmentation label for training the segmentation network.

[0033] Preferably, in step S3, a deformer segmentation network with dual independent initialization that can utilize long-range attention mechanisms is used for mutual supervision learning, including the following steps:

[0034] Step S31: Deploy two structurally identical and independently initialized deformable segmentation networks that can utilize long-range attention mechanisms for mutual learning.

[0035] Step S32: Perform collaborative training on unlabeled 3D high-resolution fetal brain images:

[0036] Step S321: Optimize network robustness through data-enhanced consistency constraints;

[0037] Step S322: Achieve network mutual learning through cross-pseudo-label supervision;

[0038] Step S323: Enhance the discriminative power of the encoder through feature contrast learning;

[0039] Step S33: Jointly optimize the loss from step S32 and output the model.

[0040] Preferably, data-enhanced consistency constraints include:

[0041] Apply weak enhancement and strong enhancement operations to the same image respectively;

[0042] Constrain the consistency of segmentation results for weak and strong augmentation outputs of the dual network.

[0043] Preferably, cross-pseudo-label supervision includes:

[0044] The prediction results of the first network are used as the supervision labels for the second network;

[0045] The prediction results of the second network are used as the supervision labels for the first network;

[0046] The supervision error is calculated using cross-entropy loss and Dice loss.

[0047] Preferably, feature contrastive learning includes:

[0048] Extract feature vectors from the encoder output layer;

[0049] Minimize the feature distance of similar anatomical structures and maximize the feature distance of different types of structures.

[0050] The present invention also provides a fetal magnetic resonance imaging brain region segmentation system, comprising:

[0051] The image processing module is used to process 2D low-resolution fetal magnetic resonance images acquired from multiple angles to generate 3D high-resolution fetal brain images.

[0052] The annotation module is used to selectively annotate the generated 3D high-resolution fetal brain image and generate complete segmentation labels through contour interpolation algorithm.

[0053] The segmentation network module is used for mutual supervision learning based on 3D high-resolution fetal brain images and generated labeled data, employing a deformable segmentation network with dual independent initialization that can utilize long-range attention mechanisms.

[0054] The data augmentation module is used to apply data augmentation consistency constraints to unlabeled 3D high-resolution fetal brain images;

[0055] The pseudo-label supervision module is used to achieve network mutual learning through cross-pseudo-label supervision;

[0056] The feature contrast learning module is used to enhance the encoder's discriminative power through feature contrast learning;

[0057] The output module is used to output the final segmentation result.

[0058] The present invention also provides a computer device, including a memory and a processor, the memory being used to store instructions and the processor being used to execute the instructions to implement the fetal magnetic resonance image brain region segmentation method as described above.

[0059] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the fetal magnetic resonance image brain region segmentation method as described above.

[0060] Therefore, the present invention employs the above-mentioned fetal magnetic resonance image brain region segmentation method, system, computer equipment and medium, which has the following beneficial technical effects: when used in actual clinical fetal brain magnetic resonance research, only a very small number of labels need to be labeled according to the process to train a fetal brain region segmenter with excellent segmentation performance, thus avoiding the serious dependence of existing supervised learning networks on labels when applied to actual clinical datasets. Attached Figure Description

[0061] Figure 1 This is a flowchart of the fetal magnetic resonance imaging brain region segmentation method of the present invention;

[0062] Figure 2 This paper compares the segmentation results obtained by the method of this invention with the segmentation results of U-Net in a real clinical scenario with very little labeled data. Detailed Implementation

[0063] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0064] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0065] Example 1

[0066] like Figure 1 As shown, the method for segmenting brain regions in fetal magnetic resonance imaging includes the following steps:

[0067] Step S1: Process the 2D low-resolution fetal MRI images (slice thickness 3-5mm) acquired in sagittal, coronal, and axial planes to generate 3D high-resolution fetal brain images (slice thickness 0.5mm), including the following steps:

[0068] The resolution of 2D low-resolution fetal magnetic resonance images acquired in sagittal, coronal and axial planes is reduced by resampling, which reduces subsequent resource consumption.

[0069] A localization convolutional neural network containing three convolutional layers and one pooling layer is used to perform preliminary brain localization on 2D low-resolution fetal MRI images, exclude a large amount of maternal tissue, and generate a segmentation mask.

[0070] Perform 3D morphological closing and opening operations sequentially on the segmentation mask;

[0071] The maximum connected component is selected and a 3D bounding cube is fitted. The 3D bounding cube is then resampled to the original resolution to obtain the final localization result.

[0072] For the final localization results, a segmentation convolutional neural network with a multi-scale Dice loss function was used to extract the fetal brain. The segmentation convolutional neural network consists of a three-layer convolutional encoder and a three-layer convolutional decoder.

[0073] ;

[0074] in, Represents the multi-scale total loss function. Represents the Dice loss function. S Indicates different scaling scales; express s Predicted labels based on scale express s The true label of scale Indicates the original predicted label, This indicates the original, authentic label.

[0075] The corrected image is registered to the selected orientation using symmetrical block matching.

[0076] Based on the registration results, Nadaraya-Watson kernel regression was used to estimate the initial 3D high-resolution fetal brain image:

[0077] ;

[0078] in, Indicates the coordinates of the point to be predicted. This represents the coordinates and values ​​of known data points. Represents the kernel function; Indicates the point to be predicted The kernel regression estimate at that location, Labels representing known data points, This indicates the number of known data points;

[0079] The final high-resolution 3D fetal brain image was obtained by iteratively solving an outlier-robust optimization problem.

[0080] ;

[0081] in, This represents the first 2D low-resolution fetal magnetic resonance image acquired in reality. k Layer data, This represents the forward operator of the data acquisition process. This represents a 3D high-resolution fetal brain image. Represents the regularization parameter. This represents the difference function.

[0082] Step S2: Selectively annotate the 3D high-resolution fetal brain image generated in step S1, and generate complete segmentation labels using a contour interpolation algorithm, including the following steps:

[0083] Step S21: Based on 3D high-resolution fetal brain images, a professional physician selects key layers according to the anatomical features of the brain regions;

[0084] Step S22: Extract the contour boundaries of two adjacent labeled key layers to obtain the original key layer contours;

[0085] Step S23: Based on the extracted contour boundaries of adjacent slices (slice A and slice B), calculate the signed distance map (SDM) for both layers:

[0086] ;

[0087] in, Indicates the outline boundary. d SDM represents the Euclidean distance; when outside the profile, it takes the negative value. Represents the x and y coordinates of the corresponding point.

[0088] Step S24: For the intermediate layer between two adjacent labeled key layers, merge the two symbol distance maps obtained in step S23 according to the inter-layer position ratio;

[0089] ;

[0090] ;

[0091] in, The proportion is tInterlayer coordinates of the intermediate layer at that time These represent the interlayer coordinates of slice A and slice B, respectively. They represent proportions respectively. t The symbolic distance plots for time and for slice A and slice B.

[0092] Step S25: Extract the zero-distance boundary from the fused symbolic distance map as the intermediate layer contour;

[0093] Step S26: Combine all the generated intermediate layer contours with the original key layer contours, perform a closing operation on the combined complete contour sequence, and output the final segmentation label for training the segmentation network.

[0094] Step S3: Based on the 3D high-resolution fetal brain image from step S1 and the segmentation labels from step S2, a deformer segmentation network with dual independent initialization and utilizing long-range attention mechanisms is used for mutual supervised learning, including the following steps:

[0095] Step S31: Deploy two structurally identical and independently initialized deformer segmentation networks that can utilize long-range attention mechanisms. and To conduct mutual learning;

[0096] Step S32: Perform collaborative training on unlabeled 3D high-resolution fetal brain images:

[0097] Step S321: Optimize network robustness through data-enhanced consistency constraints;

[0098] The same input image is subjected to strong enhancement and weak enhancement respectively. Weak enhancement includes operations such as rotation and flipping, which aims to expand the diversity of the dataset. Strong enhancement includes intensity transformation and affine transformation.

[0099] Constrain the consistency of segmentation results for weak and strong augmentation outputs of the dual network;

[0100] Step S322: Achieve network mutual learning through cross-pseudo-label supervision;

[0101] The two networks mutually generate pseudo-labels for each other for supervised training, i.e. The network's prediction for this instance serves as The network outputs labels, and vice versa. Based on this, a segmentation loss function is constructed, namely:

[0102] ;

[0103] in, Represents the cross-entropy loss function. This represents the pseudo-label generated from the output of the first network. This represents the output prediction probability of the second network; This represents the semi-supervised loss function. This represents the pseudo-label generated by the output of the second network. This represents the output prediction probability of the second network;

[0104] Step S323: Enhance the discriminative power of the encoder through feature contrast learning;

[0105] A feature contrastive loss function is applied at the encoder level to promote compactness of similar features and separation of features between classes. The contrastive loss function is in the form of Information Noise-Contrastive Estimation (InfoNCE).

[0106] ;

[0107] in, q This represents the feature vector of the query sample. Indicates positive sample features. Represents the set of negative samples. This represents the dot product of the similarity function. A coefficient representing the degree of control over the concentration of the distribution. Indicates the first The label of each negative sample, This indicates the number of negative samples.

[0108] Step S33: Jointly optimize the loss from step S32 and output the model.

[0109] The invention will be further illustrated below with specific examples.

[0110] The imaging device used was a 1.5T magnetic resonance imaging (MRI) scanner, with the following parameters: Voxel size (resolution): 0.7 × 0.7 × 3 mm 3 Field of view (FOV) 360, echo time (TE): 90-110ms, recovery time (TR): 1300-1400ms.

[0111] Eighty multi-planar 2D single-shot fast spin echo (SSFSE) images were acquired. First, brain extraction, brain localization, and iterative reconstruction were performed on this data to obtain 80 high-resolution 3D fetal brain images. Then, two images were randomly selected and subjected to brain region labeling every n rows by a professional physician. Subsequently, contour filling was used to obtain two semi-supervised labels. This training data with only two labeled images was then input into a semi-supervised network for training, resulting in the final automated fast segmenter.

[0112] like Figure 2As shown, in the case of few labels, the method is compared with the currently mainstream deep learning segmentation method U-Net (the left side represents the mainstream method, and the right side represents the method proposed in this invention). It is evident that the method of this invention significantly outperforms traditional methods when training with few labels.

[0113] Example 2

[0114] Fetal magnetic resonance imaging brain region segmentation system, including:

[0115] The image processing module is used to process 2D low-resolution fetal magnetic resonance images acquired from multiple angles to generate 3D high-resolution fetal brain images.

[0116] The annotation module is used to selectively annotate the generated 3D high-resolution fetal brain image and generate complete segmentation labels through contour interpolation algorithm.

[0117] The segmentation network module is used for mutual supervision learning based on 3D high-resolution fetal brain images and generated labeled data, employing a deformable segmentation network with dual independent initialization that can utilize long-range attention mechanisms.

[0118] The data augmentation module is used to apply data augmentation consistency constraints to unlabeled 3D high-resolution fetal brain images;

[0119] The pseudo-label supervision module is used to achieve network mutual learning through cross-pseudo-label supervision;

[0120] The feature contrast learning module is used to enhance the encoder's discriminative power through feature contrast learning;

[0121] The output module is used to output the final segmentation result.

[0122] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0123] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0124] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0125] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0126] Therefore, the present invention employs the above-mentioned fetal magnetic resonance image brain region segmentation method, system, computer equipment and medium, which can train a high-performance segmenter with only a few labels, significantly reducing labeling costs and improving segmentation efficiency. It is applicable to actual clinical fetal brain magnetic resonance research and provides efficient technical support for fetal brain development assessment.

[0127] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for segmenting brain regions in fetal magnetic resonance imaging, characterized in that, Includes the following steps: Step S1: Process the 2D low-resolution fetal magnetic resonance images acquired from multiple angles to generate 3D high-resolution fetal brain images. Step S2: Selectively perform interlayer annotation on the 3D high-resolution fetal brain image generated in step S1, and generate complete segmentation labels through contour interpolation algorithm; Step S3: Based on the 3D high-resolution fetal brain image from step S1 and the segmentation labels from step S2, a deformer segmentation network with dual independent initialization that can utilize long-range attention mechanisms is used for mutual learning. In step S2, selective interlayer annotation is performed on the 3D high-resolution fetal brain image generated in step S1, and complete segmentation labels are generated using a contour interpolation algorithm, including the following steps: Step S21: Based on 3D high-resolution fetal brain images, select key layers according to the anatomical features of brain regions; Step S22: Extract the contour boundaries of two adjacent labeled key layers to obtain the original key layer contours; Step S23: Based on the extracted contour boundaries, calculate the symbolic distance maps for the two layers respectively; Step S24: For the intermediate layer between two adjacent labeled key layers, merge the two symbol distance maps obtained in step S23 according to the inter-layer position ratio; Step S25: Extract the zero-distance boundary from the fused symbolic distance map as the intermediate layer contour; Step S26: Combine all the generated intermediate layer contours with the original key layer contours, perform a closing operation on the combined complete contour sequence, and output the final segmentation label for training the segmentation network. In step S3, a deformer segmentation network with dual independent initialization that can utilize long-range attention mechanisms is used for mutual supervision learning, including the following steps: Step S31: Deploy two structurally identical and independently initialized deformable segmentation networks that can utilize long-range attention mechanisms for mutual learning. Step S32: Perform collaborative training on unlabeled 3D high-resolution fetal brain images: Step S321: Optimize network robustness through data-enhanced consistency constraints; Step S322: Achieve network mutual learning through cross-pseudo-label supervision; Step S323: Enhance the discriminative power of the encoder through feature contrast learning; Step S33: Jointly optimize the loss from step S32 and output the model.

2. The fetal magnetic resonance imaging brain region segmentation method according to claim 1, characterized in that, In step S1, the 2D low-resolution fetal magnetic resonance images acquired from multiple perspectives are processed to generate 3D high-resolution fetal brain images, including the following steps: Resampling was performed on 2D low-resolution fetal magnetic resonance images acquired from multiple angles to reduce resolution; A localization convolutional neural network containing three convolutional layers and one pooling layer is used to perform preliminary brain localization on 2D low-resolution fetal MRI images and generate segmentation masks. Perform 3D morphological closing and opening operations sequentially on the segmentation mask; The maximum connected component is selected and a 3D bounding cube is fitted. The 3D bounding cube is then resampled to the original resolution to obtain the final localization result. For the localization results, a segmentation convolutional neural network with multi-scale Dice loss function was used to extract the fetal brain. The segmentation convolutional neural network consists of a three-layer convolutional encoder and a three-layer convolutional decoder. ； in, Represents the multi-scale total loss function. Represents the Dice loss function. S Indicates different scaling scales; express s Predicted labels based on scale express s The true label of scale Indicates the original predicted label, Indicates the original, authentic label; Bias field correction was performed on the acquired 2D low-resolution fetal magnetic resonance images; The corrected image is registered to the selected orientation using symmetrical block matching. Based on the registration results, Nadaraya-Watson kernel regression was used to estimate the initial 3D high-resolution fetal brain image: ； in, Indicates the coordinates of the point to be predicted. This represents the coordinates and values ​​of known data points. Represents the kernel function; Indicates the point to be predicted The kernel regression estimate at that location, Labels representing known data points, This indicates the number of known data points; The final high-resolution 3D fetal brain image was obtained by iteratively solving an outlier-robust optimization problem. ； in, This represents the first 2D low-resolution fetal magnetic resonance image acquired in reality. Layer data, This represents the forward operator of the data acquisition process. This represents the 3D high-resolution fetal brain image we want to obtain. Represents the regularization parameter. This represents the difference function.

3. The fetal magnetic resonance imaging brain region segmentation method according to claim 1, characterized in that, Data-enhanced consistency constraints include: Apply weak enhancement and strong enhancement operations to the same image respectively; Constrain the consistency of segmentation results for weak and strong augmentation outputs of the dual network.

4. The fetal magnetic resonance imaging brain region segmentation method according to claim 1, characterized in that, Cross-label supervision includes: The prediction results of the first network are used as the supervision labels for the second network; The prediction results of the second network are used as the supervision labels for the first network; The supervision error is calculated using cross-entropy loss and Dice loss.

5. The fetal magnetic resonance imaging brain region segmentation method according to claim 1, characterized in that, Feature contrastive learning includes: Extract feature vectors from the encoder output layer; Minimize the feature distance of similar anatomical structures and maximize the feature distance of different types of structures.

6. A fetal magnetic resonance imaging brain region segmentation system, characterized in that, A method for performing fetal magnetic resonance imaging brain region segmentation as described in any one of claims 1-5, comprising: The image processing module is used to process 2D low-resolution fetal magnetic resonance images acquired from multiple angles to generate 3D high-resolution fetal brain images. The annotation module is used to selectively annotate the generated 3D high-resolution fetal brain image and generate complete segmentation labels through contour interpolation algorithm. The segmentation network module is used for mutual supervision learning based on 3D high-resolution fetal brain images and generated labeled data, employing a deformable segmentation network with dual independent initialization that can utilize long-range attention mechanisms. The data augmentation module is used to apply data augmentation consistency constraints to unlabeled 3D high-resolution fetal brain images; The pseudo-label supervision module is used to achieve network mutual learning through cross-pseudo-label supervision; The feature contrast learning module is used to enhance the encoder's discriminative power through feature contrast learning; The output module is used to output the final segmentation result.

7. A computer device, characterized in that, It includes a memory and a processor, the memory being used to store instructions, and the processor being used to execute the instructions to implement the fetal magnetic resonance image brain region segmentation method as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the fetal magnetic resonance image brain region segmentation method as described in any one of claims 1-5.

Images