Method for automatically segmenting brain structure of multi-center applicable human brain magnetic resonance image

Abstract

The invention relates to a multi-center applicable automatic segmentation method for a brain structure of a human brain magnetic resonance image, and the segmentation method specifically comprises the steps: obtaining MRI data of different platforms, and carrying out the N4 bias field correction and image standardization processing of the MRI data; u-Net is used as a basic framework, a self-attention mechanism module is fused, and an MRI brain image structure segmentation network model suitable for multiple platforms is constructed; and inputting the processed MRI data into a segmentation network model, training the segmentation network model by using transfer learning and domain adaptation technologies, performing uncertainty estimation on an output result, and further performing model training by using screened high-quality unlabeled data so as to realize automatic segmentation of the brain structure of the human brain magnetic resonance image. Compared with the prior art, the method has the advantages that brain structures of different magnetic resonance imaging platforms can be accurately and quickly segmented, a more universal, efficient and reliable tool is provided for the field of brain image analysis, and wider application and research exploration are facilitated.

Description

Technical Field

[0001] The present invention relates to the fields of computer vision and medical imaging technology, and in particular to a multi-center applicable method for automatic brain structure segmentation of human brain magnetic resonance images. Background Art

[0002] Brain science research plays a vital role in exploring human evolution, development, and the treatment of neurological diseases. Magnetic resonance imaging (MRI), as an ideal non-invasive imaging method, can provide high-resolution and multimodal imaging of the brain. Brain region structure segmentation of MRI data is a key step in analyzing brain function and metabolic processes.

[0003] The task of fine brain segmentation is a key technology in the field of neuroscience and medicine, which aims to divide brain imaging data into different anatomical or functional regions. This task is extremely important in promoting human scientific research and clinical applications. For humans, fine brain segmentation has direct application value in clinical diagnosis and treatment, such as providing key lesion information in the diagnosis and treatment of diseases such as Alzheimer's disease, Parkinson's disease and brain tumors. These segmentation techniques not only help to better understand the structure and function of the brain, but also provide an important basis for formulating personalized treatment plans.

[0004] By focusing on the study of human brain MRI data, we can deeply explore brain function and pathological changes, and provide more accurate and reliable support for improving clinical diagnosis and treatment methods. This not only promotes the development of basic research in brain science, but also brings new hope for the treatment of neurological diseases.

[0005] The task of fine brain region structure segmentation faces many challenges, mainly due to the high complexity of brain structure and individual differences. The brain contains many subtle anatomical regions, and the boundaries between these regions are usually blurred and difficult to distinguish. The variability of brain structure between individuals further increases the difficulty of segmentation. In addition, obtaining high-quality brain imaging data requires expensive equipment and technology, while manual labeling of these data is time-consuming and susceptible to subjective bias. Therefore, developing algorithms that can automatically and accurately segment brain regions has become an important task in this field.

[0006] When performing fine brain segmentation in single-modality T1-weighted MRI images, the heterogeneity of data sources also poses a huge challenge. Differences in imaging parameters, inconsistent image quality, different standardization methods, and differences in data formats and preprocessing procedures caused by different MRI devices and scanning protocols will lead to inconsistent performance of the same anatomical region in different images, making it difficult for the segmentation algorithm to maintain stable performance. In addition, the annotation standards and methods used by different research teams may be different, which further increases the inconsistency of training data and affects the generalization ability of the segmentation model.

[0007] Currently, there are a variety of brain structure segmentation tools designed for specific MRI modalities, and the development of deep learning algorithms has also promoted the improvement of the accuracy of human brain structure segmentation. However, the performance of these methods often decreases significantly when processing data from different platforms or scanning protocols. It is still a challenge to use a single tool to accurately perform fine brain region structure segmentation on images acquired from multiple MRI platforms.

[0008] In addition, existing segmentation algorithms usually only have segmentation functions and lack the ability of self-checking and correction, which leads to the lack of guarantee of the accuracy and reliability of segmentation results. Since these algorithms cannot automatically identify and correct segmentation errors, they must rely on a large amount of manpower for subsequent inspection and correction. This reliance on manual inspection is not only extremely time-consuming, but also extremely inefficient when processing large-scale data. Each segmentation result needs to be carefully reviewed by experts to ensure that it meets the expected accuracy and precision. This process is not only cumbersome, but also easily affected by human subjective factors, resulting in inconsistent quality of segmentation results. In addition, manual inspection and correction also increase the workload, limiting the widespread use of these algorithms in practical applications. In clinical and research environments, this manual inspection and correction process not only wastes valuable human resources, but also prolongs the cycle of data processing, which may affect the progress of diagnosis and research. Especially when large amounts of data need to be processed and analyzed quickly, the limitations of this method are more obvious.

[0009] In general, the challenges faced by existing segmentation technologies are as follows: 1. Lack of flexibility and universality. Existing tools are designed for specific platforms and are not suitable for cross-platform applications. 2. Lack of automatic quality assessment mechanism. This means that in the case of incorrect segmentation, manual inspection and correction are required.

[0010] In summary, medical brain image structure segmentation technology needs to comprehensively consider the flexibility and universality of the model on different platforms, as well as automatic quality assessment and adjustment and other challenges. Therefore, how to introduce more flexible transfer learning methods and domain adaptation technology into the model to achieve more accurate and robust brain structure segmentation in a cross-domain complex data environment has become an important technical problem that needs to be solved urgently. Summary of the invention

[0011] The purpose of the present invention is to provide a multi-center applicable method for automatic brain structure segmentation of human brain magnetic resonance images in order to overcome the defects of the above-mentioned prior art. By designing a robust deep learning model, segmentation can be performed directly on the original image data, eliminating the complex preprocessing process, simplifying the use process, lowering the technical threshold, and maintaining high-precision segmentation performance.

[0012] The purpose of the present invention can be achieved by the following technical solutions:

[0013] The present invention provides a multi-center applicable method for automatic brain structure segmentation of human brain magnetic resonance images, characterized in that it comprises the following steps:

[0014] S1: Acquire MRI data from different platforms and perform N4 bias field correction and image standardization on the MRI data;

[0015] S2: Using U-Net as the basic framework and integrating the self-attention mechanism module, a MRI brain image structure segmentation network model suitable for multiple platforms is constructed;

[0016] S3: input the MRI data processed in S1 into the MRI brain image structure segmentation network model constructed in S2, and train the MRI brain image structure segmentation network model using transfer learning and domain adaptation technology;

[0017] S4: The uncertainty of the output structure of the MRI brain image structure segmentation network model in S3 is estimated, and the segmentation results are evaluated and screened. The model is further trained using the screened high-quality unlabeled data, and the trained model is used to achieve automatic segmentation of brain structures in human brain magnetic resonance images.

[0018] Furthermore, in S1, the standardization process converts the original image V into a standardized MRI image V norm , and its calculation formula is as follows:

[0019]

[0020] in, and Respectively represent the 1st and 99th percentiles of the intensity value in the input image, that is, the lower and upper limits of the intensity value; V is the original image, V norm This is the image after deviation normalization.

[0021] Furthermore, in S2, the U-Net includes five groups of symmetrical encoding blocks and decoding blocks, each encoding block is composed of two 3x3x3 convolutions, followed by a batch normalization layer, an exponential linear unit activation function and a 2x2x2 maximum pooling operation, and the stride is set to 2;

[0022] The first encoding block outputs 32 feature channels, and the number of feature channels doubles after each encoding block.

[0023] Furthermore, the interior of each decoding block consists of two 3x3x3 convolutions, followed by a batch normalization layer, an exponential linear unit activation function, and a transposed convolution layer;

[0024] In the final decoding stage, a 1x1x1 convolutional layer and a Sigmoid activation function are used to map the 32 feature channels to a 96-element probability map.

[0025] Furthermore, in S2, the self-attention mechanism module simulates the long-distance and global contextual dependencies in feature representation by adding a non-local attention layer at the center of the network model, and calculates the weighted saliency map of each position on the feature map, so as to more effectively distinguish different brain regions with different contrast characteristics and distributions.

[0026] Furthermore, in S3, the Dice loss function and adaptive batch normalization are used to train the MRI brain image structure segmentation network model, and 50 training cycles are set, the batch size is 16, and the learning rate is set to 10 in the source domain and the target domain respectively. -4 and 10 -5 ;

[0027] An adaptive batch normalization strategy is used to dynamically update the statistical parameters of the batch normalization layer during the inference phase to improve the cross-platform generalization performance of the model.

[0028] Furthermore, in S4, approximate Bayesian inference is used to estimate the uncertainty of the MRI brain image structure segmentation network model. The specific process is: introduce a dropout layer to obtain the uncertainty measure of the model when generating segmentation, keep the dropout layer activated in the inference stage, collect N independent samples through Monte Carlo sampling, and calculate the pixel level uncertainty of each sample using the following formula:

[0029]

[0030] Among them, p t It is the output probability of the sigmoid function of the final layer of the model in the inference stage; is the average prediction value of all inference results; T is the number of Monte Carlo sampling, and T=10.

[0031] Furthermore, in S4, based on uncertainty estimation, a semi-supervised process is used to screen and predict high-quality segmentation results and filter prediction results with high uncertainty and relatively unreliable to achieve quality assessment.

[0032] Furthermore, the semi-supervised process is specifically as follows: performing model evaluation on all unlabeled data, and then screening out predictions with the lowest uncertainty based on Monte Carlo quality assessment, and adding them as pseudo labels to the training set in the next round of iteration.

[0033] Furthermore, the semi-supervised process adopts a hybrid loss function, and the specific formula is as follows:

[0034]

[0035]

[0036] Among them, x i is the real label y in the current mini-batch i The total number of labeled data points is N; x j is the model parameter θ in the previous iteration t-1 The total number of pseudo-labeled data points predicted by the MCQA module is M, and the uncertainty of these data points is lower than the threshold ζ. j θ t ) represents the model's prediction of the input x with parameters θ; λ Dice is a hyperparameter that controls the weight of the Dice loss term; α(t) is used to balance the relative contribution of true labels and pseudo labels in the algorithm; T sl is the preset threshold iteration number, which controls when to start introducing pseudo labels; sl is the balance coefficient, which determines the weight of the pseudo label loss in the total loss; t represents the current number of iterations.

[0037] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0038] 1. This invention is based on deep learning technology, which is trained by a large amount of data, so that the model can capture and learn the complex features in the image data. Compared with traditional rule-based methods and toolboxes, deep learning methods have higher processing speed and stronger generalization ability. Especially when processing large-scale data sets, deep learning algorithms can significantly improve segmentation efficiency and quickly output high-quality segmentation results.

[0039] 2. The present invention does not require complicated preprocessing. Traditional MRI image segmentation usually requires complicated preprocessing steps, such as denoising and registration, to improve image quality and consistency. These preprocessing steps are not only time-consuming, but may also introduce additional errors. The present invention, by designing a robust deep learning model, can directly perform segmentation on the original image data, eliminating the complicated preprocessing process, simplifying the use process, lowering the technical threshold, and maintaining high-precision segmentation performance.

[0040] 3. The present invention has a certain ability to repair the noise in the pseudo-labels. In semi-supervised learning, the model needs to be trained with a large amount of unlabeled data and a small amount of labeled data, wherein the unlabeled data assists the training by generating pseudo-labels. However, there may be noise in the pseudo-labels, which will affect the training effect of the model. The network model in the present invention selects the pseudo-label with the highest confidence each time to add to the training data, which has a certain correction effect on the noise in the pseudo-labels to ensure the stability of the training process and the accuracy of the segmentation results, thereby improving the robustness of the model.

[0041] 4. The present invention can obtain prior knowledge through transfer learning, so that the model can achieve cross-domain and multi-platform segmentation robustness. Transfer learning technology allows the model to learn useful features and knowledge from imaging data from one field or platform, and apply this knowledge to another field or platform. This enables the present invention to achieve consistent segmentation effects on different MRI platforms. Through transfer learning, the model can quickly adapt to new data sources, significantly improve the cross-platform segmentation robustness, reduce the impact of differences in equipment and imaging parameters, and ensure the consistency and reliability of segmentation results between different platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A structural framework diagram of a multi-center applicable method for automatic brain structure segmentation of human brain magnetic resonance images;

[0043] Figure 2 A schematic diagram of a method for automatic brain structure segmentation in human brain magnetic resonance images applicable to multiple centers;

[0044] Figure 3 The segmentation results of 96 brain structures and 3D rendering schematic diagrams in Example 1;

[0045] Figure 4 This is a schematic diagram of the segmentation results of 96 brain structures from multiple centers in Example 1. DETAILED DESCRIPTION

[0046] The specific implementation methods of the present invention are described in detail below through examples. These examples are implemented on the premise of the scheme described in the present invention, and provide detailed implementation methods and specific operating processes, but the protection scope of the present invention is not limited to the following examples.

[0047] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments. Any features such as structure / module name, control mode, algorithm, process or composition ratio not clearly described in this technical solution are regarded as common technical features disclosed in the prior art.

[0048] Example 1

[0049] This embodiment provides a multi-center applicable method for automatic brain structure segmentation of human brain magnetic resonance images. Figure 1 As shown, the following steps are included:

[0050] S1: To reduce the differences between different scanning platforms, a consistent preprocessing process was performed on all datasets. Specifically, MRI data obtained from multiple MRI platforms, including Canon, GE, Lowin, Philips, Siemens, and United Imaging, were collected and bias-corrected using the N4 bias-field correction algorithm. Each image was then normalized.

[0051] The normalization process converts the original image V into a normalized MRI image V norm , and its calculation formula is as follows:

[0052]

[0053] in, and Respectively represent the 1st and 99th percentiles of the intensity value in the input image, that is, the lower and upper limits of the intensity value; V is the original image, V norm This is the image after deviation normalization.

[0054] When processing MRI data, the voxel size was adjusted to 1 mm to ensure the stability of the running results. In addition, in order to enable the network to learn invariant features from data that presents variability between different subjects and protocols, a data augmentation strategy was adopted. Transformations were applied to each input, including: random rotation (±10 degrees), random scaling (90%-110%), and random translation (up to ±10 pixels in each axis). For example, rotation and translation operations can simulate the representation of subjects in different coordinate systems and different head directions. Scaling operations can simulate the inconsistency of brain volume size that occurs in different sites or even the same subject during brain development.

[0055] S2: Using U-Net as the basic framework and integrating the self-attention mechanism module, a MRI brain image structure segmentation network model (BREN) that is suitable for multiple platforms is constructed.

[0056] like Figure 2As shown in the figure, the U-Net includes five groups of symmetrical encoding blocks and decoding blocks, which are responsible for feature compression and expansion respectively in the data feedforward process. Each encoding block consists of two 3x3x3 convolutions, followed by a batch normalization layer, an exponential linear unit (ELU) activation function and a 2x2x2 maximum pooling operation, and the stride is set to 2. The first encoding block outputs 32 feature channels, and the number of feature channels doubles after each encoding block. Each decoding block consists of two 3x3x3 convolutions, followed by a batch normalization layer, an exponential linear unit activation function and a transposed convolution layer, and in the final decoding stage, a 1x1x1 convolution layer and a Sigmoid activation function are used to map the 32 feature channels to a 96-element probability map for soft segmentation of different brain regions for each pixel. The decoding block is structurally symmetrical with the corresponding encoding block.

[0057] The cross-domain task performance of BREN is significantly enhanced by integrating a self-attention mechanism module, which simulates long-range and global contextual dependencies in feature representation by adding a non-local attention layer at the center of the network and calculates a weighted saliency map for each position on the feature map to more effectively distinguish different brain regions with different contrast characteristics and distributions.

[0058] S3: Input the processed MRI image in S1 into the MRI brain image structure segmentation network model in S2, and use transfer learning and domain adaptation technology to train the MRI brain image structure segmentation network model.

[0059] The BREN network model is trained using the Dice loss function and adaptive matrix estimation optimizer, and 50 training cycles are set, the batch size is 16, and the learning rate is set to 10- in the source domain and the target domain, respectively. 4 and 10 -5 .

[0060] In order to overcome the domain shift problem encountered by deep learning models in cross-domain applications, the adaptive batch normalization (AdaBN) strategy is used to dynamically update the statistical parameters of the batch normalization (BN) layer during the inference phase to improve the cross-platform generalization performance of the model. AdaBN dynamically updates the statistical parameters in the BN layer for a specific target domain during the inference phase, thereby improving the generalization performance of the model. In the brain region structure segmentation task, AdaBN enables the model to adapt to the differences in MRI signal intensity on different platforms without changing the architecture or hyperparameters.

[0061] S4: The uncertainty of the output structure of the MRI brain image structure segmentation network model in S3 is estimated, and the segmentation results are evaluated and screened. The model is further trained using the screened high-quality unlabeled data, and the trained model is used to achieve automatic segmentation of brain structures in human brain magnetic resonance images.

[0062] To understand the intra-rater variability exhibited in BREN and provide clinicians with valuable insights into the interpretability of the model, a dropout layer was introduced into the BREN network model to obtain a measure of the model’s uncertainty in generating segmentations.

[0063] The uncertainty of predictions in deep neural networks can be divided into two main categories: aleatory uncertainty and epistemic uncertainty. Aleatory uncertainty reflects the potential inherent noise in the input test data, and it is usually difficult to reduce this type of uncertainty by increasing the amount of training data. In contrast, epistemic uncertainty involves uncertainty in model parameters, which reflects the lack of knowledge about the optimal model. In deep learning networks, epistemic uncertainty usually arises from the lack of training data in certain areas of the input domain and decreases as the diversity of the training data distribution increases.

[0064] Approximate Bayesian inference is used to estimate the uncertainty of the MRI brain image structure segmentation network model. The specific process is as follows: a dropout layer is introduced to obtain the uncertainty measure of the model when generating segmentation, the dropout layer is kept activated during the inference phase, N independent samples are collected through Monte Carlo sampling, and the pixel-level uncertainty of each sample is calculated using the following formula:

[0065]

[0066] Among them, pt is the output probability of the sigmoid function of the final layer of the model in the inference stage; is the average prediction value of all inference results; T is the number of Monte Carlo sampling, and T=10.

[0067] Based on uncertainty estimation, a semi-supervised process is used to screen and predict high-quality segmentation results and filter out prediction results with high uncertainty and relatively unreliable to achieve quality assessment. The semi-supervised process is specifically: model evaluation is performed on all unlabeled data, and then based on Monte Carlo quality assessment (MCQA), the predictions with the lowest uncertainty are screened out and added to the training set as pseudo-labels in the next iteration.

[0068] The semi-supervised process uses a mixed loss function, the specific formula is as follows:

[0069]

[0070]

[0071] Among them, x i is the real label y in the current mini-batch i The total number of labeled data points is N; x i is the model parameter θ in the previous iteration t-1 The total number of pseudo-labeled data points predicted by the MCQA module is M. j θ t ) represents the model's prediction of the input x with parameters θ. Dice is a hyperparameter that controls the weight of the Dice loss term. α(t) is used to balance the relative contribution of true labels and pseudo labels in the algorithm.

[0072] The specific process of the relative contribution of the real label and the pseudo label in the α(t) balance algorithm is as follows: α(t) in the first stage 0 <t<T s1 In , the network is trained directly in the target domain using labeled data;

[0073] In the second stage T sl ≤t≤T epoch In , the unlabeled data with pseudo labels additionally participate in the semi-supervised learning process through the determination of the balance coefficient α(t) to contribute to the loss function;

[0074] Among them, T epoch represents the total number of training cycles, and t represents the current cycle. Set T sl =20,λ Dice =1,λ sl =0.5, and ζ is the minimum uncertainty value in each mini-batch.

[0075] like Figure 3 and Figure 4 As shown, the segmentation results of the network model in this embodiment are displayed. The results show that the BREN network model can stably segment MRI data from different platforms and instruments, and exhibits excellent segmentation effects on each platform. This advantage gives BREN obvious advantages in cross-platform applications, especially on a variety of MRI platforms such as Canon, GE, Lowin, Philips, Siemens and United Imaging, which can achieve efficient and accurate brain region segmentation. In addition, traditional toolboxes (such as segmentation methods based on registration) require a lot of time for preprocessing and manual inspection, while BREN relies on its powerful deep learning capabilities and automatic quality assessment modules to quickly complete segmentation tasks and ensure the consistency and reliability of the results, providing strong support for neuroimaging research and clinical applications.

[0076] The above description of the embodiments is to facilitate the understanding and use of the invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to these embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the present invention should be within the scope of protection of the present invention.

Claims

1. A multi-center applicable method for automatic brain structure segmentation of human brain magnetic resonance images, characterized in that: The following steps are involved: S1: Acquire MRI data from different platforms and perform N4 bias field correction and image standardization on the MRI data; S2: Using U-Net as the basic framework and integrating the self-attention mechanism module, using transfer learning and domain adaptation technology, we build a MRI brain image structure segmentation network model that is suitable for multiple platforms; S3: input the MRI data processed in S1 into the MRI brain image structure segmentation network model constructed in S2, and train the MRI brain image structure segmentation network model using transfer learning and domain adaptation technology; S4: The uncertainty of the output structure of the MRI brain image structure segmentation network model in S3 is estimated, and the segmentation results are evaluated and screened. The model is further trained using the screened high-quality unlabeled data, and the trained model is used to achieve automatic segmentation of brain structures in human brain magnetic resonance images.

2. The multi-center applicable method for automatic brain structure segmentation of human brain magnetic resonance images according to claim 1, characterized in that: In S1, the standardization process converts the original image V into a standardized MRI image V norm , and its calculation formula is as follows: in, and Respectively represent the 1st and 99th percentiles of the intensity value in the input image, that is, the lower and upper limits of the intensity value; V is the original image, V norm This is the image after deviation normalization.

3. The multi-center applicable method for automatic brain structure segmentation of human brain magnetic resonance images according to claim 1, characterized in that: In S2, the U-Net includes five groups of symmetrical encoding blocks and decoding blocks. Each encoding block consists of two 3x3x3 convolutions, followed by a batch normalization layer, an exponential linear unit activation function, and a 2x2x2 maximum pooling operation, and the stride is set to 2; The first encoding block outputs 32 feature channels, and the number of feature channels doubles after each encoding block.

4. The multi-center applicable method for automatic brain structure segmentation of human brain magnetic resonance images according to claim 3, characterized in that: Each decoding block consists of two 3x3x3 convolutions, followed by a batch normalization layer, an exponential linear unit activation function, and a transposed convolution layer; In the final decoding stage, a 1x1x1 convolutional layer and a Sigmoid activation function are used to map the 32 feature channels to a 96-element probability map.

5. The multi-center applicable method for automatic brain structure segmentation of human brain magnetic resonance images according to claim 1, characterized in that: In S2, the self-attention mechanism module simulates the long-distance and global contextual dependencies in feature representation by adding a non-local attention layer at the center of the network model, and calculates the weighted saliency map of each position on the feature map to more effectively distinguish different brain regions with different contrast characteristics and distributions.

6. The multi-center applicable method for automatic brain structure segmentation of human brain magnetic resonance images according to claim 1, characterized in that: In S3, the Dice loss function and adaptive batch normalization are used to train the MRI brain image structure segmentation network model, and 50 training cycles are set, the batch size is 16, and the learning rate is set to 10 in the source domain and the target domain respectively. -4 and 10 -5 ; An adaptive batch normalization strategy is used to dynamically update the statistical parameters of the batch normalization layer during the inference phase to improve the cross-platform generalization performance of the model.

7. The multi-center applicable method for automatic brain structure segmentation of human brain magnetic resonance images according to claim 1, characterized in that: In S4, approximate Bayesian inference is used to estimate the uncertainty of the MRI brain image structure segmentation network model. The specific process is as follows: a dropout layer is introduced to obtain the uncertainty measure of the model when generating segmentation, the dropout layer is kept activated during the inference phase, N independent samples are collected through Monte Carlo sampling, and the pixel-level uncertainty of each sample is calculated using the following formula: Among them, p t It is the output probability of the sigmoid function of the final layer of the model in the inference stage; is the average prediction value of all inference results; T is the number of Monte Carlo sampling, and T=10.

8. The multi-center applicable method for automatic brain structure segmentation of human brain magnetic resonance images according to claim 1, characterized in that: In S4, based on uncertainty estimation, a semi-supervised process is used to screen and predict high-quality segmentation results and filter out prediction results with high uncertainty and relatively unreliable to achieve quality assessment.

9. The multi-center applicable method for automatic brain structure segmentation of human brain magnetic resonance images according to claim 8, characterized in that: The semi-supervised process is specifically as follows: a model is evaluated on all unlabeled data, and then predictions with the lowest uncertainty are screened out based on Monte Carlo quality assessment, and are added to the training set as pseudo labels in the next round of iteration.

10. The multi-center applicable method for automatic brain structure segmentation of human brain magnetic resonance images according to claim 8, characterized in that: The semi-supervised process adopts a hybrid loss function, and the specific formula is as follows: Among them, x i is the real label y in the current mini-batch i The total number of labeled data points is N; x j is the model parameter θ in the previous iteration t-1 The pseudo-labeled data points predicted by the MCQA module are calculated to have an uncertainty lower than the threshold ζ, with a total number of M; f(x j θ t ) represents the model's prediction of the input x with parameters θ; λ Dice is a hyperparameter that controls the weight of the Dice loss term; α(t) is used to balance the relative contribution of true labels and pseudo labels in the algorithm; T sl is the preset threshold iteration number, which controls when to start introducing pseudo labels; sl is the balance coefficient, which determines the weight of the pseudo label loss in the total loss; t represents the current number of iterations.

Images