An artificial intelligence-based brain region automatic segmentation method

Abstract

A brain region automatic segmentation method based on artificial intelligence comprises a model training stage and a model using stage; the model training stage comprises: 1, preparing original data; 2, generating training data; the original image data is preprocessed and resampled; 3, model training; the standard template segmentation image data and the image data in the training data are input into a deep neural network model, the predicted label segmentation image data is output through the model, and the resampled label segmentation image data in the training data is calculated together to obtain a loss loss function; the model using stage comprises: obtaining a new brain MRI scanning image, preprocessing new original image data, then resampling, obtaining the data as using data, inputting the standard template segmentation image data into the trained model at the same time, and outputting the predicted label segmentation image data. The brain region segmentation recognition accuracy is higher, the model is more stable, the segmentation details are more abundant, and real-time can be performed.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence medical image processing, specifically to an automatic brain region segmentation method based on artificial intelligence. Background Technology

[0002] Medical imaging technology is the most common technical aid in the study of the brain and brain diseases. Common imaging techniques include computed tomography (CT), positron emission tomography (PET), electroencephalography (EEG), and magnetic resonance imaging (MRI). Compared with other imaging techniques, MRI is favored by researchers and clinicians due to its numerous advantages, such as being non-invasive, radiation-free, and having high resolution. The brain's ability to perform various tasks is achieved through the action of different brain regions or through cross-regional synergy. Segmenting the brain into different regions is one of the important steps in analyzing and studying brain diseases. Manual segmentation is usually the gold standard for human brain image segmentation. However, this requires manually drawing every tissue structure in the brain imaging data. This method is not only expensive, tedious, and overly dependent on the operator, but also suffers from large differences in segmentation between different doctors, low reproducibility, and inevitable bias due to human error. It is also difficult to obtain a large number of fully annotated, highly accurate medical image datasets.

[0003] Currently, brain region segmentation methods are mainly divided into two categories: traditional methods and deep learning-based methods. Traditional methods include atlas-based brain region segmentation, which transforms the segmentation problem into a registration problem and is a relatively complete system framework that integrates prior knowledge. Deep learning technology is surpassing traditional classical methods and is widely used for accurate brain region segmentation. Dolz et al. used a 3D convolutional neural network framework for MRI brain structure segmentation based on feature information at different scales, achieving effective end-to-end training; Shakeri et al. proposed a deep learning method based on fully convolutional neural networks for segmenting 2D human subcortical structures; Bernal et al. quantitatively analyzed different 2D and 3D fully convolutional neural network architectures for MRI whole-brain tissue segmentation, etc.

[0004] However, automatic brain region segmentation faces several challenges, such as the large number of functional brain regions, significant individual differences, artifacts and noise caused by imaging equipment, and blurred data edges due to partial volume effects.

[0005] Therefore, based on deep learning models and atlas-based brain region segmentation methods, we propose a deep learning algorithm for whole-brain tissue segmentation based on prior information. By introducing prior features, this method can achieve more accurate and faster segmentation results. This not only lays a foundation for subsequent medical image data analysis but also improves the classification accuracy of AI-assisted diagnosis in medical care, providing technical support for applications such as intraoperative navigation, accurate lesion location, and observation of changes in lesion sites.

[0006] Accurate brain segmentation is a crucial prerequisite for subsequent brain data analysis, surgical procedures, and lesion tracking. Over the past few years, researchers have attempted various techniques to segment the brain to achieve the most accurate results, but this remains a challenging task.

[0007] Therefore, this invention proposes a new technical solution, the method of which can quickly and accurately segment the brain into different regions. Summary of the Invention

[0008] The purpose of this invention is to provide an automatic brain region segmentation method based on artificial intelligence.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] An AI-based automatic brain region segmentation method includes a model training phase and a model usage phase; wherein,

[0011] The model training phase includes:

[0012] Step 1: Prepare raw data

[0013] At least one dataset was used, which was a brain MRI scan image dataset containing brain MRI scan images of at least 50 subjects;

[0014] Step 2: Generate training data

[0015] ① Preprocessing the raw image data in the dataset, including:

[0016] Unify all image data in the dataset to the same coordinate system;

[0017] The image data after unifying the coordinate system is linearly registered with a standard template image data one by one;

[0018] In addition, brain regions are extracted from the image data;

[0019] ② Obtaining and preprocessing the labeled segmented image data, the steps include:

[0020] First, the preprocessed image data is nonlinearly registered with the standard template image data to obtain a deformation field;

[0021] Then, the deformation field is applied to the standard template segmentation image data to obtain the label segmentation image data corresponding to the preprocessed image data;

[0022] ③ Data resampling

[0023] The preprocessed image data and its corresponding label segmented image data are resampled to the same fixed size a×b×c to obtain training data, where a, b, and c are all multiples of 16;

[0024] Step 3: Model Training

[0025] The standard template segmented image data and the image data in the training data are simultaneously input into a deep neural network model. The neural network model outputs a currently predicted label segmented image data, which is then used together with the resampled label segmented image data in the training data to calculate the loss function for the model parameter iteration, resulting in a current training model. The above operation is repeated until the loss value stabilizes. An optimal training model is selected as the final model to realize the training of the deep neural network model.

[0026] The model usage phase includes:

[0027] ① Obtain a new brain MRI scan image, and perform the preprocessing on the new raw image data of the new brain MRI scan image;

[0028] ② Perform the resampling on the preprocessed new original image data;

[0029] ③ The new original image data after resampling and preprocessing is used as the data to be used. The data to be used and the standard template segmented image data are simultaneously input into the trained deep neural network model. The neural network model outputs a predicted label segmented image data.

[0030] The relevant content in the above technical solution is explained as follows:

[0031] 1. In the above scheme, the purpose of the preprocessing is to unify different data with different acquisition parameters.

[0032] 2. In the above scheme, the standard template image data and the standard template segmentation image data belong to a template space, which is the standard template image data space for brain MRI. The template space is publicly available prior art or data, which can be obtained and used by those skilled in the art. This invention uses the BRODMANN template.

[0033] 3. In the above scheme, in step one of the model training phase, the dataset includes the ABIDE dataset and the ADHD-200 dataset.

[0034] 4. In the above scheme, in step one of the model training stage, the image data in the dataset is divided into a training group, a validation group, and a test group in a ratio of 7:2:1.

[0035] 5. In the above scheme, after the image data of the dataset mentioned in step one is used to generate training data in step two, it is divided into training group, validation group and test group in a ratio of 7:2:1.

[0036] 6. In the above scheme, after the training group data calculates the loss function in step three, the backpropagation algorithm is used to iterate the model parameters; after the validation group and the test group calculate the loss function in step three, the loss value is used as an evaluation index to evaluate the training effect.

[0037] The verification group is evaluated during the training process, while the test group is evaluated once after training is completed. The specific differences in evaluation and the evaluation principles are existing technologies and are well known and understood by those skilled in the art.

[0038] 7. In the above scheme, in step three, the optimal training model is the model with the smallest loss value after being validated by the validation group data and tested by the test group data.

[0039] 8. In the above scheme, after step three of the model training phase, the trained deep neural network model is further validated. The validation process includes:

[0040] ① Perform the preprocessing on the original image data in the verification group;

[0041] ② The image data of the preprocessed verification group are resampled as described above;

[0042] ③ The resampled data is used as the data to be used. The data to be used and the standard template segmented image data are simultaneously input into the trained deep neural network model. The neural network model outputs a predicted label segmented image data for the verification stage.

[0043] The labeled segmented image data from this verification phase is then compared with the labeled segmented image data corresponding to the preprocessed image data described in step ② of step two, to evaluate the model training process. The specific evaluation principle is existing technology and is well known and understood by those skilled in the art.

[0044] 9. The above scheme also includes testing the trained deep neural network model. The testing process includes:

[0045] ① Perform the aforementioned preprocessing on the original image data in the test group;

[0046] ② Resample the image data of the preprocessed test group;

[0047] ③ The resampled data is used as the data to be used. The data to be used and the standard template segmented image data are simultaneously input into the trained deep neural network model. The neural network model outputs a predicted label segmented image data for the test phase.

[0048] The labeled segmented image data from this testing phase is then compared with the labeled segmented image data corresponding to the preprocessed image data described in step ② of step two, to evaluate the model training results. The specific evaluation principle is existing technology and is well known and understood by those skilled in the art.

[0049] 10. In the above scheme, step one of the model training phase also includes a second set of datasets, which are used as the retest set.

[0050] 11. In the above scheme, the second set of datasets includes the PPMI dataset.

[0051] 12. In the above scheme, in step two of the model training stage, the preprocessing in ①, the acquisition of the label segmentation image data in ②, and the resampling in ③ are all performed by using ITK, Slicer, or Freesurfer medical image processing software alone or in combination.

[0052] 13. In the above scheme, in step two of the model training phase, the preprocessing in ① includes:

[0053] First, unify all image data in the dataset to the same coordinate system;

[0054] Secondly, the image data after unifying the coordinate system is linearly registered with the standard template image data one by one;

[0055] Finally, brain regions were extracted from the linearly registered image data.

[0056] 14. In the above scheme, in step two of the model training phase, the preprocessing in ① includes:

[0057] First, the image data in the dataset are unified into the same coordinate system, and brain regions are extracted from each image data.

[0058] Secondly, the image data after unifying the coordinate system is linearly registered with the standard template image data one by one, and the registration transformation matrix is ​​obtained.

[0059] Finally, the transformation matrix after linear registration is applied to the image data of the extracted brain regions.

[0060] 15. In the above scheme, in step two of the model training phase, the preprocessing in ① includes:

[0061] First, unify all image data in the dataset to the same coordinate system;

[0062] Secondly, brain regions are extracted from each of the image data after unifying the coordinate system.

[0063] Finally, the extracted brain region image data are linearly registered with the standard template image data one by one.

[0064] 16. In the above scheme, in step three of the model training stage, the neural network model is a 3D image segmentation network, including but not limited to VNet3D, UNet3D, and corresponding variants.

[0065] 17. In the above scheme, in step three of the model training phase, a generalized dice evaluation is used to evaluate the multi-class segmentation. The generalized dice loss is as follows:

[0066]

[0067] Where, r ln Let p represent the standard value of category l in the nth voxel, while p ln ω represents the corresponding predicted value. l This represents the weight of each category.

[0068] The working principle and advantages of this invention are as follows:

[0069] This invention proposes an automatic brain region segmentation method based on artificial intelligence. By introducing prior features and using currently popular frameworks, it segments the brain and obtains multiple labeled segmented regions. Compared to traditional image segmentation methods, this invention is fast, accurate, and stable. Compared to other deep learning methods, this invention does not rely on manually labeled datasets, allowing for rapid augmentation of large amounts of data and improving the accuracy of deep learning models.

[0070] Verification using a test set derived from the same source as the training set and a test set from fresh samples shows that this invention can directly and accurately divide the whole brain into multiple brain regions, exhibiting advantages such as strong segmentation ability, high stability, and strong generalization ability in brain region division.

[0071] In summary, the advantages of this invention include: higher accuracy in brain region segmentation and recognition, more stable model; richer details in brain region segmentation results; and the ability to perform real-time brain region segmentation. Attached Figure Description

[0072] Appendix Figure 1 This is a flowchart illustrating the process of generating training data during the model training phase in an embodiment of the present invention.

[0073] Appendix Figure 2 This is a flowchart illustrating the model training process during the model training phase in an embodiment of the present invention.

[0074] Appendix Figure 3 This is a flowchart illustrating the process of validating, testing, or retesting the model after training, according to an embodiment of the present invention.

[0075] Appendix Figure 4 This is a flowchart illustrating the model usage phase in an embodiment of the present invention.

[0076] Appendix Figure 5 This is a schematic diagram of the grouping of the dataset in step one of the model training stage of this embodiment of the invention. Detailed Implementation

[0077] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0078] Example: The present invention will be clearly described below with illustrations and detailed description. Any person skilled in the art who understands the examples of the present invention can make changes and modifications based on the technology taught in the present invention without departing from the spirit and scope of the present invention.

[0079] The terminology used herein is for the purpose of describing specific embodiments only and is not intended to limit the scope of this case.

[0080] The terms “include,” “including,” and “have” used in this article are all open-ended, meaning they include but are not limited to.

[0081] Unless otherwise specified, the terms used herein generally have their ordinary meaning in the context of the art, the subject matter, and the specific context. Certain terms used to describe this case will be discussed below or elsewhere in this specification to provide additional guidance to those skilled in the art in describing this case.

[0082] An AI-based automatic brain region segmentation method includes a model training phase and a model usage phase; wherein,

[0083] The model training phase includes:

[0084] Step 1: Prepare raw data

[0085] At least one dataset is used, which may be a T1-weighted brain MRI scan image dataset containing T1-weighted brain MRI scan images of at least 50 subjects;

[0086] Step 2: Generate training data (see...) Figure 1 )

[0087] ① Preprocessing the raw image data in the dataset, including:

[0088] Unify all image data in the dataset into the same coordinate system, and set the origin of the coordinate system to zero; the coordinate system can be the RAI (Right, Anterior, Inferior) coordinate system;

[0089] The image data after unifying the coordinate system is linearly registered with a standard template image data one by one;

[0090] In addition, brain regions are extracted from the image data, that is, only the brain regions are retained in the MRI images, while the skull, eyes and other regions are removed;

[0091] ② Obtaining and preprocessing the labeled segmented image data, the steps include:

[0092] First, the preprocessed image data is nonlinearly registered with the standard template image data to obtain a deformation field;

[0093] Then, the deformation field is applied to the standard template segmentation image data to obtain the label segmentation image data corresponding to the preprocessed image data;

[0094] ③ Data resampling

[0095] The preprocessed image data and its corresponding label segmented image data are resampled to the same fixed size a×b×c to obtain training data, where a, b, and c are all multiples of 16;

[0096] Step two generates training data for each image in the dataset, ultimately forming a training dataset corresponding to the dataset.

[0097] Step 3: Model Training (see...) Figure 2 )

[0098] The standard template segmented image data and the image data in the training data are simultaneously input into a deep neural network model. The neural network model outputs a currently predicted label segmented image data, which is then used together with the resampled label segmented image data in the training data to calculate the loss function for model parameter iteration. The loss function is minimized through backpropagation algorithm, and the above operation is repeated until the loss value stabilizes. An optimal training model is selected as the final model to realize the training of the deep neural network model.

[0099] The model usage phase includes (see...) Figure 4 ):

[0100] ① Obtain a new brain MRI scan image from the current patient for whom automatic brain region segmentation is required, and perform the aforementioned preprocessing on the new raw image data of the new brain MRI scan image; (The specific preprocessing steps are the same as step ① in step two of the model training phase).

[0101] ② Resample the preprocessed new original image data; (same as step ③ in step two of the model training phase, i.e., resample to a fixed size a×b×c, where a, b, and c are all multiples of 16)

[0102] ③ The new original image data after resampling and preprocessing is used as the data to be used. The data to be used and the standard template segmented image data are simultaneously input into the trained deep neural network model. The neural network model outputs a predicted label segmented image data as the automatic segmentation result of the brain region of the current patient.

[0103] Preferably, in step one of the model training phase, the dataset includes the ABIDE (Autism Brain Imaging Data Exchange) dataset and the ADHD-200 (Attention-Deficit Hyperactivity Disorder) dataset.

[0104] Preferred, such as Figure 5 As shown, in step one of the model training phase, the image data in the dataset is divided into a training group, a validation group, and a test group in a ratio of 7:2:1.

[0105] Alternatively, the image data of the dataset described in step one can be used to generate training data in step two, and then divided into a training group, a validation group, and a test group in a ratio of 7:2:1.

[0106] Following step three of the model training phase, the process also includes validating the trained deep neural network model. The validation process includes (see...). Figure 3 ):

[0107] ① Perform the preprocessing on the original image data in the validation group; (same as step ① in step two of the model training phase)

[0108] ② Resample the image data of the preprocessed validation group; (same as step ③ in step two of the model training phase, i.e., resample to a fixed size a×b×c, where a, b, and c are all multiples of 16)

[0109] ③ The resampled data is used as the data to be used. The data to be used and the standard template segmented image data are simultaneously input into the trained deep neural network model. The neural network model outputs a predicted label segmented image data for the verification stage.

[0110] The labeled segmented image data of this verification stage is compared with the labeled segmented image data corresponding to the preprocessed image data described in step ② of step two, in order to evaluate the model training process.

[0111] Preferably, the process also includes testing the trained deep neural network model, the testing process including (see...). Figure 3 ):

[0112] ① Perform the aforementioned preprocessing on the original image data in the test group; (same as step ① of step two in the model training phase).

[0113] ② Resample the image data of the preprocessed test group; (same as step ③ in step two of the model training phase, i.e., resample to a fixed size a×b×c, where a, b, and c are all multiples of 16)

[0114] ③ The resampled data is used as the data to be used. The data to be used and the standard template segmented image data are simultaneously input into the trained deep neural network model. The neural network model outputs a predicted label segmented image data for the test phase.

[0115] The labeled segmented image data from this test phase is then compared with the labeled segmented image data corresponding to the preprocessed image data described in step ② of step two, in order to evaluate the model training results.

[0116] Preferably, in step one of the model training phase, a second dataset is also included, and this second dataset is used as a retest group to further test the reliability of the trained model (see...). Figure 3 The retest group follows the same testing procedure as the test group and serves as a supplement to it, used to evaluate the generalization ability of the trained model. The second dataset includes the PPMI (Parkinson Progression Marker Initiative) dataset.

[0117] This invention uses ABIDE and ADHD-200 for model training, validation, and testing, and PPMI as fresh samples (distinct from the training data) to test the model's generalization ability. ABIDE, ADHD-200, and PPMI are all publicly available datasets.

[0118] Preferably, in step two of the model training phase, the preprocessing in ①, the acquisition of the labeled segmented image data in ②, and the resampling in ③ are all performed by using ITK, Slicer, or Freesurfer medical image processing software alone or in combination. The above three software programs are for illustrative purposes only and are not intended to limit the scope of the application.

[0119] Preferably, in step two of the model training phase, the preprocessing in ① includes:

[0120] First, unify all image data in the dataset to the same coordinate system;

[0121] Secondly, the image data after unifying the coordinate system is linearly registered with the standard template image data one by one;

[0122] Finally, brain regions were extracted from the linearly registered image data.

[0123] The second option, the preprocessing in ①, includes:

[0124] First, the image data in the dataset are unified into the same coordinate system, and brain regions are extracted from each image data.

[0125] Secondly, the image data after unifying the coordinate system is linearly registered with the standard template image data one by one, and the registration transformation matrix is ​​obtained.

[0126] Finally, the transformation matrix after linear registration is applied to the image data of the extracted brain regions.

[0127] Alternatively, the preprocessing in ① includes:

[0128] First, unify all image data in the dataset to the same coordinate system;

[0129] Secondly, brain regions are extracted from each of the image data after unifying the coordinate system.

[0130] Finally, the extracted brain region image data are linearly registered with the standard template image data one by one.

[0131] In step three of the model training phase, the neural network model is a 3D image segmentation network, including but not limited to VNet3D or UNet3D or other segmentation neural networks and their improved types, with VNet3D model being preferred.

[0132] Preferably, in step three of the model training phase, the generalized dice score (GDS) is used to evaluate multi-class segmentation. This integrates the dice losses from multiple classes, allowing for a single metric to evaluate multi-class segmentation. The generalized dice loss (GDL) is as follows:

[0133]

[0134] Where, r ln Let p represent the standard value of category l on the nth voxel (reference image), and p ln ω represents the corresponding predicted value. l This represents the weight of each category. This design aims to further increase the accuracy of training.

[0135] The technical effects of the present invention are further illustrated below with diagrams:

[0136] This invention uses T1-weighted images of 175 subjects from three databases: ABIDE, ADHD-200, and PPMI, for model training, validation, and testing. 160 subjects were from the ABIDE and ADHD-200 datasets, 10 from the ABIDE and ADHD-200 datasets, and 5 from the PPMI dataset.

[0137] To verify the effectiveness of this invention, it is compared with the following commonly used image segmentation frameworks:

[0138] Using UNet3D, VNet3D, and VoxelMorph-2 as examples, we obtain the segmentation results under different methods. The segmentation results are evaluated using the Dice coefficient, which ranges from 0 to 1. The closer the value is to 1, the better the segmentation performance.

[0139] Table 1 shows the maximum, minimum, average, and standard deviation of the four methods—UNet3D, VNet3D, VoxelMorph-2, and the method of this invention—on the ABIDE and ADHD-200 datasets.

[0140] The following results can be obtained from Table 1:

[0141] First, UNet3D's maximum DICE value can reach 0.85, which is higher than V-net, but its minimum value is the lowest, resulting in the worst stability.

[0142] Secondly, compared with the commonly used medical image segmentation models UNet3D and VNet3D, the VoxelMorph framework has better dice scores in all aspects than these two methods, which also shows the application potential of VoxelMorph in the medical image segmentation of brain regions.

[0143] The performance of the present invention is slightly better than that of VoxelMorph-2. Especially, the average dice score has increased by at least 2%, showing a relatively high advantage.

[0144] The standard deviation values from high to low are as follows: the present invention < VoxelMorph-2 = V-net < Unet3d. From the above results, it can be seen that compared with other methods, the present invention has better segmentation ability and higher stability.

[0145] Table 1 Dice values obtained under different methods on the ABIDE and ADHD-200 test sets

[0146] method Maximum Dice value Minimum Dice value Average Dice value Unet3d 0.85 0.50 0.67±0.115 V-net 0.83 0.67 0.78±0.051 VoxelMorph-2 0.86 0.69 0.80±0.051 This invention 0.87 0.73 0.82±0.034

[0147] To verify the generalization ability of the proposed model, the models trained from the ABIDE and ADHD-200 datasets are applied to fresh subjects. The fresh samples selected here are the PPMI dataset. Table 2 shows the maximum dice value, minimum dice value, average dice value, and standard deviation of Unet3d, V-net, VoxelMorph-2, and the present invention on the PPMI dataset.

[0148] Table 2 Dice values obtained under different methods on the PPMI test set

[0149] method Maximum Dice value Minimum Dice value Average Dice value Unet3d 0.84 0.61 0.70±0.079 V-net 0.78 0.70 0.74±0.024 VoxelMorph-2 0.82 0.73 0.76±0.021 This invention 0.83 0.76 0.78±0.025

[0150] It can be seen from Table 2 that although the maximum dice value of Unet3d is the highest among the four methods, its average dice value and stability are still the lowest and the worst. The VoxelMorph architecture is still better than Unet3d and V-net. The average dice values of the four brain region segmentation methods from high to low are as follows: the present invention > VoxelMorph-2 > V-net > Unet3d. The above results show that the present invention not only has high stability and strong segmentation ability, but also has high generalization ability.

[0151] The above embodiments are only used to illustrate the technical concept and characteristics of the present invention, and the purpose is to enable those who are familiar with this technology to understand the content of the present invention and implement it accordingly, and it cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be covered within the protection scope of the present invention.

Claims

1. An automatic brain region segmentation method based on artificial intelligence, characterized in that: This includes the model training phase and the model usage phase; among which, The model training phase includes: Step 1: Prepare raw data At least one dataset was used, which was a brain MRI scan image dataset containing brain MRI scan images of at least 50 subjects; Step 2: Generate training data ① Preprocessing the raw image data in the dataset, including: Unify all image data in the dataset to the same coordinate system; The image data after unifying the coordinate system is linearly registered with a standard template image data one by one; In addition, brain regions are extracted from the image data; ② Obtaining and preprocessing the labeled segmented image data, the steps include: First, the preprocessed image data is nonlinearly registered with the standard template image data to obtain a deformation field; Then, the deformation field is applied to the standard template segmentation image data to obtain the label segmentation image data corresponding to the preprocessed image data; ③ Data resampling The preprocessed image data and its corresponding label segmented image data are resampled to the same fixed size a×b×c to obtain training data, where a, b, and c are all multiples of 16; Step 3: Model Training The standard template segmented image data and the image data in the training data are simultaneously input into a deep neural network model. The neural network model outputs a currently predicted label segmented image data, which is then used together with the resampled label segmented image data in the training data to calculate the loss function for the model parameter iteration, resulting in a current training model. The above operation is repeated until the loss value stabilizes. An optimal training model is selected as the final model to realize the training of the deep neural network model. The model usage phase includes: ① Obtain a new brain MRI scan image, and perform the preprocessing on the new raw image data of the new brain MRI scan image; ② Perform the resampling on the preprocessed new original image data; ③ The new original image data after resampling and preprocessing is used as the data to be used. The data to be used and the standard template segmented image data are simultaneously input into the trained deep neural network model. The neural network model outputs a predicted label segmented image data. In step one of the model training phase, the dataset includes the ABIDE dataset and the ADHD-200 dataset; In step three of the model training phase, the neural network model is a 3D image segmentation network, including but not limited to VNet3D and UNet3D. In step three of the model training phase, the generalized dice loss is used to evaluate multi-class segmentation. The generalized dice loss is as follows: ； in, This represents the standard value of category l in the nth voxel, while This represents the corresponding predicted value. This represents the weight of each category.

2. The automatic brain region segmentation method according to claim 1, characterized in that: In step one of the model training phase, the image data in the dataset is divided into a training group, a validation group, and a test group in a ratio of 7:2:

1.

3. The automatic brain region segmentation method according to claim 1, characterized in that: After generating training data from the image data of the dataset described in step one in step two, the dataset is divided into a training group, a validation group, and a test group in a ratio of 7:2:

1.

4. The automatic brain region segmentation method according to claim 2 or 3, characterized in that: After the training group data is processed according to the loss function in step three, the model parameters are iterated using the backpropagation algorithm; after the validation group and the test group are processed according to the loss function in step three, the loss value is used as an evaluation index to assess the training effect.

5. The automatic brain region segmentation method according to claim 2 or 3, Its features are: Following step three of the model training phase, the process also includes validating the trained deep neural network model. The validation process includes: ① Perform the preprocessing on the original image data in the verification group; ② The image data of the preprocessed verification group are resampled as described above; ③ The resampled data is used as the data to be used. The data to be used and the standard template segmented image data are simultaneously input into the trained deep neural network model. The neural network model outputs a predicted label segmented image data for the verification stage. The labeled segmented image data of this verification stage is compared with the labeled segmented image data corresponding to the preprocessed image data described in step ② of step two, in order to evaluate the model training process.

6. The automatic brain region segmentation method according to claim 2 or 3, characterized in that: It also includes testing the trained deep neural network model, the testing process of which includes: ① Perform the aforementioned preprocessing on the original image data in the test group; ② Resample the image data of the preprocessed test group; ③ The resampled data is used as the data to be used. The data to be used and the standard template segmented image data are simultaneously input into the trained deep neural network model. The neural network model outputs a predicted label segmented image data for the test phase. The labeled segmented image data from this test phase is then compared with the labeled segmented image data corresponding to the preprocessed image data described in step ② of step two, in order to evaluate the model training results.

7. The automatic brain region segmentation method according to claim 1, characterized in that: Step one of the model training phase also includes a second dataset, which is used as the retest set.

8. The automatic brain region segmentation method according to claim 7, characterized in that: The second set of datasets includes the PPMI dataset.

9. The automatic brain region segmentation method according to claim 1, characterized in that: In step two of the model training phase, the preprocessing in ①, the acquisition of the labeled segmented image data in ②, and the resampling in ③ are all performed by using ITK, Slicer, or Freesurfer medical image processing software individually or in combination.

10. The automatic brain region segmentation method according to claim 1, characterized in that: In step two of the model training phase, the preprocessing mentioned in ① includes: First, unify all image data in the dataset to the same coordinate system; Secondly, the image data after unifying the coordinate system is linearly registered with the standard template image data one by one; Finally, brain regions were extracted from the linearly registered image data.

11. The automatic brain region segmentation method according to claim 1, characterized in that: In step two of the model training phase, the preprocessing mentioned in ① includes: First, the image data in the dataset are unified into the same coordinate system, and brain regions are extracted from each image data. Secondly, the image data after unifying the coordinate system is linearly registered with the standard template image data one by one, and the registration transformation matrix is ​​obtained. Finally, the transformation matrix after linear registration is applied to the image data of the extracted brain regions.

12. The automatic brain region segmentation method according to claim 1, characterized in that: In step two of the model training phase, the preprocessing mentioned in ① includes: First, unify all image data in the dataset to the same coordinate system; Secondly, brain regions are extracted from each of the image data after unifying the coordinate system. Finally, the extracted brain region image data are linearly registered with the standard template image data one by one.

Images