Brain region segmentation method and brain region segmentation system for CT brain image

Abstract

The brain region segmentation method of the CT brain image comprises the following steps: training a modality conversion model, the modality conversion model is used to convert the CT brain image obtained under the CT modality into a simulated MR brain image under the MR modality, the trained parameters of the modality conversion model include structural prior information, and the structural prior information is the structural difference between each CT brain image in the data set and the real MR brain image paired with it; after the modality conversion model is trained, a CT brain image x CT is converted into a simulated MR brain image x CT→MR by the modality conversion model; and a brain region segmentation atlas is obtained by segmenting the simulated MR brain image x CT→MR . The brain region segmentation method of the CT brain image can accurately realize the whole brain segmentation of the CT brain image. A brain region segmentation system of a CT brain image is also provided.

Description

Technical Field

[0001] This invention relates to a method for brain region segmentation of CT cranial images and a brain region segmentation system capable of performing the above-mentioned method for brain region segmentation of CT cranial images. Background Technology

[0002] The human brain can be divided into 13 regions: cerebellum, thalamus, basal ganglia, ventricles, third ventricle, fourth ventricle, brainstem, temporal lobe, frontal lobe, parietal lobe, occipital lobe, and cingulate gyrus. Currently, clinical diagnosis of conditions such as traumatic brain injury often requires the analysis and processing of cranial images. Commonly used cranial image processing methods include computed tomography (CT) and magnetic resonance imaging (MR). Due to limitations in its imaging principles, CT images present information about tissue density. Because the densities of different brain regions are very similar, and noise is introduced due to radiation dose, masking contrast differences between some brain regions, CT images often cannot obtain structural information for all brain regions. This limitation makes the localization of brain injuries somewhat difficult.

[0003] Existing CT-based cranial image analysis software, such as BrainlabiPlan, can assist in locating lesions; however, because CT images cannot provide complete information about brain structure, the software cannot support lesion localization. Traditional cranial image processing methods typically require MR images to obtain complete information about brain structure for tasks such as whole-brain segmentation. However, due to the differences between CT and MR imaging technologies, CT images cannot capture the structural information found in MR images, making whole-brain segmentation from CT images difficult to achieve. Furthermore, in emergency situations, it is often impossible to obtain both CT and MR images simultaneously. Summary of the Invention

[0004] The purpose of this invention is to provide a brain region segmentation method for CT cranial images, which can accurately achieve whole-brain segmentation of CT cranial images.

[0005] Another objective of this invention is to provide a brain region segmentation system that can accurately segment the whole brain from CT images of the brain.

[0006] This invention provides a method for brain region segmentation of CT brain images, comprising the following steps: training a modality transformation model, which is used to convert CT brain images obtained in CT modality into simulated MR brain images in MR modality. The parameters of the trained modality transformation model include structural prior information, which is the structural difference between each CT brain image in the dataset and its paired real MR brain image; after the modality transformation model is trained, a CT brain image is segmented into a simulated MR brain image using the modality transformation model. CT Converted into a simulated MR brain image x CT→MR ; and segmentation of simulated MR brain images x CT→MR A brain region segmentation map was obtained.

[0007] This invention converts CT brain images into simulated MR brain images by training the parameters of a modality transformation model that includes prior structural information, and obtains brain region segmentation maps through segmentation, thereby accurately achieving whole-brain segmentation of CT brain images.

[0008] In another illustrative embodiment of the brain region segmentation method for CT cranial images, in the step of training a modality transformation model, the modality transformation model can extract CT modality features and structural features from CT cranial images, convert the CT modality features into MR modality features, combine the structural features with structural prior information to obtain training structural prior information, and then fuse the MR modality features with the training structural prior information to obtain a simulated MR cranial image. By converting CT modality features into MR modality features and combining structural features with structural prior information, a simulated MR cranial image that closely resembles a real MR cranial image can be obtained.

[0009] In another illustrative implementation of the brain region segmentation method for CT cranial images, the step of training a modality transformation model uses an adversarial training module. The step of training a modality transformation model includes: randomly selecting a CT cranial image sample x from a dataset. i CT And input the modality transformation model, where 0 < i ≤ N, and N represents the number of CT cranial image samples in the dataset; from CT cranial image sample x i CT CT modal features and structural features are extracted. CT modal features represent the outline of the brain that can be resolved by CT equipment, while structural features represent the unique structural outline of the brain displayed by CT equipment. The CT modal features are then converted into MR modal features. The structural features are combined with prior structural information to obtain training structural features. Finally, the MR modal features and training structural features are fused to generate a simulated MR brain image. i CT→MR With the parameters of the fixed adversarial training module remaining unchanged, simulated MR cranial images x iCT→MR The input is to the adversarial training module, and the output of the adversarial training module is a D(x) i CT →MR The modality transition model optimizes its parameters, where D(x) is the value of the modality transition model. i CT→MR The value is between 0 and 1, used to determine whether the input image is a true MR image. i CT→MR The closer the value of D(x) is to 0, the closer the input image is to the distribution of the real MR image. i CT →MR The closer the value is to 1, the closer the input image is to the distribution of non-real MR images; a CT brain image sample x is randomly selected from the dataset. i+1 CT Using the optimized parameters of the modality transfer model, the CT cranial image samples x i+1 CT Converted into a simulated MR brain image x i+1 CT→MR With the parameters of the fixed modality transformation model remaining unchanged, simulated MR cranial images x i+1 CT→MR The input is to the adversarial training module, and the output of the adversarial training module is a D(x) i+1 CT→MR The parameters of the adversarial training module are then optimized, and the above steps are repeated until the optimization termination condition is met, thus completing the training of the modality conversion model. By converting CT modal features to MR modal features, combining structural features and prior structural information, and employing adversarial training on the modality conversion model, simulated MR cranial images that closely resemble real MR cranial images can be obtained.

[0010] In another illustrative implementation of the brain region segmentation method based on CT cranial images, the modality transfer model is optimized by minimizing the loss function in equation (1):

[0011]

[0012] Where N represents the number of CT cranial image samples in the dataset, 0 < i ≤ N; D(x i CT→MR The value (D(x)) is a value ranging from 0 to 1 output by the adversarial training module, used to determine whether the input image is a real MR image. i CT→MR The closer the value of D(x) is to 0, the closer the input image is to the distribution of the real MR image. i CT→MRThe closer the value is to 1, the closer the input image is to the distribution of a non-real MR image. This makes the generated simulated MR brain image closer to a real MR brain image.

[0013] In another illustrative implementation of the brain region segmentation method based on CT cranial images, the modality transformation model is optimized by minimizing the loss function in equation (2):

[0014]

[0015] Where N represents the number of CT cranial image samples in the dataset, 0 < i ≤ N; M represents the total number of pixels in a CT cranial image sample, 0 < j ≤ M. It is the j-th gray value of the CT brain image of the i-th brain image sample. It is the j-th grayscale value of the simulated MR brain image obtained by converting the i-th CT brain image sample; D(x i CT→MR The value (D(x)) is a value ranging from 0 to 1 output by the adversarial training module, used to determine whether the input image is a real MR image. i CT →MR The closer the value of D(x) is to 0, the closer the input image is to the distribution of the real MR image. i CT→MR The closer the value is to 1, the closer the input image is to the distribution of a non-real MR image. This ensures that the generated simulated MR brain image is close to a real MR brain image while maintaining consistency with CT brain images.

[0016] In another illustrative implementation of the brain region segmentation method based on CT cranial images, the parameters of the adversarial training module are optimized by minimizing the loss function in equation (3):

[0017]

[0018] Where N represents the number of CT cranial image samples in the dataset, 0 < i ≤ N; D(x i+1 CT→MR ) value and The values ​​are all output by the adversarial training module, ranging from 0 to 1, and are used to determine whether the input image is a real MR image. D(x) i+1 CT→MR ) value and The smaller the value, the closer the input image is to the distribution of the real MR image, D(x) i+1 CT→MR ) value and The closer the value is to 1, the closer the input image is to the distribution of a non-real MR image.

[0019] In another illustrative implementation of the brain region segmentation method for CT cranial images, after the modality transformation model is trained, a CT cranial image x... CT Converted into a simulated MR brain image x CT→MR The steps include: x CT The input is fed into a modality transformation model to extract CT modal features and structural features; the CT modal features are replaced with general MR modal features, where the general MR modal features are the average of the MR modal features obtained from all MR images in the inference dataset after the modality transformation model is trained; the structural features are combined with general structural prior information to obtain applied structural features, where the general structural prior information is the structural prior information obtained after the modality transformation model is trained; and the general MR modal features and applied structural features are fused to generate a simulated MR cranial image x. CT→MR This allows the trained modality conversion model to convert CT brain images into simulated MR brain images that closely resemble real MR brain images.

[0020] In another illustrative embodiment of the brain region segmentation method for CT cranial images, a simulated MR cranial image x is segmented. CT→MR The steps to obtain a brain region segmentation atlas involve using a trained segmentation model to segment simulated MR cranial images. CT→MR The training process of the segmentation model includes: extracting a real MR brain image from the dataset. The input is given to the segmentation model, where N represents the number of real MR cranial image samples in the dataset, 0 < i ≤ N; the parameters of the segmentation model are optimized; and the above steps are repeated until the optimization termination condition is met. This allows the generated simulated MR cranial images to be segmented.

[0021] In another illustrative implementation of the brain region segmentation method based on CT cranial images, in the step of optimizing the parameters of the segmentation model, the segmentation model is optimized by minimizing the loss function in equation (4):

[0022]

[0023] Where N represents the number of real MR cranial image samples in the dataset; C represents the number of brain region categories; M represents the total number of pixels in a real MR cranial image sample; y i,j,k This represents the true label of the i-th real MR cranial image sample in the k-th category at the j-th pixel position, with a value of 0 or 1. i,j,k =0 indicates that the j-th pixel of the i-th real MR cranial image sample does not belong to category k, y i,j,k =1 indicates that the j-th pixel of the i-th real MR cranial image sample belongs to the k-th category; This indicates that the segmentation model is applicable to real MR cranial images x MR The brain region segmentation results, with values ​​ranging from 0 to 1, The closer a value is to 0, the less likely the j-th pixel of the i-th real MR cranial image sample belongs to the k-th category. The closer a value is to 1, the closer the j-th pixel of the i-th real MR cranial image sample is to the k-th category. This allows the segmentation model to be trained to segment real MR cranial images, facilitating its application to the segmentation of simulated MR cranial images.

[0024] In another illustrative embodiment of the brain region segmentation method based on CT cranial images, the method further includes the step of: overlaying the brain region segmentation atlas with the CT cranial image x. CT Obtain CT images of the brain after brain region segmentation. This facilitates convenient and accurate localization of lesions from CT images of the brain.

[0025] This invention also provides a brain region segmentation system capable of performing the aforementioned method for segmenting brain regions from CT cranial images, comprising: a modality conversion module, an adversarial training module, and a segmentation module. The modality conversion module is capable of segmenting CT cranial images using a modality conversion model. CT Converted to simulated MR brain images x CT→MR The adversarial training module challenges the modality transfer model by determining whether the input image is a real MR image, thereby causing the model to generate simulated MR brain images that approximate real MR brain images. CT→MR The segmentation module can segment the generated simulated MR brain images using a segmentation model. CT→MR The brain is segmented into several regions and a brain region segmentation atlas is generated. The above brain region segmentation system can accurately achieve whole-brain segmentation from CT cranial images. Attached Figure Description

[0026] The following figures are for illustrative and explanatory purposes only and do not limit the scope of the invention, wherein:

[0027] Figure 1 This is a flowchart illustrating one implementation of a brain region segmentation method in CT cranial images;

[0028] Figure 2 This is a partial flowchart illustrating one implementation of a brain region segmentation method for CT cranial images;

[0029] Figure 3 This is a partial flowchart illustrating one implementation of a brain region segmentation method for CT cranial images;

[0030] Figure 4 This is a partial flowchart illustrating one implementation of a brain region segmentation method for CT cranial images;

[0031] Figure 5 It is a set of CT brain images and an illustrative implementation of a brain region segmentation method based on the CT brain images, resulting in corresponding simulated MR brain images and brain region segmentation atlases. Detailed Implementation

[0032] To provide a clearer understanding of the technical features, objectives, and effects of the invention, specific embodiments of the invention are now described with reference to the accompanying drawings, in which the same reference numerals denote the same parts.

[0033] In this document, “illustrative” means “serving as an example, illustration or description”, and any illustration or implementation described herein as “illustrative” should not be construed as a more preferred or advantageous technical solution.

[0034] Figure 1 This is a flowchart illustrating one implementation method of brain region segmentation in CT cranial images. Figure 5 It is a set of CT brain images and an illustrative implementation of a brain region segmentation method based on the CT brain images, resulting in corresponding simulated MR brain images and brain region segmentation atlases.

[0035] See Figure 1 The brain region segmentation method for CT cranial images includes steps S10 to S30:

[0036] S10: Train a modality conversion model to convert CT brain images obtained in CT modality into simulated MR brain images in MR modality. The parameters of the trained modality conversion model include structural prior information, which is the structural difference between each CT brain image in the dataset and its paired real MR brain image.

[0037] S20: After the modality conversion model is trained, it converts a CT brain image into a simulated MR brain image. (See CT brain image for details.) Figure 5 The figure shown in column (a) is a simulated MR brain image. Figure 5 The figure shown in column (b) is as follows.

[0038] S30: Segmenting simulated MR cranial images to obtain a brain region segmentation atlas. See [link / reference] Figure 5 The figure shown in column (c) is as follows.

[0039] This invention converts CT brain images into simulated MR brain images by training the parameters of a modality transformation model that includes prior structural information. It then obtains brain region segmentation atlases through segmentation, accurately achieving whole-brain segmentation from CT brain images, and obtaining, for example... Figure 5The brain region segmentation map shown in column (c) is divided into multiple brain regions. Each brain region can be distinguished by different colors for easy identification.

[0040] In an illustrative implementation, during the training of a modality transformation model, the model extracts CT modality features and structural features from CT brain images, converts the CT modality features into MR modality features, combines the structural features with prior structural information to obtain training structural prior information, and then fuses the MR modality features with the training structural prior information to obtain a simulated MR brain image. By converting CT modality features into MR modality features and combining them with structural features and prior structural information, a simulated MR brain image that closely resembles a real MR brain image can be obtained.

[0041] In an illustrative embodiment, the brain region segmentation method for CT cranial images may further include step S40: superimposing the brain region segmentation atlas with the original CT cranial image to obtain a brain region segmented CT cranial image, that is... Figure 5 The graph shown in column (c) is overlaid on... Figure 5 The image shown in column (a) makes it easy for doctors to visually locate the site of injury from CT brain images.

[0042] Figure 2 This is a partial flowchart illustrating one embodiment of a brain region segmentation method based on CT cranial images. In this illustrative embodiment, step S10 uses an adversarial training module for training. Step S10 includes steps S11 to S19 below, see [link to relevant documentation]. Figure 2 :

[0043] S11: Randomly select a CT cranial image sample x from a dataset. i CT Input the modality transformation model, where 0 < i ≤ N, and N represents the number of CT cranial image samples in the dataset.

[0044] Typically, a dataset stores multiple CT cranial image samples and multiple real MR cranial image samples. Each CT cranial image sample and each real MR cranial image sample can be taken from different people or from different parts of the brain. The more samples there are, the easier it is to obtain simulated MR cranial images that are closer to real MR cranial images.

[0045] S12: From CT cranial image sample x i CT The CT modal features and structural features are extracted. The CT modal features are the outlines of the brain that can be distinguished by the CT equipment, and the structural features are the unique structural outlines of the brain displayed by the CT equipment.

[0046] CT modal features and structural features can be found in [reference]. Figure 5 . Figure 5 Column (a) represents CT images of the brain. CT modal features are the outlines of the brain that can be distinguished by CT equipment, that is, the outlines of tissues that can be observed in CT images of the brain, such as the white outline of the skull and the outline of the internal gray tissue. Structural features are the structural outlines specific to the brain displayed by CT equipment, that is, the structures specific to the brain that can be observed in CT images of the brain, such as the white outline of the skull.

[0047] S13: Convert CT modal features to MR modal features.

[0048] Because CT imaging distinguishes different tissues based on their density, and since all brain regions are soft tissues with very similar densities, CT brain images suffer from poor contrast and the inability to capture the boundaries of certain tissues, or to show partial or complete tissue loss. MR imaging, on the other hand, relies on the relationship between nuclear magnetic resonance (NMR) signal intensity and the hydrogen nucleus density in the sample. Different tissues in the human body have different water content ratios, meaning they contain different numbers of hydrogen nuclei, resulting in differences in NMR signal intensity. This difference is used as a feature to distinguish different tissues. MR modal features refer to the outline of the brain that MR equipment can resolve. Therefore, compared to CT modal features, MR modal features contain richer information. See [link to CT modal features]. Figure 5 , Figure 5 Column (b) is a synthetic simulated MR brain image. Figure 5 Column (b) compared to Figure 5 Column (a) shows better contrast when the overall outline and the internal tissue outline are roughly the same. Therefore, to address the poor imaging contrast in CT cranial images, CT modal features can be converted into MR modal features through a series of nested mapping relationships.

[0049] S14: Combine structural features with prior structural information to obtain training structural features.

[0050] To address the limitations of CT brain images, which may contain the boundaries of certain tissues or lack partial or complete tissue coverage, this problem is overcome by combining structural features with prior structural information. The prior structural information consists of a set of parameters representing the structural differences between each CT brain image and its paired real MR brain image in the dataset. The modality transformation model automatically compares each input CT brain image with its paired real MR brain image, calculating the differences between the two images in the brain-specific structural contours—that is, the differences in structural contours present in the real MR brain image but not in the CT image. These parameters representing the differences are then incorporated into the extracted structural features, enriching their information and thus yielding the trained structural features.

[0051] S15: Fuse MR modal features and training structural features to generate a simulated MR brain image.i CT→MR With the parameters of the fixed adversarial training module remaining unchanged, simulated MR cranial images x i CT→MR The input is to the adversarial training module, and the output of the adversarial training module is a D(x) i CT→MR The modality transition model optimizes its parameters, where D(x) is the value of the modality transition model. i CT→MR The value is between 0 and 1, used to determine whether the input image is a true MR image. i CT→MR The closer the value of D(x) is to 0, the closer the input image is to the distribution of the real MR image. i CT→MR The closer the value is to 1, the closer the input image is to the distribution of a non-real MR image.

[0052] In one illustrative implementation, the mode transition model is optimized by minimizing the loss function in equation (1):

[0053]

[0054] in,

[0055] N represents the number of CT cranial images in the dataset, 0 < i ≤ N;

[0056] D(x i CT→MR The value (D(x)) is a value ranging from 0 to 1 output by the adversarial training module, used to determine whether the input image is a real MR image. i CT→MR The closer the value of D(x) is to 0, the closer the input image is to the distribution of the real MR image. i CT→MR The closer the value is to 1, the closer the input image is to the distribution of a non-real MR image.

[0057] Through adversarial training using the adversarial training module, the generated simulated MR brain images can closely resemble real MR brain images.

[0058] In another illustrative implementation, the mode transition model is optimized by minimizing the loss function in equation (2):

[0059]

[0060] in,

[0061] N represents the number of CT cranial images in the dataset, 0 < i ≤ N;

[0062] M represents the total number of pixels in a CT brain image sample, 0 < j ≤ M;

[0063] It is the j-th gray value of the CT brain image of the i-th brain image sample. It is the j-th gray value of the simulated MR brain image obtained by converting the i-th CT brain image sample;

[0064] D(x i CT→MR The value (D(x)) is a value ranging from 0 to 1 output by the adversarial training module, used to determine whether the input image is a real MR image. i CT→MR The closer the value of D(x) is to 0, the closer the input image is to the distribution of the real MR image. i CT→MR The closer the value is to 1, the closer the input image is to the distribution of a non-real MR image.

[0065] Through adversarial training using the adversarial training module, the generated simulated MR brain images can closely approximate real MR brain images, while constraints... It enables the generated simulated MR brain images to maintain consistency with CT brain images in terms of overall outline and internal tissue outline.

[0066] S16: Randomly select another CT cranial image sample x from the dataset. i+1 CT .

[0067] S17: Using the optimized parameters of the modality transfer model, transform the CT cranial image samples x i+1 CT Converted into a simulated MR brain image x i+1 CT→MR .

[0068] Repeat steps S12 to S15, and transfer the CT brain image sample x i+1 CT Converted to simulated MR brain images x i+1 CT→MR .

[0069] S18: With the parameters of the fixed modality transformation model unchanged, simulate MR cranial images x i+1 CT→MR The input is fed into the adversarial training module, and the adversarial training module outputs a D(x) i+1 CT→MR The values ​​are then optimized and the parameters of the adversarial training module are improved, where D(x) is the optimal value. i+1 CT→MR ) value and D(x i CT→MRThe meanings of the values ​​are the same.

[0070] In one illustrative implementation, the parameters of the adversarial training module are optimized by minimizing the loss function in equation (3):

[0071]

[0072] in,

[0073] N represents the number of CT cranial images in the dataset, 0 < i ≤ N;

[0074] D(x i+1 CT→MR ) value and The values ​​are all output by the adversarial training module, ranging from 0 to 1, and are used to determine whether the input image is a real MR image. D(x) i+1 CT→MR ) value and The closer the value is to 0, the closer the input image is to the distribution of the real MR image. i+1 CT→MR ) value and The closer the value is to 1, the closer the input image is to the distribution of a non-real MR image.

[0075] During the parameter optimization process of the adversarial training module, the input images include simulated MR cranial images. i+1 CT→MR and real MR brain images The value is a range of 0 to 1, which is the output value of the adversarial training module to the input real MR brain images. Its purpose is to train the adversarial training module to recognize real MR brain images and simulated MR brain images.

[0076] S19: Repeat the above steps until the optimization termination condition is met, and complete the training of the mode transition model.

[0077] Specifically, repeat steps S11 to S18 until the optimization termination condition is met. t This represents the value of the loss function L after t (t is a natural number) optimizations. t Satisfy L t <L t+s When s = 1, 2, 3...1000, the optimization termination condition is met, the model parameters at time t are saved, and the training of the mode transition model is completed.

[0078] By converting CT modal features into MR modal features, combining structural features with prior structural information, and employing adversarial training, it is possible to obtain simulated MR cranial images that closely resemble real MR cranial images.

[0079] Figure 3This is a partial flowchart illustrating one embodiment of a brain region segmentation method for CT cranial images. In one illustrative embodiment, step S20 includes steps S21 to S24 as follows, see [link to relevant documentation]. Figure 3 :

[0080] S21: x a CT brain image CT Input to the modality conversion model to extract CT modal features and structural features.

[0081] S22: Replace the CT modal features with general MR modal features, where the general MR modal features are the average value of the MR modal features obtained from all MR images in the inference dataset after the modality conversion model has been trained.

[0082] Once the modality transfer model is trained, it uses the trained parameters to infer the real MR brain images in the dataset, resulting in a set of MR modality features. The average value of all values ​​in the set is the general MR modality feature, which is a set of constants.

[0083] S23: Combine structural features with general structural prior information to obtain applied structural features, where the general structural prior information is the structural prior information obtained after the modality transformation model is trained, and it is a set of constants.

[0084] S24: Integrate general MR modal features and applied structural features to generate a simulated MR cranial image. CT→MR .

[0085] This method uses a trained modality conversion model to convert CT brain images into simulated MR brain images that closely resemble real MR brain images. The operation is convenient, quick, and highly accurate.

[0086] Figure 4 This is a partial flowchart illustrating one embodiment of a brain region segmentation method for CT cranial images. In one illustrative embodiment, step S30 uses a trained segmentation model to segment a simulated MR cranial image. The training process of the segmentation model includes steps S301 to S303, see [link to relevant documentation]. Figure 4 :

[0087] S301: Extract a real MR brain image from the dataset. Input to the segmentation model, where N represents the number of real MR cranial image samples in the dataset, 0 < i ≤ N;

[0088] S302: Optimize the parameters of the segmentation model.

[0089] S303: Repeat the above steps until the optimization termination condition is met.

[0090] Specifically, repeat steps S301 and S302, L t This represents the value of the loss function L after t (t is a natural number) optimizations. t Satisfy L t <L t+s When s = 1, 2, 3...1000, the optimization termination condition is met, the model parameters at time t are saved, and the training process of the segmentation model is completed. This enables the segmentation of the synthesized simulated MR brain images. The training of the segmentation model can be completed before, after, or after step S10.

[0091] In one illustrative implementation, the segmentation model in step S30 is optimized by minimizing the loss function in equation (4):

[0092]

[0093] Where N represents the number of real MR cranial image samples in the dataset; C represents the number of brain region categories; M represents the total number of pixels in a real MR cranial image sample; y i,j,k This represents the true label of the i-th real MR cranial image sample in the k-th category at the j-th pixel position, with a value of 0 or 1. i,j,k =0 indicates that the j-th pixel of the i-th real MR cranial image sample does not belong to category k, y i,j,k =1 indicates that the j-th pixel of the i-th real MR cranial image sample belongs to the k-th category; This indicates that the segmentation model is applicable to real MR cranial images x MR The brain region segmentation results, with values ​​ranging from 0 to 1, The closer a value is to 0, the less likely the j-th pixel of the i-th real MR cranial image sample belongs to the k-th category. The closer a value is to 1, the closer the j-th pixel of the i-th real MR cranial image sample is to the k-th category.

[0094] By training a segmentation model to segment real MR cranial images, brain region segmentation atlases showing the distribution of multiple brain regions were obtained. (See [link to relevant documentation]). Figure 5 Column (c) shows the brain regions. Different brain regions can be marked with different colors for easy identification.

[0095] The brain region segmentation method based on CT cranial images can also be validated using a pre-judgment model. Different brain regions will present with different symptoms; for example, cerebellar damage may cause dizziness, vomiting, ataxia, and nystagmus, while basal ganglia damage may cause limb paralysis, muscle stiffness, and convulsions. The pre-judgment model can roughly determine the brain region where the damage is located based on input symptom descriptions such as headache, dizziness, speech impairment, and muscle stiffness. If the damage localization result provided by the pre-judgment model is consistent with the damage localization result provided by the CT cranial image segmentation method, then the brain region segmentation method based on CT cranial images provided by this invention is effective; if there is a discrepancy, further judgment by a physician is required.

[0096] This invention also provides a brain region segmentation system capable of performing the aforementioned method for segmenting brain regions from CT cranial images, comprising: a modality conversion module, an adversarial training module, and a segmentation module. The modality conversion module is capable of segmenting CT cranial images using a modality conversion model. CT Converted to simulated MR brain images x CT→MR The adversarial training module challenges the modality transfer model by determining whether the input image is a real MR image, thereby causing the model to generate simulated MR brain images that approximate real MR brain images. CT→MR The segmentation module can segment the generated simulated MR brain images using a segmentation model. CT→MR The brain is segmented into several regions and a brain region segmentation map is generated.

[0097] The aforementioned brain region segmentation system can accurately segment the entire brain from CT images of the brain, and can be used to locate brain region injuries to assist doctors in making judgments.

[0098] It should be understood that although this specification is described according to various embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0099] The detailed descriptions listed above are merely specific descriptions of feasible embodiments of the present invention and are not intended to limit the scope of protection of the present invention. All equivalent implementation schemes or modifications made without departing from the spirit of the present invention, such as combinations, divisions or repetitions of features, should be included within the scope of protection of the present invention.

Claims

1. A method for segmenting brain regions in CT cranial images, characterized in that, Includes the following steps: A modality conversion model is trained to convert CT brain images obtained in CT modality into simulated MR brain images in MR modality. The parameters trained on the modality conversion model include structural prior information, which is the structural difference between each CT brain image in the dataset and its paired real MR brain image. The modality conversion model can extract CT modality features and structural features from the CT brain images, convert the CT modality features into MR modality features, combine the structural features with the structural prior information to obtain trained structural prior information, and then fuse the MR modality features with the trained structural prior information to obtain simulated MR brain images. After the modality conversion model is trained, a CT brain image is converted using the modality conversion model. Converted into a simulated MR brain image ;as well as Segmenting the simulated MR brain images A brain region segmentation map was obtained.

2. The brain region segmentation method for CT cranial images as described in claim 1, characterized in that, The steps for training a modality transfer model using an adversarial training module include: Randomly select a CT brain image sample from a dataset And input the modality conversion model, where 0 < i ≤ N, and N represents the number of CT cranial image samples in the dataset; From CT brain image samples CT modal features and structural features are extracted, wherein the CT modal features are the outline of the brain that can be distinguished by the CT equipment, and the structural features are the unique structural outline of the brain displayed by the CT equipment. The CT modal features are converted into MR modal features; The structural features are combined with the prior structural information to obtain the training structural features; By fusing the MR modal features and the training structural features, a simulated MR cranial image is generated. With the parameters of the adversarial training module fixed, the simulated MR brain images are... The input is given to the adversarial training module, and the adversarial training module outputs a... The modality transition model optimizes its parameters, where A value between 0 and 1 is used to determine whether the input image is a true MR image. The smaller the value, the closer the input image is to the distribution of a real MR image. The closer the value is to 1, the closer the input image is to the distribution of a non-real MR image; Randomly select another CT brain image sample from the dataset. ; Using the optimized parameters of the modality conversion model, the CT cranial image samples were... Converted into a simulated MR brain image ; With the parameters of the modality conversion model kept constant, the simulated MR brain image is... The input is given to the adversarial training module, and the adversarial training module outputs a... Value and optimize the parameters of the adversarial training module; and Repeat the above steps until the optimization termination condition is met, thus completing the training of the modality conversion model.

3. The brain region segmentation method for CT cranial images as described in claim 2, characterized in that, The mode transition model Optimization is performed by minimizing the loss function in equation (1): in, N represents the number of CT cranial image samples in the dataset, 0 < i ≤ N; The value output by the adversarial training module, ranging from 0 to 1, is used to determine whether the input image is a real MR image. The closer the value is to 0, the closer the input image is to the distribution of a real MR image. The closer the value is to 1, the closer the input image is to the distribution of a non-real MR image.

4. The brain region segmentation method for CT cranial images as described in claim 2, characterized in that, The mode transition model Optimization is performed by minimizing the loss function in equation (2): in, N represents the number of CT cranial image samples in the dataset, 0 < i ≤ N; M represents the total number of pixels in a CT brain image sample, 0 < j ≤ M; It is the j-th gray value of the CT brain image of the i-th brain image sample. It is the j-th gray value of the simulated MR brain image obtained by converting the i-th CT brain image sample; The output value of the adversarial training module, ranging from 0 to 1, is used to determine whether the input image is a real MR image. The closer the value is to 0, the closer the input image is to the distribution of a real MR image. The closer the value is to 1, the closer the input image is to the distribution of a non-real MR image.

5. The brain region segmentation method for CT cranial images as described in claim 2, characterized in that, The parameters of the adversarial training module are optimized by minimizing the loss function in equation (3): Where N represents the number of CT cranial image samples in the dataset, 0 < i ≤ N; All values ​​are in the range of 0 to 1, output by the adversarial training module, used to determine whether the input image is a real MR image. The closer the value is to 0, the closer the input image is to the distribution of a real MR image. The closer the value is to 1, the closer the input image is to the distribution of a non-real MR image.

6. The brain region segmentation method for CT cranial images as described in claim 1, characterized in that, After the modality conversion model is trained, a CT brain image is converted using the modality conversion model. Converted into a simulated MR brain image The steps include: A CT scan of the brain The input is fed into the modality conversion model to extract CT modal features and structural features; The CT modal features are replaced with general MR modal features, wherein the general MR modal features are the average value of the MR modal features obtained from all MR images in the inference dataset after the modality conversion model has been trained; The structural features are combined with general structural prior information to obtain applied structural features, wherein the general structural prior information is the structural prior information obtained after the modality transition model has been trained; and By fusing the general MR modal features and the application structural features, a simulated MR brain image is generated. .

7. The brain region segmentation method for CT cranial images as described in claim 1, characterized in that, The step of segmenting the simulated MR brain image to obtain a brain region segmentation atlas involves segmenting the simulated MR brain image using a trained segmentation model. The training process of the segmentation model includes: Extract a real MR brain image from the dataset. The input is given to the segmentation model, where N represents the number of real MR cranial image samples in the dataset, and 0 < i ≤ N; Optimize the parameters of the segmentation model; and Repeat the above steps until the optimization termination condition is met.

8. The brain region segmentation method for CT cranial images as described in claim 7, characterized in that, In the step of optimizing the parameters of the segmentation model, the segmentation model is optimized by minimizing the loss function in equation (4): Wherein, N represents the number of real MR cranial image samples in the dataset; C represents the number of brain region categories; M represents the total number of pixels in a real MR cranial image sample; This represents the true label of the i-th real MR cranial image sample in the k-th category at the j-th pixel position, with a value of 0 or 1. =0 indicates that the j-th pixel of the i-th real MR cranial image sample does not belong to category k. =1 indicates that the j-th pixel of the i-th real MR cranial image sample belongs to the k-th category; This indicates that the segmentation model is applicable to real MR brain images. The brain region segmentation results, with values ​​ranging from 0 to 1, The closer a value is to 0, the less likely the j-th pixel of the i-th real MR cranial image sample belongs to the k-th category. The closer a value is to 1, the closer the j-th pixel of the i-th real MR cranial image sample is to the k-th category.

9. The brain region segmentation method for CT cranial images as described in claim 1, characterized in that, The process also includes the step of: overlaying the brain region segmentation atlas with the CT cranial images. Obtain CT images of the brain after brain region segmentation.

10. A brain region segmentation system, characterized in that, A brain region segmentation method capable of performing brain images from CT cranial images as described in any one of claims 1-9, comprising: A modality conversion module that can convert CT brain images into a modality conversion model. Converted to simulated MR brain images ; An adversarial training module, by determining whether the input image is a real MR image, forms an adversarial relationship with the modality conversion model, causing the modality conversion model to generate simulated MR brain images that approximate real MR brain images. ;as well as A segmentation module that can segment generated simulated MR brain images using a segmentation model. The brain is segmented into several regions and a brain region segmentation map is generated.

Images