MRI brain image data processing method and system

Abstract

The present invention relates to an MRI brain image data processing method, comprising: preprocessing MRI brain image data to obtain a gray matter image and a white matter image that are registered into a standard space; using Box-Cox transformation to respectively standardize the obtained gray matter image and white matter image; and performing weight assignment on the standardized gray matter image and white matter image, and allocating variable computing resources. The present invention also relates to an MRI brain image data processing system. The present invention can implement balanced learning of the gray and white matter lesion features of a disease while preserving whole-brain feature information, and thus has strong robustness and high accuracy.

Description

MRI brain image data processing method and system Technical Field

[0001] The present invention relates to the technical field of image data processing, and in particular to a method and system for processing MRI brain image data. Background Art

[0002] With the rapid advancement of modern medical technology, MRI (Nuclear Magnetic Resonance Imaging) has become a key tool in neuroscience and clinical neurology. In particular, in brain neuroimaging, MRI provides a noninvasive, high-resolution means of observing neural structure and function. However, multi-center MRI data generally faces significant challenges due to the multi-center, multi-machine, and multi-parameter nature of the data. These factors include, but are not limited to, the diversity of magnetic field strength, scanning parameters, and machine manufacturers. These factors can lead to significant variability in data across centers, complicating modeling.

[0003] Deep learning has demonstrated remarkable performance in medical image analysis, including tasks such as image segmentation, classification, and regression. However, in multi-center analysis of brain MRI images, existing deep learning models often face significant challenges in accuracy and robustness due to data heterogeneity and inter-center variations. Accurate and reliable models are particularly critical for diagnosing neurodegenerative diseases such as Alzheimer's disease, but current models often fail to meet these requirements.

[0004] Due to the diverse sources of multi-center MRI data, its distribution and quality inconsistencies (such as differences in signal-to-noise ratio, resolution, and contrast), especially contrast differences, can seriously affect the generalization ability and robustness of deep learning models. In this case, traditional model training strategies often fail to obtain sufficient and generalizable feature representations. In addition, differences in scanning parameters and equipment used by different centers can also cause significant fluctuations in model performance across different datasets. These problems have greatly limited the application of deep learning on multi-center brain MRI data.

[0005] Current methods for processing multi-center MRI data include: data harmonic method, histogram matching method, and machine learning method.

[0006] Data harmonic methods, such as ComBat and Harmonization, attempt to adjust or normalize the distribution of MRI data across centers. Over-normalization or correction with data harmonic methods may eliminate or distort disease-related biological changes. Selecting appropriate parameters or features requires expertise and may require extensive tuning and validation. Adding new parameters or equipment may require re-engineering the harmonic method, making it unsuitable for future large-scale brain imaging.

[0007] Histogram matching methods match external data by setting a specific standardized histogram intensity template. Although relatively simple, histogram matching methods can lead to a bias in the model towards brighter areas and a lack of balanced consideration of gray and white matter due to the contrast differences between gray and white matter in different neurodegenerative diseases, especially those with co-occurring gray and white matter lesions.

[0008] Machine learning methods attempt to learn and adapt to differences in data distribution across multiple centers. However, traditional machine learning methods still have the risk of overfitting the training data and, like histogram matching methods, lack balanced consideration of gray matter and white matter.

[0009] Summary of the Invention

[0010] In view of this, it is necessary to provide a method and system for processing MRI brain image data.

[0011] The present invention provides a method for processing MRI brain image data, which includes the following steps: a. preprocessing MRI brain image data to obtain gray matter images and white matter images registered in a standard space; b. using Box-Cox transformation to standardize the obtained gray matter images and white matter images respectively; and c. assigning weights to the standardized gray matter images and white matter images, and allocating variable computing resources.

[0012] Preferably, the step a comprises:

[0013] First, use Realign to correct head motion and filter out images with excessive head motion.

[0014] Then, the motion-corrected images are segmented into three categories: gray matter images, white matter images, and cerebrospinal fluid images;

[0015] Next, the intensity deviation in the image due to the non-uniform magnetic field is corrected;

[0016] Finally, the rectified images are normalized to the MNI space to obtain gray matter images and white matter images registered to the standard space.

[0017] Preferably, the step b comprises:

[0018] The Box-Cox transformation is constructed using the maximum likelihood method and takes the form:

[0019] Where λ is an undetermined parameter and λ≠0; for the n voxel values ​​y1, y2, ..., y n Applying the above transformation, we get the transformed vector:

[0020] Then determine the transformation parameter λ so that y (λ) Satisfied: y (λ) =Xβ+e,e~N(0,σ 2 I) (3)

[0021] Use the maximum likelihood method to determine λ, since y (λ) Obey N(Xβ,σ 2 I) distribution, so for fixed λ, β, σ 2 The likelihood function is:

[0022] Where J is the Jacobi determinant of the transformation:

[0023] For L(β,σ 2 ), about β and σ 2 Take the derivative and set it equal to 0 to obtain β and σ 2 The maximum likelihood estimate of is: Q e (λ,y (λ) )=y (λ)′ (IX(X′X) -1 X′)y (λ) (8)

[0024] Therefore, the maximum likelihood is:

[0025] The maximum value of formula (9) is used to determine λ, and then it is substituted into formula (1) to perform Box-Cox transformation on the image.

[0026] Preferably, the step c comprises:

[0027] Gray matter image after gray matter and white matter transformation and white matter images Perform weight assignment and allocate variable computing resources through adaptive learning allocation decision-making through the back propagation algorithm. The specific formula is as follows:

[0028] Among them, α and γ are learnable parameters.

[0029] Preferably, the step c comprises:

[0030] Using a fixed weight for either gray matter image or white matter image, adaptively learning the weight of the other feature, the calculation formula is divided into the following two cases:

[0031] The present invention provides an MRI brain image data processing system, which includes a preprocessing module, a standardization module, and an allocation module, wherein: the preprocessing module is used to preprocess MRI brain image data to obtain gray matter images and white matter images registered in a standard space; the standardization module is used to use Box-Cox transformation to standardize the obtained gray matter images and white matter images respectively; and the allocation module is used to assign weights to the standardized gray matter images and white matter images and allocate variable computing resources.

[0032] Preferably, the preprocessing module is specifically used for:

[0033] First, use Realign to correct head motion and filter out images with excessive head motion.

[0034] Then, the motion-corrected images are segmented into three categories: gray matter images, white matter images, and cerebrospinal fluid images;

[0035] Next, the intensity deviation in the image due to the non-uniform magnetic field is corrected;

[0036] Finally, the rectified images are normalized to the MNI space to obtain gray matter images and white matter images registered to the standard space.

[0037] Preferably, the standardization module is specifically used for:

[0038] The Box-Cox transformation is constructed using the maximum likelihood method and takes the form:

[0039] Where λ is an undetermined parameter and λ≠0; for the n voxel values ​​y1, y2, ..., y n Applying the above transformation, we get the transformed vector:

[0040] Then determine the transformation parameter λ so that y (λ) Satisfied: y (λ) =Xβ+e,e~N(0,σ 2 I) (3)

[0041] Use the maximum likelihood method to determine λ, since y (λ) Obey N(Xβ,σ 2 I) distribution, so for fixed λ, β, σ 2 The likelihood function is:

[0042] Where J is the Jacobi determinant of the transformation:

[0043] For L(β,σ 2), about β and σ 2 Take the derivative and set it equal to 0 to obtain β and σ 2 The maximum likelihood estimate of is: Q e (λ,y (λ) )=y (λ)′ (IX(X′X) -1 X′)y (λ) (8)

[0044] Therefore, the maximum likelihood is:

[0045] The maximum value of formula (9) is used to determine λ, and then it is substituted into formula (1) to perform Box-Cox transformation on the image.

[0046] Preferably, the allocation module is specifically used to:

[0047] Gray matter image after gray matter and white matter transformation and white matter images Perform weight assignment and allocate variable computing resources through adaptive learning allocation decision-making through the back propagation algorithm. The specific formula is as follows:

[0048] Among them, α and γ are learnable parameters.

[0049] Preferably, the allocation module is further configured to:

[0050] Using a fixed weight for either gray matter image or white matter image, adaptively learning the weight of the other feature, the calculation formula is divided into the following two cases:

[0051] After processing MRI brain image data to obtain gray matter and white matter images, the present invention considers the two separately, adaptively learning the contribution ratio of gray matter and white matter in different classification tasks. While preserving the characteristics of the whole brain, it also balances the learning of gray and white matter lesion characteristics of the disease, thereby improving the classification accuracy and robustness of the model. At the same time, the present invention has strong scalability and can be applied to different brain diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] FIG1 is a flow chart of a method for processing MRI brain image data according to the present invention;

[0053] FIG2 is a hardware architecture diagram of the MRI brain image data processing system of the present invention;

[0054] FIG3 is a schematic diagram of the classification performance of an external data set according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0056] 1 , which is a flowchart of a preferred embodiment of the MRI brain image data processing method of the present invention.

[0057] Step S1: pre-process the MRI brain image data to obtain gray matter images and white matter images registered in a standard space.

[0058] This embodiment uses CAT12 software to preprocess brain structural MRI images. Specifically:

[0059] First, Realign is used to correct for excessive head movement, filtering out images with excessive head movement. In this embodiment, the standards for excessive head movement are 2 mm and 2 degrees. That is, if the head movement exceeds the standards of 2 mm and 2 degrees, the image is considered to have excessive head movement.

[0060] Then, the image after head motion correction is subjected to tissue segmentation, and the image is usually segmented into three categories: gray matter image, white matter image and cerebrospinal fluid image.

[0061] Next, the intensity deviation in the image due to magnetic field inhomogeneity is corrected.

[0062] Finally, the rectified image is normalized to the MNI space to obtain gray matter images and white matter images registered to the standard space: After stripping the skull, this embodiment finally obtains 3D gray matter images and white matter images with a size of 113×137×131.

[0063] Step S2: Use Box-Cox transformation to standardize the gray matter image and white matter image. Specifically:

[0064] The Box-Cox transformation is constructed using the standardization module. This embodiment uses the maximum likelihood method to construct the Box-Cox transformation, which is in the form of:

[0065] Where λ is an undetermined parameter and λ≠0. For n voxel values ​​y1, y2, ..., y n Applying the above transformation, we get the transformed vector:

[0066] Then determine the transformation parameter λ so that y (λ) Satisfied: y (λ) =Xβ+e,e~N(0,σ 2 I) (3)

[0067] Use the maximum likelihood method to determine λ, since y (λ) ~N(Xβ,σ 2I), that is, y (λ) Obey N(Xβ,σ 2 I) distribution. Therefore, for fixed λ, β, σ 2 The likelihood function is:

[0068] Where J is the Jacobi determinant of the transformation:

[0069] For L(β,σ 2 ), about β and σ 2 Take the derivative and set it equal to 0 to obtain β and σ 2 The maximum likelihood estimate of is: Q e (λ,y (λ) )=y (λ)′ (IX(X′X) -1 X′)y (λ) (8)

[0070] Therefore, the maximum likelihood is:

[0071] The maximum value of formula (9) is used to determine λ, and then it is substituted into formula (1) to perform Box-Cox transformation on the image.

[0072] The purpose of the Box-Cox transformation is to find the best data transformation so that the transformed data is closer to the normal distribution, reducing the impact of outliers on the model caused by MRI image scanning, and at the same time helping to reduce heteroscedasticity in the data, making the variance of the data more stable in different observations or groups, thereby reducing data interference caused by multiple centers.

[0073] Step S3: weight the standardized gray matter image and white matter image and allocate variable computing resources. Specifically:

[0074] Gray matter image after gray matter and white matter transformation and white matter images Perform weight assignment and adaptively learn allocation decisions to allocate variable computing resources through the back-propagation algorithm.

[0075] The specific formula is as follows:

[0076] Among them, α and γ are learnable parameters.

[0077] Another embodiment of the present application adaptively learns the allocation decision, uses a fixed gray matter and white matter weight, and adaptively learns another feature weight. The calculation formula is divided into the following two cases:

[0078] This step allows the model to allocate appropriate computing resources to each input example, avoiding wasting computational power on simple examples and focusing on more difficult or computationally demanding examples. Avoiding excessive computational resources on simple examples can also reduce overfitting of the model on these examples. By allocating computing resources based on data diversity, the model can better understand and learn from data from different distributions or subspaces, thereby improving the model's generalization ability.

[0079] Refer to FIG2 , which is a hardware architecture diagram of the MRI brain image data processing system 10 of the present invention. The system includes: a pre-processing module 101, a standardization module 102, and a distribution module 103.

[0080] The pre-processing module 101 is used to pre-process the MRI brain image data to obtain gray matter images and white matter images after registration in the standard space. Specifically:

[0081] In this embodiment, the pre-processing module 101 uses CAT12 software to pre-process the brain structure MRI image. Specifically:

[0082] First, Realign is used to correct for excessive head movement, filtering out images with excessive head movement. In this embodiment, the standards for excessive head movement are 2 mm and 2 degrees. That is, if the head movement exceeds the standards of 2 mm and 2 degrees, the image is considered to have excessive head movement.

[0083] Then, the image after head motion correction is subjected to tissue segmentation, and the image is usually segmented into three categories: gray matter image, white matter image and cerebrospinal fluid image.

[0084] Next, the intensity deviation in the image due to magnetic field inhomogeneity is corrected.

[0085] Finally, the rectified image is normalized to the MNI space to obtain gray matter images and white matter images registered to the standard space: After stripping the skull, this embodiment finally obtains 3D gray matter images and white matter images with a size of 113×137×131.

[0086] The standardization module 102 is used to standardize the obtained gray matter image and white matter image respectively using Box-Cox transformation. Specifically:

[0087] The standardization module 102 uses the Box-Cox method to construct the standardization module. In this embodiment, the maximum likelihood method is used to construct the Box-Cox transformation, which is in the form of:

[0088] Where λ is an undetermined parameter and λ≠0. For n voxel values ​​y1, y2, ..., yn Applying the above transformation, we get the transformed vector:

[0089] Then determine the transformation parameter λ so that y (λ) Satisfied: y (λ) =Xβ+e,e~N(0,σ 2 I) (3)

[0090] Use the maximum likelihood method to determine λ, since y (λ) ~N(Xβ,σ 2 I), that is, y (λ) Obey N(Xβ,σ 2 I) distribution. Therefore, for fixed λ, β, σ 2 The likelihood function is:

[0091] Where J is the Jacobi determinant of the transformation:

[0092] For L(β,σ 2 ), about β and σ 2 Take the derivative and set it equal to 0 to obtain β and σ 2 The maximum likelihood estimate of is: Q e (λ,y (λ) )=y (λ)′ (IX(X′X) -1 X′)y (λ) (8)

[0093] Therefore, the maximum likelihood is:

[0094] The maximum value of formula (9) is used to determine λ, and then it is substituted into formula (1) to perform Box-Cox transformation on the image.

[0095] The purpose of the Box-Cox transformation is to find the best data transformation so that the transformed data is closer to the normal distribution, reducing the impact of outliers on the model caused by MRI image scanning, and at the same time helping to reduce heteroscedasticity in the data, making the variance of the data more stable in different observations or groups, thereby reducing data interference caused by multiple centers.

[0096] The allocation module 103 is used to assign weights to the standardized gray matter image and white matter image and allocate variable computing resources. Specifically:

[0097] The allocation module 103 respectively converts the gray matter image into white matter image. and white matter images Perform weight assignment and adaptively learn allocation decisions to allocate variable computing resources through the back-propagation algorithm.

[0098] The specific formula is as follows:

[0099] Among them, α and γ are learnable parameters.

[0100] Another embodiment of the present application adaptively learns the allocation decision, uses a fixed gray matter and white matter weight, and adaptively learns another feature weight. The calculation formula is divided into the following two cases:

[0101] Through the allocation module 103, the model can allocate appropriate computing resources to each input sample, avoiding wasting computing power on simple samples and focusing on more difficult or computationally demanding samples. Avoiding excessive consumption of computing resources on simple samples can also reduce overfitting of the model on these samples. The model can allocate different computing resources based on data diversity, better understanding and learning from data from different distributions or subspaces, thereby improving the model's generalization ability.

[0102] Experimental validation demonstrates that this method reduces the random effects of random seeding by training 30 times using a multi-center, multi-machine, and multi-parameter MRI brain dataset for early diagnosis of MCI. The final results are shown in Figure 3, where OUR_acc represents the performance of the proposed model on external data, and Normal_acc represents the model's performance under normal conditions.

[0103] The present invention has higher classification accuracy (average accuracy 69.1%) and higher model training stability (standard deviation ±2.21%) on external datasets.

[0104] This paper uses a deep learning network to adaptively learn the gray / white matter contrast differences brought about by MRI brain images scanned at different centers. While ensuring the preservation of whole-brain information, it strives to improve the robustness of the model when dealing with multi-center, multi-machine, and multi-parameter MRI brain data, thereby providing more accurate and reliable services for the diagnosis of neurodegenerative diseases.

[0105] While preserving whole-brain feature information, the present invention provides adaptive gray-matter and white-matter computing resource ratios for different neurodegenerative diseases, reducing the expertise required for data preprocessing and making it more suitable for large-scale batch MRI brain imaging learning.

[0106] The present invention utilizes the Box-Cox transformation normalization module to reduce data interference caused by multiple centers, comprehensively considers the pathological characteristics of gray matter and white matter in neurodegenerative diseases, and uses deep learning adaptive technology to learn allocation decisions to allocate variable computing resources to adapt to the gray and white matter erosion preferences of different neurodegenerative diseases, thereby achieving more accurate and generalizable disease diagnosis.

[0107] Although the present invention has been described with reference to the current preferred embodiments, those skilled in the art should understand that the above-mentioned preferred embodiments are only used to illustrate the present invention and are not used to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for processing MRI brain image data, characterized in that, the method comprises the following steps: a. Preprocess the MRI brain image data to obtain gray matter images and white matter images registered to the standard space; b. Use the Box-Cox transformation to standardize the obtained gray matter images and white matter images respectively; c. Assign weights to the standardized gray matter images and white matter images and allocate variable computing resources.

2. The method for processing MRI brain image data according to claim 1, characterized in that, the step a includes: First, perform head motion correction by Realign and screen out images with excessive head motion; Then, segment the head motion-corrected images into three categories: gray matter images, white matter images, and cerebrospinal fluid images; Next, correct the intensity deviation caused by magnetic field inhomogeneity in the images; Finally, normalize the corrected images to the MNI space to obtain gray matter images and white matter images registered to the standard space.

3. The method for processing MRI brain image data according to claim 2, characterized in that, the step b includes: Construct the Box-Cox transformation using the maximum likelihood method, which has the form: where λ is a parameter to be determined and λ ≠ 0; for the n voxel values y of the MRI 1 , y 2 ,..., y n Applying the above transformation, the transformed vector is obtained: After that, the transformation parameter λ is determined such that y (λ) satisfies: y (λ) = Xβ + e, e ~ N(0, σ 2 I) (3) Determine λ using the maximum likelihood method. Since y (λ) follows an N(Xβ, σ 2 I) distribution, therefore, for fixed λ, β, σ 2 The likelihood function of where J is the Jacobi determinant of the transformation: For \(L(\beta,\sigma\) 2 ), taking the derivatives with respect to \(\beta\) and \(\sigma\) 2 and setting them equal to zero, the maximum likelihood estimates of \(\beta\) and \(\sigma\) 2 are obtained as follows: Q e (λ,y (λ) ) = y (λ)′ (I - X(X′X) -1 X′)y (λ) (8) Therefore, the maximum likelihood value is: Determine λ by finding the maximum value of formula (9) and substitute it into formula (1) to perform the Box-Cox transformation on the images.

4. The method for processing MRI brain image data according to claim 3, characterized in that, the step c includes: The gray matter image after the gray matter and white matter transformation respectively and white matter images Perform weight assignment, and adaptively learn to allocate variable computing resources for decision-making through the backpropagation algorithm. The specific formula is as follows: where α and γ are learnable parameters.

5. The method for processing MRI brain image data according to claim 3, characterized in that, the step c includes: Use one of the weights of the fixed gray matter image and the white matter image to adaptively learn the weight of the other feature. The calculation formula is divided into the following two cases:

6. An MRI brain image data processing system, characterized in that, the system includes a preprocessing module, a standardization module, and an allocation module, where: The preprocessing module is used to preprocess the MRI brain image data to obtain gray matter images and white matter images registered to the standard space; The standardization module is used to use the Box-Cox transformation to standardize the obtained gray matter images and white matter images respectively; The allocation module is used to assign weights to the standardized gray matter images and white matter images and allocate variable computing resources.

7. The MRI brain image data processing system according to claim 6, characterized in that, the preprocessing module is specifically used for: First, perform head motion correction by Realign and screen out images with excessive head motion; Then, segment the head motion-corrected images into three categories: gray matter images, white matter images, and cerebrospinal fluid images; Next, correct the intensity deviation caused by magnetic field inhomogeneity in the images; Finally, normalize the corrected images to the MNI space to obtain gray matter images and white matter images registered to the standard space.

8. The MRI brain image data processing system according to claim 7, characterized in that, the standardization module is specifically used for: The Box-Cox transformation is constructed using the maximum likelihood method and has the form: where λ is a parameter to be determined and λ≠0; for the n voxel values y of the MRI 1 , y 2 ,..., y n Applying the above transformation, the transformed vector is obtained: After that, the transformation parameter λ is determined such that y (λ) satisfies: y (λ) = Xβ + e, e ~ N(0, σ 2 I) (3) Determine λ using the maximum likelihood method. Since y (λ) follows an N(Xβ, σ 2 I) distribution, therefore, for fixed λ, β, and σ 2 the likelihood function is: where J is the Jacobi determinant of the transformation: For \(L(\beta,\sigma\) 2 ), taking the derivatives with respect to \(\beta\) and \(\sigma\) 2 and setting them equal to zero, the maximum likelihood estimates of \(\beta\) and \(\sigma\) 2 are obtained as follows: Q e (λ,y (λ) ) = y (λ)′ (I - X(X′X) -1 X′)y (λ) (8) Therefore, the maximum likelihood is: Determine λ by finding the maximum value of formula (9) and substitute it into formula (1) to perform the Box-Cox transformation on the images.

9. The MRI brain image data processing system according to claim 8, characterized in that, the allocation module is specifically used for: The gray matter image after the gray matter and white matter transformation respectively and white matter images Perform weight assignment and adaptively learn to allocate variable computing resources for decision-making through the backpropagation algorithm. The specific formula is as follows: where α and γ are learnable parameters.

10. The MRI brain image data processing system according to claim 8, characterized in that, The described distribution module is further specifically configured to: Use one of the weights of the fixed gray matter image and the white matter image to adaptively learn the weight of the other feature. The calculation formula is divided into the following two cases:

Images