A beta-PET brain template construction and quantitative analysis method based on deep learning
By constructing an Aβ-PET brain template and quantitative analysis method suitable for Chinese people through deep learning, the problem of lack of Aβ-PET brain template and insufficient quantitative analysis in the existing technology is solved, which realizes more accurate image registration and early diagnosis, and improves the efficiency of diagnosis and treatment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-02
- Publication Date
- 2026-03-10
AI Technical Summary
The lack of Aβ-PET brain templates suitable for Chinese people in current technologies leads to registration errors and insufficient data in the research and diagnosis of Aβ-neuroamyloid deposition diseases. Furthermore, existing quantitative analysis methods cannot accurately reflect the dynamic changes in the brains of the elderly.
Aβ-PET brain template was constructed using deep learning methods. Data from participants with negative Aβ-amyloid tracer images and normal PET images were obtained, and preprocessing, spatial registration, intensity normalization, data filtering, and averaging were performed to establish an Aβ-PET brain template suitable for Chinese individuals. A database of brain amyloid deposition data was also constructed, and the VoxelMorph framework model was used for image registration and quantitative analysis.
It provides ethically sound and accurate Aβ-PET brain templates, improves image registration accuracy, helps detect abnormalities early, improves diagnostic and treatment efficiency, and fills gaps in existing technologies.
Smart Images
Figure CN121639624A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical imaging technology, and in particular to a method for constructing and quantitatively analyzing Aβ-PET brain templates based on deep learning. Background Technology
[0002] Alzheimer's disease (AD) is the most common cause of dementia in the elderly, characterized by senile plaques formed by the deposition of β-amyloid (Aβ) protein in the brain and neurofibrillary tangles formed by the hyperphosphorylation of tau protein. Abnormal Aβ accumulation is considered a key initiating event in the AD pathological cascade. Therefore, accurate and early detection of Aβ deposition in vivo is crucial for the early diagnosis of AD, monitoring of disease progression, and evaluation of the efficacy of anti-Aβ drugs.
[0003] Currently, clinical research primarily relies on conventional imaging techniques, such as computed tomography (CT) and magnetic resonance imaging (MRI). CT only provides anatomical imaging and lacks amyloid deposition imaging capabilities; MRI offers both conventional structural and functional amyloid deposition imaging, providing detailed information on brain structure and function, but its effectiveness in diagnosing amyloid deposition and developmental disorders is limited. Unlike MRI, positron emission tomography (PET) imaging can diagnose amyloid deposition activity in the nervous system, allowing physicians to detect abnormal areas of amyloid deposition using PET images. Compared to MRI, PET exhibits better sensitivity and specificity in amyloid deposition detection and is superior in the quantitative analysis of amyloid deposition in PET images. Furthermore, PET offers advantages over MRI in areas such as amyloid deposition detection, neurotransmitter assessment, molecular specificity, and dynamic amyloid deposition tracking. With the development of positron emission tomography (PET) molecular imaging technology, various Aβ-specific tracers (18F-Florbetapir, 18F-Flutemetamol, 18F-Florbetaben, 11C-PiB, 18F-Flutafuranol AZD4694, etc.) have been successfully applied in clinical practice and research. Aβ-PET imaging can non-invasively and visually demonstrate the distribution and burden of Aβ in the brain, and has become a key biomarker in the diagnostic criteria for Alzheimer's disease (AD).
[0004] With the continuous advancement of brain imaging technology, brain templates play a crucial role in disease research. By registering patient images to a standard template space, a series of analytical studies can be conducted. Currently, most widely used brain templates are designed for adults, while Aβ-templates are relatively scarce and mostly based on Caucasian populations. Brain structure templates based on normal Western populations cannot accurately characterize the brain structure features of Chinese patients using Aβ-PET. The lack of specific Aβ-amyloid deposition templates based on Aβ-PET image data limits the research and diagnosis of Aβ-neuroamyloid deposition diseases.
[0005] Current brain template construction primarily relies on MRI and CT images, with relatively few studies based on PET images. Clinically, patient PET images can only be registered with MRI templates. Multimodal registration is often less effective than same-modal registration and is more complex, thus impacting diagnosis and treatment. While adult PET brain templates have been publicly released, no publicly available Aβ-PET brain templates or related research have been found. Aβ-PET brain structures differ significantly, and the brains of elderly individuals undergo degenerative changes, making registration errors and bias analysis more likely when using templates from normal adults. Therefore, constructing Aβ-PET brain templates is a critical clinical challenge that urgently needs to be addressed.
[0006] Furthermore, quantitative analysis methods for the brain provide important evidence for disease diagnosis, treatment, and research by quantifying indicators such as brain structure, function, and amyloid deposition. In the quantitative analysis of Aβ-brain PET images, constructing a dedicated PET brain template adapted to the brain deposition characteristics of Aβ-PET is the core foundation for improving image registration accuracy, while the accumulation of large-scale amyloid deposition data from normal Aβ-aged individuals is a key prerequisite for establishing a reliable quantitative analysis model.
[0007] However, current research faces two major challenges: First, existing quantitative brain analysis techniques are mainly developed for elderly individuals whose brain structures have become relatively stable. In contrast, Aβ-brain exhibits significant morphological and dynamic changes in amyloid deposition with age (such as the age-dependent differences in the gray / white matter ratio and peak amyloid deposition rate), requiring separate research. Second, a publicly available Aβ-PET brain amyloid deposition database suitable for Chinese individuals has not yet been established globally. The lack of accurate and readily available amyloid deposition data in clinical studies has long led to errors, which in turn affect the accuracy of disease diagnosis and efficacy assessment.
[0008] Therefore, based on the developmental patterns of Aβ-cerebral amyloid deposition, establishing a standardized cerebral amyloid deposition database and designing quantitative analysis methods are urgent tasks that need to be promoted. Summary of the Invention
[0009] Technical problems to be solved In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a method for constructing and quantitatively analyzing Aβ-PET brain templates based on deep learning, which solves the technical problem of how to construct Aβ-PET brain templates suitable for China and conduct quantitative analysis research based on brain templates.
[0010] Technical solution To achieve the above objectives, the main technical solutions adopted by the present invention include: This invention provides a deep learning-based method for constructing and quantitatively analyzing Aβ-PET brain templates, comprising: Preprocessed PET images, i.e., Aβ-PET images, were obtained from participants who were negative for Aβ-amyloid tracer and had normal PET images. The Aβ-PET image was spatially registered to the Aβ-MRI brain atlas MNI152 to obtain the registered Aβ-PET image. Intensity normalization and data filtering were performed on the registered Aβ-PET images to obtain the filtered normalized Aβ-PET images; The average value of the selected normalized Aβ-PET images was calculated to obtain preliminary Aβ-PET brain template images; The selected Aβ-PET images were spatially registered to the preliminary Aβ-PET brain template images, and intensity normalization, data filtering, averaging and smoothing were performed to obtain the final Aβ-PET brain template images. After spatially registering the Aβ-PET image to the final Aβ-PET brain template image and dividing the brain regions, the mean and standard deviation of each brain region and each parameter for each normal participant were calculated. By calculating the corresponding mean ± 2 × standard deviation, the normal amyloid deposition range value was obtained. The first quantitative parameter was statistically analyzed using the normal amyloid deposition range value, and the construction of the brain amyloid deposition database was completed. After spatially registering the images of Aβ-PET positive patients to be analyzed to the final Aβ-PET brain template image and dividing the brain regions, the second quantitative parameter was calculated and compared with the database for quantitative analysis.
[0011] Optionally, the Aβ-PET image space is spatially registered to the Aβ-MRI brain atlas MNI152 to obtain the registered Aβ-PET, including: Define the initial deformation field, set the similarity metric function and regularization parameters; Calculate the similarity between Aβ-PET images and Aβ-MRI brain atlas MNI152; Based on the similarity measurement results, an optimization algorithm is used to update the transformation parameters so that the similarity measurement function reaches its maximum value; Determine whether the iteration has reached the convergence condition; If convergence is achieved, the iteration stops and spatial registration is complete; otherwise, iterative optimization continues.
[0012] Optionally, the similarity between Aβ-PET images and Aβ-MRI brain atlas MNI152 is calculated, including: The similarity MI between Aβ-PET images and Aβ-MRI brain atlases MNI152 is calculated based on the mutual information method, using the following formula: ; Where p(x, y) represents the joint probability distribution at corresponding pixel (x, y) in the two images, and p(x) and p(y) represent the edge probability distributions at pixel x and y in the two images, respectively.
[0013] Optionally, the registered Aβ-PET images are subjected to intensity normalization and data filtering to obtain filtered normalized Aβ-PET images, including: The intensity of the registered Aβ-PET image is normalized according to the following formula: ; Among them, G k For the registered Aβ-PET image of participant k, i is the pixel index, G k_mean_ref I represents the average pixel value of the cerebellar reference region for the participants. k Aβ-PET image of participant k after normalization; The sum of absolute errors between the intensity-normalized Aβ-PET image and the Aβ-MRI brain atlas MNI152 is calculated using the following formula: ; Among them, I ref For normalized Aβ-MRI brain atlas MNI152, i is the pixel index, SAD k For I k with I ref The sum of the horizontal absolute errors between the images; Anomaly detection is performed on the intensity-normalized Aβ-PET image according to the following formula: ; Among them, SAD mean and SAD std These are the mean and standard deviation of SAD for all participants, respectively; if DIF k If the value is greater than 0, the participant k data is considered to have potential amyloid protein deposition abnormalities and needs to be removed. The remaining Aβ-PET images are used as the normalized Aβ-PET images after screening.
[0014] Alternatively, the average can be calculated using the following formula: ; Where i represents the pixel index, n represents the number of images used for averaging, and I k denoted as the intensity-normalized image of the k-th participant, and I represents the preliminary Aβ-PET brain template obtained by averaging.
[0015] Optionally, the method further includes: The regional anatomical differences between the final Aβ-PET brain template and the Aβ-MRI brain atlas MNI152 image were calculated based on the Mean Sum of Squared Errors (MSD) algorithm and the Normalized Product Correlation (NCC) algorithm. The calculation formula is as follows: ; ; Where I represents the final Aβ-PET brain template obtained, I ref This represents the Aβ-MRI brain atlas MNI152, where i represents the pixel index and n represents the pixel index. roi i_roi represents the number of pixels in the region of interest, and i_roi represents the pixel index within the region of interest. mean_roi and I ref_mean_roi Representing images I and I respectively ref Average pixel value within the region of interest; Among them, the smaller the MSD value, the smaller the anatomical difference between the two image regions; the larger the MSD value, the greater the anatomical difference between the two image regions; the closer the NCC value is to 1, the smaller the anatomical difference between the two image regions; the closer the NCC value is to -1, the greater the anatomical difference between the two image regions; and the closer the NCC value is to 0, the two image regions are unrelated.
[0016] Optionally, the first quantitative parameter includes the average gray value (Gray). mean Average standard intake value for SUVs mean And the maximum intake value compared to SUV max The second quantitative parameter includes the standard uptake ratio SUVr calculated by comparing it with brainstem deposition values, the z-score, and the whole-brain amyloid index.
[0017] Optionally, the standard intake ratio (SUVr), z-score, and whole-brain amyloid index are calculated according to the following formulas: ; ; ; ; ; Where R represents the individual image registered with the final Aβ-PET brain template, S represents the individual image after converting the voxel values to SUV values, i represents the pixel index, roi represents the region of interest, n represents the number of pixels, ref represents the cerebellar region of the image itself, and database represents the brain amyloid deposition database. The formula for calculating the SUV value is as follows: ; When calculating the first quantitative parameter for constructing a database of brain amyloid deposition, the individual images are images of normal participants; when calculating the second quantitative parameter for quantitative analysis, the individual images are images of Aβ-PET-positive patients to be analyzed.
[0018] Optionally, the VoxelMorph framework model can be used for image registration.
[0019] Optionally, Aβ-PET images of elderly individuals aged 60-80 years and of different genders that are negative for Aβ-amyloid tracers and have normal PET images are obtained to construct Aβ-PET brain templates for the elderly, and a database of brain amyloid deposition corresponding to the templates is obtained through quantitative analysis.
[0020] Beneficial effects The beneficial effects of this invention are: The invention constructs age- and sex-specific Aβ-PET brain templates, along with corresponding databases of brain amyloid deposition and clinical application methods. Its advantages are significant: First, it overcomes a series of problems in template construction, establishing ethically sound and reliable Aβ-PET brain templates, filling a gap in current Aβ-PET brain template research. Based on this template, better registration results can be achieved in clinical and research settings without necessarily using patient MRI images as an intermediary. Second, the method establishes an Aβ-PET brain amyloid deposition database and provides quantitative analysis methods for clinical applications, helping doctors detect abnormalities early and improving diagnostic and treatment efficiency. Finally, the invention provides software tools to enhance the applicability of the templates. Attached Figure Description
[0021] Figure 1 A flowchart illustrating a deep learning-based Aβ-PET brain template construction and quantitative analysis method provided in an embodiment of the present invention; Figure 2 A schematic diagram of the axial plane of the Aβ-PET template, one of the templates established by the deep learning-based Aβ-PET brain template construction and quantitative analysis method provided in this embodiment of the invention; Figure 3 A schematic diagram of the coronal plane of the Aβ-PET template, one of the templates established by the deep learning-based Aβ-PET brain template construction and quantitative analysis method provided in this embodiment of the invention; Figure 4 This is a sagittal view of the Aβ-PET template, one of the templates established by the deep learning-based Aβ-PET brain template construction and quantitative analysis method provided in this embodiment of the invention. Detailed Implementation
[0022] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
[0023] Constructing Aβ-PET brain templates is a critical clinical challenge that urgently needs to be addressed, and the process is extremely challenging. First, acquiring the data required for atlas construction is difficult; recruiting and collecting data from Aβ-PET subjects is more challenging, requiring consideration of safety and cooperation issues, resulting in a scarcity of high-quality data suitable for atlas construction. Second, Aβ-PET brain structure undergoes greater changes with age than in adults, and brain morphology also differs between sexes. Therefore, age and sex characteristics must be carefully considered when constructing Aβ-PET brain templates.
[0024] In quantitative analysis based on PET images, doctors primarily focus on quantitative indicators of brain regions of interest (such as the frontal lobe, temporal lobe, and parietal lobe). Commonly used brain region division methods include AAL (Anatomical Automatic Labeling), an automated brain region labeling template provided by the Montreal Neuroscience Institute in Canada, which divides the brain into 116 regions. Common quantitative indicators include the average grayscale value and the average standard uptake value (SUV). mean The standard ingress value ratio (SUVr) and Z-score are used to measure the average intensity of pixel values in an image. mean The average uptake value can be represented by the PET image, indicating the average uptake value of a specific region or the entire brain. SUVr represents the ratio of the radioactive imaging agent concentration of the tissue of interest to that of corresponding muscle or other normal tissue in the same patient. The Z-score is a standardized value typically used to assess the deviation of a measurement from the mean; when the mean is based on a database of normal individuals, this value can be used to identify abnormalities in the patient's image. Clinically, the calculation and analysis of these quantitative parameters are of great significance for the diagnosis and treatment of diseases.
[0025] To construct an Aβ-PET brain template and fill the current gap in clinical research, this invention is based on the Aβ-MRI brain atlas MNI152 and uses data from 76 normal Aβ-subjects. 18 F-Florbetapir PET images were used for image registration using the unsupervised medical image registration framework model VoxelMorph. An Aβ-PET brain template construction method was designed and implemented, ultimately yielding the template. Based on the AAL116 partitioning, the established template was further used to calculate quantitative analysis indicators, establishing a database of normal Aβ-brain amyloid deposition data. This template provides a complete quantitative amyloid deposition analysis workflow, including standardization, registration, partitioning, and statistics. Ultimately, the Aβ-PET brain template can be used for the diagnosis and analysis of diseases such as Alzheimer's disease, dementia, mild dementia, subjective cognitive decline, and suspected cognitive impairment, particularly for the quantitative analysis of these diseases using PET.
[0026] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.
[0027] Firstly, this invention is based on the Aβ-MRI brain atlas MNI152 and uses the VoxelMorph framework model to propose a method based on... 18 A deep learning-based method for constructing and quantitatively analyzing Aβ-PET brain templates from F-Florbetapir PET images.
[0028] The brain template construction process begins with preprocessing the collected PET data from normal participants, including image cropping and position correction, to make the images more suitable for template construction. Then, an Aβ-PET brain template is constructed through a series of processes, including VoxelMorph image registration, intensity normalization, data filtering, averaging, and smoothing.
[0029] In the quantitative analysis of brain amyloid deposition, the preprocessed data of normal participants are first spatially registered to the constructed template. Then, based on the AAL116 partition, quantitative analysis indicators for each brain region are calculated, and a database of brain amyloid deposition data is obtained. Based on this database, quantitative analysis and diagnosis can be performed clinically.
[0030] Reference Figure 1 This embodiment provides a method for constructing and quantitatively analyzing Aβ-PET brain templates based on deep learning, the specific steps of which include: S1. Obtain pre-processed PET images, i.e., Aβ-PET images, from rigorously validated, Aβ-negative (i.e., negative test results) healthy volunteers or cognitively normal elderly individuals.
[0031] In this embodiment, a total of 76 negative subjects were collected. 18 F-Florbetapir-PET data were used for template construction and the creation of a database of brain amyloid deposition data.
[0032] The data screening and exclusion criteria for negative subjects are as follows: (1) Age: 60-80 years old; (2) Gender: Half male and half female; (3) Imaging requirements: DICOM images of head Aβ-PET only; (4) Health status of enrolled subjects: 1) No head diseases, and the head is not affected by any disease; 2) The lesions throughout the body are relatively mild and have not caused systemic symptoms; 3) No cognitive or developmental abnormalities or mental illnesses; 4) Has never undergone brain surgery, radiotherapy, chemotherapy, or immunotherapy; (5) Exclusion criteria: 1) The disease or tumor affects multiple systems, for example, lymphoma affects the whole body; 2) The subjects had cognitive or developmental abnormalities or mental illnesses; 3) The subjects had already undergone surgical treatment, radiotherapy, chemotherapy, and immunotherapy; (6) Image quality: During the scan, the subject's head should be kept as straight as possible without any displacement; the PET image should have clear edges without artifacts or blurring. (7) Brief medical history requirements 1) Basic information: Patient's name, gender, and age; 2) Present illness: chief complaint, symptoms, and clinical diagnosis.
[0033] The collected data undergoes image preprocessing, including image cropping and position correction.
[0034] Specifically, first calculate the minimum bounding box size of the PET image, then crop out the surrounding blank areas. If the image is affected by noise, additional noise reduction processing is required. Then, the origin of the VoxelMorph framework model is corrected to ensure the standardization and consistency of the image, making it more conducive to subsequent registration operations.
[0035] For template data, multiplying the image with its brain label yields template data that retains only the brain image, which is then used for registration.
[0036] S2, spatially register the Aβ-PET image to the Aβ-MRI brain atlas MNI152 to obtain the registered Aβ-PET image.
[0037] Optionally, the unsupervised medical image registration framework model VoxelMorph can be used for image registration.
[0038] Specifically, the VoxelMorph framework model was first used to spatially register the preprocessed Aβ-PET images onto the Beijing Normal University Aβ-MRI brain atlas MNI152, so that the Aβ-PET images were mapped to a standard brain atlas space, eliminating morphological differences between subjects.
[0039] Among them, the VoxelMorph framework model is based on the Diffeomorphic transformation, which aims to align the floating image with the reference image through a smooth and reversible deformation field.
[0040] Optionally, the Aβ-PET image space is spatially registered to the Aβ-MRI brain atlas MNI152 to obtain the registered Aβ-PET image, including: Define the initial deformation field, set the similarity metric function and regularization parameters; Calculate the similarity between Aβ-PET images and Aβ-MRI brain atlas MNI152; Based on the similarity measurement results, an optimization algorithm is used to update the transformation parameters so that the similarity measurement function reaches its maximum value; Determine whether the iteration has reached the convergence condition; If convergence is achieved, the iteration stops and spatial registration is complete; otherwise, iterative optimization continues.
[0041] In VoxelMorph framework model registration, the regularization parameter is the core parameter for controlling the smoothness of the deformation field and preventing overfitting. Its setting needs to be optimized in combination with image characteristics, registration target and algorithm type.
[0042] The typical regularization term is the L2 norm of the deformation field gradient: assuming the deformation field parameters are vectors. Then the L2 norm regularization term is , where λ is the regularization parameter. In the objective function for model registration within the VoxelMorph framework, it typically takes the form: ; By minimizing this objective function, the complexity of the deformation field can be controlled by limiting the size of the deformation field parameters while optimizing the similarity metric.
[0043] The optimization algorithms include gradient descent and conjugate gradient methods.
[0044] The updated transformation parameters are deformation field parameters, typically displacement vector field parameters and velocity field parameters. The entire update process is iterative. In each iteration, a similarity metric is calculated based on the current deformation field parameters, and then the deformation field parameters are updated based on the gradient of the objective function. This process is repeated until the convergence condition is met.
[0045] The convergence condition is that the change in the objective function value is less than a preset threshold, the change in the deformation field parameter is less than a preset threshold, or the maximum number of iterations is reached.
[0046] Optionally, the similarity between Aβ-PET images and Aβ-MRI brain atlas MNI152 is calculated, including: The similarity MI between Aβ-PET images and Aβ-MRI brain atlases MNI152 is calculated based on the mutual information method, using the following formula: ; Where p(x, y) represents the joint probability distribution at corresponding pixel (x, y) in the two images, and p(x) and p(y) represent the edge probability distributions at pixel x and y in the two images, respectively.
[0047] The performance of the VoxelMorph framework model is highly dependent on the similarity measurement function, with mutual information (MI) being a common method. Mutual information, as a similarity measurement method, treats two images as two random variables and measures their similarity by calculating their mutual information value. A higher mutual information value indicates a higher correlation between the two images, resulting in better registration. This measurement method has the advantage of being insensitive to image grayscale variations and noise, making it very suitable for image registration.
[0048] The formula for calculating mutual information is expressed as follows: ; In the case of continuous random variables, summation is replaced by double definite integrals: ; The output of this step is the participant's Aβ-PET image spatially registered with the Beijing Normal University Aβ-MRI brain atlas MNI152, which is in the standard brain atlas space and can be used for the subsequent template construction operation.
[0049] S3. Perform intensity normalization and data filtering on the registered Aβ-PET image to obtain the filtered normalized Aβ-PET image.
[0050] Optionally, the registered Aβ-PET images are subjected to intensity normalization and data filtering to obtain filtered normalized Aβ-PET images, including: The registered Aβ-PET images are normalized according to the following formula: ; Among them, G k For the registered Aβ-PET image of participant k, i is the pixel index, G k_mean_ref I represents the average pixel value of the cerebellar reference region for the participants. k The image shows the normalized Aβ-PET image of participant k.
[0051] The intensity normalization method for Aβ-MRI brain atlas MNI152 is the same as above.
[0052] The sum of absolute errors between the intensity-normalized Aβ-PET image and the Aβ-MRI brain atlas MNI152 is calculated using the following formula: ; Among them, I ref For normalized Aβ-MRI brain atlas MNI152, i is the pixel index, SAD k For I k with I ref The sum of the horizontal absolute errors between the images; Anomaly detection is performed on the intensity-normalized Aβ-PET image according to the following formula: ; Among them, SAD mean and SAD std These are the mean and standard deviation of SAD for all participants, respectively; if DIF k If the value is greater than 0, the participant k data is considered to have potential amyloid protein deposition abnormalities and needs to be removed. The remaining Aβ-PET images are used as the normalized Aβ-PET images after screening.
[0053] Considering the potential for inconsistent pixel intensity ranges among participants' Aβ-PET images due to variations in the image acquisition process, intensity normalization was performed in this step to ensure that all participants' images were within a comparable intensity range. The method employed was to divide all voxel values of the registered participant's Aβ-PET image by the average pixel value of a reference region. The reference region was the participant's own cerebellum (using the AAL116 partition for region division; alternatively, the alairach partition, Brodmann partition, and Harvard-Oxford Atlas could also be selected).
[0054] To ensure the health and normality of the participant dataset, a data screening step was performed after data spatial registration and intensity normalization. The method used was the Sum of Absolute Differences (SAD) algorithm.
[0055] After calculating the SAD of all participants, a statistical method is used to detect anomalies. A relatively simple parametric anomaly detection model assumes that the samples follow a univariate normal distribution, and a data point is considered an anomaly if its difference from the mean is greater than three standard deviations.
[0056] S4. The average value of the selected normalized Aβ-PET images is calculated to obtain preliminary Aβ-PET brain template images.
[0057] Alternatively, the average can be calculated using the following formula: ; Where i represents the pixel index, n represents the number of images used for averaging, and I k denoted as the intensity-normalized image of the k-th participant, and I represents the preliminary Aβ-PET brain template obtained by averaging.
[0058] S5. The selected Aβ-PET images are spatially registered to the preliminary Aβ-PET brain template images, and intensity normalization, data filtering, averaging and smoothing are performed to obtain the final Aβ-PET brain template images.
[0059] The specific steps for spatially registering the selected Aβ-PET images to the preliminary Aβ-PET brain template images, and then performing intensity normalization, data filtering, averaging, and smoothing to obtain the final Aβ-PET brain template images are the same as those in the above method and will not be repeated here.
[0060] The final PET brain template construction process is largely the same as the preliminary PET brain template construction process, both involving four steps: image and template spatial registration, intensity normalization, data filtering, and averaging. However, there are also some differences, as follows: In the image space registration process of the final brain template construction, the preliminary PET brain template obtained in the above steps is used as the registration reference template image, and the method still adopts the VoxelMorph framework model.
[0061] The template used is a preliminary PET brain template. Therefore, when performing image screening for deviation calculation, the degree of deviation between the intensity-normalized image and the preliminary PET brain template should be calculated.
[0062] The final template image obtained by averaging needs to undergo additional smoothing to reduce the noise caused by individual anatomical differences and to conform to the Gaussian distribution assumption, thus providing a basis for subsequent use. The smoothing parameter is determined based on parameters such as the spatial resolution of the collected participants' Aβ-PET images, and is set to full width at half height (FWHM) = [4 4 4].
[0063] After this step, the Aβ-PET brain template is finally constructed.
[0064] S6. After spatially registering the Aβ-PET image to the final Aβ-PET brain template image and dividing the brain regions, calculate the mean and standard deviation of each brain region and each parameter for each normal participant. By calculating the corresponding mean ± 2 × standard deviation, the normal amyloid deposition range value is obtained. The first quantitative parameter is statistically analyzed using the normal amyloid deposition range value, thus completing the construction of the brain amyloid deposition database.
[0065] The normal amyloid deposition range was obtained by calculating the mean ± 2 × standard deviation for each brain region and each parameter in all normal participants.
[0066] S7. After spatially registering the images of Aβ-PET positive patients to be analyzed to the final Aβ-PET brain template image and dividing the brain regions, calculate the second quantitative parameter for clinical auxiliary diagnosis and compare it with the brain amyloid deposition database to complete the quantitative analysis.
[0067] Optionally, the first quantitative parameter includes the average gray value (Gray). mean Average standard intake value for SUVs mean And the maximum intake value compared to SUV max The second quantitative parameter is calculated by comparing it with the brainstem deposition value, including the standard uptake ratio SUVr, the z-score, and the whole brain amyloid index.
[0068] Optionally, the standard intake ratio (SUVr), z-score, and whole-brain amyloid index are calculated according to the following formulas: ; ; ; ; ; Where R represents the individual image registered with the final Aβ-PET brain template, S represents the individual image after converting the voxel values to SUV values, i represents the pixel index, ROI represents the region of interest, n represents the number of pixels, REF represents the cerebellar region of the image itself, database represents the brain amyloid deposition database, and the whole brain amyloid index is obtained by weighted averaging or calculating a comprehensive index based on the values of multiple regions of interest (ROIs) in the cerebral cortex. The formula for calculating the SUV value is as follows: ; When calculating the first quantitative parameter for constructing a database of brain amyloid deposition, the individual image is an image of a normal participant; when calculating the second quantitative parameter for quantitative analysis, the individual image is an Aβ-PET image to be analyzed.
[0069] Optionally, the method further includes: The regional anatomical differences between the final Aβ-PET brain template and the Aβ-MRI brain atlas MNI152 image were calculated based on the Mean Sum of Squared Errors (MSD) algorithm and the Normalized Product Correlation (NCC) algorithm. The calculation formula is as follows: ; ; Where I represents the final Aβ-PET brain template obtained, I ref This represents the Aβ-MRI brain atlas MNI152, where i represents the pixel index and n represents the pixel index. roi i_roi represents the number of pixels in the region of interest, and i_roi represents the pixel index within the region of interest. mean_roi and I ref_mean_roi Representing images I and I respectively ref Average pixel value within the region of interest; Among them, the smaller the MSD value, the smaller the anatomical difference between the two image regions; the larger the MSD value, the greater the anatomical difference between the two image regions; the closer the NCC value is to 1, the smaller the anatomical difference between the two image regions; the closer the NCC value is to -1, the greater the anatomical difference between the two image regions; and the closer the NCC value is to 0, the two image regions are unrelated.
[0070] This invention provides an evaluation method for the constructed template to ensure its accuracy and reliability. The template is compared with the Beijing Normal University MRI template MNI152, which serves as the research baseline, to calculate the regional anatomical differences between the two template images. The region segmentation label uses AAL116 partitions (in addition to AAL partitions, Talairach, Brodmann, and Harvard-Oxford Atlas partitions can also be selected). The calculation metrics employ the Mean Square Differences (MSD) algorithm and the Normalized Cross Correlation (NCC) algorithm.
[0071] In clinical applications, patient images are first spatially registered to a pre-constructed template, using the VoxelMorph framework model as a registration method. Then, the registered images are segmented into brain regions based on AAL partitions (in addition to AAL partitions, Talairach, Brodmann, and Harvard-Oxford Atlas partitions can also be used, ensuring consistency with the partitioning method used to construct the database), and the quantitative parameter value Gray is calculated. mean SUV mean SUVr and Z-score; finally, the calculated values are compared with the corresponding template amyloid deposition values in the brain amyloid deposition database to determine the degree of abnormality in the patient's brain region. Gray mean SUV mean If the SUVr value falls outside the normal amyloid deposition range in the brain amyloid deposition database, it indicates an abnormality in the brain region. The Z-score is analyzed using a threshold: if 1.96 ≤ |Z-score| < 2.58, it indicates a mild abnormality; if |Z-score| ≥ 2.58, it indicates a moderate to severe abnormality. Through these analyses, the abnormalities in the brain regions of patients can be identified, aiding in clinical diagnosis.
[0072] This invention can be implemented as a software template, including functions such as image uploading, template registration, quantitative report generation, and result analysis. It can be installed as a plug-in into third-party software to improve the applicability of the template.
[0073] The invention ultimately resulted in the construction and quantitative analysis of a database of brain amyloid protein deposition data corresponding to the template. This invention also provides quantitative analysis methods for clinical applications and supporting software tools. Figure 2 , Figure 3 and Figure 4 The image shows a schematic diagram of the three orientations of the template.
[0074] In a second aspect, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed, implements the deep learning-based Aβ-PET brain template construction and quantitative analysis method described in any of the first aspects above.
[0075] Thirdly, embodiments of the present invention provide a storage device, including a storage medium and a processor, wherein the storage medium stores a computer program, and when the program is executed by the processor, it implements the deep learning-based Aβ-PET brain template construction and quantitative analysis method described in any of the first aspects above.
[0076] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0077] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, then this invention should also include these modifications and variations.
[0078] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of the present invention.
Claims
1.A method for constructing and quantitative analysis of Aβ-PET brain template based on deep learning, characterized in that, The method comprises the following steps: obtaining pre-processed PET images of Aβ-amyloid tracer negative and PET image normal participants, i.e. Aβ-PET images; spatially registering the Aβ-PET images to an Aβ-MRI brain atlas MNI152 to obtain registered Aβ-PET images; performing intensity normalization processing and data screening on the registered Aβ-PET images to obtain screened normalized Aβ-PET images; performing average calculation on the screened normalized Aβ-PET images to obtain a preliminary Aβ-PET brain template image; spatially registering the screened Aβ-PET images to the preliminary Aβ-PET brain template image and performing intensity normalization processing, data screening, average calculation and smoothing processing to obtain a final Aβ-PET brain template image; after spatially registering the Aβ-PET images to the final Aβ-PET brain template image and dividing the brain regions, calculating the average value and standard deviation of each brain region and each parameter of each normal participant, obtaining the normal amyloid deposition range value by calculating the corresponding average value ± 2×standard deviation, and statistically analyzing the first quantitative parameter according to the normal amyloid deposition range value to complete the construction of the brain amyloid deposition data library; after spatially registering the Aβ-PET positive patient images to be analyzed to the final Aβ-PET brain template image and dividing the brain regions, calculating the second quantitative parameter and comparing it with the library for quantitative analysis. 2.The deep learning based Aβ-PET brain template construction and quantitative analysis method according to claim 1, characterized in that, spatially registering the Aβ-PET images to the Aβ-MRI brain atlas MNI152 to obtain registered Aβ-PET, comprising: defining an initial deformation field, setting a similarity measure function and a regularization parameter; calculating the similarity between the Aβ-PET images and the Aβ-MRI brain atlas MNI152; updating the transformation parameters using an optimization algorithm according to the similarity measure result to maximize the similarity measure function; judging whether the iteration reaches a convergence condition; if the convergence is reached, stopping the iteration and completing the spatial registration; otherwise, continuing the iteration optimization. 3.The deep learning based Aβ-PET brain template construction and quantitative analysis method according to claim 2, characterized in that, calculating the similarity between the Aβ-PET images and the Aβ-MRI brain atlas MNI152, comprising: calculating the similarity MI between the Aβ-PET images and the Aβ-MRI brain atlas MNI152 based on the mutual information method, and the formula is as follows: ； wherein p(x, y) represents the joint probability distribution of the corresponding pixel points (x, y) of the two images, and p(x) and p(y) represent the marginal probability distributions of the pixel points x and y of the two images, respectively. 4.The deep learning based Aβ-PET brain template construction and quantitative analysis method of claim 3, characterized in that, performing intensity normalization processing and data screening on the registered Aβ-PET images to obtain screened normalized Aβ-PET images, comprising: performing intensity normalization processing on the registered Aβ-PET images according to the following formula: ； where G k is the registered Aβ-PET image of the participant k, i is the pixel index, G k_mean_ref is the pixel mean of the cerebellum reference region of the participant k, I k is the normalized Aβ-PET image of the participant k; calculating the absolute error sum of the Aβ-PET images after intensity normalization processing and the Aβ-MRI brain atlas MNI152 according to the following formula, and the formula is as follows: ； where I ref is the normalized Αβ-MRI brain atlas MNI152, i is the pixel index, SAD k is I k is the image level absolute error sum between I ref and I performing anomaly detection on the Aβ-PET images after intensity normalization processing according to the following formula: ； where SAD mean and SAD std are the mean and standard deviation of SAD for all participants, respectively; if the value of DIF k is greater than 0, then the data of participant k is considered to have potential amyloid deposition abnormality and is rejected, and the remaining Aβ-PET images are taken as the normalized Aβ-PET images after screening. 5.The deep learning based Aβ-PET brain template construction and quantitative analysis method of claim 4, characterized in that, performing average calculation according to the following formula: ； where i represents the pixel index, n represents the number of images used for averaging, I k represents the kth participant intensity normalized image, I represents the preliminary Aβ-PET brain template obtained by averaging. 6.The deep learning based Aβ-PET brain template construction and quantitative analysis method of claim 5, wherein, The method further comprises: The regional anatomical difference between the final Aβ-PET brain template and the MNI152 image of Aβ-MRI brain atlas is calculated based on the mean square difference algorithm MSD and the normalized cross-correlation algorithm NCC, and the calculation formula is: ； ； where I represents the final Aβ-PET brain template constructed, I ref represents the Aβ-MRI brain atlas MNI152, i represents the pixel index, n roi represents the number of pixels in the region of interest, i_roi represents the pixel index in the region of interest, I mean_roi and I ref_mean_roi respectively represent the images I and I ref the average value of pixels in the region of interest; Wherein, the smaller the MSD value, the smaller the regional anatomical difference of two images; the larger the MSD value, the larger the regional anatomical difference of two images; the closer the NCC value to 1, the smaller the regional anatomical difference of two images; the closer the NCC value to-1, the larger the regional anatomical difference of two images; the closer the NCC value to 0, the two images are not related. 7.The deep learning based Aβ-PET brain template construction and quantitative analysis method according to claim 6, characterized in that, said first quantitative parameters comprise a mean gray value Gray mean , a mean standard uptake value SUV mean and a maximum uptake value ratio SUV max ; said second quantitative parameters comprise a standard uptake value ratio SUVr, a z-score value Z-score and an amyloid index Amyloid Index calculated in comparison with the brainstem deposition value. 8.The deep learning based Aβ-PET brain template construction and quantitative analysis method of claim 7, wherein, The standard uptake value ratio SUVr, the z-score value Z-score and the whole brain amyloid index AmyloidIndex are calculated according to the following formula: ； ； ； ； ； Wherein, R represents the individual image registered with the final Aβ-PET brain template constructed, S represents the individual image after converting the voxel value to the SUV value, i represents the pixel index, roi represents the region of interest, n represents the number of pixels, ref represents the cerebellum region of the image itself, and database represents the brain amyloid deposition data library; Wherein, the calculation formula of the SUV value is: ； In the case of using the first quantitative parameter to construct the brain amyloid deposition data library, the individual image is the image of a normal participant; in the case of using the second quantitative parameter for quantitative analysis, the individual image is the Aβ-PET positive patient image to be analyzed. 9.The deep learning based Aβ-PET brain template construction and quantitative analysis method of claim 8, wherein, Image registration is performed using the VoxelMorph framework model. 10.The deep learning based Aβ-PET brain template construction and quantitative analysis method of claim 9, wherein, Aβ-PET images of old people of 60-80 age groups and different genders, which are negative for Aβ-amyloid tracer and normal for PET images, are obtained to construct an Aβ-PET brain template of old people, and a brain amyloid deposition data library corresponding to the template is quantitatively analyzed.