An a beta-pet brain image standardization uptake rate quantitative analysis method combined with artificial intelligence algorithm

Abstract

This invention relates to the field of brain imaging technology, specifically a method for quantitative analysis of standardized uptake rates in Aβ-PET brain images combined with artificial intelligence algorithms. The method involves inputting pre-processed single-modality Aβ-PET images into a pre-trained generative model to generate high-resolution images; inputting the synthesized structural magnetic resonance images into a pre-trained brain region segmentation model to obtain a segmentation mask containing multiple brain regions; and extracting brain region signals from the Aβ-PET images based on the segmentation mask to calculate the standardized uptake ratio of each brain region. The advantages of this invention are: converting single-modality Aβ-PET images into high-resolution synthesized structural magnetic resonance images, obtaining clear anatomical structural information in subsequent analysis; achieving high-precision analysis relying solely on single-modality images, overcoming the limitations of existing technologies that require the acquisition and registration of multi-modality images, and significantly simplifying data acquisition requirements and the complexity of subsequent processing.

Description

Technical Field

[0001] This invention relates to the field of brain imaging technology, and in particular to a quantitative analysis method for standardized uptake rate of Aβ-PET brain images that combines artificial intelligence algorithms. Background Technology

[0002] In the study of neurodegenerative diseases, particularly Alzheimer's disease (AD), abnormal deposition of β-amyloid (Aβ) protein is one of the key pathological markers. Aβ-PET (positron emission tomography) imaging is currently the authoritative non-invasive imaging method for detecting Aβ deposition in the living brain. This technique uses a radioactive tracer to specifically bind to Aβ plaques, providing a direct visual representation of their distribution in the brain, thus offering crucial information for pathological assessment and disease progression monitoring.

[0003] To achieve accurate quantitative analysis of Aβ deposition, statistical analysis of Aβ-PET images based on brain region-based standardized uptake values ​​(SUVs) or standardized uptake ratios (SUVRs) has become standard practice in clinical and research settings. However, this process faces significant technical challenges. First, the inherently low spatial resolution of Aβ-PET images (typically 4-5 mm voxels) leads to blurred boundaries of brain tissue structures, particularly poor contrast between gray and white matter. Directly segmenting brain regions on these low-resolution PET images results in significant boundary recognition errors, severely impacting the accuracy of subsequent SUV value statistics within specific brain regions.

[0004] To overcome the insufficient resolution of PET images, current mainstream technologies, such as those based on FreeSurfer or SPM software, typically rely on high-resolution structural magnetic resonance imaging (MRI) as a reference for brain region segmentation. Specifically, this involves first acquiring 3D-T1 structural MRI images of the same patient, then aligning the PET images to MRI space using complex image registration algorithms, utilizing the clear anatomical contrast of MRI to achieve precise brain region segmentation, and finally mapping the segmentation results back to the PET images for parameter extraction. While this method improves segmentation accuracy to some extent, it also introduces new limitations: firstly, it requires additional MRI data acquisition, which not only increases the patient's examination time, financial burden, and potential discomfort (such as claustrophobia), but also complicates the examination process; secondly, the registration and segmentation of multimodal images often require significant manual intervention and parameter adjustments, resulting in low analysis efficiency and difficulty in ensuring consistency of results between different centers or operators.

[0005] Therefore, existing technologies suffer from a contradiction between "high-precision segmentation relying on multimodal data" and "insufficient accuracy of single-modal PET data analysis," and there is an urgent need for a method that can achieve high-precision, automated brain region segmentation and quantitative analysis using only Aβ-PET single-modal images. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a quantitative analysis method for standardized uptake rate of Aβ-PET brain images that combines artificial intelligence algorithms, effectively solving the deficiencies of the prior art.

[0007] To achieve the above objectives, one aspect of the present invention provides a method for quantitative analysis of standardized uptake rate in Aβ-PET brain images combined with artificial intelligence algorithms, comprising the following steps:

[0008] S1. Input the preprocessed single-modal Aβ-PET image into the pre-trained generative model to generate high-resolution synthetic structural magnetic resonance images.

[0009] S2. Input the synthesized structural magnetic resonance image into a pre-trained brain region segmentation model to obtain a segmentation mask containing multiple brain regions;

[0010] S3. Based on the segmentation mask, extract brain region signals from the Aβ-PET image and calculate the standardized uptake ratio of each brain region.

[0011] Preferably, any of the above schemes further includes a preprocessing step before the Aβ-PET image is input to the generation model. The preprocessing step includes: spatial standardization and intensity standardization of the Aβ-PET image.

[0012] The spatial standardization refers to registering Aβ-PET images to a standard space using a rigid registration method.

[0013] The intensity standardization includes attenuation correction of Aβ-PET images and calibration of standardized uptake values ​​based on injection dose and body weight.

[0014] Preferably, the generative model is a 3D conditional generative adversarial network, whose generator adopts a densely connected U-Net structure to input low-resolution Aβ-PET images and output high-resolution synthetic structural magnetic resonance images; its discriminator adopts the 3DPatchGAN architecture.

[0015] Preferably, the training loss function of the generative model is a combined loss function, including adversarial loss, pixel-level L1 loss, and structural similarity loss.

[0016] Preferably, in any of the above schemes, the generative model is trained using a dataset containing at least 500 pairs of paired Aβ-PET images and concurrent 3D-T1 structural magnetic resonance images, and the training process includes a two-stage training strategy.

[0017] Preferably, the brain region segmentation model is a cascaded 3D U-Net segmentation framework, including a primary network and a refined network; the primary network is used for coarse segmentation of the whole brain, gray matter, white matter, and cerebrospinal fluid; and the refined network is used for fine segmentation of specific brain regions.

[0018] Preferably, in any of the above schemes, the training loss function of the brain region segmentation model is a combination of minimizing the multi-scale Dice loss and the boundary Hausdorff distance loss.

[0019] Preferably, in any of the above schemes, after obtaining the segmentation mask in step S2, the step further includes performing a 3D morphological closing operation on the segmentation edges for optimization.

[0020] Preferably, in any of the above schemes, the calculation of the standardized uptake ratio of each brain region includes the following steps:

[0021] 1) The signal from the original Aβ-PET image is mapped to the space of the synthesized structural magnetic resonance image through spatial transformation;

[0022] 2) The standardized uptake values ​​of voxels in each brain region are extracted using the segmentation mask;

[0023] 3) Using the segmented cerebellar gray matter region in the synthesized structural magnetic resonance image as a reference region, calculate the standardized uptake ratio of the target brain region.

[0024] Preferably, in any of the above schemes, the method further includes a result visualization step, which includes: setting a positive threshold for each brain region based on data from a healthy control group, and fusing the standardized uptake ratio of each brain region with the anatomical structure of the synthetic structural magnetic resonance image in the form of a heatmap.

[0025] The present invention has the following advantages:

[0026] 1. This method for quantitative analysis of Aβ-PET brain image standardized uptake rate, which combines artificial intelligence algorithms, converts single-modality Aβ-PET images into high-resolution synthetic structural magnetic resonance images, thereby obtaining clear anatomical information in subsequent analysis. This approach enables high-precision analysis based solely on single-modality images, overcoming the limitations of existing technologies that require the acquisition and registration of multimodal images (PET combined with MRI), and significantly simplifying data acquisition requirements and the complexity of subsequent processing.

[0027] 2. This Aβ-PET brain image standardized uptake ratio quantitative analysis method, which combines artificial intelligence algorithms, uses synthesized high-resolution structural magnetic resonance images for brain region segmentation. Compared with segmentation directly on low-resolution, low-contrast original PET images, this method can more accurately identify the boundaries of different brain tissues. This technological improvement effectively reduces boundary recognition errors in the segmentation process, providing a more accurate spatial localization basis for subsequent calculation of standardized uptake ratios based on brain regions, thereby improving the accuracy of the entire quantitative analysis process.

[0028] 3. This Aβ-PET brain image standardized uptake rate quantitative analysis method, which combines artificial intelligence algorithms, integrates the entire process from image input, cross-modal generation, automated segmentation to parameter calculation, and constructs an end-to-end automated processing system. This solution eliminates the image registration, parameter adjustment and other steps that rely heavily on manual intervention in traditional methods. It not only significantly shortens the processing time and improves the analysis efficiency, but also minimizes the subjective bias introduced by human operation, and ensures the consistency and repeatability of the analysis method. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0030] Figure 2 This is a schematic diagram of paired PET-MR data preprocessing according to the present invention;

[0031] Figure 3 This is a schematic diagram of the high-resolution structural MRI generated based on low-resolution PET in this invention;

[0032] Figure 4 This is a schematic diagram of brain structure segmentation based on structural MRI according to the present invention;

[0033] Figure 5 This is a schematic diagram illustrating the quantitative parameter calculation and report generation of the present invention;

[0034] Figure 6 This is a schematic diagram of the fusion display of the SUVR thermal image and the synthetic MRI anatomical structure of the present invention;

[0035] Figure 7 This is a schematic diagram of the abnormal brain region of the present invention. Detailed Implementation

[0036] The present invention will be further described below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0037] like Figures 1 to 7 As shown, a method for quantitative analysis of standardized uptake rate in Aβ-PET brain images combined with artificial intelligence algorithms includes the following steps:

[0038] S1. Construction and preprocessing of multimodal training datasets

[0039] First, a dataset was constructed to train the artificial intelligence model. At least 500 pairs of Aβ-PET images and their corresponding 3D-T1 structural magnetic resonance (MRI) images were collected. To ensure the validity of the data alignment, the time interval between the PET scan and the MRI acquisition for the same subject was required to be no more than 7 days.

[0040] The collected pairing data is preprocessed, including:

[0041] Spatial normalization: A rigid registration algorithm is used to register Aβ-PET images to their contemporaneous 3D-T1 structural MRI image space to eliminate spatial inconsistencies caused by head movement and differences in scanning position.

[0042] Strength standardization:

[0043] Attenuation correction was performed on Aβ-PET images, and standardized uptake values ​​(SUVs) were calculated based on the injection dose and subject weight to complete intensity calibration.

[0044] N4 bias field correction was applied to 3D-T1 structural MRI images, and white matter signal intensity was normalized.

[0045] Data augmentation: To improve the model's generalization ability, data augmentation operations are performed on the preprocessed paired data, including elastic deformation, rotation within ±15°, and injection of random noise within a signal-to-noise ratio (SNR) range of 20-40dB, thereby expanding the training dataset.

[0046] S2. Training of a structural MRI generative model based on conditional generative adversarial networks (cGANs)

[0047] A 3D conditional generative adversarial network is constructed to generate high-resolution synthetic structural MRI images from low-resolution Aβ-PET images.

[0048] Network Architecture: The generator employs a densely connected U-Net structure, taking preprocessed low-resolution Aβ-PET images (voxel size approximately 4 mm³) as input and outputting high-resolution (1 mm³ voxel) synthetic structural MRI images. The discriminator uses a 3DPatchGAN architecture to distinguish the authenticity of local regions between the generated images and real MRI images.

[0049] Loss function: The training loss function L_total of the model is a weighted combination of multiple losses, and its expression is:

[0050]

[0051] in, The Wasserstein distance-based adversarial loss is used for stable training; The L1 loss is applied at the pixel level to ensure the overall consistency between the generated image and the target image; The structural similarity loss is used to preserve the structural information of the image. Gradient penalty (λ=10) is used during training to further stabilize the adversarial training process.

[0052] Training strategy: A two-stage training strategy is adopted. First, with The generator is pre-trained separately for 20,000 steps using the learning rate. Subsequently, with the generator parameters fixed, the generator and discriminator are jointly trained and adversarially fine-tuned for a total of no less than 50,000 training iterations until the model converges.

[0053] S3. Training of a brain region segmentation model based on cascaded 3DU-Net

[0054] A cascaded 3DU-Net segmentation framework is constructed for automated brain region segmentation of generated high-resolution synthetic structural MRI images.

[0055] Network structure: The framework includes a primary segmentation network and a refined segmentation network.

[0056] Primary network: responsible for coarse segmentation of the whole brain and the three major tissue categories of gray matter, white matter, and cerebrospinal fluid.

[0057] The refined network uses the segmentation results from the previous stage as input to perform refined segmentation of multiple specific brain regions of interest, such as the hippocampus, prefrontal cortex, and posterior cingulate cortex. During training, labels obtained from segmentation of real 3D-T1 MRI images using specialized software (such as FreeSurfer) are used as the gold standard.

[0058] Loss Function: The segmentation model is trained to minimize the combination of multi-scale Dice loss and boundary Hausdorff distance loss, with a weight ratio of 0.6:0.4. This loss function is designed to simultaneously optimize region overlap and boundary fitting accuracy. The Adam optimizer is used, with an initial learning rate set to... Conduct training.

[0059] S4. Single-modal Aβ-PET image processing and quantitative analysis workflow

[0060] For a single Aβ-PET image to be analyzed, process it according to the following steps:

[0061] S4.1 Input Image Preprocessing: Spatial standardization (registration to standard template space) and intensity standardization (attenuation correction and SUV calibration) are performed on the Aβ-PET images to be analyzed. The processing method is the same as that for training data preprocessing.

[0062] S4.2 Synthetic MRI Generation: The preprocessed Aβ-PET image is input into the trained generative model, which automatically performs cross-modal feature mapping and outputs a high-resolution (1 mm³) synthetic structural MRI image.

[0063] S4.3 Automated Brain Region Segmentation: Input the synthetic structural MRI images generated in step 2) into the trained cascaded 3DU-Net segmentation model, which automatically outputs binary segmentation masks containing each target brain region. To further optimize the segmentation results, a 3D morphological closing operation (using 3x3x3 spherical structural elements) can be performed on the edges of the segmentation mask to eliminate voxel-level jagged artifacts.

[0064] S4.4 Calculation of PET signal extraction to normalized uptake ratio (SUVR):

[0065] a) Spatial mapping: The signal of the original Aβ-PET image is resampled to the space of the synthetic structure MRI image generated in step 2) through inverse spatial transformation.

[0066] b) Signal extraction: Using the segmentation mask obtained in step 3), extract the standardized uptake values ​​of all voxels in the target brain region and calculate their average value (SUVmean).

[0067] c) Ratio Calculation: Using the segmented cerebellar gray matter region from the synthetic structural MRI image as the reference area, its SUVmean is calculated. The standardized uptake ratio (SUVR) of the target brain region is calculated using the following formula:

[0068] SUVR = (SUVmean of target brain region) / (SUVmean of cerebellar gray matter region)

[0069] S5. Result Visualization and Output: Based on a database of healthy controls, a positive SUVR threshold is set for each brain region (e.g., SUVR ≥ 1.21 can be set as positive). The calculated SUVR values ​​for each brain region are encoded in heatmap form and fused with the anatomical image of the synthetic structural MRI generated in step 2) for 3D rendering, generating an interactive Aβ deposition distribution map that supports multi-planar (coronal, sagittal, axial) reconstruction. Simultaneously, a structured report table containing the name of each brain region, SUVmean, and SUVR value is generated (e.g., ...). Figure 5 (as shown), for further analysis.

[0070] Example 1:

[0071] This embodiment provides a complete implementation process for a quantitative analysis method of standardized uptake rate in Aβ-PET brain images that combines artificial intelligence algorithms.

[0072] 1. Data preparation and model training phase

[0073] a. Training Dataset Construction: 520 pairs of paired Aβ-PET images and concurrent 3D-T1 structural MRI images were collected. Each image pair underwent preprocessing, including: rigid registration (6 degrees of freedom) to align the Aβ-PET images to their corresponding MRI space; attenuation correction and normalized uptake (SUV) calibration based on injection dose and body weight were applied to the PET images; and N4 bias field correction was applied to the MRI images. Subsequently, random elastic deformation (maximum displacement field ±3 voxels) and ±12-degree rotation were applied to the preprocessed paired data for data augmentation, constructing the final training set.

[0074] b. Generative Model Training: A 3D Conditional Generative Adversarial Network (GAN) is constructed. The generator is a U-Net with densely connected blocks, taking low-resolution Aβ-PET images (4×4×4 mm³ voxels) as input and outputting high-resolution synthetic MRI images (1×1×1 mm³ voxels). The discriminator is a 3DPatchGAN. Training is performed using a dataset containing at least 500 pairs of augmented paired data, with the following loss function combinations:

[0075] Where Ladv is the Wasserstein GAN loss with gradient penalty (λ=10). Training employs a two-stage strategy, first using... The learning rate pre-trained generator was trained for 20,000 steps, and then...

[0076] The generator and discriminator were trained together using a learning rate for 50,000 steps.

[0077] c. Segmentation Model Training: Constructing a cascaded 3D U-Net segmentation framework. The primary network performs coarse segmentation of the whole brain, gray matter, white matter, and cerebrospinal fluid; the refinement network refines the primary segmentation results, segmenting 90 brain regions including the hippocampus and prefrontal cortex. The model is trained using real MRI and its gold standard segmentation labels. The loss function is the sum of the multi-scale Dice loss (weight 0.6) and the boundary Hausdorff distance loss (weight 0.4), using the Adam optimizer (initial learning rate 5×). Train until convergence.

[0078] 2. Single-modal PET image analysis stage

[0079] For a new, unique Aβ-PET image, process it as follows:

[0080] a. Preprocessing: Spatial normalization (registration to MNI152 standard space) and intensity normalization (attenuation correction and SUV calibration) are performed on the PET image.

[0081] b. Synthetic MRI generation: The preprocessed PET image is input into the trained generative model, and the model automatically outputs a high-resolution synthetic 3D-T1 structure MRI image.

[0082] c. Brain region segmentation: The synthesized MRI images are input into the trained cascaded segmentation model, which automatically obtains binary segmentation masks for each brain region. A 3×3×3 voxel morphological closing operation is performed on the segmentation mask to smooth the boundaries.

[0083] d. Quantitative parameter calculation: The SUV values ​​of the original PET images are mapped to the synthetic MRI space through spatial transformation. SUV values ​​of all voxels within each brain region (e.g., the left hippocampus) are extracted using a segmentation mask, and their average value (SUVmean) is calculated. The SUVmean of the segmented cerebellar gray matter region from the synthetic MRI is calculated using the same reference region. The formula is:

[0084]

[0085] Calculate the standardized uptake ratio for each brain region.

[0086] e. Results Visualization: Based on a healthy control group database, a positive threshold for SUVR was set for each brain region (e.g., global cortical mean ≥1.21). The calculated SUVR values ​​for each brain region were overlaid on the anatomical background of the synthetic MRI in the form of a pseudo-color heatmap to generate a three-dimensional interactive Aβ deposition distribution map, and a quantitative report table containing the name of each brain region, SUVmean, and SUVR was output.

[0087] Example 2:

[0088] This embodiment provides an implementation method that focuses on efficient and stable treatment in a clinical setting. Its core process is consistent with that of Embodiment 1 and the claims, but some technical details have been optimized to meet the needs of rapid treatment.

[0089] 1. Model Training and Deployment

[0090] a. Lightweight Generative Model: The generative model employs a computationally more efficient lightweight generator, with approximately 70% of the parameters of the standard U-Net, to improve inference speed. The training data consists of 600 paired datasets, and the loss function remains the same. , and The model uses a combination of methods, but employs an end-to-end one-time training strategy with a cosine annealing learning rate and a total training iteration of 80,000 steps to ensure model accuracy.

[0091] b. Integrated Segmentation Model: An optimized, non-cascaded 3DU-Net model is used to segment all target brain regions in the whole brain in one go. During training, this model uses the same combination of multi-scale Dice loss and boundary loss as in Example 1, but the network depth is adjusted to achieve a balance between segmentation speed and accuracy. After the model is trained, it is integrated with the generative model into an end-to-end inference pipeline.

[0092] 2. Integrated analysis process

[0093] a. Integrated Preprocessing and Inference: The spatial and intensity normalization steps of Aβ-PET images are integrated with the inference processes of the generation and segmentation models into a single automated script. Users only need to provide the raw PET image files, and the system automatically completes all steps from registration and SUV calibration to generating synthetic MRI and segmenting brain regions.

[0094] b. Rapid Parameter Calculation and Report Generation: After obtaining the segmentation mask, the system directly extracts the SUV values ​​of each brain region in the original PET image space (after SUV calibration) using the mapped segmentation mask, and calculates the SUVmean and SUVR with the cerebellar gray matter as the reference area. This step avoids additional resampling operations, further accelerating the process. The system has a built-in standardized positive threshold and automatically highlights brain regions with SUVR exceeding the threshold in the report.

[0095] c. Output: The entire processing can be completed within 5 minutes (depending on hardware). It automatically generates a structured PDF report containing a list of SUVR values ​​for key brain regions and a two-dimensional synthetic MRI axial slice image annotated with anomalous deposition areas, providing intuitive evidence for rapid clinical assessment.

[0096] Summary of Examples: Both examples above implement the core steps described in claim 1: obtaining synthetic MRI from single-modal Aβ-PET images through a generative model, and then segmenting and calculating SUVR based on the synthetic MRI. Example 1 details the complete technical details, including complex model training; Example 2 depicts an optimized scheme focusing on deployment efficiency and process integration. Both are within the scope of protection defined by the material description and claims, and are sufficiently disclosed to be implemented.

[0097] In summary, this invention converts single-modality Aβ-PET images into high-resolution synthetic structural magnetic resonance images, thereby obtaining clear anatomical information in subsequent analysis. This approach achieves high-precision analysis relying solely on single-modality images, overcoming the limitations of existing technologies that require the acquisition and registration of multi-modality images. It significantly simplifies data acquisition requirements and the complexity of subsequent processing. By utilizing synthetic high-resolution structural magnetic resonance images for brain region segmentation, compared to segmentation directly on low-resolution, low-contrast raw PET images, this method can more accurately identify the boundaries of different brain tissues. This technology... The improvement effectively reduced boundary recognition errors during segmentation, providing a more accurate spatial localization basis for subsequent calculation of standardized uptake ratios based on brain regions. This improved the accuracy of the entire quantitative analysis process. The system integrates the entire process from image input, cross-modal generation, automated segmentation to parameter calculation, and constructs an end-to-end automated processing system. This solution eliminates the steps of image registration and parameter adjustment that rely heavily on manual intervention in traditional methods. It not only significantly shortens processing time and improves analysis efficiency, but also minimizes subjective biases introduced by human operation, ensuring the consistency and repeatability of the analysis method.

[0098] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An Aβ-PET brain image standardized uptake ratio quantitative analysis method combined with an artificial intelligence algorithm, characterized in that: Includes the following steps: S1. Input the preprocessed single-modal Aβ-PET image into the pre-trained generative model to generate high-resolution synthetic structural magnetic resonance images. S2. Input the synthesized structural magnetic resonance image into a pre-trained brain region segmentation model to obtain a segmentation mask containing multiple brain regions; S3. Based on the segmentation mask, extract brain region signals from the Aβ-PET image and calculate the standardized uptake ratio of each brain region.

2. The AI algorithm-integrated Aβ-PET brain image standardized uptake ratio quantification analysis method according to claim 1, characterized by: The Aβ-PET image input generation model includes a preprocessing step, which includes spatial normalization and intensity normalization of the Aβ-PET image. The spatial standardization refers to registering Aβ-PET images to a standard space using a rigid registration method. The intensity standardization includes attenuation correction of Aβ-PET images and calibration of standardized uptake values ​​based on injection dose and body weight.

3. The method for quantitative analysis of standardized uptake rate of Aβ-PET brain images combined with artificial intelligence algorithms according to claim 1, characterized in that: The generative model is a 3D conditional generative adversarial network. Its generator adopts a densely connected U-Net structure to take low-resolution Aβ-PET images as input and output high-resolution synthetic structural magnetic resonance images. Its discriminator adopts the 3DPatchGAN architecture.

4. The method for quantitative analysis of standardized uptake rate of Aβ-PET brain images combined with artificial intelligence algorithms according to claim 3, characterized in that: The training loss function of the generative model is a combined loss function, which includes adversarial loss, pixel-level L1 loss, and structural similarity loss.

5. The method for quantitative analysis of standardized uptake rate of Aβ-PET brain images combined with artificial intelligence algorithms according to claim 4, characterized in that: The generative model is trained using a dataset containing at least 500 pairs of paired Aβ-PET images and concurrent 3D-T1 structural magnetic resonance images. The training process includes a two-stage training strategy.

6. The method for quantitative analysis of standardized uptake rate of Aβ-PET brain images combined with artificial intelligence algorithms according to claim 1, characterized in that: The brain region segmentation model is a cascaded 3D U-Net segmentation framework, including a primary network and a refined network; the primary network is used for coarse segmentation of the whole brain, gray matter, white matter, and cerebrospinal fluid; the refined network is used for fine segmentation of specific brain regions.

7. The method for quantitative analysis of standardized uptake rate of Aβ-PET brain images combined with artificial intelligence algorithms according to claim 6, characterized in that: The training loss function of the brain region segmentation model is a combination of minimizing the multi-scale Dice loss and the boundary Hausdorff distance loss.

8. The method for quantitative analysis of standardized uptake rate of Aβ-PET brain images combined with artificial intelligence algorithms according to claim 1, characterized in that: After obtaining the segmentation mask in step S2, the method further includes performing a 3D morphological closing operation on the segmentation edges for optimization.

9. The method for quantitative analysis of standardized uptake rate of Aβ-PET brain images combined with artificial intelligence algorithms according to claim 1, characterized in that: The calculation of the standardized uptake ratio for each brain region includes the following steps: 1) The signal from the original Aβ-PET image is mapped to the space of the synthesized structural magnetic resonance image through spatial transformation; 2) The standardized uptake values ​​of voxels in each brain region are extracted using the segmentation mask; 3) Using the segmented cerebellar gray matter region in the synthesized structural magnetic resonance image as a reference region, calculate the standardized uptake ratio of the target brain region.

10. The method for quantitative analysis of standardized uptake rate of Aβ-PET brain images combined with artificial intelligence algorithms according to claim 1, characterized in that: The method further includes a result visualization step, which includes: setting a positive threshold for each brain region based on data from a healthy control group, and fusing the standardized uptake ratio of each brain region with the anatomical structure of the synthetic structural magnetic resonance image in the form of a heatmap.

Images