An automated quantitative and intelligent analysis method based on Aβ-PET brain imaging

Abstract

The present invention proposes an automated quantitative and intelligent analysis method based on Aβ-PET brain imaging, including: preprocessing the Aβ-PET raw data to obtain brain parenchyma data; automatically quantitatively analyzing the brain parenchyma data to calculate voxel SUVMean and voxel SUVR; calculating the average SUVR of each brain region based on the brain structure mask to achieve brain region-level quantification; generating a quantitative parameter map based on voxel SUVMean, voxel SUVR and the average SUVR of each brain region; constructing a standard reference table; using a trained multimodal 3D convolutional neural network to classify the quantitative parameter map, and obtaining the top 5% contribution voxels as suspected abnormal voxels; judging brain region abnormalities based on the standard reference table and suspected abnormal voxels, and realizing automatic positioning of abnormal brain regions. Through the scheme of the present invention, quantitative analysis of Aβ-PET raw data is realized, and while providing voxel-level quantitative data, accurate quantitative parameters based on anatomical partitions are provided and visual output of abnormal area images are output, which provides useful help for clinicians' diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to an automated quantitative and intelligent analysis method based on Aβ-PET brain images. Background Art

[0002] As a traditional functional imaging method, positron emission tomography (PET) has been widely used in the clinical diagnosis of neurodegenerative diseases including Alzheimer's disease, frontotemporal dementia, Lewy body dementia, etc. In recent years, with the popularization of PET imaging equipment, many hospitals have purchased PET imaging equipment.

[0003] Alzheimer's disease (AD) is a common neurodegenerative disease. According to the amyloid protein (Aβ) cascade hypothesis, the deposition of Aβ in the brain is the initiation and core event of AD disease pathology. Aβ-based PET brain imaging can intuitively display the deposition and distribution of Aβ in brain tissue, providing an effective means for early diagnosis, differential diagnosis and follow-up of AD. In recent years, in the field of Aβ-PET, China has obtained NMPA approval for marketing and is routinely used in clinical practice. The research and development, evaluation system, and clinical application of Aβ-PET-related imaging agents have made rapid progress, and relevant expert consensus has been released, "Expert Consensus on the Application of Amyloid Protein PET Imaging in the Diagnosis of Alzheimer's Disease".

[0004] Visual analysis, quantitative analysis, and semi-quantitative analysis are required for Aβ-PET film reading. Visual analysis is performed by comparing the uptake of imaging agents in the gray matter of the cerebral cortex and the adjacent white matter areas. Through visual analysis, each patient can be evaluated and classified as "amyloid PET positive" or "amyloid PET negative". Since visual analysis is relatively subjective and semi-quantitative analysis is being standardized, quantitative and semi-quantitative analysis of PET images is currently being performed in clinical practice. In the more widely used static imaging of the plateau phase, the SUVr obtained by dividing the mean of standardized uptake value (SUVmean) of the volume of interest (VOI) by the SUVmean of the reference area is often used as a semi-quantitative measurement method. In addition to SUVr, Centiloid can also be used as a PET quantification unit, which linearly scales and standardizes the quantitative value to 0~100 to achieve comparability of results between different imaging agents, data acquisition, and analysis methods.

[0005] At present, for most hospitals equipped with PET imaging equipment, doctors do not have much experience in judging abnormal areas of brain Aβ-PET. The existing visual analysis or quantitative and semi-quantitative analysis methods are highly dependent on the doctor's own experience. Visual analysis based on the naked eye has great subjective differences and low accuracy. The quantitative analysis method based on the volume of interest (VOI) is very time-consuming for the selection and delineation of VOI, and it is also unable to avoid the subjective differences in the selection and delineation of artificial areas, making it difficult to output objective and consistent diagnostic conclusions between doctors. The regional identification of abnormal Aβ deposition and the fine anatomical division can help to more accurately assess the condition, but this is difficult for inexperienced doctors to do. Summary of the invention

[0006] The present invention is based on the above-mentioned problems and proposes an automated quantitative and intelligent analysis method based on Aβ-PET brain images. Through the scheme of the present invention, quantitative analysis of Aβ-PET raw data is realized, and while providing voxel-level quantitative data, accurate quantitative parameters based on anatomical partitions are provided and images of abnormal areas are visually output, which provides useful assistance to clinicians in diagnosis.

[0007] In view of this, one aspect of the present invention proposes an automated quantitative and intelligent analysis method based on Aβ-PET brain imaging, comprising:

[0008] The Aβ-PET raw data were preprocessed to obtain the first brain parenchyma data;

[0009] Performing automatic quantitative analysis on the first brain parenchyma data to calculate voxel SUVMean and voxel SUVR;

[0010] Based on the predefined brain structure mask, the average SUVR of each brain region was calculated to achieve brain region-level quantification;

[0011] Generate a quantitative parameter map according to the voxel SUVMean, the voxel SUVR and the average SUVR of each brain region;

[0012] construct standard reference tables;

[0013] The trained multimodal 3D convolutional neural network was used to classify the quantitative parameter map, and the top 5% contribution voxels were obtained as suspected abnormal voxels;

[0014] The abnormality of the brain region is determined according to the standard reference table and the suspected abnormal voxels, so as to automatically locate the abnormal brain region.

[0015] Optionally, the step of preprocessing the Aβ-PET raw data to obtain the first brain parenchyma data includes:

[0016] Registering the PET image data in the Aβ-PET raw data with a standard brain template to achieve spatial standardization;

[0017] Obtaining a meningeal mask from the standard brain template;

[0018] The skull data and background data are removed by using the meningeal mask, the brain parenchyma data are extracted, and the brain parenchyma data are subjected to data standardization processing to obtain the first brain parenchyma data.

[0019] Optionally, the step of registering the PET image data in the Aβ-PET raw data with a standard brain template to achieve spatial standardization includes:

[0020] Select the standard brain template;

[0021] Preprocessing the standard brain template and the PET image data;

[0022] Select the registration algorithm;

[0023] Using the selected registration algorithm, the preprocessed standard brain template, and the PET image data, the affine and nonlinear transformation matrices that best match the two images are calculated through multiple iterative optimizations;

[0024] transforming the PET image data using the nonlinear transformation matrix;

[0025] The matching accuracy after registration is evaluated by image superposition and reference marker error.

[0026] When the matching accuracy reaches a first preset matching accuracy, the registered individual PET image is output.

[0027] Optionally, the step of constructing a standard reference table includes:

[0028] Collect Aβ-PET data of a preset number of healthy controls;

[0029] The Aβ-PET data of the healthy control group were automatically quantitatively analyzed, and the mean value and standard deviation of SUVR in each brain region in the healthy control group were calculated.

[0030] Optionally, the classifying the quantitative parameter map using a trained multimodal 3D convolutional neural network includes:

[0031] Multimodal quantitative parameter maps were collected for individuals with Alzheimer's disease and Aβ-negative normal controls;

[0032] Label the multimodal quantitative parameter maps as abnormal and normal to generate a labeled training set;

[0033] Mapping the labeled training set into a three-dimensional space consistent with the PET image to form a multi-channel quantitative parameter input, each channel being a reference quantitative parameter value;

[0034] The multi-channel quantitative parameter input is used as the input source to input the preset 3D convolutional neural network model for training. The network will learn the feature representation function for distinguishing abnormalities and normality in the multi-modal quantitative parameter map to obtain the first recognition model;

[0035] The quantitative parameter map is classified using the first recognition model, and the voxel with the largest contribution is located using the CAM heat map as the suspected abnormal voxel.

[0036] Optionally, the step of automatically performing quantitative analysis on the first brain parenchyma data to calculate voxel SUVMean and voxel SUVR comprises:

[0037] According to the first brain parenchyma data, counting the radioactivity concentration value of each voxel;

[0038] The voxel SUVMean is obtained by averaging all voxels in the brain parenchyma;

[0039] Select the reference area and calculate the SUVMean of the voxels in the reference area;

[0040] Calculate the test voxel SUV value and divide it by the reference area SUVMean to obtain the voxel SUVR;

[0041] The above calculations are repeated for all brain parenchymal voxels, and the voxel SUVMean and voxel SUVR of each voxel are output.

[0042] Optionally, the step of calculating the average SUVR of each brain region based on a predefined brain structure mask to achieve brain region-level quantification includes:

[0043] generating a three-dimensional mask of each brain region using the standard brain template;

[0044] matching the three-dimensional mask to the PET image space by registration;

[0045] The PET image space was divided into brain regions according to the matched three-dimensional mask, and the SUVR of each voxel was counted and averaged;

[0046] The average SUVR value of each brain region was output as a quantitative indicator.

[0047] Optionally, the step of generating a quantitative parameter map according to the voxel SUVMean, the voxel SUVR and the average SUVR of each brain region comprises:

[0048] Mapping the voxel SUVMean, the voxel SUVR and the average SUVR of each brain region into the same three-dimensional image space;

[0049] Resampling the voxel SUVMean data to generate a three-dimensional voxel SUVMean distribution map;

[0050] Resampling the voxel SUVR data to generate a three-dimensional voxel SUVR distribution map;

[0051] Assigning the average SUVR of each brain region to the corresponding three-dimensional brain region mask area;

[0052] The three-dimensional voxel SUVMean distribution map, the three-dimensional voxel SUVR distribution map and the assigned three-dimensional brain area mask area are presented to generate a distribution map of quantitative parameters in the three-dimensional brain space to obtain the quantitative parameter map.

[0053] Optionally, the step of determining the abnormality of the brain region according to the standard reference table and the suspected abnormal voxels to automatically locate the abnormal brain region includes:

[0054] According to the brain region mask, the proportion of suspected abnormal voxels in each brain region was counted;

[0055] The average SUVR of each brain region was calculated;

[0056] The mean and standard deviation of SUVR in each brain region were obtained according to the standard reference table;

[0057] Calculate Z = (average SUVR of brain region - mean) / standard deviation;

[0058] A first threshold is set, and when the Z value is higher than the first threshold, the corresponding brain region is identified as an abnormal region.

[0059] Optionally, the step of generating a three-dimensional mask of each brain region using the standard brain template comprises:

[0060] Reading the template image and the contour annotation data of each brain region from the standard brain template respectively;

[0061] The labeled data is copied into a three-dimensional matrix consistent with the pixels of the template image;

[0062] Assign the pixel value of the corresponding brain area to 1 and the other parts to 0 to generate a three-dimensional matrix of the labeled mask;

[0063] Repeat the above process for each brain region to generate a separate mask;

[0064] All brain region masks were superimposed to generate a comprehensive mask;

[0065] Read the boundary data of each brain region according to the template image coordinates;

[0066] Generate the correspondence between the comprehensive mask and the template structure;

[0067] Save the integrated mask in image format.

[0068] The technical solution of the present invention is adopted, and the automated quantitative and intelligent analysis method based on Aβ-PET brain images includes: preprocessing the Aβ-PET raw data to obtain the first brain parenchyma data; automatically quantitatively analyzing the first brain parenchyma data to calculate the voxel SUVMean and voxel SUVR; calculating the average SUVR of each brain region based on the predefined brain structure mask to achieve brain region-level quantification; generating a quantitative parameter map according to the voxel SUVMean, the voxel SUVR and the average SUVR of each brain region; constructing a standard reference table; using the trained multimodal 3D convolutional neural network to classify the quantitative parameter map, and obtaining the top 5% contribution voxels as suspected abnormal voxels; judging brain region abnormalities according to the standard reference table and the suspected abnormal voxels, and realizing automatic positioning of abnormal brain regions. Through the solution of the present invention, quantitative analysis of Aβ-PET raw data is realized, and while providing voxel-level quantitative data, accurate quantitative parameters based on anatomical partitions are provided and visual output of abnormal area pictures are output, which provides useful help for clinicians' diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a flow chart of an automated quantitative and intelligent analysis method based on Aβ-PET brain imaging provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0070] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0071] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0072] The terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units that are not listed, or optionally includes other steps or units inherent to these processes, methods, products or devices.

[0073] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0074] Refer to the following Figure 1 To describe an automated quantification and intelligent analysis method based on Aβ-PET brain imaging provided according to some embodiments of the present invention.

[0075] like Figure 1 As shown, one embodiment of the present invention provides an automated quantitative and intelligent analysis method based on Aβ-PET brain imaging, comprising:

[0076] The Aβ-PET raw data were preprocessed to obtain the first brain parenchyma data;

[0077] Performing automatic quantitative analysis on the first brain parenchyma data to calculate voxel SUVMean and voxel SUVR;

[0078] Based on the predefined brain structure mask, the average SUVR of each brain region was calculated to achieve brain region-level quantification;

[0079] The previous step is to calculate voxel-level quantitative parameters (such as SUVmean, etc.) for the entire brain parenchyma data, but each voxel is analyzed separately, and the result lacks anatomical distinction; in order to achieve anatomical positioning, it is necessary to use brain structure information for quantitative analysis; in this step, a predefined brain structure mask is used to achieve this; the brain structure mask is generated based on a standard brain template, and the outlines of each brain region are drawn. This mask can be used to distinguish different brain regions in the brain image, and the voxel SUVR values ​​in each brain region are counted, and the average SUVR of the brain region is calculated, thus achieving quantitative analysis from voxel to brain region. Brain region-level quantification aims to give the degree of uptake of each brain region to facilitate quantitative diagnosis. Therefore, this step uses brain region masks to achieve a higher level of anatomical quantitative analysis based on the voxel-level quantification in the previous step.

[0080] Generate a quantitative parameter map according to the voxel SUVMean, the voxel SUVR and the average SUVR of each brain region;

[0081] construct standard reference tables;

[0082] In this step, the standard reference table refers to a reference database obtained by analyzing a large number of standard Aβ-PET image data, which records the average SUVR value of each brain region (such as amygdala, supplementary motor area, etc.), the standard deviation of the average SUVR value of each brain region, etc. The method for obtaining this table is: collect Aβ-PET image data of a preset number of healthy control individuals who have been diagnosed as normal; automatically quantitatively process these images to obtain the SUVR value of each brain region; count all standard SUVR values ​​of each brain region to obtain the average value and standard deviation; save the average SUVR value and standard deviation of each brain region as a database as a reference value. This standard reference table records the average distribution of SUVR in each brain region in the normal population, which can be used as an important reference for subsequent judgment of whether an individual brain region is abnormal. It is obtained through a large number of samples and reflects the quantitative law at the population level.

[0083] The trained multimodal 3D convolutional neural network was used to classify the quantitative parameter map, and the top 5% contribution voxels were obtained as suspected abnormal voxels;

[0084] In this step, the multi-source quantitative parameter map of the object to be detected (such as a patient) is input into the network using the trained 3D convolutional neural network model; the network will classify the map and predict whether it is AD or normal; at the same time, the network can output the contribution of each voxel to the final classification result; the contribution indicates the degree of quantitative abnormality of the voxel; all voxels are sorted from high to low according to the contribution, and the top 5% of the voxels are extracted as suspected abnormal voxels; these voxels have the largest quantitative deviation and are more likely to be real abnormal voxels. In other words, a small number of voxels in the quantitative parameter map that contribute the most to the classification result are identified through the network as the initial location of the abnormal voxel area, providing clues for subsequent detection.

[0085] The abnormality of the brain region is determined according to the standard reference table and the suspected abnormal voxels, so as to automatically locate the abnormal brain region.

[0086] The technical solution of this embodiment is used to achieve quantitative analysis of Aβ-PET raw data, provide voxel-level quantitative data, provide accurate quantitative parameters based on anatomical partitions and visualize output abnormal area images, which provides useful assistance to clinicians in diagnosis.

[0087] In some possible embodiments of the present invention, the step of preprocessing the Aβ-PET raw data to obtain the first brain parenchyma data includes:

[0088] Registering the PET image data in the Aβ-PET raw data with a standard brain template to achieve spatial standardization;

[0089] Obtaining a meningeal mask from the standard brain template;

[0090] In this step, the meningeal mask refers to a binary image mask generated based on a standard brain template. The standard brain template usually contains quality contours of various structures in the brain. By processing the template image, the boundary lines of cerebrospinal fluid (CSF) and non-brain tissues such as the skull can be obtained, and then this CSF-skull boundary line is converted into a binary mask image; in this binary mask image, the values ​​of the cerebrospinal fluid and skull parts are 1, and the values ​​of the brain parenchyma part are 0. This binary mask based on the standard brain template that depicts the boundary of CSF and skull is called a meningeal mask. The meningeal mask can be used to perform dot multiplication on the original PET image to be processed, so that the PET image can be automatically segmented into two parts, the brain and the brain, according to the mask, so as to remove the skull and background and extract the sub-image data containing only the brain parenchyma signal.

[0091] The skull data and background data are removed by using the meningeal mask, the brain parenchyma data are extracted, and the brain parenchyma data are subjected to data standardization processing to obtain the first brain parenchyma data.

[0092] In this embodiment, some outliers may appear during the process of extracting brain parenchyma data, which requires additional processing:

[0093] 1. High outlier points

[0094] Possible reasons: scanning error, reading error, etc.

[0095] Processing method: Statistically calculate the SUV value distribution of each voxel in the brain parenchyma data; use the 99.7% quantile in the data distribution as the upper threshold; points above the threshold are identified as abnormal and replaced by the threshold.

[0096] 2. Low signal value point

[0097] Possible reasons: insufficient scanning or defect in the edge of brain parenchyma, etc.

[0098] Processing method: Calculate the lower limit of the SUV value distribution and use the 0.3% quantile as the lower threshold; points below the threshold are identified as abnormal and replaced by the threshold.

[0099] 3. Obvious outlier areas

[0100] Possible causes: foreign objects such as hair gaps.

[0101] Processing method: Calculate the SUV difference between each voxel and its 26 nearest neighboring voxels; identify areas with too large a difference (such as outside 3 standard deviations) as outliers; selectively replace or delete after manual inspection.

[0102] The above methods can effectively identify and correct various types of outliers in the extraction process, ensuring the reliability of subsequent quantitative analysis. At the same time, manual review is required to ensure that no valid information is lost.

[0103] In some possible embodiments of the present invention, the step of registering the PET image data in the Aβ-PET raw data with a standard brain template to achieve spatial standardization includes:

[0104] Select a standard brain template (commonly used standard brain templates include the MNI152 standard template, which records the average morphological information of the brain structure of a large number of normal participants);

[0105] Preprocessing the standard brain template and the PET image data (such as detrending, intensity standardization, etc., so that the means and variances of the two image data tend to be consistent);

[0106] Select the registration algorithm (registration algorithms include Syn complementary method (SyN), Jackson method and other nonlinear registration algorithms);

[0107] Using the selected registration algorithm, the preprocessed standard brain template, and the PET image data, through multiple iterative optimizations, affine and nonlinear transformation matrices that best match the two images are calculated (according to the registration algorithm, the registration transformation matrix is ​​iteratively optimized so that the three-dimensional pixels of the PET image and the standard brain template reach a first preset matching value in morphology);

[0108] Using the nonlinear transformation matrix to transform the PET image data (i.e., applying the calculated transformation matrix to the individual PET image to achieve image space and morphological matching of the standard brain template);

[0109] The matching accuracy after registration is evaluated by image superposition and reference marker error.

[0110] When the matching accuracy reaches the first preset matching accuracy, the registered individual PET image is output (that is, the individual PET image matched with the standard brain template, which provides a basis for subsequent quantitative analysis and brain region segmentation).

[0111] In this embodiment, it is very important to select a suitable algorithm during the registration process. Generally speaking, the registration algorithm can be selected and determined based on the following points:

[0112] Image type: Different image types require different matching levels and different algorithms. For example, similar images (PET-PET) can use Euclidean transformation, while different images (PET-MRI) require nonlinear transformation.

[0113] Image quality: Image resolution, contrast, etc. affect the matching effect. Do not use complex algorithms for low-quality images.

[0114] Transformation complexity: the degree of matching required, simple registration (flip, shear) or multi-parameter nonlinear matching required.

[0115] Running time: Avoid complex and time-consuming algorithms when clinical diagnostic response speed is required.

[0116] Algorithm accuracy: By testing and matching the same image multiple times, compare the reproducibility and average error of each algorithm and select the one with higher accuracy.

[0117] Reference research experience: search for similar image registration literature and select recognized algorithms with proven effectiveness.

[0118] The above points can serve as important basis for selecting an appropriate registration algorithm. In this embodiment, Aβ-PET is registered with the standard brain template. Considering that the image types are the same and high precision is required, nonlinear algorithms such as Syn can be selected, and their accuracy on the test images can be evaluated to ensure the matching effect of the algorithm.

[0119] In this embodiment, the following points need to be considered when selecting a standard brain template:

[0120] Anatomical detail: Choose a template with clear anatomical details to better represent various parts of the brain.

[0121] Template spatial matching: A template with a high degree of spatial matching with the PET scan is selected for registration.

[0122] Individual differences: The elderly and sick people should choose the elderly template to reduce the impact of structural differences.

[0123] Template contains structure: contains the main brain area mask structures required by the project, such as hippocampus, amygdala, etc.

[0124] Frequency of clinical application: Select public templates that are widely used in clinical practice to reduce duplication of work.

[0125] Software Support: Select standard templates that are compatible with the analysis software and provide technical support.

[0126] Quality Control: Choose open source templates that are provided by reputable organizations and have undergone strict quality control.

[0127] Commonly used standard brain templates include:

[0128] MNI152 standard template: spatial resolution 1mm, fine structure, most widely used.

[0129] Hammers template: reflects common brain structural changes in the elderly.

[0130] Taking the above conditions into consideration, selecting a public standard brain template with the highest matching degree of anatomical structure and spatial requirements is conducive to the reuse of tools.

[0131] The process of this embodiment realizes automatic and accurate three-dimensional registration of individual PET images with standard templates.

[0132] In some possible implementations of the present invention, the step of constructing a standard reference table includes:

[0133] Collect Aβ-PET data of a preset number of healthy controls;

[0134] The Aβ-PET data of the healthy control group were automatically quantitatively analyzed, and the mean value and standard deviation of SUVR in each brain region in the healthy control group were calculated.

[0135] In some possible implementations of the present invention, the method of classifying the quantitative parameter map using a trained multimodal 3D convolutional neural network includes:

[0136] Collect multimodal quantitative parameter maps of individuals in patients with Alzheimer's disease and Aβ-negative normal control groups (multimodal quantitative parameter maps contain three-dimensional spatial data of multiple quantitative mappings, such as different channels such as SUV and SUVR);

[0137] Label the multimodal quantitative parameter maps as abnormal and normal to generate a labeled training set;

[0138] Mapping the labeled training set into a three-dimensional space consistent with the PET image to form a multi-channel quantitative parameter input, each channel being a reference quantitative parameter value;

[0139] The multi-channel quantitative parameter input is used as the input source to input the preset 3D convolutional neural network model for training. The network will learn the feature representation function for distinguishing abnormalities and normality in the multimodal quantitative parameter map (the goal is to correctly classify the parameter map as AD or normal), and obtain the first recognition model;

[0140] The quantitative parameter map is classified using the first recognition model, and the voxel with the largest contribution is located using the CAM heat map as the suspected abnormal voxel.

[0141] This embodiment can also optimize network parameters through back propagation to improve classification accuracy and realize the medical image classification task based on multi-source quantitative information. In short, using known diagnosed Alzheimer's disease (AD) and normal quantitative parameters as samples, the 3D convolutional neural network classification model is trained to automatically detect AD features implicit in multimodal quantification.

[0142] In this embodiment, the 3D convolutional neural network performs deep learning in combination with the correlation between the reference quantitative parameters, thereby improving the accuracy of anomaly detection.

[0143] In some possible implementations of the present invention, the step of automatically performing quantitative analysis on the first brain parenchyma data to calculate voxel SUVMean and voxel SUVR includes:

[0144] According to the first brain parenchyma data, the radioactivity concentration value of each voxel is counted (usually expressed as a standardized uptake value SUV);

[0145] The voxel SUVMean is obtained by averaging all voxels in the brain parenchyma;

[0146] Select the reference area and calculate the SUVMean of the voxels in the reference area;

[0147] Calculate the test voxel SUV value and divide it by the reference area SUVMean to obtain the voxel SUVR;

[0148] The above calculations are repeated for all brain parenchymal voxels, and the voxel SUVMean and voxel SUVR of each voxel are output.

[0149] In this embodiment, the reference region refers to a region with a stable uptake signal that does not change with time, disease, race, etc. In Aβ-PET, cerebellar gray matter, whole cerebellum, pons, white matter, and a composite region including cerebellum and white matter are all commonly used reference regions.

[0150] In an embodiment of the present invention, a neural network can be used to screen the most suitable reference area for the current application case from the commonly used reference areas such as cerebellar gray matter, whole cerebellum, pons, white matter, and composite areas including cerebellum and white matter. The specific method includes: collecting a large number of Aβ-PET scan samples, manually marking possible multiple reference areas; statistically analyzing the uptake signals of each reference area in different samples to obtain its degree of variation; taking the uptake value of each area as input, and the sample label as its stability; building a classification neural network model to distinguish high and low stability; training the network to learn the stable characteristics of the reference area and perform classification prediction; testing the accuracy of each area on the validation set to reflect its stability; selecting one or several areas with the highest accuracy as the best reference area candidate; further verifying the prospect area under a larger sample size; calculating the similarity of the candidate area between different populations as a reference; and comprehensively analyzing and screening out the most suitable reference area for subsequent research. In this way, a more comprehensive and optimized reference area selection scheme can be obtained by using a neural network to perform end-to-end screening of the reference area and considering a large amount of clinical data.

[0151] Through the scheme of this embodiment, SUVMean, which quantitatively reflects the overall brain parenchymal uptake level, can be used; the voxel-level quantitative index SUVR reflects the relative uptake level of each region; automatic high-throughput calculation avoids manual positioning and labeling errors, and provides important quantitative input for subsequent analysis such as brain region analysis and abnormality detection; this automatic calculation method can effectively quantitatively analyze PET data and provide a quantitative basis for clinical evaluation.

[0152] In some possible embodiments of the present invention, the step of calculating the average SUVR of each brain region based on a predefined brain structure mask to achieve brain region-level quantification includes:

[0153] generating a three-dimensional mask of each brain region using the standard brain template;

[0154] Matching the three-dimensional mask to the PET image space by registration (i.e. obtaining a transformation result of the three-dimensional mask in the PET image coordinate system);

[0155] The PET image space was divided into brain regions according to the matched three-dimensional mask, and the SUVR of each voxel was counted and averaged;

[0156] The average SUVR value of each brain region was output as a quantitative indicator.

[0157] Through this embodiment, quantitative analysis is performed based on the anatomical structure of the brain, thereby improving the anatomical significance of quantification; it is more advanced than voxel-level quantification, with each brain region representing the overall level average, reducing the impact of noise; it provides a quantitative reference for subsequent abnormal judgments of brain regions and guides quantitative diagnosis; it is complementary to voxel-level quantification, and the combination can fully reflect quantitative information at different levels of the brain; automatic and efficient calculation reduces artificial variables, which is conducive to large-scale research and clinical applications; this method can achieve multi-level quantitative analysis and improve the accuracy and scope of application of quantitative results.

[0158] In some possible implementations of the present invention, the step of generating a quantitative parameter map according to the voxel SUVMean, the voxel SUVR and the average SUVR of each brain region comprises:

[0159] Mapping the voxel SUVMean, the voxel SUVR and the average SUVR of each brain region into the same three-dimensional image space (consistent with the PET image space);

[0160] Resampling the voxel SUVMean data to generate a three-dimensional voxel SUVMean distribution map;

[0161] Resampling the voxel SUVR data to generate a three-dimensional voxel SUVR distribution map;

[0162] Assigning the average SUVR of each brain region to the corresponding three-dimensional brain region mask area;

[0163] The three-dimensional voxel SUVMean distribution map, the three-dimensional voxel SUVR distribution map and the assigned three-dimensional brain area mask area are presented to generate a distribution map of quantitative parameters in the three-dimensional brain space to obtain the quantitative parameter map.

[0164] Through this embodiment, the spatial distribution of quantitative information in various parts of the brain is intuitively displayed, and quantitative evaluation is performed in combination with quantitative values, providing a quantitative basis for subsequent correction, hypothesis testing, etc.; generating a quantitative parameter map can intuitively and clearly display the quantitative characteristics, provide important assistance for subsequent quantitative analysis and result interpretation, and improve analysis efficiency.

[0165] In some possible implementations of the present invention, the step of determining the abnormality of a brain region based on the standard reference table and the suspected abnormal voxels to automatically locate the abnormal brain region includes:

[0166] According to the brain region mask, the proportion of suspected abnormal voxels in each brain region was counted;

[0167] The average SUVR of each brain region was calculated;

[0168] The mean and standard deviation of SUVR in each brain region were obtained according to the standard reference table;

[0169] Calculate Z = (average SUVR of brain region - mean) / standard deviation;

[0170] A first threshold is set, and when the Z value is higher than the first threshold, the corresponding brain region is identified as an abnormal region.

[0171] Through this embodiment, areas of the brain that may be abnormal can be automatically located; based on population standards, an assessment is given as to whether each brain region is abnormal, providing a directional reference for subsequent abnormal diagnosis, reducing rough misjudgments, and improving positioning accuracy; the results are moved from voxels to brain regions, which aids understanding; and the efficiency and reliability of automatic analysis and quantitative diagnosis are greatly improved.

[0172] In some possible implementations of the present invention, the step of generating a three-dimensional mask of each brain region using the standard brain template includes:

[0173] Reading the template image and the contour annotation data of each brain region from the standard brain template respectively;

[0174] The labeled data is copied into a three-dimensional matrix consistent with the pixels of the template image;

[0175] Assign the pixel value of the corresponding brain area to 1 and the other parts to 0 to generate a three-dimensional matrix of the labeled mask;

[0176] Repeat the above process for each brain region to generate a separate mask;

[0177] All brain region masks were superimposed to generate a comprehensive mask;

[0178] Read the boundary data of each brain region according to the template image coordinates;

[0179] Generate the correspondence between the comprehensive mask and the template structure;

[0180] Save the integrated mask in image format.

[0181] In this embodiment, the template structure is used to annotate data, generate a three-dimensional 0-1 mask consistent with the template pixels, and superimpose and annotate each brain region to form a complete brain region mask, which provides an important tool for subsequent brain region segmentation and quantitative analysis.

[0182] In some possible implementations of the present invention, after the three-dimensional mask is generated, the main steps of performing brain region segmentation and quantitative analysis are as follows:

[0183] PET data standardization and mask registration; perform registration conversion from PET to template space; enable the generated mask to be segmented and used in PET image space; perform PET brain region segmentation according to the mask; perform logical AND operation on the mask pixels of the masked PET image pixels; obtain independent PET sub-image data after cutting each brain region; perform quantitative calculation on each brain region sub-image; calculate the average SUV of brain region voxels; calculate the average SUVR of brain region, etc.; output quantitative parameters of each brain region; visualize the distribution of quantitative parameters of each brain region; perform statistical analysis to evaluate the quantitative level of different brain regions.

[0184] In this embodiment, the use of masks can automatically and efficiently segment the brain regions of PET images, thereby realizing quantitative feature extraction of each brain region. This provides an important reference for quantitative judgment of the degree of lesions in each region and solves the problem of accurately locating each brain region.

[0185] In some possible implementations of the present invention, the following further includes:

[0186] Enhanced PET image standardization:

[0187] In addition to registration and spatial normalization, deep learning methods are used for image normalization, namely: a large number of PET images and standard MRI image pairs are collected and annotated; a 3D convolutional autoencoder is constructed to convert PET images into standard MRI space, realizing nonlinear standardized spatial conversion between different imaging modalities;

[0188] Fully automatic brain segmentation:

[0189] Construct a 3D convolutional neural network with PET images as input; use a soft maximization loss function for end-to-end training to achieve fully automatic and high-precision brain structure segmentation without relying on preset templates;

[0190] Multi-feature cascade positioning:

[0191] In addition to quantitative parameters, other features such as image texture are combined; deep learning methods are used for multi-feature cascade analysis to achieve a unified analysis framework of quantitative parameters and image features, thereby improving the accuracy and completeness of abnormal area positioning;

[0192] Analysis of abnormal propagation mechanism:

[0193] Collect a large amount of long-term dynamic data from case follow-up and healthy control samples; establish a nonlinear spatiotemporal model based on deep learning; analyze the regional propagation pattern of Aβ deposition and explore the mechanism of disease evolution.

[0194] The technical solution provided in this embodiment not only realizes automation but also improves the accuracy of quantitative analysis, and can explore the temporal and spatial laws of disease changes, which is helpful for in-depth exploration of the pathogenesis.

[0195] In some possible implementations of the present invention, the main steps of implementing fully automatic brain region segmentation include:

[0196] Collect a large number of sample pairs containing PET images and corresponding high-precision MRI image segmentation labels;

[0197] Standardize and preprocess image data (core transformation, filtering, etc.);

[0198] Construct a fully convolutional network based on 3D CNN, with PET images as input and brain region segmentation results as output;

[0199] The network adopts a residual structure to increase the network depth, and adds an attention mechanism to improve spatial positioning capabilities;

[0200] Use the soft maximization loss function commonly used in image segmentation as the optimization objective;

[0201] Perform end-to-end training on network-related parameters to make the output results more consistent with the annotations;

[0202] Predict the test set data and output the brain region probability distribution of each voxel;

[0203] Obtain the final quantitative results based on probability, and evaluate and compare them with the annotation results;

[0204] Input the PET case to be tested for automatic segmentation prediction;

[0205] Directly output images of each brain region to provide reference for downstream quantification and abnormality detection;

[0206] Continuously collect new data for model migration and fine-tuning;

[0207] Improve the generalization ability and stability of the model on clinical data.

[0208] In this embodiment, the end-to-end deep learning method can achieve fully automatic high-precision brain area segmentation, no longer relying on preset templates, and has a wider application.

[0209] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0210] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0211] In the several embodiments provided in the present application, it should be understood that the disclosed device can be implemented in other ways. For example, the device embodiments described above are only schematic, such as the division of the above-mentioned units, which is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of devices or units can be electrical or other forms.

[0212] The units described above as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0213] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0214] If the above-mentioned integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a memory, including a number of instructions to enable a computer device (which can be a personal computer, server or network device, etc.) to perform all or part of the steps of the above-mentioned methods of each embodiment of the present application. The aforementioned memory includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or CD-ROM and other media that can store program codes.

[0215] A person skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable memory, and the memory can include: a flash drive, a read-only memory (English: Read-Only Memory, abbreviated as: ROM), a random access memory (English: Random Access Memory, abbreviated as: RAM), a magnetic disk or an optical disk, etc.

[0216] The embodiments of the present application are introduced in detail above. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method and core idea of ​​the present application. At the same time, for general technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

[0217] Although the present invention is disclosed as above, the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions without departing from the spirit and scope of the present invention, and can make various changes and modifications, including the combination of the above-mentioned different functions and implementation steps, including software and hardware implementation methods, all of which are within the scope of protection of the present invention.

Claims

1. An automated quantitative and intelligent analysis method based on Aβ-PET brain imaging, characterized in that: include: The Aβ-PET raw data were preprocessed to obtain the first brain parenchyma data; Performing automatic quantitative analysis on the first brain parenchyma data to calculate voxel SUVMean and voxel SUVR; Based on the predefined brain structure mask, the average SUVR of each brain region was calculated to achieve brain region-level quantification; Generate a quantitative parameter map according to the voxel SUVMean, the voxel SUVR and the average SUVR of each brain region; construct standard reference tables; The trained multimodal 3D convolutional neural network was used to classify the quantitative parameter map, and the top 5% contribution voxels were obtained as suspected abnormal voxels; Judging brain region abnormalities according to the standard reference table and the suspected abnormal voxels, and automatically locating abnormal brain regions; Wherein, the classifying of the quantitative parameter map using the trained multimodal 3D convolutional neural network includes: Multimodal quantitative parameter maps were collected for individuals with Alzheimer's disease and Aβ-negative normal controls; Label the multimodal quantitative parameter maps as abnormal and normal to generate a labeled training set; Mapping the labeled training set into a three-dimensional space consistent with the PET image to form a multi-channel quantitative parameter input, each channel being a reference quantitative parameter value; The multi-channel quantitative parameter input is used as the input source to input the preset 3D convolutional neural network model for training. The network will learn the feature representation function for distinguishing abnormalities and normality in the multi-modal quantitative parameter map to obtain the first recognition model; Using the first recognition model to classify the quantitative parameter map, and using the CAM heat map to locate the voxel with the largest contribution as the suspected abnormal voxel; The step of generating a quantitative parameter map according to the voxel SUVMean, the voxel SUVR and the average SUVR of each brain region includes: Mapping the voxel SUVMean, the voxel SUVR and the average SUVR of each brain region into the same three-dimensional image space; Resampling the voxel SUVMean data to generate a three-dimensional voxel SUVMean distribution map; Resampling the voxel SUVR data to generate a three-dimensional voxel SUVR distribution map; Assigning the average SUVR of each brain region to the corresponding three-dimensional brain region mask area; Generating a distribution map of quantitative parameters in a three-dimensional brain space according to the three-dimensional voxel SUVMean distribution map, the three-dimensional voxel SUVR distribution map and the assigned three-dimensional brain area mask area, and obtaining the quantitative parameter map; The step of determining the abnormality of the brain region according to the standard reference table and the suspected abnormal voxels to automatically locate the abnormal brain region includes: According to the brain region mask, the proportion of suspected abnormal voxels in each brain region was counted; The average SUVR of each brain region was calculated; The mean and standard deviation of SUVR in each brain region were obtained according to the standard reference table; Calculate Z = (average SUVR of brain region - mean) / standard deviation; A first threshold is set, and when the Z value is higher than the first threshold, the corresponding brain region is identified as an abnormal region.

2. The automated quantitative and intelligent analysis method based on Aβ-PET brain imaging according to claim 1, characterized in that: The step of preprocessing the Aβ-PET raw data to obtain the first brain parenchyma data comprises: Registering the PET image data in the Aβ-PET raw data with a standard brain template to achieve spatial standardization; Obtaining a brain template mask from the standard brain template; The skull data and background data are removed by using the brain template mask, the brain parenchyma data are extracted, and the brain parenchyma data are subjected to data standardization processing to obtain the first brain parenchyma data.

3. The automated quantitative and intelligent analysis method based on Aβ-PET brain imaging according to claim 2, characterized in that: The step of registering the PET image data in the Aβ-PET raw data with a standard brain template to achieve spatial standardization includes: Select the standard brain template; Preprocessing the standard brain template and the PET image data; Select the registration algorithm; Using the selected registration algorithm, the preprocessed standard brain template, and the PET image data, the affine and nonlinear transformation matrices that best match the two images are calculated through multiple iterative optimizations; transforming the PET image data using the nonlinear transformation matrix; The matching accuracy after registration is evaluated by the reference marker error; When the matching accuracy reaches a first preset matching accuracy, the registered individual PET image is output.

4. The automated quantitative and intelligent analysis method based on Aβ-PET brain imaging according to claim 3, characterized in that: The step of constructing the standard reference table comprises: Collect Aβ-PET data of a preset number of healthy controls; The Aβ-PET data of the healthy control group were automatically quantitatively analyzed, and the mean value and standard deviation of SUVR in each brain region in the healthy control group were calculated.

5. The automated quantitative and intelligent analysis method based on Aβ-PET brain imaging according to claim 4, characterized in that: The step of automatically performing quantitative analysis on the first brain parenchyma data to calculate voxel SUVMean and voxel SUVR comprises: According to the first brain parenchyma data, counting the radioactivity concentration value of each voxel; The voxel SUVMean is obtained by averaging all voxels in the brain parenchyma; Select the reference area and calculate the SUVMean of the voxels in the reference area; Calculate the test voxel SUV value and divide it by the reference area SUVMean to obtain the voxel SUVR; The above calculations are repeated for all brain parenchymal voxels, and the voxel SUVMean and voxel SUVR of each voxel are output.

6. The automated quantitative and intelligent analysis method based on Aβ-PET brain imaging according to claim 5, characterized in that: The step of calculating the average SUVR of each brain region based on the predefined brain structure mask to achieve brain region-level quantification includes: generating a three-dimensional mask of each brain region using the standard brain template; matching the three-dimensional mask to the PET image space by registration; The PET image space was divided into brain regions according to the matched three-dimensional mask, and the SUVR of each voxel was counted and averaged; The average SUVR value of each brain region was output as a quantitative indicator.

7. The automated quantitative and intelligent analysis method based on Aβ-PET brain imaging according to claim 6, characterized in that: The step of generating a three-dimensional mask of each brain region using the standard brain template comprises: Reading the template image and the contour annotation data of each brain region from the standard brain template respectively; The labeled data is copied into a three-dimensional matrix consistent with the pixels of the template image; Assign the pixel value of the corresponding brain area to 1 and the other parts to 0 to generate a three-dimensional matrix of the labeled mask; Repeat the above process for each brain region to generate a separate mask; All brain region masks were superimposed to generate a comprehensive mask; Read the boundary data of each brain region according to the template image coordinates; Generate the correspondence between the comprehensive mask and the template structure; Save the integrated mask in image format.

Images