Alzheimer's disease risk assessment method and device based on multi-modal medical images

Abstract

This application provides a method and device for Alzheimer's disease risk assessment based on multimodal medical imaging. The method includes: acquiring and registering PET and CT images of the same subject; segmenting a skull region mask based on the CT images and dividing it into multiple anatomical sub-regions using a standard skull partitioning template; extracting radiomics features of each sub-region from the PET images, including robust statistical features that overcome partial volume effects and signal spillover, and texture features that characterize spatial heterogeneity; combining all features into a feature vector, inputting it into a preset AD risk assessment model, and obtaining the risk assessment result. This application can stably and repeatedly extract biologically significant skull Aβ signals from mixed PET signals that are contaminated by thin skull layers, spillover from adjacent brain tissue, and severe noise, and then use them for personalized Alzheimer's disease risk assessment to effectively improve the accuracy of the assessment.

Description

Technical Field

[0001] This application relates to the fields of medical image processing and Alzheimer's disease auxiliary diagnosis technology, and in particular to Alzheimer's disease risk assessment methods and equipment based on multimodal medical images. Background Technology

[0002] Alzheimer's disease (AD) is a neurodegenerative disease characterized primarily by the abnormal deposition of amyloid-β (Aβ) protein. In the field of AD, positron emission tomography (PET) imaging typically refers to Aβ-PET imaging. Aβ-PET imaging can visualize the Aβ (Amyloid-β, β-amyloid protein) load and its spatial distribution in vivo, making it an important tool for AD clinical assessment and research. Currently, commonly used PET quantitative indicators in clinical practice and research include the Standardized Uptake Value (SUV) and the Standardized Uptake Value Ratio (SUVR). These measures standardize the radioactive uptake of the target area by selecting a stable reference region to reduce biases caused by different scanning conditions and individual differences.

[0003] In traditional systems, quantitative analysis of Aβ-PET is almost entirely focused on the cerebral cortex and some subcortical nuclei, forming a relatively mature analytical workflow. The closest existing technologies to this application can be summarized into two categories: the first category is a brain region-based Aβ-PET quantitative and predictive model scheme, using cortical SUVR or radiomics features as the main input, combined with structural MRI indicators and clinical information to construct a predictive model; the second category is observational studies targeting abnormal uptake in non-brain tissues. These studies have found abnormal radioactive uptake in the skull bone marrow region of AD and related populations, and proposed a potential biological link between "skull bone marrow-meningeal immune response-brain".

[0004] However, directly applying existing brain region Aβ-PET quantitative methods to the skull region faces the following technical obstacles: First, the skull structure is thin and discontinuous, and the limited PET resolution leads to significant volume effects, causing the skull signal to be mixed with adjacent tissues; second, the skull is adjacent to highly radioactive brain tissue, resulting in scattering / spillover effects and strong background interference, leading to false positives or regional bias; third, the overall skull PET signal is low and noise-sensitive, and traditional single intensity indicators have poor reproducibility; finally, the lack of an automated, standardized segmentation and zoning system for the skull structure makes related studies difficult to reproduce and scale up. These technical obstacles have prevented the reliable utilization of skull region signals for a long time. Summary of the Invention

[0005] In view of this, embodiments of this application provide a method and device for Alzheimer's disease risk assessment based on multimodal medical imaging, in order to eliminate or improve one or more defects existing in the prior art.

[0006] One aspect of this application provides a method for Alzheimer's disease risk assessment based on multimodal medical imaging, including: Acquire multimodal medical imaging data containing positron emission tomography (PET) images and computed tomography (CT) images of the same subject, and register the PET images and the CT images. Based on the density characteristics of bone tissue in the CT image, a skull region mask is segmented, and a standard skull partitioning template is used to divide the skull region mask into multiple anatomically defined skull sub-regions. Radiomic features of each of the skull sub-regions are extracted from the PET images; wherein the radiomic features include robust statistical features to overcome partial volume effects and signal spillover, and texture features to characterize spatial distribution heterogeneity. The extracted radiomics features are combined into a feature vector, and the feature vector is input into a preset Alzheimer's disease risk assessment model to obtain the Alzheimer's disease risk assessment results of the subject.

[0007] In some embodiments of this application, the registration of the PET image and the CT image includes: The PET images and the CT images are resampled to a uniform spatial resolution; Rigid body or affine registration is performed on the PET image and the CT image with uniform spatial resolution to align the spatial coordinate systems of the PET image and the CT image.

[0008] In some embodiments of this application, segmenting the skull region mask based on the density features of bone tissue in the CT image includes: Voxels in the CT images with density values ​​equal to or higher than a preset bone density threshold are marked as bone tissue to generate an initial bone mask; A three-dimensional morphological closing operation is performed on the initial bone mask to fill the holes inside the initial bone mask and connect the adjacent bone tissue voxels to obtain a morphologically complete bone mask. Three-dimensional connected component analysis was performed on the complete bone mask to identify the independent bone structure image regions that are spatially unconnected. Based on the three-dimensional spatial position and shape of each independent bony structure image region, the bony structure image region corresponding to the cranial cavity is selected from each independent bony structure image region to generate a skull region mask.

[0009] In some embodiments of this application, dividing the skull region mask into multiple anatomically defined skull sub-regions using a standard skull partitioning template includes: The CT image is spatially registered with a preset standard skull partition template, and the transformation parameters for transforming the spatial coordinate system of the CT image to the spatial coordinate system of the standard skull partition template are calculated. Based on the transformation parameters, the partition labels in the standard skull partition template are inversely mapped to the spatial coordinate system of the CT image to generate an individual partition label map; wherein, the position of each voxel in the individual partition label map is consistent with the position of the corresponding voxel in the CT image; In the individual partition labeling diagram, voxels whose spatial location coincides with the voxels within the skull region mask are identified as source voxels to be labeled. Each source voxel carries a partition label, which is then assigned to a target voxel in the same spatial location within the skull region mask, thereby dividing the skull region mask into multiple skull sub-regions.

[0010] In some embodiments of this application, the extraction of radiomics features of each of the skull sub-regions from the PET image includes: For each of the aforementioned cranial sub-regions, the following processing is performed: From the voxel intensity values ​​of the PET image corresponding to the skull sub-region, a robust statistic for overcoming partial volume effects and signal overflow is calculated; Furthermore, the voxel intensity values ​​of the PET image corresponding to the skull sub-region are subjected to grayscale discretization processing to obtain discretized grayscale values; wherein, the grayscale discretization processing includes: linearly or non-linearly mapping continuous PET intensity values ​​to a predetermined number of discrete grayscale levels; A texture matrix is ​​constructed based on the discretized gray values, and texture features for characterizing spatial distribution heterogeneity are calculated from the texture matrix. The texture matrix includes at least one of a gray-level co-occurrence matrix, a gray-level run-length matrix, a gray-level size region matrix, and a neighborhood gray-level dependency matrix. The texture features include at least one of contrast, entropy, a homogeneity index for representing the uniformity of gray levels, and a correlation index for representing the linear dependence of gray levels of adjacent pixels.

[0011] In some embodiments of this application, the robust statistical features include at least one of the median, quantile, and interquartile range of the voxel intensity values ​​of the PET image corresponding to the cranial sub-region.

[0012] In some embodiments of this application, multiple anatomically defined cranial subregions are pre-divided based on at least three different anatomical bone blocks among the frontal, temporal, parietal, and occipital bones.

[0013] In some embodiments of this application, the step of combining the extracted radiomics features into a feature vector and inputting the feature vector into a preset Alzheimer's disease risk assessment model to obtain the Alzheimer's disease risk assessment result of the subject includes: Following a fixed anatomical order corresponding to the standard skull partition template, the radiomics features extracted from all the skull sub-regions are stitched together to form a feature vector; The feature vector is input into the Alzheimer's disease risk assessment model so that the Alzheimer's disease risk assessment model outputs the corresponding predicted value; wherein, the Alzheimer's disease risk assessment model is obtained by pre-training a random forest model based on the multimodal medical imaging data of each historical subject and the corresponding longitudinal cognitive change labels; the training includes: optimizing the hyperparameters of the random forest model using grid search or random search strategies and evaluating the performance of the random forest model using K-fold cross-validation; The predicted value is mapped to an Alzheimer's disease risk assessment result that characterizes the subject's future cognitive decline risk and is output, wherein the Alzheimer's disease risk assessment result includes a continuous risk score or a discrete risk level.

[0014] Another aspect of this application provides an Alzheimer's disease risk assessment system based on multimodal medical imaging, including: The preprocessing module is used to acquire multimodal medical image data containing positron emission tomography (PET) images and computed tomography (CT) images of the same subject, and to register the PET images and the CT images. The skull segmentation module is used to segment a skull region mask based on the density features of bone tissue in the CT image, and to divide the skull region mask into multiple anatomically defined skull sub-regions using a standard skull partitioning template. The feature extraction module is used to extract radiomics features of each of the skull sub-regions from the PET image; wherein the radiomics features include robust statistical features to overcome partial volume effects and signal overflow, and texture features to characterize spatial distribution heterogeneity. The risk prediction module is used to combine the extracted radiomics features into a feature vector and input the feature vector into a preset Alzheimer's disease risk assessment model to obtain the Alzheimer's disease risk assessment result of the subject.

[0015] A third aspect of this application provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned Alzheimer's disease risk assessment method based on multimodal medical images.

[0016] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the described Alzheimer's disease risk assessment method based on multimodal medical images.

[0017] The fifth aspect of this application provides a computer program product comprising a computer program that, when executed by a processor, implements the described Alzheimer's disease risk assessment method based on multimodal medical images.

[0018] This application provides a multimodal medical imaging-based Alzheimer's disease risk assessment method. Utilizing the high-resolution bone tissue imaging capabilities of CT images, it precisely segments the skull region mask, physically isolating signal spillover interference from adjacent brain tissue. Through multimodal image registration, PET functional information is accurately mapped to the CT anatomical space, ensuring that the extracted signals originate from the actual skull region, obtaining a pure and complete skull Aβ signal source, fundamentally solving the problem of signal contamination with adjacent tissues. For the first time, a standard skull partitioning template is introduced, dividing an individual skull into multiple anatomically defined sub-regions through spatial mapping. This makes skull signals from different subjects and research centers comparable, effectively establishing a reproducible quantitative analysis system for the skull, addressing the lack of standardized segmentation and partitioning. This paper addresses the problem of Aβ contamination in PET signals. It employs robust statistical features (such as median and quantiles) insensitive to extreme values ​​and noise, replacing traditionally contaminated single intensity indicators. Furthermore, it introduces texture features to capture the spatial distribution patterns of Aβ deposition, increasing the information dimension. This effectively overcomes the impact of partial volume effects and signal spillover on quantitative results, significantly improving the repeatability and stability of skull signals and resolving the issue of poor repeatability of single intensity indicators. Based on the aforementioned pure, standardized, and robust skull features, a risk assessment model is constructed that can predict the future cognitive decline risk of subjects at baseline, providing a new risk assessment dimension independent of traditional brain region Aβ analysis. This model can transform ignored background signals into clinically predictive biomarkers, offering a new technical path for personalized Alzheimer's disease risk assessment. Therefore, the method in this application can stably and reproducibly extract biologically significant skull Aβ signals from mixed PET signals contaminated by thin skull layers, spillover from adjacent brain tissue, and severe noise, and can then be used for personalized Alzheimer's disease risk assessment to effectively improve assessment accuracy.

[0019] Additional advantages, objectives, and features of this application will be set forth in part in the description which follows, and will in part become apparent to those skilled in the art upon review of the following description, or may be learned by practice of the application. The objectives and other advantages of this application can be realized and obtained by means of the structures specifically pointed out in the specification and drawings.

[0020] Those skilled in the art will understand that the purposes and advantages that can be achieved with this application are not limited to those specifically described above, and that the above and other purposes that this application can achieve will be more clearly understood from the following detailed description. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, do not constitute a limitation thereof. The components in the drawings are not drawn to scale but are merely for illustrating the principles of this application. For ease of illustration and description of certain parts of this application, corresponding portions in the drawings may be enlarged, i.e., may appear larger relative to other components in an exemplary device actually manufactured according to this application. In the drawings: Figure 1 This is a schematic diagram of the first process of an Alzheimer's disease risk assessment method based on multimodal medical imaging according to an embodiment of this application.

[0022] Figure 2 This is a schematic diagram of the second process of the Alzheimer's disease risk assessment method based on multimodal medical imaging in one embodiment of this application.

[0023] Figure 3 This is a schematic diagram of the structure of an Alzheimer's disease risk assessment system based on multimodal medical imaging according to one embodiment of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and their descriptions are used to explain this application, but are not intended to limit it.

[0025] It should also be noted that, in order to avoid obscuring this application with unnecessary details, only the structures and / or processing steps closely related to the solution according to this application are shown in the accompanying drawings, while other details that are not closely related to this application are omitted.

[0026] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0027] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0028] In the following description, embodiments of the present application will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0029] It should be noted that in the traditional system, the quantitative analysis of Aβ-PET is almost entirely concentrated in the cerebral cortex and some subcortical nuclei, such as the frontal lobe, parietal lobe, temporal lobe, posterior cingulate cortex / precuneus, striatum, etc., and a relatively mature analysis process has been formed: image preprocessing (resampling, smoothing, registration, partial volume correction, etc.), cortical or whole-brain ROI extraction, calculation of SUV / SUVR, and correlation analysis or risk prediction modeling with cognitive scales, diagnostic labels, etc.

[0030] The closest existing technologies to this application can be summarized into two categories: one is a quantitative and predictive model based on Aβ-PET of brain regions, and the other is an observational study targeting abnormal uptake in non-brain tissues. The first category uses cortical AβSUVR (standardized uptake ratio) or brain region texture / radiomics features as primary inputs, combined with structural MRI indicators (such as hippocampal volume and cortical thickness) and clinical information to construct a predictive model for cognitive decline or disease progression. Its key steps are: extracting regions of interest (ROIs) based on brain region segmentation templates, calculating the SUV (standardized uptake value), SUVR, or texture features of the ROI, and then conducting predictive modeling. The second category has identified abnormal radioactive uptake in the skull bone marrow region of Alzheimer's disease (AD) and related populations, and proposed a potential biological link between "skull bone marrow-meningeal immune response-brain." However, this type of method lacks standardized definitions of skull ROIs and automated segmentation procedures, and has not yet developed reproducible quantitative indicators for the skull.

[0031] In other words, although existing research suggests the possible presence of aberrant uptake signals in the skull region related to Alzheimer's disease (AD), directly applying existing brain region Aβ-PET quantitative methods (such as SUVR and texture analysis) to the skull region presents significant technical and engineering challenges, resulting in the long-term inability to reliably utilize these signals. Specifically, these challenges include: 1) The skull structure is thin and discontinuous, and the limited PET resolution leads to a significant partial volume effect. The skull cortex and bone marrow structures are distributed in thin layers or cavities. The limited spatial resolution of PET causes the skull signal to be mixed with the adjacent tissue, making the voxel-level SUV / SUVR unstable.

[0032] 2) The skull is adjacent to highly radioactive brain tissue, resulting in scattering / spillover effects and strong background interference. The skull is close to the cerebral cortex and meninges, and high uptake signals from the brain parenchyma may "spill" into the skull voxels, leading to false positives or regional bias; traditional brain region pathways typically do not consider this type of proximity interference.

[0033] 3) Skull PET signals are generally low and sensitive to noise, and traditional single intensity indicators have poor repeatability. Directly calculating indicators such as average SUVR or maximum value may be affected by noise points, scanning artifacts, and individual bone thickness differences, making it difficult to establish stable quantitative indicators.

[0034] 4) Lack of automated and standardized segmentation and partitioning system for skull structure. Current analyses focus on brain tissue and lack automated segmentation and anatomically consistent partitioning methods that treat the skull as an independent analytical object, making it difficult to reproduce and scale up studies.

[0035] Based on this, in order to stably and repeatedly extract pure and biologically significant skull Aβ signals from heavily contaminated mixed signals and use them for individualized risk assessment, embodiments of this application provide an Alzheimer's disease risk assessment method based on multimodal medical imaging, an Alzheimer's disease risk assessment system based on multimodal medical imaging for executing the method, an electronic device, a computer-readable storage medium, and a computer program product.

[0036] The following examples will provide a detailed description.

[0037] Based on this, embodiments of this application provide a method for Alzheimer's disease risk assessment based on multimodal medical images, which can be implemented by an Alzheimer's disease risk assessment system based on multimodal medical images. See [link to relevant documentation]. Figure 1 The Alzheimer's disease risk assessment method based on multimodal medical imaging specifically includes the following: Step 100: Acquire multimodal medical image data containing positron emission tomography (PET) images and computed tomography (CT) images of the same subject, and register the PET images and the CT images.

[0038] It should be noted that the subject needs to be injected with an Aβ tracer (e.g., , After an appropriate acquisition time, PET and CT images can be obtained by scanning with a PET / CT scanner. To ensure the accuracy of subsequent analysis, spatial alignment of the PET and CT images is necessary. This embodiment uses rigid body registration or affine registration methods to register the PET image to the spatial coordinate system of the CT image. Before registration, the two images can be resampled to a uniform voxel size (e.g., 1mm × 1mm × 1mm) and spatial resolution to eliminate device differences. The goal of registration is to maximize the mutual information or normalized mutual information between the two images, thereby obtaining registered PET and CT images that share the same spatial coordinate system.

[0039] Step 200: Based on the density characteristics of bone tissue in the CT image, segment the skull region mask, and use a standard skull partition template to divide the skull region mask into multiple anatomically defined skull sub-regions.

[0040] In step 200, the specific implementation process of segmenting the skull region mask can be as follows: Since bone tissue in CT images has high density (high CT value), bone structure can be initially extracted through threshold segmentation. In this embodiment, a preset bone density threshold T is used. bone (For example, CT value ≥ 200 HU), density values ​​≥ T in CT images bone Voxels are labeled as bone tissue to generate an initial bone mask. To eliminate holes and noise that may be generated by threshold segmentation, a three-dimensional morphological closing operation is performed on the initial bone mask (e.g., using a spherical structuring element with a radius of 2 voxels) to fill the internal holes and connect adjacent bone tissue voxels, resulting in a morphologically complete bone mask. Subsequently, three-dimensional connected component analysis is performed on the morphologically complete bone mask to identify spatially unconnected independent bony structure image regions. These regions may include the skull, mandible, cervical vertebrae, teeth, etc. Based on the three-dimensional spatial location and morphological features of each independent bony structure image region (e.g., located at the image center, largest volume, surrounding brain tissue, etc.), the bony structure image regions corresponding to the cranial cavity are selected, i.e., the skull region mask.

[0041] The specific implementation process of dividing the skull into sub-regions can be as follows: To achieve standardized analysis, this embodiment pre-defines a standard skull partition template in a standard anatomical space (e.g., MNI space). This template divides the skull into eight anatomically defined sub-regions: left frontal bone, right frontal bone, left temporal bone, right temporal bone, left parietal bone, right parietal bone, left occipital bone, and right occipital bone. The CT image (or skull region mask) obtained from the aforementioned segmentation is spatially registered with the standard skull partition template to obtain transformation parameters (e.g., rigid body transformation matrix or affine transformation matrix) from the individual CT image space to the standard template space. Based on the inverse transformation of these transformation parameters, the partition labels in the standard template are mapped back to the individual CT image space, generating an individual partition label map aligned with the CT image space. The position of each voxel in this label map is consistent with the corresponding voxel position in the CT image, and each voxel is assigned a label value (1 to 8) representing its respective skull sub-region. Then, the individual partition label map is spatially overlaid with the obtained skull region mask: for each voxel in the skull region mask, the label value of the voxel at the same spatial location in the individual partition label map is found, and this label value is assigned to the corresponding voxel in the skull region mask. Finally, the skull region mask is divided into multiple skull sub-region masks corresponding to the partition labels, and each sub-region is composed of voxels with the same partition label.

[0042] Step 300: Extract radiomics features of each of the skull sub-regions from the PET image; wherein the radiomics features include robust statistical features to overcome partial volume effects and signal overflow, and texture features to characterize spatial distribution heterogeneity.

[0043] It is understandable that in step 300, the following processing can be performed on each of the cranial sub-regions defined in step 200: (1) Extracting robust statistical features: A set of robust statistics is calculated from the voxel intensity values ​​of the PET image corresponding to the sub-region. Since the skull signal is affected by partial volume effect and spillover from adjacent brain tissue, the traditional mean is easily disturbed by extreme values. Therefore, this embodiment prefers to use noise-insensitive indicators such as the median, quantiles (e.g., the 10th percentile P10, the 90th percentile P90), and interquartile range (IQR). These statistics can reflect the overall Aβ load level of the sub-region and have good repeatability.

[0044] (2) Texture Feature Extraction: To capture the spatial distribution pattern of Aβ deposition within the skull, texture features are further extracted. First, the PET intensity values ​​within the sub-region are discretized using grayscale, that is, continuous intensity values ​​are mapped to a predetermined number of discrete grayscale levels (e.g., the intensity range is linearly divided into 64 grayscale levels). Based on the discretized grayscale values, various texture matrices are constructed, such as the Gray-Level Co-occurrence Matrix (GLCM), Gray-Level Run-Length Matrix (GLRLM), Gray-Level SizeZone Matrix (GLSZM), and Neighboring Gray-Level Dependence Matrix (NGLDM). A series of texture features are calculated from these matrices, including but not limited to contrast, entropy, homogeneity, correlation, energy, and inverse difference moment. These features can reflect spatial heterogeneity information such as local grayscale changes, texture roughness, and periodicity.

[0045] After the above processing, each skull sub-region yields a multidimensional feature vector F containing robust statistical and textural features. k (Where k represents the sub-region number).

[0046] Step 400: Combine the extracted radiomics features into a feature vector, and input the feature vector into a preset Alzheimer's disease risk assessment model to obtain the Alzheimer's disease risk assessment results of the subject.

[0047] Understandably, in step 400, the feature vectors of the eight cranial sub-regions obtained in step 300 are first concatenated according to a fixed anatomical order corresponding to the standard cranial partition template (e.g., left frontal bone, right frontal bone, left temporal bone, right temporal bone, left parietal bone, right parietal bone, left occipital bone, right occipital bone) to form the overall feature vector F=[F1,F2,...,F8] of the subject. To eliminate the influence of different feature dimensions, the feature vector can be standardized (e.g., Z-score standardization, making the mean of each feature dimension 0 and the variance 1).

[0048] In this embodiment, the Alzheimer's disease risk assessment model employs a random forest model. This model is pre-trained by collecting PET and CT images of historical subjects, along with corresponding longitudinal cognitive change labels (e.g., changes in the Mini-Mental State Examination (MMSE) during follow-up, or labels indicating whether mild cognitive impairment (MCI) progressed to AD). Steps 100 to 300 are performed for each historical subject to generate feature vectors. These feature vectors are then used as input, with the longitudinal cognitive change labels serving as the supervised target, to train the random forest model. During training, grid search or random search can be used to optimize hyperparameters (e.g., number of trees, maximum depth), and K-fold cross-validation is used to evaluate model performance and select the optimal model.

[0049] For the current subject, the feature vector obtained in step 300 is input into the trained random forest model, and the model outputs a predicted value (e.g., the predicted MMSE decline score, or the probability of belonging to the progression category). This predicted value is mapped to an Alzheimer's disease risk assessment result that is easy to interpret clinically, such as a continuous risk score linearly scaled to 0-100, or divided into low-risk, medium-risk, and high-risk levels according to a preset threshold, and the result is output.

[0050] As described above, the Alzheimer's disease risk assessment method based on multimodal medical imaging provided in this application utilizes the high-resolution bone tissue imaging capability of CT images to accurately segment the skull region mask, physically isolating signal spillover interference from adjacent brain tissue. Through multimodal image registration, PET functional information is precisely mapped to the CT anatomical space, ensuring that the extracted signals originate from the real skull region, obtaining a pure and complete skull Aβ signal source, fundamentally solving the problem of signal mixing with adjacent tissues. For the first time, a standard skull partition template is introduced, dividing an individual skull into multiple anatomically defined sub-regions through spatial mapping, making skull signals from different subjects and research centers comparable, thereby effectively establishing a reproducible quantitative analysis system for the skull, addressing the lack of standardized segmentation. This paper addresses the issues of segmentation and partitioning systems by employing robust statistical features (such as median and quantiles) insensitive to extreme values ​​and noise to replace traditional, easily contaminated single intensity indicators. It also introduces texture features to capture the spatial distribution patterns of Aβ deposition, increasing the information dimension and effectively overcoming the impact of partial volume effects and signal spillover on quantitative results. This significantly improves the repeatability and stability of skull signals, resolving the problem of poor repeatability of single intensity indicators. Based on the aforementioned pure, standardized, and robust skull features, a risk assessment model is constructed that can predict the future cognitive decline risk of subjects at baseline, providing a new risk assessment dimension independent of traditional brain region Aβ analysis. This model can transform ignored background signals into clinically predictive biomarkers, providing a new technical path for personalized Alzheimer's disease risk assessment. Therefore, the method in this application can stably and reproducibly extract biologically significant skull Aβ signals from mixed PET signals contaminated by thin skull layers, spillover from adjacent brain tissue, and severe noise, and can then be used for personalized Alzheimer's disease risk assessment to effectively improve assessment accuracy.

[0051] It should also be noted that prior to the filing of this application, those skilled in the art, when faced with the technical problem of using skull Aβ signals for AD risk assessment, would find it difficult to conceive of the technical solution combining PET and CT images described in this application. Even if they could conceive of it, they would face a series of insurmountable technical obstacles in the specific implementation process. A detailed analysis follows: (1) Skull signals have long been considered background noise rather than the object of analysis: In the field of AD imaging research, a deep-seated technical bias exists: all standardized procedures for PET image analysis prioritize the elimination of non-brain tissue interference. Whether it's skull dissection in image preprocessing, the design of partial volume correction algorithms, or the construction of standard spatial templates (such as AAL and MNI atlases), non-brain structures like the skull, scalp, and meninges are treated as irrelevant signals that must be removed. Through long-term research practice, those skilled in the art have been trained to automatically ignore signal changes in the skull region, never considering it a potential analytical target. This technical bias leads to the instinctive attribution of abnormal uptake in the skull region to tracer non-specific binding or image artifacts, rather than biologically significant pathological signals.

[0052] (2) The traditional role of CT images in PET analysis limits their new applications: In current PET / CT clinical practice, the core functions of CT images are strictly limited to two: first, attenuation correction of PET images to improve quantitative accuracy; and second, providing rough anatomical localization to assist in determining the anatomical source of abnormal PET uptake. Those skilled in the art have long held the mindset that CT serves PET quantification and localization, never considering CT as an independent, precise anatomical analysis tool, let alone utilizing CT's high-resolution bone tissue imaging capabilities to actively extract and purify skull PET signals, which are considered background noise. In other words, CT, in the traditional system, is a supporting role to PET, rather than an analytical object that collaborates with it on an equal footing.

[0053] (3) Inertial application of radiomics technology pathways: Radiomics, as a high-throughput feature extraction method, has established a mature application paradigm in fields such as oncology and neurodegenerative diseases. Typically, it involves delineating lesions or regions of interest on high-resolution anatomical images (such as CT and MRI) or extracting regional features based on standardized brain atlases on functional images (such as PET). When faced with the need to analyze skull signals, the most natural reaction for those skilled in the art is to directly apply existing radiomics workflows to skull regions—that is, manually or semi-automatically delineating skull ROIs on PET images and then calculating their SUV / SUVR or conventional radiomics features. This technological inertia leads them to try things in the wrong direction, without considering the need to introduce CT images and construct a completely new analytical system specifically for skull characteristics.

[0054] Even if someone skilled in the art were to break free from conventional thinking and realize that CT images might be needed to assist in skull signal analysis, the following intractable technical challenges would still exist in the actual implementation process: (4) The challenge of precise segmentation of the skull structure: While CT images can clearly display bone tissue, the skull, as a complex three-dimensional shell enclosing brain tissue, faces multiple challenges in automatic segmentation: the skull and adjacent skeletal structures (such as teeth, mandible, and cervical vertebrae) have similar density characteristics on CT images, making thresholding segmentation ineffective in distinguishing them; the skull itself contains natural openings (such as the superior orbital fissure and foramen ovale) and sutures, which simple segmentation algorithms may misclassify as background or fractures; the diploic layer between the inner and outer plates of the skull has a low density, potentially leading to voids within the segmented mask. To obtain a clean, complete skull region mask suitable for subsequent quantitative analysis, a multi-step refined processing workflow including morphological repair, connected component analysis, and spatial location filtering is required, which is far beyond the capabilities of conventional thresholding segmentation in this field.

[0055] (5) Accuracy and consistency requirements for multimodal image registration: Precisely placing PET signals within a skull mask segmented by CT requires sub-voxel-level registration accuracy between PET and CT images. However, PET images have low spatial resolution and high noise levels, making them susceptible to local optima during registration with CT images, leading to registration errors. While such errors might be tolerable in routine brain region analysis, for the thin-layered skull structure only a few millimeters thick, even a shift of 1-2 voxels can cause a large amount of signal from adjacent brain tissue or scalp to be mixed into the extracted PET signal, rendering subsequent analysis meaningless. Even if those skilled in the art recognized the need for registration, they would find it difficult to anticipate the extreme importance of registration accuracy for skull analysis, let alone design a technical solution to ensure registration reliability.

[0056] (6) The dilemma of correcting partial volume effect and spillover effect: The PET signal problem of the skull is a fundamental physical challenge. Even with a precise skull mask obtained via CT, the PET signal of each voxel within the mask remains a mixture of the skull's own signal and signals spilling over from adjacent brain tissue. Traditional partial volume correction algorithms (such as the Muller-Gartner method and the geometric transfer matrix method) typically assume clear tissue boundaries and regular regional shapes, which have limited effectiveness for the skull, an extremely thin and complex structure. Even those skilled in the art who recognize the need for correction often struggle to find effective correction methods suitable for the skull. This application offers a novel solution by bypassing physical correction and instead employing statistical features (such as median and quantiles) robust to mixed signals and texture features capable of capturing spatial patterns. This roundabout strategy is far beyond what conventional thinking in the field could conceive.

[0057] (7) Lack of a standardized analysis system: Even if all the aforementioned technical problems could be solved to obtain a set of skull signal features from a particular subject, how to compare these features with the results of other studies, and how to ensure the comparability of data from different subjects and centers, requires a standardized skull partitioning system. However, existing brain atlases do not contain any skull partitioning information, and those skilled in the art lack available standardized tools for skull analysis. This application is the first to create a standard skull partitioning template and establish a mapping process from individual space to standard space, which is itself a pioneering work with value far exceeding conventional image registration or ROI definition.

[0058] In summary, the Alzheimer's disease risk assessment method based on multimodal medical imaging described in this application is not a simple combination of existing technologies, but a systematic solution designed to address the specific technical challenge of skull Aβ signal extraction and utilization. This solution comprises several interconnected technical steps: 1) CT image import and registration: providing accurate anatomical benchmarks for signal extraction; 2) Refined segmentation process: ensuring a clean and complete skull analysis object; 3) Standard partition template construction and mapping: establishing a reproducible standardized analysis framework; 4) Robust statistics and texture feature extraction: extracting biologically significant information from physically inseparable mixed signals; 5) Predictive model based on skull features: transforming the above signals into a clinically valuable risk assessment tool.

[0059] The solution to each of the above steps relies on specific technical means, and the combination of these technical means cannot be naturally deduced by a person skilled in the art when faced with the vague suggestion that skull signals may be useful.

[0060] To further eliminate the original spatial inconsistencies between PET and CT images caused by factors such as scanning parameters and subject displacement, and to provide a reliable spatial benchmark for subsequent precise skull segmentation based on CT, this application provides a method for Alzheimer's disease risk assessment based on multimodal medical imaging, see [link to relevant documentation]. Figure 2 Step 100 of the Alzheimer's disease risk assessment method based on multimodal medical imaging specifically includes the following: Step 110: Acquire multimodal medical imaging data containing positron emission tomography (PET) images and computed tomography (CT) images of the same subject.

[0061] Step 120: Resample the PET image and the CT image to a uniform spatial resolution.

[0062] Specifically, positron emission tomography (PET) and computed tomography (CT) images of the same subject are resampled to ensure they have the same voxel size and spatial resolution. Since PET and CT images are typically acquired by different devices, their voxel sizes may differ (e.g., PET images commonly have voxel sizes of 2mm × 2mm × 2mm, while CT images have 0.5mm × 0.5mm × 0.5mm). Direct registration would lead to inaccurate spatial correspondence. Therefore, interpolation algorithms (such as linear interpolation or cubic spline interpolation) are used to resample the two images to a uniform resolution, for example, isotropic 1mm × 1mm × 1mm. The resampled PET and CT images maintain consistency in grid size and physical space, laying the foundation for subsequent registration.

[0063] Step 130: Perform rigid body or affine registration on the PET image and the CT image with uniform spatial resolution to align the spatial coordinate systems of the PET image and the CT image.

[0064] Specifically, rigid registration or affine registration is performed on resampled PET and CT images with uniform spatial resolution. Rigid registration only allows translation and rotation, suitable for correcting displacement caused by slight subject movements; affine registration additionally allows scaling and shearing, which can further correct for minor scale differences between different devices. This embodiment preferably uses maximization based on mutual information as a similarity measure, and iteratively solves for the optimal spatial transformation parameters through optimization algorithms (such as gradient descent) to accurately align the spatial coordinate systems of the PET and CT images. After registration, each voxel in the PET image corresponds strictly in spatial position to the corresponding voxel in the CT image, meaning the two images share the same spatial coordinate system.

[0065] As can be seen from the above description, the Alzheimer's disease risk assessment method based on multimodal medical images provided in this application ensures precise alignment of PET and CT images by resampling to unify resolution and performing rigid body / affine registration. This allows the high-resolution anatomical information of the CT images to accurately guide the extraction of PET signals, effectively improving the accuracy of subsequent skull segmentation and feature extraction.

[0066] To further automatically and accurately segment clean skull regions from CT images, removing non-skull bony structures such as the mandible, teeth, and cervical spine, and overcoming the defects of incomplete skull masking and holes after threshold segmentation, this application provides an Alzheimer's disease risk assessment method based on multimodal medical imaging, see [link to relevant documentation]. Figure 2 Step 200 of the Alzheimer's disease risk assessment method based on multimodal medical imaging specifically includes the following: Step 210: Mark voxels in the CT image whose density values ​​are equal to or higher than a preset bone density threshold as bone tissue to generate an initial bone mask.

[0067] The bone density threshold is a critical CT value used to distinguish bone tissue from non-bone tissue in CT images, typically set based on the Hounsfield Unit (HU). The initial bone mask is a binary 3D image where voxels with a value of 1 represent bone tissue and voxels with a value of 0 represent non-bone tissue.

[0068] In step 210, the preset bone mineral density threshold T bone The CT value is set to 200 HU. Iterating through all voxels in the CT image, if a voxel's CT value is ≥200 HU, its value at the corresponding position in the initial bone mask is set to 1; otherwise, it is set to 0. This yields a binary image B0 containing only candidate bone tissue regions. This mask may contain all high-density bony structures such as the skull, mandible, cervical vertebrae, and teeth. Due to noise and partial volume effects, there may be holes inside the skull, and the edges may be discontinuous.

[0069] Step 220: Perform a three-dimensional morphological closing operation on the initial bone mask to fill the holes inside the initial bone mask and connect the adjacent bone tissue voxels to obtain a morphologically complete bone mask.

[0070] It should be noted that 3D morphological closing is an image processing operation that first dilates a binary image and then performs erosion to fill small holes, connect neighboring regions, and smooth boundaries. The structuring element is a 3D template used to define the shape of the neighborhood for the dilation and erosion operations; it is commonly a sphere or a cube.

[0071] In step 220, a spherical structural element S with a radius of 2 voxels can be used. r Perform a three-dimensional closing operation on the initial bone mask B0, denoted as B1=Close(B0,S rSpecifically, B0 is first expanded to connect adjacent bone tissue voxels and fill small pores; then, erosion is performed to restore the boundary expansion caused by the expansion while maintaining connectivity. In the resulting bone mask B1, the pores in the skull region are filled, fractures are bridged, and the edges are smoother, forming a morphologically complete bone mask. However, this mask may still contain multiple independent bony structures (such as the skull and mandible not yet separated).

[0072] Step 230: Perform three-dimensional connected component analysis on the complete bone mask to identify the independent bone structure image regions that are not spatially connected.

[0073] Understandably, 3D connected component analysis is an algorithm used to label all interconnected sets of voxels in a 3D binary image, with each connected set assigned a unique identifier. For connectivity, a 26-neighborhood (i.e., the 26 adjacent positions of a voxel) is typically used to define whether a voxel is connected.

[0074] In step 230, a 26-neighborhood connectivity method can be used to label the morphologically complete bone mask B1 with three-dimensional connected components. The algorithm traverses all voxels, grouping interconnected voxels into the same connected region and assigning a unique label (e.g., 1, 2, 3, ...) to each region. The output is a label map, where the value of each voxel represents the connected region number it belongs to. These regions are spatially unconnected independent bone structure image regions, which may include the main skull, mandible, cervical vertebrae, teeth, etc.

[0075] In one or more embodiments of this application, the bony structure image region may also be referred to simply as a bone block.

[0076] Step 240: Based on the three-dimensional spatial position and shape of each of the independent bony structure image regions, select the bony structure image regions corresponding to the cranial cavity from each of the independent bony structure image regions to generate a skull region mask.

[0077] It should be noted that spatial location refers to the coordinate range of each connected region in three-dimensional space, such as the location of the center of gravity and the bounding box. Morphological features are the geometric attributes of each connected region, such as volume, surface area, and degree of matching with the anatomical model of the cranial cavity. The image regions corresponding to the bony structures of the cranial cavity refer to those image regions that anatomically constitute the skull (i.e., the bone structures that surround and protect the brain tissue).

[0078] In step 240, filtering can be performed based on prior anatomical knowledge of the skull. First, the volume of each individual bony structure image region is calculated, and the center position of its spatial bounding box is determined. The skull is typically located slightly above the center of the image, has the largest volume, and is a near-spherical shell in shape. Specific filtering rules are as follows: (1) Retain the largest connected region (usually the main body of the skull).

[0079] (2) For other areas, judge according to their center of gravity coordinates: if the center of gravity is located below the main body of the skull (such as a small z coordinate), it is considered as the mandible or cervical vertebra and is removed; if the center of gravity is located in front of or to the side of the main body of the skull and the volume is small, it is considered as a tooth or cheekbone and is removed in the same way.

[0080] (3) Optionally, the selected regions can be matched with a pre-constructed cranial anatomical template to further improve accuracy.

[0081] (4) Finally, all the remaining connected regions (usually only the main body of the skull) are merged to generate a pure skull region mask B. skull This mask contains only the bony structures corresponding to the cranial cavity, eliminating all non-cranial bone fragments, thus laying a precise anatomical foundation for subsequent standardized partitioning and feature extraction.

[0082] As can be seen from the above description, the Alzheimer's disease risk assessment method based on multimodal medical images provided in this application obtains a complete and pure skull region mask through a refined process of threshold processing, three-dimensional morphological closing operation, connected component analysis and spatial morphology screening. This can significantly improve the accuracy and robustness of segmentation and provide a reliable foundation for subsequent standardized partitioning and signal extraction.

[0083] To further divide the segmented skull region mask according to a standardized anatomical partitioning system, making it into multiple comparable and reproducible analytical units, an Alzheimer's disease risk assessment method based on multimodal medical imaging is provided in an embodiment of this application, see [link to relevant documentation]. Figure 2 Step 200 of the Alzheimer's disease risk assessment method based on multimodal medical imaging further includes the following: Step 250: Spatial registration of the CT image with a preset standard skull partition template is performed, and transformation parameters for transforming the spatial coordinate system of the CT image to the spatial coordinate system of the standard skull partition template are calculated.

[0084] It should be noted that the standard skull partitioning template is a skull partitioning atlas predefined in a standard anatomical space (such as the Montreal Neurological Institute MNI space), dividing the skull into multiple anatomically significant subregions. In this embodiment, the template divides the skull into eight subregions: left frontal bone, right frontal bone, left temporal bone, right temporal bone, left parietal bone, right parietal bone, left occipital bone, and right occipital bone, with each region assigned a unique partitioning label (e.g., 1 to 8). Spatial registration is the process of aligning an individual image (CT image) with a standard template, solving for the optimal spatial transformation relationship by optimizing the spatial similarity between the two images. Transformation parameters are mathematical parameters describing the spatial transformation relationship. For rigid body transformations, these include three translation parameters and three rotation parameters; for affine transformations, they also include scaling and shearing parameters, typically represented as a 4×4 transformation matrix.

[0085] In this method, an individual CT image can be used as the registration source image, and a standard skull partition template can be used as the registration target image. A registration method based on mutual information is employed, using an optimization algorithm (such as gradient descent) to iteratively solve for the optimal spatial transformation parameter T, ensuring that the transformed CT image and the standard template achieve optimal spatial alignment. This transformation parameter T defines the mapping relationship from the spatial coordinate system of the individual CT image to the spatial coordinate system of the standard template; that is, for any point x in the individual CT image, its corresponding point in the standard template space is y=T(x). Because CT images contain rich anatomical details, using them as the registration reference yields higher registration accuracy than using a binarized skull mask.

[0086] Step 260: Based on the transformation parameters, the partition labels in the standard skull partition template are reverse mapped to the spatial coordinate system of the CT image to generate an individual partition label map; wherein, the position of each voxel in the individual partition label map is consistent with the position of the corresponding voxel in the CT image.

[0087] Inverse mapping is the process of mapping information from the standard template space back to the individual CT image space using the inverse transform T-1 of the transform parameter T. The individual partition label map is a three-dimensional image with the same spatial dimensions and voxel grid as the CT image, where the value of each voxel represents the skull partition label to which that location belongs.

[0088] In step 260, the inverse transform T-1 of the transform parameter T is first calculated, which defines the mapping from the standard template space to the individual CT image space. Then, for each voxel position x in the individual CT image space, its corresponding point y′=T-1(x) in the standard template space is found through the inverse transform T-1. Since y′ is usually not an integer voxel coordinate, the nearest neighbor interpolation method is used to obtain the partition label value L of the y′ position. std(y′) is assigned as a label value, and this label value is assigned to voxel x in the individual CT image space. After traversing all voxels, an individual partition label map that is completely spatially aligned with the CT image is obtained. In this label map, each voxel location is assigned a label value (1 to 8) representing its standard cranial partition, and voxels in non-cranial regions can be assigned a label of 0.

[0089] Step 270: In the individual partition label map, voxels whose spatial location coincides with the voxels in the skull region mask are identified as source voxels to be labeled.

[0090] Understandably, source voxels refer to those voxels in the individual partition label map that happen to be located within the spatial area covered by the skull region mask; their partition labels will be used to label the skull mask. Spatial overlap refers to the voxel positions in two images having exactly the same three-dimensional coordinates.

[0091] Specifically, in step 270, the individual partition label map generated in step 260 can be spatially superimposed with the previously obtained skull region mask. For each voxel with a value of 1 in the skull region mask (i.e., a voxel belonging to the skull), the voxels at the same coordinate positions in the individual partition label map are located. These located individual partition label map voxels are the source voxels to be labeled, and each source voxel carries a label value representing its respective skull partition.

[0092] Step 280: Assign the partition label carried by each source voxel to the target voxel with the same spatial position in the skull region mask, thereby dividing the skull region mask into multiple skull sub-regions.

[0093] Assigning labels refers to the process of copying or assigning the partition label value of a source voxel to a target voxel. At the data level, this manifests as updating the values ​​of the corresponding voxels in the skull region mask. The target voxel is the voxel in the skull region mask to be assigned a partition label. A skull sub-region is a connected set of voxels in the skull region mask that have been assigned the same partition label after partitioning, corresponding to an anatomically defined skull region.

[0094] In step 280, for each identified source voxel, its partition label value (e.g., 3, representing the occipital bone) is obtained, and this label value is assigned to the target voxel at the same spatial location in the skull region mask. Specifically, a new 3D array can be created to store the divided skull sub-region masks, with the same size as the skull region mask. All voxels with a value of 1 in the skull region mask are traversed, and the label value of the corresponding source voxel is written to the same position in the new array. After processing, the binary information in the new array, which originally only indicated "whether it is a skull," is replaced with specific partition label information: voxels with a value of 1 constitute the left frontal bone sub-region, voxels with a value of 2 constitute the right frontal bone sub-region, and so on. Thus, the skull region mask is successfully divided into multiple anatomically defined skull sub-regions corresponding to the standard skull partition template.

[0095] As can be seen from the above description, the Alzheimer's disease risk assessment method based on multimodal medical images provided in this application can achieve standardized and automated zoning of individual skulls through the process of registering CT images with standard templates, generating individual zoning label maps through inverse mapping, and then superimposing them with the skull mask space. This ensures the comparability of skull sub-regions among different subjects and lays the foundation for subsequent robust radiomics feature extraction.

[0096] To further extract stable and information-rich quantitative indicators from PET signals of cranial sub-regions, and to overcome the shortcomings of single intensity indicators being sensitive to noise and unable to reflect the spatial heterogeneity of Aβ distribution, this application provides an Alzheimer's disease risk assessment method based on multimodal medical imaging. (See also...) Figure 2 Step 300 in the Alzheimer's disease risk assessment method based on multimodal medical imaging specifically includes the following: For each of the aforementioned cranial sub-regions, the following processing is performed: Step 310: Calculate robust statistics from the voxel intensity values ​​of the PET image corresponding to the skull sub-region to overcome partial volume effects and signal overflow.

[0097] Among them, the cranial sub-regions refer to multiple anatomically defined regions after being divided by the standard cranial partition template. In this embodiment, there are eight sub-regions (left frontal bone, right frontal bone, left temporal bone, right temporal bone, left parietal bone, right parietal bone, left occipital bone, and right occipital bone).

[0098] Robust statistics are statistical indicators that are insensitive to outliers, noise, and extreme data points, and can more stably reflect the central tendency and dispersion of data. Partial volumetric effects occur because the spatial resolution of PET images is limited, and a single voxel may contain signals from multiple tissues (such as the skull and adjacent brain tissue). Signal spillover is the phenomenon where signals from highly radioactive brain tissue diffuse into surrounding less radioactive areas (such as the skull).

[0099] In step 310, for each skull sub-region, the corresponding set of voxel intensity values ​​from the PET image is extracted. The robust statistical features include at least one of the median, quantiles, and interquartile range of the voxel intensity values ​​of the PET image corresponding to the skull sub-region. Considering that skull signals are affected by partial volume effects and spillover effects, and that traditional means are easily affected by extreme values, this embodiment preferentially adopts the following robust statistical measures: (1) Median: The value in the middle after all intensity values ​​are sorted, and is not affected by a few extreme values.

[0100] (2) Quantiles: For example, the 10th percentile (P10) and the 90th percentile (P90) represent the low end and high end of the intensity distribution, respectively, and can reflect the distribution range of the signal.

[0101] (3) Interquartile Range (IQR): The difference between the 75th percentile and the 25th percentile, which measures the dispersion of the middle 50% of the data.

[0102] These robust statistics can effectively reduce the interference of noise points and spillover signals, and more accurately reflect the overall Aβ load level of the sub-region.

[0103] Step 320: Perform grayscale discretization on the voxel intensity values ​​of the PET image corresponding to the skull sub-region to obtain discretized grayscale values; wherein, the grayscale discretization includes: linearly or non-linearly mapping continuous PET intensity values ​​to a predetermined number of discrete grayscale levels.

[0104] It's important to note that grayscale discretization is the process of converting continuously valued image intensities into a finite number of discrete levels, and it's a prerequisite for texture feature calculation. Grayscale levels are the number of intensity levels after discretization, typically set to 8, 16, 32, 64, etc. Linear mapping uniformly divides the original intensity range into several intervals, with each interval mapped to a grayscale level. Nonlinear mapping performs non-uniform division based on intensity distribution characteristics (such as histogram equalization), ensuring that each grayscale level contains approximately the same number of voxels.

[0105] In step 320, the minimum and maximum values ​​of the intensity of all PET voxels within the skull sub-region can be obtained first to determine the intensity range. A linear mapping method is used to uniformly divide the continuous intensity values ​​into 64 discrete gray levels. Specifically, let the minimum intensity value be I. min The maximum value is I max Then the width of each gray level is Δ=(I max - I min ) / 64. For the original intensity value I, the formula for calculating its discretized gray level g is: ,in This indicates rounding down. If a non-linear mapping is used, the number of voxels in each gray level can be made approximately equal based on the histogram equalization principle, thus enhancing texture contrast. After discretization, the original PET intensity value of each voxel in this sub-region is replaced with the corresponding discrete gray level (0 to 63).

[0106] And step 330: construct a texture matrix based on the discretized gray values, and calculate texture features from the texture matrix to characterize spatial distribution heterogeneity.

[0107] The texture matrix includes at least one of a gray-level co-occurrence matrix, a gray-level run-length matrix, a gray-level size region matrix, and a neighborhood gray-level dependency matrix; the texture features include at least one of contrast, entropy, a homogeneity index for representing the degree of gray-level uniformity, and a correlation index for representing the degree of linear dependence of gray levels of adjacent pixels.

[0108] Specifically, the Gray-Level Co-occurrence Matrix (GLCM) is used to calculate the probability of gray levels i and j occurring simultaneously at a specific direction and distance. The Gray-Level Run-Length Matrix (GLRLM) is used to calculate the length and number of continuous voxels (runs) with the same gray level. The Gray-Level Size Zone Matrix (GLSZM) is used to calculate the size and number of connected regions (size zones) with the same gray level. The Neighboring Gray-Level Dependence Matrix (NGLDM) is used to calculate the dependency relationship between each voxel and the gray level differences of its neighboring voxels.

[0109] Texture features are quantitative indicators calculated from the texture matrix and are used to describe the spatial structural characteristics of an image, such as roughness, uniformity, and periodicity.

[0110] In one example, the calculation process of the texture features in step 230 is illustrated using the Gray-Level Co-occurrence Matrix (GLCM): (1) First, determine the parameters of GLCM: the directions are usually considered as 13 directions in 3D space (such as (1,0,0), (0,1,0), (0,0,1) and the diagonal direction), and the distance is usually set to 1 voxel. For each direction, count the occurrence of all adjacent voxel pairs (gray levels i and j) to form a 64×64×64 co-occurrence matrix. Normalize the co-occurrence matrix of each direction to obtain the probability P(i,j) of each gray level pair.

[0111] (2) Based on the normalized GLCM, calculate the following texture features: Contrast: This measures the intensity of local grayscale changes; a larger value indicates a clearer texture and a more dramatic change.

[0112] Entropy: This measures the randomness and complexity of image texture; a larger value indicates a more complex and uneven texture.

[0113] Homogeneity: This measures the uniformity of local gray levels; a larger value indicates a more uniform texture and smoother changes.

[0114] Correlation: It measures the degree of linear dependence of gray values ​​of adjacent pixels; the larger the value, the stronger the directionality of gray value distribution.

[0115] Similarly, if the gray run length matrix (GLRLM) is used, features such as short run advantage, long run advantage, and gray level non-uniformity can be calculated; if the gray size matrix (GLSZM) is used, features such as small cell advantage, large cell advantage, and regional percentage can be calculated.

[0116] Then, steps 310 to 330 are repeated for each skull sub-region to obtain the feature vector F of each sub-region. k (Including robust statistical features and texture features), for example: for eight skull sub-regions, eight feature vectors are obtained respectively. Each F k It includes multiple robust statistics (such as median, P10, P90, IQR) and multiple texture features (such as contrast, entropy, homogeneity, correlation). These feature vectors will be used for subsequent concatenation and model input.

[0117] As can be seen from the above description, the Alzheimer's disease risk assessment method based on multimodal medical images provided in this application can significantly improve the stability and information dimension of skull Aβ signal representation by extracting robust statistics that are insensitive to noise and multi-type texture features that are sensitive to spatial patterns, thus providing more discriminative input features for subsequent risk assessment models.

[0118] To further integrate radiomics features from multiple cranial subregions and construct a stable and reliable predictive model using historical data to output personalized clinical risk assessment results, this application provides a method for Alzheimer's disease risk assessment based on multimodal medical imaging, see [link to relevant documentation]. Figure 2 Step 400 in the Alzheimer's disease risk assessment method based on multimodal medical imaging specifically includes the following: Step 410: Following a fixed anatomical order corresponding to the standard skull partition template, the radiomics features extracted from all the skull sub-regions are stitched together to form a feature vector.

[0119] It should be noted that the fixed anatomical order is a pre-determined arrangement of the skull sub-regions, consistent with the region order in the standard skull partitioning template, ensuring that the feature vectors of each subject correspond to the same anatomical region at the same location. In this embodiment, the standard skull partitioning template divides the skull into eight sub-regions, with the fixed anatomical order as follows: left frontal bone, right frontal bone, left temporal bone, right temporal bone, left parietal bone, right parietal bone, left occipital bone, and right occipital bone. The splicing refers to concatenating the feature vectors of multiple sub-regions end-to-end to form a longer one-dimensional vector. For example, if M features are extracted from each sub-region, the total length of the spliced ​​feature vector is 8 × M.

[0120] In step 410, feature vectors F1, F2, ..., F8 have been extracted from the eight cranial sub-regions, respectively. k The dataset contains M feature values, including robust statistical features (such as median, P10, P90, IQR) and texture features (such as contrast, entropy, homogeneity, and correlation). Following the fixed anatomical order described above, F1, F2, ..., F8 are concatenated to obtain the overall feature vector F = [F1, F2, ..., F8] for the subject. To further eliminate the influence of different feature dimensions, F can be standardized, for example, using Z-score standardization, so that the mean of each feature dimension on the training set is 0 and the variance is 1. The standardized parameters (mean, standard deviation) are fitted from the training set and applied to the current subject.

[0121] Step 420: Input the feature vector into the Alzheimer's disease risk assessment model so that the Alzheimer's disease risk assessment model outputs the corresponding predicted value; wherein, the Alzheimer's disease risk assessment model is obtained by pre-training a random forest model based on the multimodal medical imaging data and corresponding longitudinal cognitive change labels of each historical subject; the training includes: optimizing the hyperparameters of the random forest model using grid search or random search strategies and evaluating the performance of the random forest model using K-fold cross-validation.

[0122] Understandably, the Alzheimer's disease risk assessment model can employ a Random Forest model, an ensemble learning algorithm that improves prediction accuracy and stability by constructing multiple decision trees and combining their outputs. The predicted value is the raw result output by the model for the current subject's feature vector. For regression tasks, the predicted value can be a continuous numerical value (e.g., a predicted decrease in MMSE score); for classification tasks, the predicted value can be the probability of belonging to a certain category (e.g., the probability of progressing to AD). Longitudinal cognitive change labels are indicators of cognitive assessment changes obtained by historical subjects during follow-up, such as changes in MMSE scores, changes in CDR-SB scores, or binary labels indicating whether progress has been made from Mild Cognitive Impairment (MCI) to Alzheimer's disease. Grid Search is an optimization method that systematically traverses combinations of hyperparameters, exhaustively exploring all possible combinations within a predefined parameter space, evaluating the performance of each parameter group through cross-validation, and selecting the optimal combination. Random search refers to randomly sampling several sets of parameters within a predefined parameter space for testing. It is more efficient than grid search, especially suitable for large parameter spaces. K-fold cross-validation involves randomly dividing the training data into K equal parts, taking one part as the validation set and the remaining K-1 parts as the training set for model training and validation. This process is repeated K times, and the average performance metric is used as the model evaluation result. K=5 or 10 is commonly used.

[0123] During the model training phase, PET images, CT images, and corresponding longitudinal cognitive change labels of historical subjects can be collected. For each historical subject, the aforementioned registration, segmentation, partitioning, feature extraction, and stitching steps are performed to obtain the training feature vector F. train and its corresponding label y train (e.g., MMSE decrease value or MCI→AD label). This embodiment uses a random forest model as the base learner. During training, grid search is used to optimize hyperparameters: a hyperparameter search space is set, for example, the number of trees can be [100, 200, 500], the maximum depth can be [5, 10, None], and the minimum number of samples required for node splitting can be [2, 5, 10], etc. Five-fold cross-validation is used to evaluate the performance of each hyperparameter combination. The mean squared error (MSE) is used as the evaluation metric for regression tasks, and the area under the curve (AUC) is used as the evaluation metric for classification tasks. The hyperparameter combination with the best average performance in cross-validation is selected, and the model is retrained on all training data to obtain the final model.

[0124] In the prediction phase: for the current subject, their feature vector F is input into a pre-trained random forest model. The model consists of multiple decision trees, and for regression tasks, the average of the output values ​​of each decision tree is used as the final predicted value. For classification tasks, the average of the class probabilities output by each decision tree is used as the final predicted probability.

[0125] Step 430: Map the predicted value to an Alzheimer's disease risk assessment result that characterizes the subject's future cognitive decline risk and output it, wherein the Alzheimer's disease risk assessment result includes a continuous risk score or a discrete risk level.

[0126] The continuous risk score is a score within a numerical range, such as 0-100, with higher scores indicating a greater risk of cognitive decline. The discrete risk level can be divided into several categories based on preset thresholds, such as low risk, medium risk, and high risk.

[0127] In step 430, if the model output is a continuous predicted value (such as the predicted MMSE decrease score), it can be linearly scaled to map to a risk score of 0-100. For example, let the minimum MMSE decrease value on the training set be y. min The maximum value is y max Then the current subject's risk score is Score = 100 × ( -y min ) / y max -y min The risk score can be further divided into low risk (0-30), medium risk (31-60), and high risk (61-100) based on clinical experience by setting thresholds. If the model output is a classification probability (such as the probability of progression to AD), this probability value can be directly multiplied by 100 to obtain the risk score (0-100). Risk levels are then determined based on preset thresholds (e.g., probability <0.3 for low risk, 0.3-0.7 for medium risk, and >0.7 for high risk). Finally, the calculated risk score or risk level is output as the Alzheimer's disease risk assessment result, for example, displayed on a report interface or stored in a database for clinicians' reference.

[0128] As can be seen from the above description, the Alzheimer's disease risk assessment method based on multimodal medical images provided in this application ensures the structured and reproducible nature of the model input by splicing feature vectors in a fixed anatomical order, and obtains a prediction model with strong generalization ability by using random forest combined with hyperparameter optimization and cross-validation. The model output is mapped to an intuitive risk score or level, which can achieve seamless transformation from image features to clinical decision support.

[0129] From a software perspective, this application also provides a system for performing all or part of the Alzheimer's disease risk assessment method based on multimodal medical images, specifically a system for performing such assessments. See [link to relevant documentation]. Figure 3 The Alzheimer's disease risk assessment system based on multimodal medical imaging specifically includes the following components: The preprocessing module 10 is used to acquire multimodal medical image data containing positron emission tomography (PET) images and computed tomography (CT) images of the same subject, and to register the PET images and the CT images.

[0130] The skull segmentation module 20 is used to segment a skull region mask based on the density characteristics of bone tissue in the CT image, and to divide the skull region mask into multiple anatomically defined skull sub-regions using a standard skull partitioning template.

[0131] The feature extraction module 30 is used to extract radiomics features of each of the skull sub-regions from the PET image; wherein the radiomics features include robust statistical features to overcome partial volume effects and signal overflow, and texture features to characterize spatial distribution heterogeneity.

[0132] The risk prediction module 40 is used to combine the extracted radiomics features into a feature vector and input the feature vector into a preset Alzheimer's disease risk assessment model to obtain the Alzheimer's disease risk assessment result of the subject.

[0133] The embodiments of the Alzheimer's disease risk assessment system based on multimodal medical images provided in this application can be used to execute the processing flow of the embodiments of the Alzheimer's disease risk assessment method based on multimodal medical images in the above embodiments. Its functions will not be repeated here, but can be referred to the detailed description of the embodiments of the Alzheimer's disease risk assessment method based on multimodal medical images above.

[0134] The Alzheimer's disease risk assessment system based on multimodal medical imaging can perform the risk assessment on either a server or a client device. The choice can be made based on the processing power of the client device and the limitations of the user's usage scenario. This application does not impose any limitations on this. If all operations are performed on the client device, the client device may further include a processor for the specific processing of the Alzheimer's disease risk assessment based on multimodal medical imaging.

[0135] The aforementioned client device may have a communication module (i.e., a communication unit) that can communicate with a remote server to achieve data transmission with the server. The server may include a server on the task scheduling center side; in other implementation scenarios, it may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, a server cluster consisting of multiple servers, or a distributed server structure.

[0136] The server and the client device can communicate using any suitable network protocol, including those not yet developed as of the date of this application. Such network protocols may include, for example, TCP / IP, UDP / IP, HTTP, HTTPS, etc. Furthermore, such network protocols may also include RPC (Remote Procedure Call Protocol) and REST (Representational State Transfer Protocol) protocols used on top of the aforementioned protocols.

[0137] As can be seen from the above description, the Alzheimer's disease risk assessment system based on multimodal medical imaging provided in this application embodiment can stably and repeatedly extract biologically significant skull Aβ signals from mixed PET signals that are contaminated by thin layers of skull, spillover from adjacent brain tissue, and severe noise pollution, and then use them for individualized Alzheimer's disease risk assessment to effectively improve the accuracy of the assessment.

[0138] To further illustrate the above embodiments, this application also provides a specific application example of an Alzheimer's disease risk assessment method based on multimodal medical imaging, executed using an Alzheimer's disease risk assessment system based on multimodal medical imaging. Specifically, it is an automated analysis and risk assessment method for skull regions in PET or PET-CT images. The objectives of this application example include: 1) establishing a reproducible automatic skull segmentation and anatomical partitioning technique, transforming the skull from an "ignored background" into a "quantifiable analytical object"; 2) proposing a stable strategy for quantitative Aβ and radiomics feature extraction of the skull, addressing the physical and anatomical characteristics of skull PET imaging, overcoming instability caused by noise, thin-layer structures, and proximity interference; 3) constructing a predictive model based on the features of eight skull regions, predicting the cognitive change trend of subjects over a future period under baseline PET data conditions, and outputting an individualized progression risk score; 4) supplementing the limitations of traditional brain region Aβ-PET without changing existing scanning equipment and clinical procedures, thereby improving the individualized prediction capability of AD.

[0139] In this application example, the method is executed as follows: PET or PET-CT images are standardized and preprocessed; the skull region is automatically segmented in three-dimensional space and further subdivided into multiple functionally relevant sub-regions; Aβ-related radiomics features are extracted from these skull sub-regions; a random forest model is constructed based on the radiomics features of eight skull regions to predict future cognitive changes in the subject and output an individualized progress score. Within the framework of this application example, skull segmentation can employ various deep learning, template registration, or intensity-based traditional algorithms; radiomics features can be generated from hand-engineered features, autoencoder features, or deep networks; and the prediction model can employ statistical learning, machine learning, or deep learning structures.

[0140] The overall process of the Alzheimer's disease risk assessment method based on multimodal medical imaging is as follows: S1. Image Input and Preprocessing; S2. Automatic segmentation and quality control of the skull region; S3. Skull spatial standardization and anatomical partitioning (forming multiple skull sub-region masks). S4. Quantitative analysis of Aβ signals in subregions of the skull and extraction of radiomics features; S5. Feature standardization and feature vector construction; S6. Train or invoke the prediction model to output the cognitive decline prediction value and the individualized progress score.

[0141] In the above application example, the Alzheimer's disease risk assessment system based on multimodal medical imaging, in addition to the preprocessing module, skull segmentation module, feature extraction module, and risk prediction module mentioned above, also includes an image input module. Specifically: 1) Image input module: Receives PET or PET-CT data; 2) Preprocessing module: performs resampling, registration, smoothing, etc. 3) Skull segmentation module: Outputs skull mask and sub-region masks; 4) Feature extraction module: Outputs quantitative and radiomics features of each sub-region; 5) Risk prediction module: Outputs predicted values / risk scores based on the trained model.

[0142] The key steps of this application example are described in detail below.

[0143] (a) Image preprocessing The input image can be PET or PET-CT. In the case of pure PET: intensity / edge-based segmentation, template registration, or deep learning segmentation networks are used to directly locate the skull region from PET. In the case of MRI+PET: MRI bone / soft tissue separation or synthetic CT-assisted segmentation is used. Traditional methods such as different thresholding strategies, adaptive thresholding, image segmentation, or active contouring are used to replace the closing operation process.

[0144] Preprocessing includes, but is not limited to: Resampling: Resampling PET (and CT) images to a uniform voxel size and spatial resolution; Registration: Perform rigid or affine registration between PET and CT to place them in the same coordinate system; Smoothing and Denoising: Gaussian smoothing or median filtering is applied to the PET image to reduce noise and artifacts; Intensity normalization and partial volume correction: used to improve the quantitative stability of thin-layer structures.

[0145] (II) Automatic Segmentation and Partitioning of the Skull Region 1. CT-based skull segmentation Let the three-dimensional CT image be 3D PET images are Preprocessed output after registration , Utilizing the high-density characteristics of bone tissue in CT scans, a bone threshold is set. Construct the initial bone mask: right Perform 3D morphological closing operations and connected component analysis: in The structural element is used. Further, the most connected components and components located within the skull space are retained, while non-skull bone blocks such as teeth and cervical vertebrae are removed to obtain the skull mask. .

[0146] 2. Skull partition template and spatial mapping Predefined skull partition template in standard space The skull is divided into eight regions: left frontal bone, right frontal bone, left temporal bone, right temporal bone, left parietal bone, right parietal bone, left occipital bone, and right occipital bone. Let the labels take values. It is understandable that, instead of being limited to eight partitions, four, six, ten or more partitions may be used; continuous parameterized partitioning (e.g., stratification along the sagittal / coronal direction) or adaptive partitioning based on anatomical landmarks may be used.

[0147] right or Rigid body / affine registration is performed with a standard template to obtain the transformation. make: And define region labels in the subject space: Thus, the masks for each region are obtained: in , indicating different cranial regions.

[0148] (III) Quantitative analysis of Aβ in subregions of the skull and extraction of radiomic features 1. Quantitative indicators of the skull region Considering the low signal strength of the skull, its sensitivity to noise, and interference from adjacent brain tissue, this application example preferably calculates robust statistical indices at the sub-region level, for example: Average intake, maximum intake, median, interquartile range, skewness, kurtosis; Quantile statistics (such as P10 / P90) are used to reduce the impact of extreme noise points; Calculate the standardized SUVR index by selecting reference areas (cerebellar gray matter / whole cerebellum / white matter, etc.): 2. Radiomics features (texture + intensity + heterogeneity) For each sub-region Discretize the internal PET strength (gray volume scaled to a fixed number of grades) This involves constructing the Gray-Level Co-occurrence Matrix (GLCM), Gray-Level Run-Length Matrix (GLRLM), Gray-Level Region Size Matrix (GLSZM), and Neighborhood Gray-Level Dependency Matrix (NGTDM / GLDM). Based on these, texture features (contrast, homogeneity, entropy, correlation, run length, region size distribution, etc.) are calculated and combined with first-order statistical features to construct sub-region feature vectors. Finally, the features of the eight sub-regions are spliced ​​together in a fixed order: And on Standardize the data (mean 0, variance 1) and use it as model input.

[0149] In addition to manual image omics, deep features can be extracted using autoencoders / self-supervised learning; multi-scale filtering features (LoG, wavelet), gradient features, or sparse representation features can be added; and cross-regional features (left-right differences, inter-regional correlations) can be added to reflect spatial patterns.

[0150] (iv) Prediction Model and Reasoning Process 1. Tag Definition and Output Construct a supervised learning task based on follow-up data: Regression task: Predict changes in cognitive scale values ​​(e.g., decrease in MMSE, increase in CDR-SB). Classification task: Predict whether progress will occur (e.g., MCI→AD, or cognitive decline reaching a preset threshold).

[0151] The model output is: Predicted value (Continuous) or category probability (Classification); Map the output to an individualized progress score. For example, it can be linearly scaled to 0–100 and categorized into low / medium / high risk based on thresholds. The output can be a continuous risk score, a binary / multi-class risk level, or a prediction curve over time.

[0152] 2. Random Forest Model Structure This application example preferably uses the random forest model, which is composed of... It consists of several decision trees. Each tree generates a training subset based on bootstrap sampling of the training samples, and randomly selects candidate features from the feature subset for optimal splitting at each node split, thereby improving diversity and reducing overfitting.

[0153] Integration method: Regression: Take the average of the outputs of each tree; Classification: Output the probability of each tree by voting or taking the average probability.

[0154] Among them, logistic regression, SVM, XGBoost, neural networks, etc. can be used to replace random forest; multimodal fusion models can be constructed by combining clinical variables or brain region features (with skull features as the core input).

[0155] 3. Training steps and parameters The training process includes: 1) Data partitioning: Divide the samples into training and validation sets, or use K-fold cross-validation (K=5 or 10). 2) Feature preprocessing: Fit standardized parameters to the training set and apply them to the training / validation set; 3) Hyperparameter optimization: number of trees Grid / random search is performed based on factors such as maximum depth, number of selectable features per split, and minimum number of samples per leaf node. 4) Model evaluation: Regression uses mean squared error, correlation coefficient, etc.; classification uses AUC, accuracy, sensitivity / specificity, etc. 5) Model solidification: Save the optimal parameter model for inference.

[0156] 4. Reasoning process For new subjects, input PET or PET-CT, repeat S1-S5 to obtain feature vector F, input it into the trained model to obtain predicted output and progression score, and the results output module provides risk level and visual interpretation (optional: display the contribution of the most important skull region features).

[0157] In other words, this application provides the following application examples: 1) Automatic skull segmentation and non-skull bone removal methods, including technical solutions for obtaining skull masks based on CT thresholds, morphological closing operations, connected component screening, and spatial range constraints; and corresponding quality control and correction strategies (optional). 2) Skull anatomical partitioning and spatial mapping mechanisms, including the construction of standard spatial skull partition templates, registration of subject space and standard space, and steps for forming multiple skull sub-region masks (not limited to eight regions, but can be summarized as multiple sub-regions). 3) Aβ quantification and radiomics feature extraction strategies based on skull characteristics. This includes robust statistical quantification, reference area normalization (SUVR), grayscale discretization, texture matrix construction and high-order feature calculation; and a process of concatenating features from multiple sub-regions to form an overall feature vector and standardizing it. 4) A longitudinal cognitive decline prediction model and progression score generation based on skull features. This includes: input data (skull feature vector) — output (cognitive change prediction / risk probability) — score mapping and risk grading; the model can be a random forest or an equivalent learning model.

[0158] Based on this, the Alzheimer's disease risk assessment method and system based on multimodal medical imaging provided in this application example have the following beneficial effects: 1) The skull is treated as an independent structure for quantifiable analysis. Existing systems ignore skull signals. This application example establishes a complete link from segmentation to partitioning, quantification to modeling, making the "background signal" a "usable indicator".

[0159] 2) Address key technical obstacles related to skull imaging characteristics to improve repeatability. Measures such as CT-guided segmentation, connected domain removal of non-skull structures, standard spatial partitioning, robust statistics, and (optional) edge processing overcome quantitative instability caused by the thin-layer structure of the skull, low signal-to-noise ratio, and neighbor spillover interference.

[0160] 3) Introducing cranial radiomics to form a new dimension of Aβ phenotype. Unlike single-region SUVR, this application example uses multi-type texture and heterogeneity features to describe the spatial pattern of cranial Aβ, providing a new perspective for understanding AD heterogeneity and immune-related mechanisms.

[0161] 4) Achieving baseline-period prediction of future cognitive changes and obtaining unexpected technical effects. This application example predicts longitudinal cognitive changes through skull features, providing a new risk assessment path independent of traditional brain region analysis. Its predictive performance and correlation patterns cannot be reasonably expected by those skilled in the art based on the premise of "possible abnormal skull uptake".

[0162] 5) The project is feasible and scalable. The workflow can be applied to existing PET-CT examination data without relying on additional hardware, and has clinical translation potential; at the same time, it can be scaled to different tracers, different partitioning strategies and different machine learning models.

[0163] This application also provides an electronic device, which may include a processor, a memory, a receiver, and a transmitter. The processor is used to execute the Alzheimer's disease risk assessment method based on multimodal medical images mentioned in the above embodiments. The processor and memory can be connected via a bus or other means, taking a bus connection as an example. The receiver can be connected to the processor and memory via wired or wireless means.

[0164] The processor can be a central processing unit (CPU). The processor can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0165] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the Alzheimer's disease risk assessment method based on multimodal medical images in the embodiments of this application. The processor executes various functional applications and data processing by running the non-transitory software programs, instructions, and modules stored in the memory, thereby implementing the Alzheimer's disease risk assessment method based on multimodal medical images in the above method embodiments.

[0166] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the processor, etc. Furthermore, the memory may include high-speed random access memory and non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0167] The one or more modules are stored in the memory, and when executed by the processor, they execute the Alzheimer's disease risk assessment method based on multimodal medical imaging in the embodiment.

[0168] In some embodiments of this application, the user equipment may include a processor, a memory, and a transceiver unit. The transceiver unit may include a receiver and a transmitter. The processor, memory, receiver, and transmitter may be connected via a bus system. The memory is used to store computer instructions, and the processor is used to execute the computer instructions stored in the memory to control the transceiver unit to send and receive signals.

[0169] As one implementation method, the functions of the receiver and transmitter in this application can be implemented by transceiver circuits or dedicated transceiver chips, and the processor can be implemented by dedicated processing chips, processing circuits or general-purpose chips.

[0170] As another implementation approach, the server provided in this application embodiment can be implemented using a general-purpose computer. That is, the program code implementing the processor, receiver, and transmitter functions is stored in memory, and the general-purpose processor implements the processor, receiver, and transmitter functions by executing the code in memory.

[0171] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned Alzheimer's disease risk assessment method based on multimodal medical images. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0172] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the aforementioned Alzheimer's disease risk assessment method based on multimodal medical images.

[0173] In the research and development and implementation of the technical solutions involved in this application, all user personal information (if any) was processed in strict accordance with the principles of legality, legitimacy, necessity, and good faith. Specifically, the relevant data was obtained through one or more of the following compliant methods: 1) Before collecting users' personal information, the purpose, method, and scope of the collection have been clearly communicated to the users, and the users' individual and explicit authorization and consent have been obtained; 2) The personal data used comes from publicly available datasets permitted by laws and regulations, and the personal data has undergone necessary anonymization or de-identification processing during use to ensure that no specific individual can be identified and the information is irretrievable; 3) The use of personal data is limited to the technical research and development, model training and verification purposes described in this application, and strict technical and management measures have been taken to protect data security and prevent information leakage, abuse and unauthorized access.

[0174] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. The programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave.

[0175] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0176] In this application, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0177] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to the embodiments of this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for Alzheimer's disease risk assessment based on multimodal medical imaging, characterized in that, include: Acquire multimodal medical imaging data containing positron emission tomography (PET) images and computed tomography (CT) images of the same subject, and register the PET images and the CT images. Based on the density characteristics of bone tissue in the CT image, a skull region mask is segmented, and a standard skull partitioning template is used to divide the skull region mask into multiple anatomically defined skull sub-regions. Radiomic features of each of the skull sub-regions are extracted from the PET images; wherein the radiomic features include robust statistical features to overcome partial volume effects and signal spillover, and texture features to characterize spatial distribution heterogeneity. The extracted radiomics features are combined into a feature vector, and the feature vector is input into a preset Alzheimer's disease risk assessment model to obtain the Alzheimer's disease risk assessment results of the subject.

2. The Alzheimer's disease risk assessment method based on multimodal medical imaging according to claim 1, characterized in that, The registration of the PET image and the CT image includes: The PET images and the CT images are resampled to a uniform spatial resolution; Rigid body or affine registration is performed on the PET image and the CT image with uniform spatial resolution to align the spatial coordinate systems of the PET image and the CT image.

3. The Alzheimer's disease risk assessment method based on multimodal medical imaging according to claim 1, characterized in that, The process of segmenting the skull region mask based on the density features of bone tissue in the CT image includes: Voxels in the CT images with density values ​​equal to or higher than a preset bone density threshold are marked as bone tissue to generate an initial bone mask; A three-dimensional morphological closing operation is performed on the initial bone mask to fill the holes inside the initial bone mask and connect the adjacent bone tissue voxels to obtain a morphologically complete bone mask. Three-dimensional connected component analysis was performed on the complete bone mask to identify the independent bone structure image regions that are spatially unconnected. Based on the three-dimensional spatial position and shape of each independent bony structure image region, the bony structure image region corresponding to the cranial cavity is selected from each independent bony structure image region to generate a skull region mask.

4. The Alzheimer's disease risk assessment method based on multimodal medical imaging according to claim 1, characterized in that, The method of dividing the skull region mask into multiple anatomically defined skull sub-regions using a standard skull partitioning template includes: The CT image is spatially registered with a preset standard skull partition template, and the transformation parameters for transforming the spatial coordinate system of the CT image to the spatial coordinate system of the standard skull partition template are calculated. Based on the transformation parameters, the partition labels in the standard skull partition template are inversely mapped to the spatial coordinate system of the CT image to generate an individual partition label map; wherein, the position of each voxel in the individual partition label map is consistent with the position of the corresponding voxel in the CT image; In the individual partition labeling diagram, voxels whose spatial location coincides with the voxels within the skull region mask are identified as source voxels to be labeled. Each source voxel carries a partition label, which is then assigned to a target voxel in the same spatial location within the skull region mask, thereby dividing the skull region mask into multiple skull sub-regions.

5. The Alzheimer's disease risk assessment method based on multimodal medical imaging according to claim 1, characterized in that, The extraction of radiomics features of each of the skull sub-regions from the PET image includes: For each of the aforementioned cranial sub-regions, the following processing is performed: From the voxel intensity values ​​of the PET image corresponding to the skull sub-region, a robust statistic for overcoming partial volume effects and signal overflow is calculated; Furthermore, the voxel intensity values ​​of the PET image corresponding to the skull sub-region are subjected to grayscale discretization processing to obtain discretized grayscale values; wherein, the grayscale discretization processing includes: linearly or non-linearly mapping continuous PET intensity values ​​to a predetermined number of discrete grayscale levels; A texture matrix is ​​constructed based on the discretized gray values, and texture features for characterizing spatial distribution heterogeneity are calculated from the texture matrix. The texture matrix includes at least one of a gray-level co-occurrence matrix, a gray-level run-length matrix, a gray-level size region matrix, and a neighborhood gray-level dependency matrix. The texture features include at least one of contrast, entropy, a homogeneity index for representing the uniformity of gray levels, and a correlation index for representing the linear dependence of gray levels of adjacent pixels.

6. The Alzheimer's disease risk assessment method based on multimodal medical imaging according to claim 1 or 5, characterized in that, The robust statistical features include at least one of the median, quantile, and interquartile range of the voxel intensity values ​​of the PET image corresponding to the cranial sub-region.

7. The Alzheimer's disease risk assessment method based on multimodal medical imaging according to claim 1, characterized in that, Multiple anatomically defined cranial subregions are pre-divided based on at least three different anatomical bone blocks in the frontal, temporal, parietal, and occipital bones.

8. The Alzheimer's disease risk assessment method based on multimodal medical imaging according to claim 1, characterized in that, The step of combining the extracted radiomics features into a feature vector and inputting the feature vector into a preset Alzheimer's disease risk assessment model to obtain the Alzheimer's disease risk assessment result of the subject includes: Following a fixed anatomical order corresponding to the standard skull partition template, the radiomics features extracted from all the skull sub-regions are stitched together to form a feature vector; The feature vector is input into the Alzheimer's disease risk assessment model so that the Alzheimer's disease risk assessment model outputs the corresponding predicted value; wherein, the Alzheimer's disease risk assessment model is obtained by pre-training a random forest model based on the multimodal medical imaging data of each historical subject and the corresponding longitudinal cognitive change labels; the training includes: optimizing the hyperparameters of the random forest model using grid search or random search strategies and evaluating the performance of the random forest model using K-fold cross-validation; The predicted value is mapped to an Alzheimer's disease risk assessment result that characterizes the subject's future cognitive decline risk and is output, wherein the Alzheimer's disease risk assessment result includes a continuous risk score or a discrete risk level.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the Alzheimer's disease risk assessment method based on multimodal medical imaging as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the Alzheimer's disease risk assessment method based on multimodal medical images as described in any one of claims 1 to 8.

Images