Atherosclerosis model based on nano-molecule magnetic resonance imaging and machine learning algorithm, nano-targeting probe and application

Abstract

The invention belongs to the technical field of biological medicine and electronic information technology, and discloses an atherosclerosis model based on nano-molecule magnetic resonance imaging and a machine learning algorithm, a nano-targeting probe and application of the atherosclerosis model based on the nano-molecule magnetic resonance imaging and the machine learning algorithm. A high-resolution MRI image capable of specifically displaying foam macrophage distribution in the plaque is obtained; then, image omics features are extracted from the images, a machine learning model which is constructed based on the features and verified by pathological labels is adopted for analysis, and finally an objective and quantitative plaque vulnerability score is output. According to the model disclosed by the invention, an objective and quantitative risk score can be output by fusing nano-targeted molecular imaging and machine learning, and accurate and reliable quantification of plaque vulnerability is realized. According to the invention, through systematic integration of targeted molecular imaging and machine learning, significant technical progress is brought.

Description

Technical Field

[0001] This invention belongs to the fields of biomedicine and electronic information technology, and in particular to an atherosclerosis model, nano-targeting probe and its application based on nanomolecular magnetic resonance imaging and machine learning algorithms. Background Technology

[0002] Rupture of vulnerable atherosclerotic plaques is a major contributing factor to acute cardiovascular and cerebrovascular events such as ischemic stroke. The key factor determining plaque vulnerability is its intrinsic molecular pathological activity, particularly the degree of foam macrophage infiltration within the plaque. Currently, most clinically routine techniques, such as magnetic resonance imaging (MRI) or computed tomography (CT) angiography, are used to assess plaque vulnerability. Preclinical studies are exploring the use of nano-targeted probes to image the content of foam macrophages within plaques at the molecular level; and utilizing machine learning techniques to extract massive amounts of patient image features to establish radiomics models for assessing plaque vulnerability. Details are as follows: 1. Clinically routine techniques such as Time-of-Flight Magnetic Resonance Angiography (TOF-MRA), High-Resolution Vessel Wall Imaging (HR-VWI), and CT angiography primarily assess plaque vulnerability by acquiring images of the patient's carotid arteries, intracranial arteries, and coronary arteries. Radiologists then determine plaque stability based on macroscopic indicators such as the degree of luminal stenosis, plaque size, and morphological characteristics shown in the images. However, these techniques have the following limitations: (1) Poor consistency and reproducibility of assessment: The judgment of results is highly dependent on the subjective experience of physicians. Different observers have significant differences in their evaluation of the same image (e.g., low interpretation consistency coefficient), resulting in poor reproducibility of results.

[0003] (2) Inability to obtain key molecular information: Relying on physicians’ visual observation can only provide macroscopic morphological information, and cannot specifically obtain key imaging information that determines plaque vulnerability, such as foam macrophage infiltration, resulting in insufficient biological basis for risk assessment and limited accuracy.

[0004] (3) Lack of quantitative output indicators: The lack of objective and quantitative analysis indicators makes it difficult to quantify the vulnerability of plaques and accurately stratify the risks, thus failing to meet the needs of personalized medicine.

[0005] 2. Molecular-level imaging of the content of foam macrophages within plaques was performed using nano-targeted probes, as detailed below: (1) Construction of nano-targeting probes: A nanoscale contrast agent is used. This nanomaterial itself has physical properties suitable for MRI (such as Gd-based, Fe-based), CT, optics, or ultrasound. The nanoscale effect and high specific surface area can produce significant signal amplification. Based on this, the modularity of nanomaterials is further utilized to modify the surface of the nanomaterials with targeting elements (such as peptides, antibodies, or fragments thereof) that can specifically recognize foam macrophage surface markers (such as scavenger receptor SR-AI), thereby realizing the probe's active targeting capability for key pathological progressions of atherosclerosis.

[0006] (2) Imaging and signal acquisition: The above-mentioned nano-targeting probes were injected into the model animals. The probes bind to specific receptors on the surface of foam macrophages through their targeting elements, thereby achieving their accumulation in the plaque sites. Subsequently, scanning was performed using appropriate imaging equipment (such as MRI scanners, CT scanners, fluorescence imaging systems, and ultrasound). Due to the signal amplification effect of nanomaterials, the probes produce stronger imaging signals in the areas where foam macrophages are accumulated compared to the surrounding tissues.

[0007] (3) Image interpretation: Researchers usually conduct empirical and semi-quantitative analysis of foam macrophage content based on the intensity and range of imaging signals. This method relies on manual qualitative analysis and cannot establish a precise and stable correlation between imaging signals, foam macrophage content and plaque vulnerability. Therefore, the assessment of plaque vulnerability remains at the level of empirical inference and has limited accuracy.

[0008] While this technical solution successfully elevates the imaging level from macroscopic morphology to the molecular level, it has certain limitations in information interpretation and quantification output. The main drawbacks are as follows: (1) Subjective output of results and lack of objective quantitative standards: The final output of this scheme depends on the professional knowledge and experience of the observer. Different researchers may have different interpretations of signal intensity and range in the same image, resulting in poor reproducibility of the results and limiting its application as a reliable biomarker in precision medicine.

[0009] (2) Insufficient depth of information mining, remaining at the "visual level": The human eye has limited ability to interpret images and cannot extract a large number of deep-level quantitative features closely related to pathophysiological states from complex image data. Therefore, this technical solution is difficult to truly interpret the effective information contained in molecular images that is highly correlated with the vulnerability of plaques, and thus it is difficult to establish an accurate and stable mapping relationship between imaging signals and plaque vulnerability.

[0010] (3) Difficulty in accurately monitoring subtle changes: In the evaluation of treatment efficacy, it is difficult to distinguish between “significant improvement” and “slight improvement” based on subjective interpretation, thus making it difficult to meet the clinical needs of dynamic and accurate efficacy evaluation.

[0011] 3. Utilize machine learning techniques to extract massive amounts of patient image features and establish a radiomics model to determine plaque vulnerability. This existing technical solution represents a class of radiomics analysis methods based on conventional medical imaging (such as non-targeted MRI and CT). Its technical solution typically includes the following steps: (1) Image acquisition and preprocessing: Acquire standard clinical medical imaging data of patients. These images usually focus on showing the anatomical structure of blood vessels, the degree of luminal stenosis and the morphological characteristics of plaques, and cannot specifically target a molecular pathological target related to plaque vulnerability.

[0012] (2) Feature extraction: Using radiomics software platforms, hundreds to thousands of quantitative image features are extracted in high throughput from manually or semi-automatically delineated patch areas. These features typically include first-order features, shape features, texture features, etc.

[0013] (3) Model Development and Training: The extracted radiomics features are correlated with clinical endpoints (such as subsequent stroke events) or validated pathological criteria (such as plaque vulnerability determined from specimens obtained after endarterectomy) to construct a supervised machine learning model. Subsequently, the high-dimensional features are reduced in dimensionality using feature selection methods (such as removing redundancy and selecting features with strong discriminative power). Finally, a model for plaque vulnerability classification or risk prediction is trained using machine learning algorithms (such as logistic regression, support vector machine, random forest, etc.).

[0014] Although this technical solution incorporates objective quantitative analysis and machine learning models, its core effectiveness is limited by the informational constraints of the raw image data itself, and it mainly suffers from the following drawbacks: (1) Low dimensionality of original image information leads to insufficient predictive performance of the model: The conventional or non-targeted enhanced images used in this scheme mainly reflect the macroscopic physical characteristics of plaques (such as size, shape, physical density of components, etc.), but cannot specifically capture key molecular pathological activities directly related to plaque vulnerability (e.g., the degree of infiltration of foam macrophages, a core inflammatory marker). This results in a fundamental bottleneck in the accuracy of risk stratification of the model constructed from features extracted from such images with insufficient information dimensionality.

[0015] This invention demonstrates through comparative experiments that the radiomics model constructed based on clinically routine non-targeted gadolinium contrast agents (Gallicarpain®) enhanced MRI images exhibits significantly lower performance in predicting plaque vulnerability compared to the method described in this invention. To verify the advancement of this invention, it was compared with a model constructed based on clinically routine non-targeted gadolinium contrast agents (Gallicarpain®) enhanced MRI images. Specifically, on the independent validation set, the area under the receiver operating characteristic (AUC) of the Gallicarpain® model was only 0.56, while the AUC of the model described in this invention was as high as 0.87. Furthermore, as... Figure 10 As shown, other key indicators also demonstrate the superiority of this invention: the accuracy, sensitivity, and specificity of the Galerkin® model are only 65%, 60%, and 68%, respectively, while the accuracy, sensitivity, and specificity of the nano-targeted probe model in this invention reach 83%, 85%, and 80%, respectively. This fully proves that the model constructed based on targeted molecular images with higher information abundance has achieved a fundamental leap in performance.

[0016] (2) The imaging features have a weak correlation with pathobiology, resulting in poor interpretability of model decisions: Since the extracted radiomics features originate from macroscopic morphological images, their correlation with the underlying specific pathobiological processes is often indirect and speculative, lacking clear biological mechanism support. This makes the model a "black box" disconnected from the core pathology, and its decision-making basis is difficult for clinicians to understand and trust, severely limiting its application in precision medicine decision-making that requires clear biological evidence.

[0017] In summary, there is an urgent need for a new related method or technology to meet the usage requirements. Summary of the Invention

[0018] The purpose of this invention is to overcome the shortcomings of the prior art and provide an atherosclerosis model, a nano-targeting probe, and its applications based on nanomolecular magnetic resonance imaging and machine learning algorithms.

[0019] The technical solution adopted by this invention to solve its technical problem is: A nano-targeting probe for targeting foam macrophages, the preparation method of the nano-targeting probe includes the following steps: Freshly prepared 24 mM chloroauric acid aqueous solution and 46.1 mg / mL glutathione aqueous solution were added together to deionized water and magnetically stirred at 1000 rpm for 20 minutes at room temperature to obtain a colorless and transparent GSH-Au complex solution; wherein the volume ratio of chloroauric acid aqueous solution: glutathione aqueous solution: deionized water was 12.6:3.0:130. Take the above GSH-Au complex solution and mix it with 1 mg / mL PP1 peptide aqueous solution. Place the mixture in 10 mM phosphate-buffered saline (PBS) at pH 7.4 and incubate at 37°C with shaking for 2 hours at a shaking frequency of 60 times / min to obtain the GSH-Au-PP1 complex solution; wherein the volume ratio of GSH-Au complex solution to PP1 peptide aqueous solution is 1.0:0.1. Diethylenetriaminepentaacetic dianhydride and the GSH-Au-PP1 complex solution prepared in the previous step were dissolved together in 10 mM PBS buffer at pH 8.5 and reacted at 25°C for 2 hours. Subsequently, gadolinium trichloride hexahydrate was added to the reaction system, and the pH of the system was adjusted to 6.5 with 0.1 M sodium acetate solution, and the reaction continued. The ratio of diethylenetriaminepentaacetic dianhydride:GSH-Au-PP1 complex solution:PBS buffer:gadolinium trichloride hexahydrate (mg:mg:mL:mg) was 400:400:10:200. After the reaction was complete, the mixed solution was centrifuged at 4°C and 14,000 rpm for 30 minutes. The supernatant was discarded, and the precipitate was collected, which is the foam macrophage-targeting nanoprobe, i.e., the nano-targeting probe.

[0020] Furthermore, the nano-targeting probe possesses the following physicochemical properties: ① Morphology and stability: Transmission electron microscopy images showed that the nano-targeted probe was uniformly spherical with good dispersibility and stability. Dynamic light scattering analysis determined its hydrated particle size to be approximately 6.37 ± 0.16 nm, polydispersity index (PDI) to be 0.18, and zeta potential to be -4.07 ± 0.61 mV. ② Relaxation performance: The longitudinal relaxation rate of this nano-targeted probe is as high as 20.18 mM. -1 ·s -1 This indicates that it is a highly efficient T1-weighted MRI contrast agent; ③ Targeting efficacy: In vitro cell confocal imaging and quantitative fluorescence intensity analysis showed that the nano-targeting probe can efficiently and specifically recognize and bind to foam macrophages, and its binding efficiency to foam macrophages is higher than that of ordinary macrophages.

[0021] An atherosclerosis model based on nanomolecular magnetic resonance imaging and machine learning algorithms, utilizing the nano-targeting probes described above, is proposed. The atherosclerosis model first uses nano-targeting probes targeting foam macrophages to perform enhanced MR imaging to obtain high-resolution MRI images that specifically show the distribution of foam macrophages within plaques. Subsequently, radiomics features are extracted from the images, and a machine learning model constructed based on these features and validated with pathological labels is used for analysis. Finally, an objective and quantitative plaque vulnerability score is output.

[0022] Furthermore, it includes the following steps: (1) Image acquisition: Acquire MRI data of the target region after enhanced scanning by nano-targeting probes targeting foam macrophages; (2) Generation of three-dimensional reconstruction and quantitative distribution map: In order to realize the visualization and quantitative analysis of the spatial distribution of foam macrophages, the MRI data is converted into a three-dimensional reconstruction quantitative distribution map of foam macrophages in the target area through image segmentation and three-dimensional reconstruction technology; (3) Radiomics feature extraction: Extract radiomics features of patch vulnerability from the three-dimensional reconstructed quantitative distribution map; (4) Model building and risk assessment: Based on the radiomics features obtained in step (3), a machine learning model is built and applied to output a quantitative plaque vulnerability score.

[0023] Furthermore, step (1) includes the following steps: Images and data were acquired using MR imaging sequences and high-resolution vessel wall imaging technology to obtain isotropic vessel wall images with a high contrast-to-noise ratio (CNR), thereby clearly displaying plaque morphology and suppressing interference from intraluminal blood flow signals.

[0024] Furthermore, the generation of the three-dimensional reconstructed quantitative distribution map in step (2) includes the following steps: 1) Image segmentation: To improve the continuity of segmentation boundaries and suppress image noise, the original MRI image was subjected to Gaussian smoothing filtering before segmentation. Specifically, a Gaussian kernel G with a standard deviation σ = 0.3 mm was used. σ The image I is obtained by convolving it with the original image. smooth The preprocessed MR imaging data is then processed to generate masks for the target blood vessels (vascular mask) and foam macrophage-rich plaques (plaque mask), respectively. The specific formulas are as follows:

[0025] Among them: I smooth Let I be the smoothed image, G be the Gaussian kernel function, (x',y',z') be the coordinates of the convolution domain, and (x,y,z) be the spatial coordinates of any voxel (3D pixel) in the smoothed image. The segmentation process consists of two parts: ① Blood vessel segmentation: By tracing the boundary of the blood vessel lumen, a blood vessel mask reflecting the overall three-dimensional geometry of the target blood vessel is generated; the segmentation is achieved through manual, semi-automatic, or fully automatic algorithms; ② Patch Segmentation: Based on the signal intensity enhanced by the nano-targeted probe, patch regions enriched with foam macrophages are identified and segmented. A semi-automatic segmentation algorithm is used, which uses a CNR of at least 3 times the baseline value (CNR ≥ 3) as the growth threshold, supplemented by a region growing algorithm for expansion. Seed points are selected based on the location of the maximum signal intensity in the image. If multiple maximum signal intensity values ​​exist, the point with the largest connected region area is selected as the initial seed point. For each neighboring voxel (i, j, k) in the current region, if its intensity value I(i, j, k) satisfies the following formula, the voxel is included in the growth region: The specific formula is as follows:

[0026] Where I(i, j, k) is the signal intensity of the voxel to be determined; I seed The signal strength of the seed point; CNR is the contrast-to-noise ratio (preset threshold ≥3); σ background The standard deviation of the background noise is obtained by manually delineating one or more regions of interest (ROIs) in the signal-free air regions of the image, calculating the standard deviation of the intensity of all pixels within the delineated ROI, and using this as σ. background The value; 2) Quantitative distribution map generation: The segmented two-dimensional blood vessel mask and plaque mask are input into the three-dimensional rendering engine. Through volumetric rendering technology, the above-mentioned blood vessel and plaque masks are reconstructed into three-dimensional models. Subsequently, the three-dimensional model of the plaque is spatially fused with the three-dimensional model of the target blood vessel to finally generate the quantitative distribution map of foam macrophages. This distribution map is a three-dimensional visualization model that can intuitively and three-dimensionally show the specific spatial location, morphology and distribution range of plaques rich in foam macrophages, i.e., enhanced signal regions, in the target blood vessel wall.

[0027] Furthermore, in step (3), radiomics analysis methods are used to extract radiomics features from the segmented patch regions in high throughput. Before feature extraction, image grayscale normalization and resampling are required to ensure data consistency.

[0028] Furthermore, the construction and application of the machine learning model in step (4) includes the following steps: 1) Training Data Preparation and Preprocessing: A training dataset labeled with the gold standard pathology data was constructed. Each sample includes a plaque radiomics feature vector extracted from nano-targeted probe-enhanced MRI images, and its corresponding binary classification label. The binary classification label is obtained based on the staining results of the gold standard pathological sections, and the calculation formula is as follows:

[0029] Among them: macrophage area was obtained by CD68 immunohistochemical staining quantitative analysis; lipid area was obtained by Oil Red O staining quantitative analysis; smooth muscle cell area was obtained by α-SMA immunohistochemical staining quantitative analysis; collagen area was obtained by Masson trichrome staining quantitative analysis. After obtaining the pathological vulnerability index of each plaque, it is classified as "vulnerable" or "stable" according to a preset threshold value; Before model training, the radiomics features are standardized and preprocessed to eliminate dimensional differences and improve model stability and generalization ability. The Z-score standardization method is used, and the calculation formula is as follows:

[0030] Where X is the original radiomics feature value (unprocessed initial feature data); μ is the mean of the radiomics feature across all training samples (a global statistic that needs to be calculated based on the entire training set); and σ is the standard deviation of the radiomics feature across all training samples (a global statistic that, along with the mean, is calculated based on the same training set).

[0031] 2) Feature Selection: To select the most relevant and stable subset of features from the extracted features, a multi-stage feature selection strategy is adopted, specifically including: ① Stability Filtering: Calculate the intragroup correlation coefficient (ICC) of all initial features and remove features that are below a preset stability threshold. The ICC calculation formula is as follows:

[0032] Among them, MS between For mean square between groups, MS within The mean square of the group is given, and k is the number of times the measurement was repeated.

[0033] ② Collinearity Dimensionality Reduction: Analyze the pairwise correlations between the remaining features, remove redundant features, and calculate collinearity using the Pearson correlation coefficient (with a collinearity threshold of |r|>0.9). The specific formula is as follows:

[0034] Among them, X i and Y i These are the i-th observations of the two features to be analyzed; and The mean of two characteristic observations; n is the total number of observations; ③ Discriminative selection: Using methods based on statistical tests or machine learning embedding, a predetermined number of features most relevant to patch vulnerability are selected to form the final modeling feature subset; 3) Model training, validation, and score generation ① Model Training: Using the standardized feature subset and its corresponding labels, train a binary classification machine learning model; input the final feature subset, divide it into training and validation sets according to a preset ratio, and fine-tune the key parameters of the model using hyperparameter optimization methods to maximize the model's performance on the validation set; for example, the formula for calculating the linear prediction value of the logistic regression model is:

[0035] Where z is the linear prediction value, β0 is the intercept, and β i X is the weight coefficient corresponding to the i-th feature. norm, i Let i be the i-th standardized feature value; ② Model performance validation: The trained model needs to be validated on an independent validation set to prove its diagnostic efficacy; ③ Risk score generation and classification rules: The trained model can process the standardized features of new samples and output a continuous probability value between 0 and 1. This value is defined as a quantitative plaque vulnerability score. The score is calculated by mapping the linear prediction values ​​obtained by the model to a probability, that is, by using the Sigmoid function to map the linear prediction values ​​to the predicted probability of vulnerable plaques. The calculation formula is as follows:

[0036] Where e is the natural constant (Euler number), approximately equal to 2.71828; z is the linear prediction value obtained through the first step. To translate the predicted probabilities into clinical decisions, a binary classification rule needs to be established. This rule is established as follows: Based on the receiver operating characteristic curve of the model on the validation set, the optimal classification cutoff value, i.e., the Youden index (sensitivity + specificity - 1), is selected. Based on this cutoff value, the following classification rule is established: if the nano-AML score of a plaque is not lower than this value, it is classified as a "vulnerable plaque"; if the score is lower than this value, it is classified as a "stable plaque," thus enabling the assessment of the vulnerability of atherosclerotic plaques.

[0037] The application of the atherosclerosis model described above in assessing the vulnerability of atherosclerotic plaques.

[0038] An auxiliary diagnostic model utilizing the atherosclerosis model based on nanomolecular magnetic resonance imaging and machine learning algorithms as described above, the auxiliary diagnostic model comprising: (1) Data acquisition interface: used to read MRI imaging data in DICOM format from MRI scanners or medical image storage and transmission systems; (2) Image processing engine: Calls the image segmentation algorithm and the three-dimensional reconstruction routine to process the MRI imaging data and generate a quantitative distribution map of foam macrophages in the target area; (3) Feature calculation unit: Based on the predefined image omics feature set, automatically extract quantitative features from the quantitative distribution map and output them as structured data; (4) Core of risk assessment: Load and run the trained machine learning model, receive feature data, calculate and output a quantitative plaque vulnerability score; Alternatively, it may include (5) a report generator: integrating and visualizing the above analysis results in the user interface.

[0039] The advantages and positive effects of this invention are as follows: 1. To address the shortcomings of existing plaque vulnerability assessment methods, which rely on physician subjective experience leading to poor repeatability and lack specific detection capabilities for key molecular pathological information such as foam macrophages, resulting in insufficient accuracy, this invention provides a quantitative assessment model. This model, by integrating nano-targeted molecular imaging and machine learning, can output an objective and quantitative risk score, achieving accurate and reliable quantification of plaque vulnerability. This invention brings significant technological advancements through the systematic integration of targeted molecular imaging and machine learning.

[0040] 2. The objectivity and repeatability of the assessment are significantly improved by the present invention: The risk score provided by the present invention is an objective quantitative indicator, which can reduce the heavy reliance on the physician's subjective experience in the process of image interpretation. Figure 5 The advantages of using nanomolecular targeting probes in this invention have been systematically demonstrated through multiple experimental data. For example... Figure 5 The imaging indicators (vascular wall area and plaque area) measured based on this invention are highly consistent with pathological results, with a correlation coefficient r>0.9 and p<0.001, and are significantly superior to the traditional contrast agent PlusX®, demonstrating the advantage of nanomolecular imaging in accurately reflecting real anatomical structures. Figure 5 The CNR values ​​(e and f) used to evaluate the imaging effect of this invention were highly positively correlated with the expression level of the key target SR-AI within the plaque (Pearson correlation coefficient R = 0.77–0.85). In contrast, the CNR of the Galerxin® control group showed a weaker correlation with SR-AI expression levels (R = 0.24–0.62). The correlation coefficient of the imaging signal of this invention was significantly higher than that of the control group. These results indicate that the CNR value of the nano-targeted probe-enhanced imaging of this invention is more closely associated with changes in SR-AI expression levels, suggesting superior specificity in identifying key pathological processes in foam macrophages.

[0041] Figure 5g,h indicates that the vessel wall area and plaque area detected from MRI images enhanced by nano-targeted probes correlate better with H&E staining measurements than with Galerix®. Furthermore, Figure 13 The Brand-Altmann analysis showed that, based on MRI images enhanced by the nano-targeted probes of this invention, the diagnostic results of the two observers were highly consistent (small mean difference and narrow 95% concordance threshold, p>0.05); in contrast, the diagnoses based on existing technology (MRI images enhanced by GaleSym®) showed poor consistency (large mean difference and wide 95% concordance threshold, p<0.05). This demonstrates that the present invention can effectively improve the stability and reproducibility of image interpretation.

[0042] 3. This invention significantly enhances diagnostic accuracy and molecular specificity: This invention utilizes nano-targeted probes to acquire imaging data that specifically reflects key pathological features of plaque vulnerability (the degree of foam macrophage infiltration). The nano-AML model constructed accordingly (i.e., a diagnostic model integrating nano-targeted probe-enhanced MR imaging and machine learning algorithms, hereinafter referred to as the nano-AML model) achieved excellent diagnostic efficacy on the validation set. Figure 10 As shown, its AUC is as high as 0.87, with sensitivity and specificity reaching 85% and 80%, respectively. Furthermore, its performance is significantly superior to the contrast model constructed based on clinically routine contrast agents (Gallicon®) enhanced MRI images, whose AUC is only 0.56 (p<0.05). Simultaneously, the image features extracted from the images of this invention show a higher correlation with the pathological gold standard (pathological vulnerability index). Figure 14 The p-value was higher than that of models constructed using clinically commonly used Galerix® enhanced MRI images (p<0.05). Figure 15 (p>0.05), demonstrating that the present invention can improve diagnostic accuracy at the molecular level by combining nano-targeted probes with machine learning technology.

[0043] 4. This invention achieves precise quantification and dynamic monitoring of plaque vulnerability: The nano-AML score (hereinafter referred to as nano-AML score) output by this invention is a continuous probability value, enabling precise quantification of plaque vulnerability, rather than a simple "yes / no" classification. This allows it to sensitively monitor subtle changes in plaques during treatment. Figure 12 As shown, in the treatment trial, the nano-AML score of the plaque significantly decreased from 0.821 (vulnerable) at baseline to 0.581 (stable) at follow-up, providing an objective and effective quantitative tool for efficacy assessment and precision medicine decision-making.

[0044] 5. This invention combines nano-targeted probes for identifying foam macrophages with HR-VWI technology; extracts stable and highly discriminative radiomics features from molecular images; and implements a process for outputting a quantitative risk score based on a machine learning model. This objective and quantitative assessment tool provides a new technical means for vulnerability studies of atherosclerotic plaques, preclinical evaluation of treatment effects, and future exploration of precision medicine. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of an overall process of the method of the present invention; Figure 2 The following are performance characterization diagrams of the nano-targeting probe in Example 1 of the present invention: a) transmission electron microscopy image, b) hydrated particle size, c) ultraviolet-visible spectra of the nano-targeting probe after 1 day and 1 week of storage and its aqueous solution photograph (upper right inset), d) relaxation efficiency characterization. Figure 3 The images show a) in vitro confocal imaging and b) average fluorescence intensity (scale bar = 20 μm) of macrophages, foam macrophages, and foam macrophages whose surface receptors were blocked by excessive blocking agent after co-incubation with nano-targeting probes for 6 h. Figure 4 This image shows a comparison of the in vivo enhanced imaging effects of the nano-targeted probe of this invention and a conventional clinical contrast agent (Gallicon®). a) A TOF-MRA image showing the location of plaque-induced stenosis in the left common carotid artery of a rabbit model of atherosclerosis fed a high-fat diet for 8 weeks; b) HR-VWI comparison images of atherosclerotic rabbits before and after injection of the conventional clinical contrast agent Gallicon® and the nano-targeted probe of this invention (both injected at a dose of 0.04 mmol Gd / kg body weight). Boxes indicate the areas of plaque progression, and arrows indicate specific plaque locations. Figure 5This invention provides a verification diagram of the correlation and reliability between nano-targeted probe-enhanced imaging and pathological results. Specifically: a) a schematic diagram of in vivo HR-VWI assessment of atherosclerotic rabbit plaque progression; b) a schematic diagram of segmentation of vessel area, lumen area, plaque area, and vessel wall area used for imaging analysis; c) axial HR-VW images of different plaque stages after intravenous injection of Galex® and nano-targeted probe (0.04 mmol Gd / kg body weight, n=3 biologically independent rabbits per group); and d) quantitative analysis of plaque-specific region of interest (CNR) values; e) SR-AI staining of plaque tissue sections showing SR-AI expression levels at different plaque stages (n=3 biologically independent rabbits per group, scale bar=500 μm); f) analysis of the correlation between SR-AI expression in plaques at different stages and the CNR of Galex® and nano-targeted probe using a simple linear regression model; and g) H&E staining images of the left common carotid artery at different plaque stages (scale bar=1000 μm). h) The vessel wall area and plaque area detected by HR-VWI enhanced by nano-targeted probes showed better correlation with H&E staining results than those of Galerix®. Figure 6 This is a schematic diagram of the three-dimensional image segmentation and reconstruction process in this invention; it sequentially illustrates the key processing steps for generating a three-dimensional quantitative distribution map from the original MRI image: a) the original HR-VWI image enhanced by a nano-targeted probe, including vessel segmentation (left image) and the three-dimensional structure of the target vessel extracted through image segmentation; and (right image) the foam macrophage-rich plaque region identified and segmented based on signal intensity; b) the final generated three-dimensional quantitative distribution map of foam macrophages that integrates vessel and plaque information. This process ensures the accuracy and reliability of the data upon which subsequent radiomics feature extraction is based; Figure 7 This is a quantitative distribution map of high-risk plaques (foam macrophage-rich plaques) generated by three-dimensional reconstruction in Example 1 of the present invention; wherein, a) the original MRI image of the carotid artery plaques of the rabbit atherosclerosis model obtained, b) the distribution map of foam macrophage-rich plaques generated by three-dimensional reconstruction, and c) Oil Red O staining; Figure 8 This is a flowchart of the radiomics feature selection and machine learning model construction process in this invention. This invention employs a multi-stage feature selection strategy to build a high-performance prediction model: First, a large number of initial radiomics features are extracted in high throughput from the patch region; then, through three steps—stability filtering, collinearity reduction, and discriminative selection—a subset of key features most relevant to and stable patch vulnerability is selected; finally, this feature subset is used to train the machine learning model, and its performance is evaluated using a validation set, ultimately outputting a quantitative patch vulnerability score. Figure 9The nine radiomics features and the nodal plot selected in Example 1 of this invention are shown below. Figure 10 This is a performance comparison chart between the nano-AML model and the model based on Galerfin® in this invention; the AUC of the Galerfin® model on the independent validation set is only 0.56, while the AUC of the model in this invention is as high as 0.87; furthermore, as Figure 10 Other key indicators also demonstrate the superiority of this invention: the accuracy, sensitivity, and specificity of the Galerxin® model are only 65%, 60%, and 68%, respectively, while the accuracy, sensitivity, and specificity of the nano-targeted probe model in this invention reach 83%, 85%, and 80%, respectively. Figure 11 This is a schematic diagram of an exemplary software architecture for the evaluation method described in this invention; Figure 12 This invention presents the dynamic monitoring results of atherosclerosis treatment based on animal models. Specifically: a) a comparison of nano-AML score (left) and pathological vulnerability index (right) of representative case 1 (scale bar = 500 μm); b) classification of case 1 as a "vulnerable plaque" based on nano-AML score (0.821, >0.681); c) high consistency between the pathological vulnerability index classification result (1.165, >1.024, determined as a vulnerable plaque) and the nano-AML score result of case 1; d) a comparison of nano-AML score (left) and pathological vulnerability index (right) of representative case 2 (scale bar = 500 μm); e) classification of case 2 as a "stable plaque" based on nano-AML score (0.397, <0.681); g) high consistency between the pathological vulnerability index classification result (0.847, <1.024, determined as a stable plaque) and the nano-AML score result of case 2.

[0046] Figure 13 This is a consensus graph of observer analysis in the Brand-Altmann analysis of this invention: the two observers had significant differences in their image judgments for the Galerxin® group (p < 0.05), but their judgments on the image results for the nano-targeted probe group were more consistent, with no significant difference (p > 0.05). Figure 14This is a correlation analysis of the radiomics features of the nano-targeted probe group in this invention with the pathological vulnerability index, and an evaluation of its plaque classification effect. Specifically, a) the heatmap shows a significant positive / negative correlation between typical radiomics features extracted from the nano-targeted probe-enhanced images and the pathological gold standard (pathological vulnerability index), demonstrating that these features have clear biological significance; b) the box plot further shows that the extracted radiomics features show a statistically significant difference (p<0.05) between pathologically confirmed "vulnerable" and "stable" plaques, highlighting its strong discriminative ability. Figure 15 This is a correlation analysis of radiomics features and pathological vulnerability index in the Galerkin® group of this invention, and an evaluation of the plaque classification effect. Among them, a) the heatmap shows that the radiomics features extracted from Galerkin® enhanced images have no significant correlation with the pathological vulnerability index, indicating that the features are weakly associated with core pathobiological processes; b) the box plot shows that the feature has no significant difference between "vulnerable" and "stable" plaques (p>0.05), confirming that features based on non-targeted images are insufficient in judging plaque vulnerability. Detailed Implementation

[0047] The present invention will be further described below with reference to the embodiments. The following embodiments are descriptive and not limiting, and should not be used to limit the scope of protection of the present invention.

[0048] The various experimental operations involved in the specific embodiments are all conventional techniques in the field. For parts not specifically annotated in this document, those skilled in the art can refer to various commonly used reference books, scientific and technological documents or related instructions and manuals prior to the filing date of this invention to carry out the operations.

[0049] This invention aims to construct a diagnostic model that integrates nanoprobe-assisted machine learning (hereinafter referred to as the nano-AML model) with nano-targeted probe-enhanced MR imaging and machine learning algorithms. The model first utilizes nanoprobes targeting foam macrophages for enhanced MR imaging to obtain high-resolution MRI images that specifically display the distribution of foam macrophages within plaques. Then, radiomics features are extracted from these images, and a machine learning model based on these features and validated using pathological labels (the gold standard) is employed for analysis. Finally, an objective and quantitative plaque vulnerability score (nano-AML score) is output. This risk score can determine the vulnerability of atherosclerotic plaques, thereby achieving precise risk stratification.

[0050] The core process of this method includes four key steps. See also Figure 1 It shows the overall flowchart of the method of the present invention.

[0051] 1. Image Acquisition: MRI data of the target region was acquired after enhancement scanning using a nano-targeting probe targeting foam macrophages. The nano-targeting probe is a nanoscale contrast agent capable of specifically recognizing markers on the surface of foam macrophages. It possesses an ultra-small hydration kinetic particle size and a high longitudinal relaxation rate (r1) to ensure good biocompatibility and efficient T1-weighted imaging contrast. For example, a PP1 peptide-functionalized gold-based nanocluster probe has a hydration kinetic particle size of approximately 6.37 nm and a longitudinal relaxation rate (r1) of approximately 20.18 mM. -1 ·s -1 It can specifically recognize the scavenger receptor SR-AI. For a specific method for preparing and characterizing the nanotargeting probe (e.g., a PP1 peptide-functionalized gold-based nanocluster probe), please refer to the detailed description in Example 1 of this specification.

[0052] The images and data are acquired through MR imaging sequences. In this invention, high-resolution vascular wall imaging (HR-VWI) is preferably used, which can effectively suppress blood flow signals and is one of the core black-blood imaging techniques to obtain isotropic, high CNR vascular wall images, thereby clearly displaying plaque morphology and suppressing interference from intraluminal blood flow signals.

[0053] 2. Three-dimensional reconstruction and quantitative distribution map generation: To achieve visualization and quantitative analysis of the spatial distribution of foam macrophages, this invention uses image segmentation and three-dimensional reconstruction techniques to convert the MR imaging data into a quantitative distribution map of foam macrophages within the target region. Specifically, the generation of the quantitative distribution map includes the following steps: (1) Image segmentation: To improve the continuity of segmentation boundaries and suppress image noise, Gaussian smoothing filtering was performed on the original MRI image before segmentation. Specifically, a Gaussian kernel with a standard deviation σ = 0.3 mm was used. σ The image is convolved with the original image to obtain the smoothed image (I). smooth The preprocessed MR imaging data is then processed to generate masks for the target blood vessels (vessel mask) and masks for foam macrophage-rich plaques (plaque mask), respectively. The specific formulas are as follows:

[0054] Among them: I smooth Let I be the smoothed image, G be the Gaussian kernel function, (x',y',z') be the coordinates of the convolution domain, and (x,y,z) be the spatial coordinates of any voxel (3D pixel) in the smoothed image.

[0055] The segmentation process consists of two parts: ① Vessel Segmentation: By tracing the vascular lumen boundary, a vascular mask reflecting the overall three-dimensional geometry of the target vessel (e.g., the carotid artery) is generated. Segmentation can be achieved manually, semi-automatically, or fully automatically. This step can be performed using the segmentation module of a medical image processing platform commonly used in the field (such as 3D Slicer, ITK-SNAP, or MITK software). For example, manual segmentation allows the operator to directly sketch the vascular lumen boundary on continuous axial images. Semi-automatic segmentation, based on manually sketching a few key slices, utilizes the platform's interpolation algorithm (such as linear interpolation) or contour propagation algorithm to generate the complete three-dimensional structure. Fully automatic segmentation can employ commonly used deep learning-based segmentation models (such as U-Net, V-Net, or their variants). These models need to be pre-trained using a dataset with labeled vascular boundaries.

[0056] Preferably, high-precision segmentation is achieved using manual initialization combined with a level set algorithm. Specifically, an initial contour (seed region) is first manually set in the image, and then the contour is evolved using a level set function, so that it automatically conforms to the luminal boundary of the blood vessel under the drive of features such as image gradients. The level set algorithm can be implemented through the built-in functions of the image processing platform (e.g., the Level Tracing module in 3D Slicer software) or by calling programming libraries (such as ITK, VTK). This method yields a continuous and smooth three-dimensional structural model of the target blood vessel.

[0057] ② Patch Segmentation: Based on the signal intensity enhanced by the nano-targeted probe, patch regions enriched with foam macrophages are identified and segmented. In one embodiment of the invention, a semi-automatic segmentation algorithm is employed. This algorithm uses a CNR value not less than 3 times the baseline value (CNR ≥ 3) as the growth threshold criterion, and is supplemented by a region growing algorithm for expansion. Seed points are selected based on the location of the maximum signal intensity in the image. If multiple maximum signal intensity values ​​exist, the point with the largest connected region area is preferentially selected as the initial seed point. For each neighboring voxel (i,j, k) in the current region, if its intensity value I(i,j, k) satisfies the following formula, the voxel is included in the growth region.

[0058] The specific formula is as follows:

[0059] Where I(i, j, k) is the signal intensity of the voxel to be determined; I seed The signal strength of the seed point; CNR is the specific noise ratio (preset threshold ≥ 3); σ backgroundThe standard deviation of background noise is obtained by manually delineating one or more Regions of Interest (ROIs) in the signal-free air region of the image, calculating the standard deviation of the intensity of all pixels within the delineated ROI, and using this as σ. background The value of .

[0060] (2) Quantitative Distribution Map Generation: The segmented two-dimensional vessel mask and plaque mask are input into the three-dimensional rendering engine. Using volumetric rendering technology, the vessel and plaque masks are reconstructed into three-dimensional models. Subsequently, the three-dimensional model of the plaque is spatially fused with the three-dimensional model of the target vessel to finally generate the quantitative distribution map of foam macrophages. This distribution map is a three-dimensional visualization model that can intuitively and three-dimensionally display the specific spatial location, morphology, and distribution range of plaques rich in foam macrophages (i.e., enhanced signal regions) in the target vessel wall.

[0061] The aforementioned segmentation and 3D reconstruction steps are executed by an image processing system containing corresponding algorithm modules. Specifically, in this embodiment, the image segmentation and 3D reconstruction functions are implemented using a medical image processing and analysis software platform (i.e., 3D Slicer software, version 5.22, https: / / www.slicer.org). The resulting quantitative distribution map provides a precise, quantifiable, and isotropic 3D spatial data foundation for subsequent high-throughput extraction of radiomics features. Figure 6 This demonstrates the process of 3D image segmentation and reconstruction. Figure 7 A quantitative distribution map of high-risk plaques (foam macrophage-rich plaques) generated by three-dimensional reconstruction in Example 1 is shown.

[0062] 3. Radiomics Feature Extraction: From the reconstructed 3D distribution map, radiomics features of patch vulnerability are extracted. This step utilizes radiomics analysis methods to extract a large number of quantitative image features in high throughput from the segmented patch regions. Before feature extraction, image grayscale normalization and resampling (e.g., 1.0 × 1.0 × 1.0 mm) are required. 3 (Volumetrics) to ensure data consistency.

[0063] 4. Model Construction and Risk Assessment: Based on the radiomics features extracted from the 3D reconstructed distribution map, a machine learning model is constructed and applied to output a quantitative patch vulnerability score (nano-AML score). The construction and application of the machine learning model includes the following steps: (1) Training Data Preparation and Preprocessing: A training dataset labeled with the gold standard pathology data is constructed. Each sample includes a plaque radiomics feature vector extracted from nano-targeted probe-enhanced MRI images, and its corresponding binary classification label (determined as "vulnerable" or "stable" based on the pathological vulnerability index). Preferably, the plaque classification label is obtained based on the staining results of the gold standard pathological sections. The calculation formula is as follows:

[0064] Among them: macrophage area was obtained by CD68 immunohistochemical staining quantitative analysis; lipid area was obtained by Oil Red O staining quantitative analysis; smooth muscle cell area was obtained by α-SMA immunohistochemical staining quantitative analysis; and collagen area was obtained by Masson trichrome staining quantitative analysis.

[0065] After obtaining the pathological vulnerability index of each plaque, it is classified as "vulnerable" or "stable" according to a preset threshold value (e.g., ≥1.024 is vulnerable, <1.024 is stable).

[0066] Before model training, the radiomics features are standardized preprocessed to eliminate dimensional differences and improve model stability and generalization ability. Preferably, the Z-score standardization method is used, and the calculation formula is as follows:

[0067] Where X is the original radiomics feature value (unprocessed initial feature data); μ is the mean of the radiomics feature across all training samples (a global statistic that needs to be calculated based on the entire training set); and σ is the standard deviation of the radiomics feature across all training samples (a global statistic that, along with the mean, is calculated based on the same training set).

[0068] (2) Feature selection: To select the most relevant and stable subset of features from the extracted features, this invention adopts a multi-stage feature selection strategy, specifically including: ① Stability Filtering: Calculate the intragroup correlation coefficient (ICC) for all initial features and remove features with an ICC below a preset stability threshold (e.g., ICC < 0.8). The ICC calculation formula is as follows:

[0069] Among them, MS between For mean square between groups, MS within The mean square of the group is given, and k is the number of times the measurement was repeated.

[0070] ② Collinearity Dimensionality Reduction: Analyze the pairwise correlations between the remaining features and remove redundant features. For example, use the Pearson correlation coefficient to calculate collinearity (the collinearity threshold is |r|>0.9). The specific formula is as follows:

[0071] Among them, X i and Y i These are the i-th observations of the two features to be analyzed; and The mean of two characteristic observations; n is the total number of observations.

[0072] ③ Discriminative selection: Using methods based on statistical tests (such as ANOVA) or machine learning embedding methods (such as L1 regularization of logistic regression), a predetermined number of features most relevant to patch vulnerability are selected to form the final feature subset for modeling. Figure 8 The feature selection process is demonstrated.

[0073] (3) Model training, validation and score generation ① Model Training: Using the standardized feature subset and its corresponding labels, train a binary classification machine learning model. Input the final feature subset, divide it into training and validation sets according to a preset ratio, and fine-tune the key parameters of the model using hyperparameter optimization methods to maximize the model's performance on the validation set. For example, the formula for calculating the linear prediction value of a logistic regression model is:

[0074] Where z is the linear prediction value, β0 is the intercept, and β i X is the weight coefficient corresponding to the i-th feature. norm, i Let be the i-th standardized feature value.

[0075] ② Model Performance Validation: The trained model needs to be validated on an independent validation set to demonstrate its diagnostic efficacy. Validation should comprehensively evaluate multiple metrics, including AUC, sensitivity, and specificity. The constructed model should significantly outperform the baseline model built based on non-targeted contrast agents or conventional imaging in the above metrics.

[0076] ③ Risk score generation and classification rules: The trained model can process the standardized features of new samples and output a continuous probability value between 0 and 1, which is defined as a quantitative plaque vulnerability score (nano-AML score). The score is calculated by probabilistically mapping the linear predictions obtained by the model; that is, using the Sigmoid function to map the linear predictions to the predicted probability of vulnerable plaques. The calculation formula is as follows:

[0077] Where e is the natural constant (Euler number), approximately equal to 2.71828; z is the linear prediction value obtained through the first step.

[0078] To translate the predicted probabilities into clinical decisions, a binary classification rule needs to be established. This rule is established as follows: based on the Receiver Operating Characteristic Curve (ROC curve) of the model on the validation set, the optimal classification cutoff value that maximizes the Youden index (sensitivity + specificity - 1) is selected. Based on this cutoff value, the following classification rule is established: if the nano-AML score of a plaque is not lower than this value, it is classified as a "vulnerable plaque"; if the score is lower than this value, it is classified as a "stable plaque".

[0079] Furthermore, based on the evaluation method described above, those skilled in the art can deploy it as an auxiliary diagnostic tool. See also Figure 11 This illustrates a feasible software architecture diagram. The tool can be developed using a general programming environment (such as Python) and corresponding algorithm libraries (such as ITK for image segmentation and 3D reconstruction, PyRadiomics for feature extraction, and scikit-learn for model loading and scoring), implementing the specific steps of the method of this invention as collaboratively runnable software modules. An exemplary modular design may include: (1) Data acquisition interface: used to read MR imaging data in DICOM format from MRI scanners or medical image storage and communication systems (PACS).

[0080] (2) Image processing engine: Calls the image segmentation algorithm (e.g., level set-based blood vessel segmentation, region growth-based plaque segmentation) and three-dimensional reconstruction routine as described in the specific implementation to process the MR imaging data and generate a quantitative distribution map of foam macrophages in the target area.

[0081] (3) Feature calculation unit: Based on the predefined image omics feature set, it automatically extracts quantitative features from the quantitative distribution map and outputs them as structured data (such as JSON format).

[0082] (4) Risk assessment core: Load and run the machine learning model (such as nano-AML model) trained according to the method described in this invention, receive feature data, calculate and output a quantitative plaque vulnerability score (such as nano-AML score).

[0083] (5) (Optional) Report Generator: Integrates and visualizes the above analysis results in the user interface, including quantitative distribution plots, nano-AML scores, and plaque classification conclusions.

[0084] The above modules can be integrated through standard software architectures (such as service-oriented architectures or simple script call chains) and deployed on workstations, servers, or cloud computing platforms, thereby providing convenient software support for implementing the evaluation method of this invention.

[0085] Specifically, the relevant preparation and testing methods are as follows: Example 1: Implementation of a plaque vulnerability assessment method based on nanomolecular imaging and machine learning in an animal model (1) Preparation and characterization of nano-targeting probes Preparation steps: Take 12.6 mL of freshly prepared chloroauric acid aqueous solution (HAuCl4, 24 mM) and 3.0 mL of glutathione aqueous solution (GSH, 46.1 mg / mL), and add them together to 130 mL of deionized water. Stir magnetically at 1000 rpm for 20 minutes at room temperature to obtain a colorless and transparent GSH-Au complex solution.

[0086] Take 1.0 mL of the above GSH-Au complex solution (60 mg / mL) and mix it with 0.1 mL of PP1 peptide aqueous solution (1 mg / mL). Place the mixture in 10 mM phosphate-buffered saline (PBS) at pH 7.4 and incubate at 37°C with shaking for 2 hours (shaking frequency of 60 times / min) to obtain the GSH-Au-PP1 complex solution.

[0087] 400 mg of diethylenetriaminepentaacetic dianhydride (DTPA dianhydride) and 400 mg of the GSH-Au-PP1 complex solution prepared in the previous step were dissolved together in 10 mL of PBS buffer (10 mM, pH 8.5) and reacted at 25 °C for 2 hours. Subsequently, 200 mg of gadolinium trichloride hexahydrate (GdCl3·6H2O) was added to the reaction system, and the pH of the system was adjusted to 6.5 with 0.1 M sodium acetate solution, and the reaction continued.

[0088] After the reaction was complete, the mixed solution was centrifuged at 4°C and 14,000 rpm for 30 minutes. The supernatant was discarded, and the precipitate was collected, which is the foam macrophage-targeting nanoprobe, i.e., the nanotargeting probe. The nanotargeting probe was redispersed and stored in 10 mM, pH 7.4 PBS buffer for later use.

[0089] The nano-targeting probes prepared by the above optimization method possess the following physicochemical properties (which are the basis for their function): ① Morphology and stability: Transmission electron microscope images ( Figure 2 a) shows that the nano-targeting probe is uniformly spherical, with good dispersibility and stability. Dynamic light scattering ( Figure 2 (b) The hydrated particle size was determined to be approximately 6.37 ± 0.16 nm, the polydispersity index (PDI) was 0.18, and the zeta potential was -4.07 ± 0.61 mV. These results indicate that the nano-targeting probe belongs to the range of ultra-small nanomaterials, and its surface potential is near neutral, which is beneficial for reducing non-specific adsorption in biological organisms.

[0090] ② Relaxation properties: such as Figure 2 As shown in d, the longitudinal relaxation rate of this nano-targeted probe is as high as 20.18 mM. -1 ·s -1 This indicates that it is a highly efficient T1-weighted MRI contrast agent.

[0091] ③ Targeting efficacy: In vitro cell confocal imaging ( Figure 3 a) and quantitative fluorescence intensity analysis ( Figure 3 (b) This demonstrates that the nano-targeting probe can efficiently and specifically recognize and bind to foam macrophages, with a higher binding efficiency to foam macrophages than to ordinary macrophages. Furthermore, this specific binding can be completely competitively inhibited by excess free PP1 peptide (target blocking agent), confirming that the nano-targeting probe achieves specific targeting of the scavenger receptor SR-AI on the surface of foam macrophages through high-affinity ligand-receptor (PP1-SR-AI) interaction.

[0092] (2) Image acquisition Animal model construction: The animal model of atherosclerosis was constructed using the widely accepted method of "balloon injury combined with a high-fat diet," and the specific steps are as follows: ① Male New Zealand white rabbits weighing 2.5 ± 0.2 kg were purchased from Beijing Huafukang Biotechnology Co., Ltd. The animals were acclimatized in the facility for at least one week before the experiment, and their health was observed and confirmed to be good.

[0093] ② Starting one week before balloon injury surgery, rabbits should be fed a restricted high-fat diet (150 grams per day, containing 0.75 grams of cholesterol, 15 grams of lard, and 15 grams of egg yolk powder).

[0094] ③ Balloon injury was performed on the left common carotid artery of the rabbit to induce endothelial ablation. The procedure is briefly described as follows: The animal was anesthetized by injecting 40 mg / kg sodium pentobarbital via the marginal ear vein; a midline incision was made in the neck to expose the left common carotid artery; a small incision was made in the external carotid artery, and a balloon catheter was inserted into the left common carotid artery; the balloon was inflated with water, and after dilating the blood vessels, it was slowly pulled back to the carotid bifurcation. This procedure was repeated three times to ensure complete endothelial ablation.

[0095] ④ After surgery, continue feeding the animals the above-mentioned high-fat diet to promote the formation of atherosclerotic plaques.

[0096] Animal model imaging: A New Zealand rabbit model of atherosclerosis was selected, which underwent carotid endothelial injury via balloon injury combined with a high-fat diet for 8 weeks. First, a clinically accepted TOF-MRA sequence was used to obtain vascular anatomy. The imaging parameters were: repetition time (TR) 20 ms, echo time (TE) 3.5 ms, flip angle 20°, and field of view 160 × 160 mm. 2 The matrix was 256×256 with a slice thickness of 0.8 mm. Subsequently, the foam macrophage-targeting nanoprobe prepared in Example 1 was injected via the ear vein (injection dose was 0.04 mmol / kg body weight, calculated as gadolinium). Ten hours post-injection, a high-resolution black-blood imaging sequence, commonly used in clinical practice, was used for scanning. This example employed the IR-SPACE sequence on the Siemens MRI platform, with the following key parameters: repetition time 1020 ms, echo time 15 ms, and field of view 200×134 mm. 2 MRI scans were performed using a matrix of 384×258 pixels, a slice thickness of 0.5 mm, and an acquisition time of 13 minutes and 16 seconds to effectively suppress blood flow signals and obtain isotropic high-resolution images of the vessel wall.

[0097] As can be seen from the comparison, the nano-targeted probe is effective in the narrow area shown in the TOF-MRA image ( Figure 4 Arrow a) produced specific signal enhancement ( Figure 4 (b arrow). Compared to conventional non-targeted contrast agents (Gallicon®), the nano-targeted probe of this invention exhibits more significant and specific signal enhancement in the plaque region. Figure 4 (As indicated by arrow b and box). This demonstrates that the probe can actively target and accumulate in plaques infiltrated by foam macrophages in vivo, specifically enhancing the MR signal of foam macrophages, laying a solid foundation for obtaining high CNR and high specificity foam macrophage-related imaging data.

[0098] (3) Three-dimensional reconstruction and quantitative distribution map generation The acquired DICOM format MR imaging data was imported into an image processing system (3D Slicer software, version 5.22, https: / / www.slicer.org) with image segmentation and 3D reconstruction capabilities. The process of generating a quantitative distribution map is as follows: ① Vessel segmentation: On continuous axial slices, the carotid artery lumen boundary is delineated by manual tracking combined with a level set algorithm, generating a vascular mask that reflects the three-dimensional geometry of the target vessel.

[0099] ② Plaque Segmentation: Based on the high signal enhanced by nano-targeted probes, a semi-automatic segmentation algorithm is adopted: first, seed points are set in the core of the plaque, and then a region growing algorithm is applied to expand the plaque with a CNR ≥ 3 as the threshold, thereby generating a plaque mask of foam macrophage-enriched plaques. A Gaussian smoothing filter (σ = 0.3 mm) is used for preprocessing before segmentation.

[0100] ③ Quantitative Distribution Map Generation and Data Analysis: The segmented vascular mask and plaque mask are reconstructed into 3D vascular models and 3D plaque models respectively using volumetric rendering technology. These two 3D models are then spatially fused to generate a 3D spatial distribution map (i.e., a quantitative distribution map) of the foam macrophages. Figure 6 , Figure 7 The process of segmentation and the steps involved in generating a three-dimensional quantitative distribution map are demonstrated.

[0101] (4) Image feature extraction Feature extraction was performed using the Deepwise Multimodal Research Platform (version 2.6.0). First, the images underwent normalization preprocessing: B-spline interpolation was used to resample the voxels to 1.0 × 1.0 × 1.0 mm. 3 According to the formula The image grayscale values ​​are normalized, where, I original The values ​​represent the original voxel grayscale values, and μ and σ are the mean and standard deviation of all voxel grayscale values ​​in the entire patch region, respectively. Subsequently, PyRadiomics was used to extract thousands of initial radiomics features from the patch region.

[0102] (5) Model building and risk assessment ① Training Data and Feature Preprocessing: This embodiment uses nano-targeted probe-enhanced HR-VWI data from 300 rabbit atherosclerotic plaques confirmed by pathological vulnerability index for training. First, 1204 initial radiomics features were extracted from the quantitative distribution map of each plaque using PyRadiomics. Subsequently, all features were Z-score standardized, i.e., for each feature, the mean of that feature across all samples was subtracted and then divided by the standard deviation.

[0103] ② Feature Selection: The initial features are selected through a three-step process: a) Stability filtering (ICC < 0.8); b) Collinearity reduction (|Pearson r| > 0.9); c) Discriminant selection: The analysis of variance (ANOVA) F-values ​​between the non-redundant features and the gold standard label are calculated and ranked in descending order. This embodiment uses the "k-optimal feature selection" method, directly selecting the top 9 features with the highest F-values ​​to form the final feature subset used for modeling. These nine key radiomics features include: square_glcm_JointEntropy, original_firstorder_RootMeanSquared, wavelet-LLL_firstorder_Maximum, square_glcm_JointEnergy, exponential_glszm_LowGrayLevelZoneEmphasis, square_glrlm_RunEntropy, wavelet-LLL_firstorder_TotalEnergy, square_glszm_ZoneEntropy, and wavelet-LHL_glrlm_ShortRunLowGrayLevelEmphasis. These features encompass first-order statistics and multi-dimensional texture information such as the gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), and gray-level region size matrix (GLSZM), enabling in-depth mining of the complex heterogeneity of patches in molecular images.

[0104] ③ Model Training: For example, using a logistic regression classifier. Input the final feature subset and divide it into training and validation sets in a ratio (e.g., 8:2). Optimize the regularization strength C (search range, e.g., [0.001, 0.01, 0.1, 1, 10, 100]) using a grid search method to maximize the AUC of the validation set.

[0105] Based on the nine key radiomics features selected above and their corresponding training data, this embodiment trains a logistic regression model. The standardized preprocessed values ​​of the nine radiomics features and their corresponding feature weight coefficients are then substituted into the linear prediction formula. The final linear prediction value of plaque vulnerability is obtained from this. In the formula, X... norm1 To X norm9 β1 to β9 are the corresponding feature weight coefficients obtained from model training; β0 is the bias term of the model.

[0106] Therefore, the specific formula for the linear prediction value z of patch vulnerability determined in this embodiment is as follows ( Figure 9 ):

[0107] ④ Model performance verification: The trained model exhibited excellent diagnostic performance on the independent validation set. For example... Figure 10 As shown in Figure a, the model achieved an AUC of 0.87, a sensitivity of 0.75, a specificity of 1, an accuracy of 0.857, and a precision of 1 on the validation set. To demonstrate the advantages of this invention, we compared it with a model constructed based on clinically routine contrast agents (Gallicon®) enhancing MRI images. Figure 10 b) Comparison. The results show that the AUC of the Galerxin® model is 0.538, the sensitivity is 0.6, the specificity is 769, the accuracy is 0.722, and the precision is 0.5. Therefore, the diagnostic performance of the model of this invention is superior to that of the Galerxin® model, confirming that molecular imaging data based on nano-targeted probes can effectively improve the discriminative ability of the model.

[0108] Example 2: Application of this quantitative model in monitoring the therapeutic effect of animal models This embodiment aims to demonstrate the unique capabilities and clinical application value of the plaque vulnerability quantitative assessment method provided by the present invention in dynamically and accurately monitoring the effects of drug treatment.

[0109] (1) Monitoring method: A New Zealand rabbit model of atherosclerosis (constructed by balloon injury of carotid artery endothelium combined with high-fat diet for 8 weeks) (n=8) was selected and randomly divided into treatment group (n=4, given a specific dose of atorvastatin calcium (or other equivalent statin) (2.5 mg / kg / day) for 8 weeks) and control group (n=4, given an equal volume of 0.9% saline. To evaluate the treatment effect, the following complete imaging procedure was performed on the model animals before and after drug treatment: foam macrophage-targeting nano-targeting probe (0.04 mmol / kg as Gd) prepared in Example 1 was injected into the marginal ear vein; 10 hours after injection, MRI scan was performed using HR-VWI sequence; the acquired MRI data were used to obtain the quantitative nano-AML score of each plaque by applying the image segmentation, three-dimensional reconstruction, feature extraction and risk score calculation method described in Example 1 of this invention (or the claims).

[0110] (2) Results and analysis: In a representative case, the specific plaque score of a model rabbit in the treatment group decreased from 0.821 (vulnerable) at baseline to 0.581 (stable) at follow-up, achieving a category conversion from "vulnerable" to "stable" (based on the preset classification cutoff value of 0.681).

[0111] (3) Pathological Verification: After treatment, the animals were euthanized and tissue samples were collected for standard histopathological analysis. Specifically, H&E staining, Masson's trichrome staining, Oil Red O staining, and CD68 immunohistochemical staining were performed on the plaque tissue sections to assess the overall morphology, collagen content, lipid composition, and macrophage infiltration of the plaques, respectively. Subsequently, the staining results were analyzed according to the pathological vulnerability index calculation formula defined in the "Specific Embodiments" of this invention to calculate the pathological vulnerability index of each plaque, which was then used as the gold standard for evaluating plaque vulnerability.

[0112] The results showed that the pathological vulnerability index of the plaque decreased from a baseline simulated value of 1.165 to 0.847 after follow-up, with reduced lipid components and foam cell infiltration, and increased collagen content. This result is highly consistent with the decreasing trend of the nano-AML score. Figure 12 This biologically confirms the reliability of the evaluation results of this invention. Figure 12 The results of the treatment efficacy assessment are presented. This embodiment demonstrates that the model in this invention has quantitative and objective characteristics, effectively overcoming the shortcomings of traditional imaging methods in efficacy assessment that rely on subjective interpretation and are difficult to quantify subtle changes. It provides an unprecedented technical means for the precise clinical assessment of the efficacy of anti-atherosclerotic drugs.

[0113] In summary, the present invention has the following advantages: (1) Advantages of the nano-targeted probe: To objectively assess the reliability of image interpretation, the Brand-Altmann analysis, a standard statistical method for assessing inter-observer consistency, was used to test the consistency of judgments between two radiologists. The results showed that for MRI images enhanced by the nano-targeted probe of this invention, the consistency in physicians' judgments regarding the presence of plaques was high, with a mean difference of 0.05 and a 95% consensus threshold of (-0.037, 0.027), and the difference was not statistically significant (p>0.05). In contrast, for MRI images enhanced by conventional clinical contrast agents (Gallicon®), the consistency in judgment was poor, with a mean difference increasing to 0.18 and a wider 95% consensus threshold range (-0.196, 0.024), and the difference was statistically significant (p<0.05). This demonstrates that the nano-targeted probe-enhanced images provided by this invention can reduce subjective differences in physician interpretation, thereby improving the stability and repeatability of diagnostic results.

[0114] (2) The radiomics features obtained by the method described in this invention have higher biological relevance. Analysis using Pearson correlation coefficients shows a higher correlation between the radiomics features extracted from MRI images enhanced by nano-targeted probes and the gold standard for pathology (pathological vulnerability index). Figure 14 a). Further independent samples t-tests were used for intergroup comparisons, and the expression levels of these features also showed statistically significant differences between pathologically confirmed "vulnerable" and "stable" plaques (p<0.05). Figure 14 b). The above statistical analyses can all be performed using standard scientific computing software in the field (such as SPSS, R, or the Python SciPy library). The above results demonstrate that the features extracted by this invention can directly and effectively reflect the pathobiological changes closely related to plaque vulnerability.

[0115] In contrast, using the same statistical methods and software analysis, radiomics features extracted from existing technologies (Galos® enhanced MRI images) generally showed weaker correlations with the pathological vulnerability index. Figure 15 a), and there was no significant difference in expression levels between stable and vulnerable plaques (p>0.05, Figure 15 (b) This directly confirms the shortcomings of existing technologies: their information based on non-targeted imaging cannot effectively reflect the core molecular pathological states that determine plaque vulnerability. In summary, the above comparative results strongly demonstrate that this invention, through molecularly targeted imaging, provides subsequent machine learning models with input features that have clearer biological significance and stronger diagnostic efficacy.

[0116] Although embodiments of the invention have been disclosed for illustrative purposes, those skilled in the art will understand that various substitutions, variations, and modifications are possible without departing from the spirit and scope of the invention and the appended claims. Therefore, the scope of the invention is not limited to the contents disclosed in the embodiments.

Claims

1. A nano-targeting probe for targeting foam macrophages, characterized in that: The preparation method of the nano-targeting probe includes the following steps: Freshly prepared 24 mM chloroauric acid aqueous solution and 46.1 mg / mL glutathione aqueous solution were added together to deionized water and magnetically stirred at 1000 rpm for 20 minutes at room temperature to obtain a colorless and transparent GSH-Au complex solution; wherein the volume ratio of chloroauric acid aqueous solution: glutathione aqueous solution: deionized water was 12.6:3.0:

130. Take the above GSH-Au complex solution and mix it with 1 mg / mL PP1 peptide aqueous solution. Place the mixture in 10 mM phosphate-buffered saline (PBS) at pH 7.4 and incubate at 37°C with shaking for 2 hours at a shaking frequency of 60 times / min to obtain the GSH-Au-PP1 complex solution; wherein the volume ratio of GSH-Au complex solution to PP1 peptide aqueous solution is 1.0:0.

1. Diethylenetriaminepentaacetic dianhydride and the GSH-Au-PP1 complex solution prepared in the previous step were dissolved together in 10 mM PBS buffer at pH 8.5 and reacted at 25°C for 2 hours. Subsequently, gadolinium trichloride hexahydrate was added to the reaction system, and the pH of the system was adjusted to 6.5 with 0.1M sodium acetate solution, and the reaction continued. The ratio of diethylenetriaminepentaacetic dianhydride:GSH-Au-PP1 complex solution:PBS buffer:gadolinium trichloride hexahydrate (mg:mg:mL:mg) was 400:400:10:

200. After the reaction was complete, the mixed solution was centrifuged at 4°C and 14,000 rpm for 30 minutes. The supernatant was discarded, and the precipitate was collected, which is the foam macrophage-targeting nanoprobe, i.e., the nano-targeting probe.

2. The nanotargeting probe according to claim 1, characterized in that: The nano-targeting probe has the following physicochemical properties: ① Morphology and stability: Transmission electron microscopy images showed that the nano-targeted probe was uniformly spherical with good dispersibility and stability. Dynamic light scattering analysis determined its hydrated particle size to be approximately 6.37 ± 0.16 nm, polydispersity index (PDI) to be 0.18, and zeta potential to be -4.07 ± 0.61 mV. ② Relaxation performance: The longitudinal relaxation rate of this nano-targeted probe is as high as 20.18 mM. -1 ·s -1 This indicates that it is a highly efficient T1-weighted MRI contrast agent; ③ Targeting efficacy: In vitro cell confocal imaging and quantitative fluorescence intensity analysis showed that the nano-targeting probe can efficiently and specifically recognize and bind to foam macrophages, and its binding efficiency to foam macrophages is higher than that of ordinary macrophages.

3. An atherosclerosis model based on nanomolecular magnetic resonance imaging and machine learning algorithms using the nano-targeting probe as described in claim 1 or 2, characterized in that: The atherosclerosis model first uses nano-targeting probes that target foam macrophages to perform enhanced MR imaging to obtain high-resolution MRI images that specifically show the distribution of foam macrophages within plaques. Then, radiomics features are extracted from the images, and a machine learning model based on these features and validated by pathological labels is used for analysis, ultimately outputting an objective and quantitative plaque vulnerability score.

4. The atherosclerosis model according to claim 3, characterized in that: Includes the following steps: (1) Image acquisition: Acquire MRI data of the target region after enhanced scanning by nano-targeting probes targeting foam macrophages; (2) Generation of three-dimensional reconstruction and quantitative distribution map: In order to realize the visualization and quantitative analysis of the spatial distribution of foam macrophages, the MRI data is converted into a three-dimensional reconstruction quantitative distribution map of foam macrophages in the target area through image segmentation and three-dimensional reconstruction technology; (3) Radiomics feature extraction: Extract radiomics features of patch vulnerability from the three-dimensional reconstructed quantitative distribution map; (4) Model building and risk assessment: Based on the radiomics features obtained in step (3), a machine learning model is built and applied to output a quantitative plaque vulnerability score.

5. The atherosclerosis model according to claim 3, characterized in that: Step (1) includes the following steps: Images and data were acquired through MR imaging sequences, using high-resolution vascular wall imaging technology to obtain isotropic vascular wall images with a high contrast-to-noise ratio, thereby clearly displaying plaque morphology and suppressing interference from intraluminal blood flow signals.

6. The atherosclerosis model according to claim 3, characterized in that: The generation of the three-dimensional reconstructed quantitative distribution map in step (2) includes the following steps: 1) Image segmentation: To improve the continuity of segmentation boundaries and suppress image noise, the original MRI image was subjected to Gaussian smoothing filtering before segmentation. Specifically, a Gaussian kernel G with a standard deviation σ = 0.3 mm was used. σ The image I is obtained by convolving it with the original image. smooth The preprocessed MRI imaging data is then processed to generate masks for the target blood vessels (vascular masks) and masks for foam macrophage-rich plaques (plaque masks), respectively. The specific formulas are as follows: ; Among them: I smooth Let I be the smoothed image, G be the Gaussian kernel function, (x',y',z') be the coordinates of the convolution domain, and (x,y,z) be the spatial coordinates of any voxel (3D pixel) in the smoothed image. The segmentation process consists of two parts: ① Blood vessel segmentation: By tracing the boundary of the blood vessel lumen, a blood vessel mask reflecting the overall three-dimensional geometry of the target blood vessel is generated; the segmentation is achieved through manual, semi-automatic, or fully automatic algorithms; ② Patch Segmentation: Based on the signal intensity enhanced by the nano-targeted probe, patch regions enriched with foam macrophages are identified and segmented. A semi-automatic segmentation algorithm is used, which uses a contrast-to-noise ratio (CNR) of at least 3 times the baseline value (CNR ≥ 3) as the growth threshold criterion, supplemented by a region growing algorithm for expansion. Seed points are selected based on the location of the maximum signal intensity in the image. If multiple maximum signal intensity values ​​exist, the point with the largest connected region area is selected as the initial seed point. For each neighboring voxel (i, j, k) in the current region, if its intensity value I(i, j, k) satisfies the following formula, the voxel is included in the growth region: The specific formula is as follows: ; Where I(i, j, k) is the signal intensity of the voxel to be determined; I seed σ represents the signal strength of the seed point; CNR is the contrast-to-noise ratio, with a preset threshold ≥ 3; background The standard deviation of background noise is obtained by manually delineating one or more Regions of Interest (ROIs) in the signal-free air region of the image, calculating the standard deviation of the intensity of all pixels within the delineated ROI, and using this as σ. background The value; 2) Quantitative distribution map generation: The segmented two-dimensional blood vessel mask and plaque mask are input into the three-dimensional rendering engine. Through volumetric rendering technology, the above-mentioned blood vessel and plaque masks are reconstructed into three-dimensional models. Subsequently, the three-dimensional model of the plaque is spatially fused with the three-dimensional model of the target blood vessel to finally generate the quantitative distribution map of foam macrophages. This distribution map is a three-dimensional visualization model that can intuitively and three-dimensionally show the specific spatial location, morphology and distribution range of plaques rich in foam macrophages, i.e., enhanced signal regions, in the target blood vessel wall.

7. The atherosclerosis model according to claim 3, characterized in that: In step (3), radiomics analysis is used to extract radiomics features in high throughput from the segmented patch regions. Before feature extraction, image grayscale normalization and resampling are required to ensure data consistency.

8. The atherosclerosis model according to any one of claims 3 to 7, characterized in that: Step (4) involves the construction and application of the machine learning model, including the following steps: 1) Training Data Preparation and Preprocessing: A training dataset labeled with the gold standard pathology data was constructed. Each sample includes a plaque radiomics feature vector extracted from nano-targeted probe-enhanced MRI images, and its corresponding binary classification label. The binary classification label is obtained based on the staining results of the gold standard pathological sections, and the calculation formula is as follows: ; Among them: macrophage area was obtained by CD68 immunohistochemical staining quantitative analysis; lipid area was obtained by Oil Red O staining quantitative analysis; smooth muscle cell area was obtained by α-SMA immunohistochemical staining quantitative analysis; collagen area was obtained by Masson trichrome staining quantitative analysis. After obtaining the pathological vulnerability index of each plaque, it is classified as "vulnerable" or "stable" according to a preset threshold value; Before model training, the radiomics features are standardized and preprocessed to eliminate dimensional differences and improve model stability and generalization ability. The Z-score standardization method is used, and the calculation formula is as follows: ; Where X represents the original radiomics feature value, i.e., the unprocessed initial feature data; μ represents the mean of the radiomics feature across all training samples, i.e., the global statistic, which needs to be calculated based on the entire training set; σ represents the standard deviation of the radiomics feature across all training samples, i.e., the global statistic, which, along with the mean, is calculated based on the same training set. 2) Feature Selection: To select the most relevant and stable subset of features from the extracted features, a multi-stage feature selection strategy is adopted, specifically including: ① Stability Filtering: Calculate the intragroup correlation coefficient (ICC) of all initial features and remove features that are below a preset stability threshold. The ICC calculation formula is as follows: ; Among them, MS between For mean square between groups, MS within The mean square of the group is given, and k is the number of measurement repetitions. ② Collinearity Dimensionality Reduction: Analyze the pairwise correlations between the remaining features, remove redundant features, and calculate collinearity using the Pearson correlation coefficient. The collinearity threshold is |r| > 0.9, and the specific formula is as follows: ; Among them, X i and Y i These are the i-th observations of the two features to be analyzed; and The mean of two characteristic observations; n is the total number of observations; ③ Discriminative selection: Using methods based on statistical tests or machine learning embedding, a predetermined number of features most relevant to patch vulnerability are selected to form the final modeling feature subset; 3) Model training, validation, and score generation ① Model Training: Using the standardized feature subset and its corresponding labels, train a binary classification machine learning model; input the final feature subset, divide it into training and validation sets according to a preset ratio, and fine-tune the key parameters of the model using hyperparameter optimization methods to maximize the model's performance on the validation set; for example, the formula for calculating the linear prediction value of the logistic regression model is: ; Where z is the linear prediction value, β0 is the intercept, and β i X is the weight coefficient corresponding to the i-th feature. norm, i Let i be the i-th standardized feature value; ② Model performance validation: The trained model needs to be validated on an independent validation set to prove its diagnostic efficacy; ③ Risk score generation and classification rules: The trained model can process the standardized features of new samples and output a continuous probability value between 0 and 1. This value is defined as a quantitative plaque vulnerability score. The score is calculated by mapping the linear prediction values ​​obtained by the model to a probability, that is, by using the Sigmoid function to map the linear prediction values ​​to the predicted probability of vulnerable plaques. The calculation formula is as follows: ; Where e is the natural constant (Euler number), approximately equal to 2.71828; z is the linear prediction value obtained through the first step. To translate the predicted probabilities into clinical decisions, a binary classification rule needs to be established. This rule is established as follows: Based on the receiver operating characteristic curve of the model on the validation set, the optimal classification cutoff value, i.e., the Youden index (sensitivity + specificity - 1), is selected. Based on this cutoff value, the following classification rule is established: if the nano-AML score of a plaque is not lower than this value, it is classified as a "vulnerable plaque"; if the score is lower than this value, it is classified as a "stable plaque," thus enabling the assessment of the vulnerability of atherosclerotic plaques.

9. The application of the atherosclerosis model as described in any one of claims 3 to 7 in the assessment of the vulnerability of atherosclerotic plaques.

10. An auxiliary diagnostic model utilizing the atherosclerosis model based on nanomolecular magnetic resonance imaging and machine learning algorithms as described in any one of claims 3 to 7, characterized in that: The auxiliary diagnostic model includes: (1) Data acquisition interface: used to read MR imaging data in DICOM format from MRI scanners or medical image storage and transmission systems; (2) Image processing engine: Calls the image segmentation algorithm and the three-dimensional reconstruction routine to process the MR imaging data and generate a quantitative distribution map of foam macrophages in the target area; (3) Feature calculation unit: Based on the predefined image omics feature set, automatically extract quantitative features from the quantitative distribution map and output them as structured data; (4) Core of risk assessment: Load and run the trained machine learning model, receive feature data, calculate and output a quantitative plaque vulnerability score; Alternatively, it may include (5) a report generator: integrating and visualizing the above analysis results in the user interface.

Images