A three-high risk re-stratification method and system based on abdominal MRI combined phenotype

Abstract

The application discloses a three-high risk re-stratification method and system based on abdominal MRI combined phenotypes. The method takes the DICOM sequence and the associated image report of the same abdominal MRI examination as the processing object, extracts the report fat phenotype and the DICOM quantitative fat phenotype from the processing object respectively, maps to a unified potential fat phenotype space through protocol-specific calibration, and outputs the combined phenotype and the fusion confidence through the gating fusion. The BMI conditional residual risk characteristics are generated based on the reference object data. The risk adjustment amount is obtained by combining the baseline risk model and the image incremental model, and the up-regulation, maintenance or down-regulation result relative to the original risk level is output according to the gating rule. Through the multi-source calibration fusion and the safety gating mechanism, the accuracy and reliability of the three-high risk re-stratification are improved.

Description

Technical Field

[0001] This invention relates to the fields of medical image data processing, medical text structuring, body composition quantitative analysis, and computer-aided risk assessment, and particularly to a method and system for risk restratification of hypertension, hyperlipidemia, and hyperglycemia based on the combined phenotypes of visceral fat, subcutaneous fat, and hepatic fat in abdominal MRI. Background Technology

[0002] Hypertension, hyperglycemia, and hyperlipidemia are often collectively referred to as the "three highs." In this invention, the risk of the "three highs" refers to at least one of the risks of hypertension, hyperglycemia, and hyperlipidemia. Existing risk assessments for the "three highs" typically establish rules or predictive models based on structured clinical information such as blood pressure, blood glucose, blood lipids, body mass index (BMI), age, sex, medical history, and medication history, and assign risk levels such as low, medium, or high risk accordingly. While these methods can reflect general clinical risk, they generally do not directly utilize the fat distribution information contained in abdominal MRI. BMI can represent the relationship between total body weight and height, but it is difficult to represent the differences in fat distribution within the abdominal cavity, subcutaneous abdominal wall, and liver parenchyma. Therefore, individuals with similar BMIs may have significantly different levels of visceral fat, subcutaneous fat, and liver fat, and thus different metabolic risks.

[0003] Abdominal MRI (Magnetic Resonance Imaging) can record information related to visceral fat, subcutaneous fat, and liver fat. Imaging reports are readily available in batches, but are often presented qualitatively with terms like fatty liver, abdominal obesity, and increased abdominal fat, often including negative, uncertain, and degree descriptions. DICOM (Digital Imaging and Communications in Medicine) images can provide quantitative information such as VAT (Visceral Adipose Tissue), SAT (Subcutaneous Adipose Tissue), VAT / SAT, and liver PDFF (Proton Density Fat Fraction), but this is affected by MRI scanning protocols, sequence type, echo time, water-lipid separation quality, anatomical coverage, and image quality. Directly weighting report grades with continuous DICOM values ​​yields inconsistent interpretations across different sources; and inputting reports and DICOM as independent risk evidence into a risk model may amplify reporting conventions and image measurement errors. Therefore, it is necessary to first calibrate, merge, and evaluate the confidence level of the reporting source and the DICOM source at the phenotypic measurement layer, and then use the merged phenotype for risk restratification.

[0004] Furthermore, a single target plane cannot always simultaneously cover both the lumbar spine plane used for VAT and SAT measurements and the upper abdominal plane used for liver fat measurements. Different MRI sequences and different PDFF generation protocols also differ in their physical meaning, noise characteristics, and comparability of liver fat phenotypes; phenotypes that have not been identified and calibrated according to protocols should not be used interchangeably. Existing methods often directly stitch image features with clinical risk models, which can easily lead to ambiguity in the meaning of image increments and original risk levels, and lack mechanisms to prevent inappropriate risk downsizing when there are conflicts between reports and images, missing key sequences, or insufficient segmentation quality. These problems make it difficult to stably use abdominal MRI fat phenotypes for risk restratification of hypertension, hyperlipidemia, and hyperglycemia. Summary of the Invention

[0005] This invention aims to provide a high-risk restratification scheme based on abdominal MRI fat phenotype, which can uniformly utilize image reports and DICOM quantitative information. The scheme first performs protocol-specific calibration, unified potential phenotype mapping, and gated fusion of report information and DICOM information at the phenotype measurement layer. Then, the baseline risk without image fat phenotype is probabilistically corrected by the image increment model. This solves the problems of inconsistent report and DICOM scales, phenotypic incomparability due to MRI protocol and sequence differences, body mass index masking abnormal fat distribution, unclear definition of image increment, and erroneous downscaling due to low confidence data.

[0006] To address the aforementioned technical problems, this invention provides a method for risk restratification of hypertension, hyperlipidemia, and hyperglycemia performed by an electronic device. The method includes: acquiring the DICOM sequence of the same abdominal MRI examination of the subject, the image report precisely associated with the DICOM sequence, the clinical characteristics of hypertension, hyperlipidemia, and hyperglycemia, and the original risk level or baseline risk probability; extracting the reported fat phenotype and report confidence from the image report; locating a first target region and a second target region from the DICOM sequence based on the MRI scanning protocol and sequence type, and generating the DICOM fat phenotype and sub-phenotype image quality confidence; establishing an MRI protocol-specific calibration mapping based on image report-DICOM paired data and labeled DICOM data, mapping the reported fat phenotype and DICOM fat phenotype to a unified latent fat phenotype space, and outputting a joint phenotype vector and fusion confidence through a gated fusion model; generating BMI conditional residual risk features based on reference subject data; combining the baseline risk model and the image increment model to obtain the adjustment amount for the risks of hypertension, hyperlipidemia, and hyperglycemia, and outputting the up-adjustment, maintenance, or down-adjustment results according to gating rules.

[0007] In the above technical solution, the structured extraction of the image report includes: performing chapter recognition, medical entity recognition, entity attribute recognition, negation recognition, and uncertainty recognition on the image report; mapping at least one of the following descriptions—abdominal obesity, increased visceral fat, increased subcutaneous fat, elevated liver fat fraction, elevated PDFF, in-phase and out-of-phase signal attenuation, abnormal water-lipid signal ratio, and fatty liver severity—to a preset fat phenotype dictionary; determining the value of the report's fat phenotype, report confidence level, and evidence text location based on the chapter in which the description is located, the negation state, the uncertainty state, the severity level, the report template source, and the entity recognition probability, wherein a report without fat-related content is recorded as a missing state rather than a negative state. The evidence text location is mainly used for evidence tracing of the output results and can also serve as a basis for manual review when necessary.

[0008] In the above technical solution, the image report-DICOM paired data includes the DICOM sequence, image report, examination number, unique identifier of the study case, examination time, MRI scan protocol, sequence type, magnetic field strength, echo time, aqueous phase map, lipid phase map, in-phase map, out-of-phase map, availability of PDFF map, manual annotation of the fat phenotype in the report, and at least a portion of the DICOM manual annotation results; the DICOM manual annotation results include at least one of the following: VAT region, SAT region, abdominal wall boundary, liver region, liver region of interest, and spleen region or spleen region of interest used when adopting the liver-spleen MRI signal ratio scheme.

[0009] In the above technical solution, the reference object data includes first reference object data and second reference object data. The first reference object data is used to establish a fat phenotype reference distribution, a body mass index conditional expectation model, or an image phenotype extraction model. The second reference object data is image report-DICOM paired data, used to establish a report extraction model, a protocol-specific calibration mapping, a gated fusion model, and a local population calibration model. During the model building phase, the first reference object data and the second reference object data are trained and calibrated respectively. The model parameters are not updated during the inference phase of the object to be processed.

[0010] In the above technical solution, the first target area includes the level of the third lumbar vertebra, the level between the third and fourth lumbar vertebrae, the umbilicus plane, or a continuous abdominal volume covering the above levels, used to measure VAT and SAT; the second target area includes the level of the upper abdomen where the liver is located, the three-dimensional region of the liver, the MRI proton density fat fraction map, or, when using the liver-spleen MRI signal ratio scheme, the level where the spleen is located or the three-dimensional region of the spleen, used to measure the liver fat phenotype, wherein the first target area and the second target area may be the same or different.

[0011] In the above technical solution, the DICOM fat phenotype vector includes at least one of VAT area or volume, SAT area or volume, total abdominal fat area or volume, VAT / SAT ratio, and liver fat phenotype. The DICOM fat phenotype vector is generated based on the Dixon water-lipid separation sequence, fat fraction map, PDFF map, fat suppression sequence, or in-phase and out-of-phase images of abdominal MRI. The VAT and SAT phenotypes are determined by the fat candidate region, body surface contour, and abdominal wall boundary of the first target region, and the liver fat phenotype is determined by the liver PDFF, fat fraction, or water-lipid signal ratio of the second target region. When multiple MRI sequences or multiple fat quantification maps exist, the system determines the primary sequence based on sequence integrity, spatial coverage, registration quality, artifact degree, and protocol priority, and uses the remaining sequences as sources for consistency verification or missing information supplementation.

[0012] In the above technical solution, the unified potential fat phenotype space is used to express the common meaning of the same fat abnormality under both the reported source and the DICOM source, including the calibrated fat phenotype abnormality probability, standardized continuous phenotype, quantile rank, or a combination of the above variables; the reported phenotype calibration converts qualitative or semi-quantitative reported descriptions into the reported source's fat phenotype abnormality probability, quantile rank, or potential continuous variable; the DICOM continuous phenotype standardization performs z-score standardization, quantile mapping, or protocol-level calibration on at least one of VAT, SAT, VAT / SAT, and liver fat phenotypes according to MRI sequence type, scanning protocol, target level, unit of measurement, and population reference distribution; the gated fusion does not directly and equally weight the reported phenotype and DICOM phenotype, but outputs the fused phenotype and fusion confidence based on the reported confidence, DICOM image quality confidence, source missing mask, and source conflict identifier; when the reported phenotype and DICOM phenotype are inconsistent in direction, the difference exceeds the conflict threshold, or the quality of any source is insufficient, the fusion confidence is reduced and a verification identifier is generated.

[0013] In the above technical solution, the gating rules include: when the risk adjustment reaches the upward threshold and the fusion confidence of the corresponding core phenotype reaches the confidence threshold, the original risk level is increased; when the risk adjustment reaches the downward threshold, the fusion confidence reaches the confidence threshold, there are no missing key clinical features and core imaging phenotypes, the report does not conflict with DICOM, and the downward restriction zone is not triggered, the original risk level is allowed to be decreased; when the upward or downward conditions are not met, the original risk level is maintained; the downward restriction zone includes being diagnosed with the corresponding three hypertension diseases, extreme blood pressure or blood sugar or blood lipid values, currently using related treatment drugs, previous cardiovascular and cerebrovascular events, significantly abnormal renal function, and other clinical hard rejection conditions preset by the system.

[0014] Another aspect of the present invention provides a risk restratification system for hypertension, hyperlipidemia, and hyperglycemia based on combined abdominal MRI phenotypes, comprising: a data acquisition unit for acquiring the DICOM sequence of the same abdominal MRI examination of the subject, the image report associated with the DICOM sequence, the clinical characteristics of hypertension, hyperlipidemia, and hyperglycemia, and the original risk level or baseline risk probability; a report structuring unit for extracting the report fat phenotype and report confidence level from the image report; and an image phenotype extraction unit for locating a first target region and a second target region from the DICOM sequence based on the MRI scanning protocol and sequence type, and generating the DICOM fat phenotype and sub-phenotype. The system includes: a model image quality confidence level; a calibration fusion unit, used to establish an MRI protocol-specific calibration mapping based on image report-DICOM paired data and labeled DICOM data, mapping the reported fat phenotype and the DICOM fat phenotype to a unified latent fat phenotype space, and outputting a joint phenotype vector and fusion confidence level through a gated fusion model; a residual generation unit, used to generate BMI conditional residual risk features based on reference object data; and a risk restratification unit, used to combine the baseline risk model and the image increment model to obtain the risk adjustment amount for the three highs (hypertension, hyperlipidemia, and hyperglycemia), and output the risk level adjustment results of up, maintain, or down according to gating rules.

[0015] In another aspect, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the aforementioned method for restratifying high-risk individuals based on combined abdominal MRI phenotypes.

[0016] In another aspect, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for restratifying high-risk individuals based on combined abdominal MRI phenotypes.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: First, reports and DICOM are no longer simply averaged numerically, nor are they used as independent evidence of repeated risk. Instead, they are first mapped to a unified potential fat phenotype space through paired hospital data, making qualitative texts and continuous measurements comparable. Second, the first target area is used for abdominal VAT and SAT measurements, and the second target area is used for liver PDFF or other MRI liver fat phenotype measurements, avoiding forcibly binding different anatomical coverage requirements to the same level. Third, the system establishes protocol-specific processing paths for MRI sequences such as Dixon water-lipid separation, in-phase and out-of-phase, and PDFF, and can simultaneously complete phenotypic reference, report-DICOM fusion, and local population calibration using large-scale MRI reference cohorts and hospital-paired data. Fourth, BMI conditional residuals are used to represent the degree of fat abnormality relative to people of the same body mass index, age and sex, so that people with similar body mass index but different fat distributions can be distinguished. Fifth, baseline risk and imaging increments are explicitly distinguished, which facilitates the interpretation of the additional contribution of abdominal MRI fat phenotype to the original clinical risk; Sixth, missing, conflicting, and downgraded exclusion gate controls can reduce the inappropriate risk downgrade caused by low-confidence results. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the process for a risk restratification method for hypertension, hyperlipidemia, and hyperglycemia based on abdominal MRI combined with phenotype, provided in an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of the protocol-specific DICOM fat phenotype extraction method provided in the embodiments of the present invention.

[0021] Figure 3 This is a schematic diagram of the three-high-risk re-stratification system structure provided in an embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram of the risk level adjustment and confidence level gating method provided in the embodiments of the present invention.

[0023] Figure 5 This is a schematic diagram of the electronic device structure provided in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0025] Unless otherwise specified, the subject to be processed in this invention can be a subject who has undergone abdominal MRI examination and has the clinical characteristics of the "three highs" (hypertension, hyperglycemia, and hyperlipidemia); the risk of the "three highs" can correspond to the risk of hypertension, hyperglycemia, and hyperlipidemia respectively, or it can correspond to the combined risk of at least one of the "three highs".

[0026] Example 1

[0027] Reference Figure 1 The method in this embodiment includes steps 110 to 170.

[0028] In step 110, the data acquisition unit acquires the DICOM sequence, imaging report, clinical characteristics of hypertension, hyperlipidemia, and hyperglycemia (the "three highs") corresponding to the same abdominal MRI examination of the subject, as well as the original risk level or baseline risk probability generated by the external risk assessment module. The DICOM sequence and imaging report are precisely associated through the examination number, unique identifier of the study instance, unique identifier of the sequence instance, and examination time. Preferably, the imaging report and DICOM sequence originate from the same abdominal MRI examination. The clinical characteristics of hypertension, hyperlipidemia, and hyperglycemia may include blood pressure, fasting blood glucose, glycated hemoglobin, triglycerides, low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, body mass index, age, sex, medical history, and medication history.

[0029] In step 120, the report structuring unit extracts the reported fat phenotype, report confidence level, and evidence text location from the imaging report. The system first identifies sections such as imaging findings, diagnostic opinions, and comparative descriptions, and then identifies fat-related entities, attributes, denial states, uncertain states, and severity levels. No obvious fatty liver seen in the report is recorded as a denial of liver fat abnormalities; mild fatty liver is considered as mild liver fat with uncertainty; and increased intra-abdominal fat is recorded as increased visceral fat with longitudinal variation. When the report does not mention fat-related content, the system records the corresponding report phenotype as missing, rather than as no abnormality.

[0030] In step 130, the image phenotype extraction unit processes the DICOM sequence according to the MRI scanning protocol and sequence type. The system reads the MRI sequence type identifier, sequence description, image orientation, slice thickness, pixel spacing, echo time, reconstruction method, water-lipid separation identifier, and PDFF image identifier from the DICOM header information. For abdominal VAT and SAT, the system determines the third lumbar vertebra level, the level between the third and fourth lumbar vertebrae, the umbilicus level, or a continuous abdominal volume containing the above levels as the first target region. For liver fat, the system determines the upper abdominal region where the liver is located, the three-dimensional region of the liver, or the MRI PDFF image as the second target region. The first and second target regions can be different, and the system generates DICOM fat phenotype vectors and sub-phenotype image quality confidence scores, respectively.

[0031] In step 140, the joint phenotypic fusion unit invokes a protocol-specific calibration map. This calibration map is established based on image report-DICOM paired data and labeled DICOM data, and includes report phenotypic calibration, DICOM continuous phenotypic normalization, and source conflict determination. Report rank is converted into calibrated probabilities, quantile ranks, or latent continuous variables. DICOM quantification values ​​are z-score normalized, quantile mapped, or protocol-stratified according to MRI sequence type, scanning protocol, target slice, unit, and population reference distribution. The gated fusion model receives the report phenotype, DICOM phenotype, report confidence, image quality confidence, source identifier, missing mask, and conflict identifier, and outputs a unified latent fat phenotype and fusion confidence. This step completes the fusion of report and DICOM at the phenotypic level, and subsequent risk models only use the fused joint phenotype and its quality information.

[0032] In step 150, the residual risk generation unit establishes a conditional expectation model based on the reference object data. For the j-th fat phenotype, the conditional expectation model takes body mass index (BMI), age, and sex as inputs to obtain the conditional expectation value shown in equation (1). The conditional expectation value satisfies equation (1): .

[0033] Based on the difference between the actual value and the conditional expected value of the fusion phenotype and the residual scaling parameter in the reference object data, the standardized residual is obtained, which satisfies equation (2), where ε is a preset positive number: , in To integrate actual phenotypic values, For the residual scaling parameters in the reference object data, The value is a preset positive number. The risk characteristics of the body mass index residuals may include at least one of the following: VAT-standardized residuals, VAT / SAT-standardized residuals, liver fat-standardized residuals, maximum positive residuals, combined outlier counts, and residual direction consistency.

[0034] In step 160, the risk restratification unit calculates the baseline risk and the image increment. This step inputs the clinical features of hypertension, hyperlipidemia, and hyperglycemia (excluding image-associated fat phenotype) into the baseline risk model to obtain the baseline risk probability. It then inputs the combined phenotype vector, body mass index residual risk features, missing mask, source identifier, and fusion confidence into the image increment model to obtain the image risk increment relative to the baseline risk probability. Based on the baseline risk probability and the image risk increment, it obtains the corrected risk probability and the risk adjustment amount relative to the original risk level. For the k-th risk task, the baseline risk model outputs the baseline risk probability based on the clinical features excluding image-associated fat phenotype. The image increment model outputs the image increment based on the combined phenotype, residual risk features, missing mask, source identifier, and fusion confidence. The system obtains the corrected risk probability according to equation (3) and uses equation (4) or equation (5) as the risk adjustment amount. Thus, the image information only interprets the increment relative to the existing clinical risk, rather than retraining an ambiguous overall risk level.

[0035] The baseline risk model and the image increment model are established for the risks of hypertension, hyperglycemia, and hyperlipidemia, respectively. For the k-th risk task, the baseline risk model outputs the baseline risk probability, the image increment model outputs the image increment, the corrected risk probability satisfies equation (3), and the risk adjustment satisfies equation (4) or equation (5). , , , The training labels are generated based on diagnosis, testing, medication and follow-up records within a preset observation window. Confirmed cases, cases lacking baseline exclusion criteria, and cases with insufficient follow-up are subject to level adjustment training, censoring, or exclusion from the new risk training set according to preset rules.

[0036] In step 170, the system performs gated output. The system outputs an upward adjustment when the risk adjustment reaches the upward threshold and the core phenotype fusion confidence reaches the confidence threshold. The system outputs a downward adjustment only when the risk adjustment reaches the downward threshold, the fusion confidence reaches the confidence threshold, there are no missing key clinical features and core imaging phenotypes, the report does not conflict with DICOM, and the downward adjustment contraindication is not triggered. In other cases, the system maintains the original risk level. The system also outputs the contribution of key features, the location of the report evidence text, DICOM checks and sequence identifiers, the target-level index, the model version, the calibration version, and the review identifier.

[0037] Example 2

[0038] Based on the same inventive concept as in Embodiment 1, this embodiment further explains the model construction and data division of labor.

[0039] In this embodiment, the model building phase uses both first and second reference object data. The first reference object data can be a large-scale abdominal Dixon MRI cohort, including fields such as water-lipid separation DICOM, VAT volume, abdominal subcutaneous fat volume, and liver fat percentage. This type of data is primarily used to establish an MRI fat phenotype reference distribution, a BMI conditional expectation model, and an optional image phenotype extraction model. The second reference object data is hospital real-world report-DICOM paired data, used to train the report structured model, the protocol-specific DICOM phenotype calibration model, the gated fusion model, and the local population calibration model.

[0040] The basic unit for hospital-paired data is a single MRI examination. Each sample includes at least the DICOM sequence from the same examination, imaging report, examination time, MRI scan protocol, sequence type, magnetic field strength, echo time, availability of aqueous phase image, lipid phase image, in-phase image, out-of-phase image, or PDFF image, age, sex, body mass index, blood pressure, blood glucose, glycated hemoglobin, blood lipids, medications, previous diagnosis, follow-up diagnosis, and follow-up tests. Imaging reports are precisely matched to DICOM by examination number or study instance, not just by rough matching using patient number and date.

[0041] Hospital-matched data can be categorized into three types of annotation. The first type is report annotation, where text locations such as fatty liver, increased abdominal fat, abdominal obesity, no fatty liver, and suspected mild fatty liver are manually marked, along with negative, uncertain, and degree indications. The second type is DICOM gold standard annotation, where VAT, SAT, abdominal wall boundaries, liver ROI, and spleen ROI are manually annotated. The third type is fusion validation annotation, where experts determine whether there are source conflicts based on DICOM quantification and reports. These three types of annotation are used for report extraction, image quantification, and fusion confidence modeling, respectively.

[0042] To avoid information leakage, reports can be used as sources for phenotypic extraction and fusion during the model building phase. However, during the risk model training phase, the risk model only receives fusion phenotypes, fusion confidence scores, missing masks, and source identifiers generated by cross-fitting. For each training object, its fusion phenotype is generated by a calibration or fusion model that has not seen that object before. When a pending object enters the inference phase, all model parameters and calibration parameters are fixed, and the object is not used to update the model.

[0043] Example 3

[0044] Based on the same inventive concept as in Example 1, this embodiment further illustrates the extraction of protocol-specific DICOM phenotypes.

[0045] Reference Figure 2 In step 210, the system performs quality checks, protocol identification, sequence selection, and target region localization on the MRI DICOM. The system can select Dixon aqueous phase, Dixon fat phase, in-phase, out-of-phase, fat fraction map, or PDFF map as the sequence to be analyzed. The system performs position normalization, spatial resampling, intensity normalization, and sequence registration, and records the processed version. In step 220, the system segments the body surface contour, abdominal wall boundary, VAT, and SAT within the first target region. For MRI Dixon sequences, the system can generate fat candidate regions based on the fat phase, aqueous phase, or fat fraction map, and divide the candidate regions into VAT and SAT based on the abdominal wall boundary. The system excludes the examination bed, external air, intestinal gas, bone marrow fat, and non-target structures located within solid organs.

[0046] In step 230, the system segments the liver in the second target region or obtains the liver region from the MRI PDFF image. The MRI path preferentially uses the PDFF image; if there is no PDFF image but there are water map W and lipid map F, the fat fraction can be calculated according to formula (6) to replace the phenotype. The system excludes large blood vessels, bile ducts, focal lesions and the marginal region of the liver, and generates a confidence score of liver fat quality based on the liver segmentation quality, PDFF image integrity, water-lipid separation artifacts and sequence coverage.

[0047] .

[0048] In step 240, the system calculates the area or volume of VAT, the area or volume of SAT, the area or volume of TAT, the VAT / SAT ratio, and the hepatic steatosis phenotype. Area-based and volume-based phenotypes can be used selectively or both can be retained. In step 250, the system generates an image quality confidence score for each specific phenotype. The quality confidence score can be composed of DICOM integrity, target region coverage, localization score, segmentation probability, abdominal wall boundary topology score, effective ROI proportion, artifact score, and area or volume reasonableness score.

[0049] Example 4

[0050] Based on the same inventive concept as in Example 1, this embodiment further illustrates the fusion of the report with the DICOM potential phenotypic.

[0051] In this embodiment, a unified latent fat phenotype space is used to express the common meaning of the same fat abnormality from different sources. For hepatic steatosis, the reports of no, mild, moderate, and severe fatty liver can be mapped to the probability, quantile rank, or standardized latent variable of hepatic steatosis. PDFF, fat fraction, or water-lipid signal ratio in DICOM are also mapped to the same latent variable. For VAT and SAT, DICOM typically provides continuous quantitative values, and reports usually provide abnormality indications or missing information. Therefore, reports are mainly used for missing information supplementation, conflict identification, and assistance with low-quality images, rather than being weighted and added to DICOM quantitative values.

[0052] Gated fusion models can employ logistic regression, gradient boosting trees, or other machine learning models capable of outputting calibration probabilities and confidence scores. Model inputs include reported calibration values, DICOM calibration values, reported confidence scores, DICOM quality confidence scores, MRI protocol identifiers, source missing masks, and conflict features; model outputs include the fusion phenotype. and fusion confidence When DICOM quality is high and the report is consistent with the DICOM direction, the fusion result prioritizes retaining the quantitative information of DICOM and increases the fusion confidence. When DICOM is missing or of insufficient quality but the report provides clear evidence, the report source can be used as a supplement but the confidence level is reduced. When the report explicitly denies fat abnormalities but DICOM shows obvious abnormalities, or when the report indicates obvious abnormalities but DICOM shows low risk and the image quality is sufficient, the system reduces the fusion confidence level and generates a verification label.

[0053] The calibration mapping and gated fusion model were trained using paired hospital data. This design treats reports and DICOM as two observational sources of the same potential phenotype, rather than two independent sources of risk evidence. Therefore, the risk model does not amplify reporting bias because reports are both involved in phenotype extraction and serve as independent clinical evidence input during the inference phase, and the image increment can be limited to the additional contribution of the fused fat phenotype relative to existing clinical risk.

[0054] Example 5

[0055] Based on the same inventive concept as in Example 1, this example further explains risk labels, baseline risks, and image increments.

[0056] In this embodiment, the three high-risk tasks are modeled separately. Hypertension risk labels can be generated based on diagnostic records, blood pressure measurements, and antihypertensive medication usage records within the observation window. Hyperglycemia risk labels can be generated based on diagnoses of diabetes or abnormal glucose metabolism, fasting blood glucose, glycated hemoglobin, and hypoglycemic medication usage records. Hyperlipidemia risk labels can be generated based on dyslipidemia diagnoses, triglycerides, low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, and lipid-lowering medication usage records. Subjects with corresponding diseases or treatments already present at baseline in the training data can be used for existing level adjustment training or excluded from the newly emerging risk training set. Subjects with insufficient follow-up are processed according to censoring rules.

[0057] Reference Figure 4 In step 410, the baseline risk model uses only clinical features of the non-image-related fat phenotype and outputs the baseline risk probability p0. If the external risk assessment module has already provided low, medium, and high risk levels, the system can map these levels to baseline risk probabilities using a preset calibration table or calibration model. In step 420, the image incremental model only learns the incremental contributions of fused fat phenotype, BMI residual, missing mask, source identifier, and fusion confidence to the baseline risk. The final corrected risk p1 is used to remap the risk level. In step 430, confidence, missing, conflict, and downregulation exclusion zones are read.

[0058] In one implementation, the risk adjustment amount satisfies equation (7): .

[0059] In step 440, the gating and threshold are compared. In step 451, if Δ is greater than or equal to the threshold, the threshold is raised. And the core phenotypic fusion confidence level is greater than or equal to The system raises the risk level. In step 453, if Δ is less than or equal to the lowering threshold... Furthermore, if the fusion confidence level meets the requirements, key features are complete, there are no source conflicts, and there are no downgrade restrictions, the system downgrades the risk level. In step 452, if the conditions for upgrading or downgrading are not met, the original risk level is maintained. , and It can be configured separately for hypertension, hyperglycemia, and hyperlipidemia, and the calibration, sensitivity, specificity, and net weight classification improvement indicators can be determined through validation set.

[0060] Example 6

[0061] Based on the same inventive concept as in Embodiment 1, this embodiment further explains the handling of missing, conflicting, and downgraded restricted areas.

[0062] When DICOM is temporarily inaccessible but the image report is accessible, the system can generate a temporary phenotype from the report source, but the DICOM source features are marked as missing, and automatic downgrading is prohibited. When the DICOM arrives later, the system re-executes image phenotype extraction, potential phenotype fusion, and risk restratification, saving the before and after versions in the audit log. When the report does not describe fat-related content but the DICOM quality meets the requirements, the system does not interpret the lack of description in the report as normal, but only uses the DICOM phenotype and its quality confidence level to generate a fused phenotype. When the report and DICOM clearly conflict, the system maintains the original risk level or only allows upgrading, does not perform automatic downgrading, and outputs a recommendation for review. Downgrading is prohibited for diagnosed corresponding hypertension, hyperlipidemia, and hyperglycemia, extreme blood pressure, extreme blood sugar, significant dyslipidemia, current use of corresponding treatment drugs, previous cardiovascular and cerebrovascular events, significant renal function abnormalities, and other clinically pre-set hard rejection conditions. This mechanism is used to ensure that a low-fat image phenotype does not offset a clear clinical risk.

[0063] Example 7

[0064] Based on the same inventive concept as any of the foregoing embodiments, this embodiment provides a high-risk, high-demand, and high-risk re-stratification system. (Refer to...) Figure 3 The high-risk restratification system of this embodiment includes a data acquisition unit 310, a report structuring unit 320, a protocol-specific image phenotype extraction unit 330, a paired data calibration and potential phenotype fusion unit 340, a BMI conditional residual generation unit 350, a baseline risk and image increment restratification unit 360, and an interpretation, auditing and review management unit 370.

[0065] The data acquisition unit 310 is used to acquire DICOM sequences, image reports, clinical characteristics of hypertension, hyperlipidemia, and hyperglycemia, and original risk levels from image archiving and communication systems, electronic medical record systems, testing systems, or scientific research data platforms.

[0066] The report structure unit 320 is used to extract the report fat phenotype, confidence level, and evidence text location.

[0067] The protocol-specific image phenotype extraction unit 330 is used to complete DICOM quality checks, protocol identification, target region localization, segmentation, and phenotype calculation, namely, generating DICOM fat phenotype vectors and sub-phenotype image quality confidence scores according to MRI sequence type and scanning protocol.

[0068] The paired data calibration and latent phenotype fusion unit 340 is used to call the calibration mapping and gated fusion model trained from hospital paired data. That is, based on the image report-DICOM paired data and labeled DICOM data, the report phenotype and DICOM phenotype are mapped to a unified latent fat phenotype space and gated fusion is performed.

[0069] The BMI conditional residual generation unit 350 is used to call the conditional expectation model trained from the reference object data to generate body mass index residual risk features.

[0070] The baseline risk and image increment re-stratification unit 360 is used to generate baseline risk probability, image risk increment, corrected risk probability and risk adjustment amount, and output the results of up-adjustment, maintenance or down-adjustment according to gating rules.

[0071] The Explanation, Audit, and Review Management Unit 370 is used to output key feature contributions, evidence sources, model versions, calibration versions, processing logs, and review identifiers.

[0072] Example 8

[0073] Based on the same inventive concept as any of the foregoing embodiments, this embodiment provides an electronic device. (Refer to...) Figure 5 The electronic device may include a processor 510, a communication interface 520, a memory 530, and a communication bus 540. The processor 510 may be a central processing unit, a graphics processing unit, a digital signal processor, an application-specific integrated circuit (ASIC), or a combination thereof. The memory 530 may include random access memory, read-only memory, solid-state memory, disk storage, or a distributed storage system. The communication interface 520 is used to communicate with hospital information systems, image archiving and communication systems, testing systems, data platforms, or cloud model services. When the computer program stored in the memory 530 is executed by the processor 510, the electronic device performs the high-risk re-stratification method according to the embodiments of the present invention. The computer program may also be stored in a non-transitory computer-readable storage medium and, when executed by the processor, implement the steps of the methods described in the above embodiments.

[0074] Example 9

[0075] Based on the same inventive concept as any of the foregoing embodiments, this embodiment provides a non-transitory computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the three-high-risk re-stratification method in this embodiment of the invention. The storage medium may be ROM, RAM, magnetic disk, optical disk, solid-state drive (SSD), or other non-transitory storage medium of any form. When the computer program stored thereon is executed by a processor, it causes the processor to perform the method steps described in any embodiment of the invention.

[0076] In summary, this invention provides a method and system for risk restratification of hypertension, hyperlipidemia, and hyperglycemia based on the combined phenotypes of visceral fat, subcutaneous fat, and hepatic fat from abdominal MRI. The method uses a single abdominal MRI examination as the processing object, acquiring the abdominal MRI DICOM sequence, image reports precisely associated with the DICOM sequence, clinical characteristics of hypertension, hyperlipidemia, and hyperglycemia, and the original risk level or baseline risk probability. Reported fat phenotypes and DICOM quantitative fat phenotypes are generated from the image reports and MRI DICOM sequences, respectively. Based on image report-DICOM paired data and labeled DICOM data, MRI protocol-specific calibration is performed on the report level and continuous DICOM values, and mapped to a unified potential fat phenotype space. At the phenotype measurement layer, gated fusion is performed based on source confidence, image quality, missing status, and conflict status to obtain a combined phenotype vector and fusion confidence. Residual risk features relative to the conditional expectation value of body mass index are generated based on reference data. Finally, image increments are superimposed on the baseline risk without image phenotypes to obtain the risk adjustment amount for hypertension, hyperlipidemia, and hyperglycemia, and the risk level adjustment results are output according to confidence, missing, conflict, and down-adjustment exclusion zone gating rules. This invention reduces misjudgments caused by qualitative descriptions in reports, differences in MRI protocols, and body mass index masking abnormal fat distribution through a closed-loop process of first calibrating the fusion phenotype and then correcting for risk increments.

[0077] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for risk restratification of hypertension, hyperlipidemia, and hyperglycemia based on abdominal MRI combined with phenotype, characterized in that, include: Obtain the DICOM sequence from the same abdominal MRI examination of the subject, the imaging report associated with the DICOM sequence, the clinical characteristics of hypertension, hyperlipidemia, and hyperglycemia, and the original risk level; The image report is subjected to structured extraction to obtain the report fat phenotype vector and report confidence level; A first target region for abdominal fat measurement and a second target region for liver fat measurement are determined from the DICOM sequence, and DICOM fat phenotype vectors and sub-phenotype image quality confidence scores are generated respectively. Based on the image report-DICOM paired data, a protocol-specific calibration mapping is formed, which maps the reported fat phenotype vector and the DICOM fat phenotype vector to a unified latent fat phenotype space. Then, gating fusion is performed according to the report confidence, sub-phenotype image quality confidence, missing mask and conflict marker to obtain the joint phenotype vector and fusion confidence. Based on reference object data, the conditional expected value of at least one adipose phenotype in the joint phenotype vector is determined under the conditions of body mass index, age and sex, and the body mass index residual risk characteristics are calculated based on the deviation between the actual value and the conditional expected value. The clinical features of hypertension, hyperlipidemia, and hyperglycemia that do not include the image fat phenotype are input into the baseline risk model to obtain the baseline risk probability. The joint phenotype vector, body mass index residual risk features, and fusion confidence are input into the image increment model to obtain the image risk increment. The risk adjustment amount is obtained based on the baseline risk probability and the image risk increment. Based on the aforementioned risk adjustment amount and gating rules, the risk level adjustment result is output relative to the original risk level, either by increasing, maintaining, or decreasing.

2. The method for risk restratification of hypertension, hyperlipidemia, and hyperglycemia based on abdominal MRI combined with phenotype as described in claim 1, characterized in that, Structured extraction of the image report includes: The image report is subjected to chapter recognition, medical entity recognition, entity attribute recognition, negation recognition, and uncertainty recognition. Map at least one of the following descriptions to a predefined fat phenotype dictionary: abdominal obesity, increased visceral fat, increased subcutaneous fat, elevated liver fat fraction, and degree of fatty liver. Based on the chapter in which the description is located, the negative state, the uncertain state, the degree level, and the entity recognition probability, the value of the reported fat phenotype vector and the report confidence level are determined; wherein, when the image report does not describe fat-related content, the corresponding fat phenotype is recorded as missing.

3. The method for risk restratification of hypertension, hyperlipidemia, and hyperglycemia based on abdominal MRI combined with phenotype as described in claim 1, characterized in that, The first target area includes the level of the third lumbar vertebra, the level between the third and fourth lumbar vertebrae, the umbilical plane, or a continuous abdominal volume covering the above levels; The second target area includes the upper abdominal layer where the liver is located, the three-dimensional region of the liver, or the MRI proton density fat fraction map.

4. The method for risk restratification of hypertension, hyperlipidemia, and hyperglycemia based on abdominal MRI combined phenotype according to claim 1, characterized in that, The unified potential fat phenotype space is used to express the common meaning of the same fat abnormality under both the reported source and the DICOM source, including the calibrated fat phenotype abnormality probability, standardized continuous phenotype or quantile rank; The protocol-specific calibration mapping includes: The qualitative or semi-quantitative descriptions in the reported adipose phenotype vectors are converted into adipose phenotype abnormality probabilities or potential continuous variables from the reported sources. The DICOM fat phenotype vectors were standardized according to MRI sequence type, scanning protocol, target slice, unit of measurement, and population reference distribution. The gated fusion outputs the fused phenotype and fusion confidence based on the report confidence, the sub-phenotype image quality confidence, the missing mask, and the conflict identifier; When the reported fat phenotype vector and the DICOM fat phenotype vector have inconsistent directions or the difference exceeds the conflict threshold, the conflict flag is set to valid and the fusion confidence is reduced; when the reported confidence or the confidence of the sub-phenotype image quality is lower than the preset threshold, the fusion confidence is reduced and a verification flag is generated.

5. The method for risk restratification of hypertension, hyperlipidemia, and hyperglycemia based on abdominal MRI combined phenotype according to claim 1, characterized in that, The risk characteristics of the body mass index residuals are determined by the following formula: For any adipose phenotype, a conditional expectation model is established based on reference data, with body mass index, age, and gender as inputs, to obtain the conditional expectation value; based on the difference between the actual value and the conditional expectation value of the joint phenotype vector and the residual scaling parameter in the reference data, the standardized residual is obtained. The risk characteristics of the body mass index residuals include at least one of the following: visceral fat standardized residuals, visceral fat to subcutaneous fat ratio standardized residuals, liver fat standardized residuals, maximum positive residuals, combined outlier counts, and residual direction consistency.

6. The method for risk restratification of hypertension, hyperlipidemia, and hyperglycemia based on abdominal MRI combined with phenotype as described in claim 1, characterized in that, The baseline risk model and the incremental imaging model were established for the risks of hypertension, hyperglycemia, and hyperlipidemia, respectively. For the k-th risk task, the baseline risk model outputs the baseline risk probability. The image increment model outputs the image increment. The corrected risk probability satisfies ; The risk adjustment amount satisfies .

7. The method for risk restratification of hypertension, hyperlipidemia, and hyperglycemia based on abdominal MRI combined with phenotype according to claim 1, characterized in that, The gating rules include: When the risk adjustment amount reaches the upward adjustment threshold and the fusion confidence level reaches the confidence threshold, the original risk level is increased. When the risk adjustment amount reaches the downgrade threshold, the fusion confidence level reaches the confidence threshold, the missing mask indicates no missing data, the conflict identifier indicates no conflict, and the downgrade restricted area is not triggered, the original risk level is downgraded. Otherwise, maintain the original risk level.

8. A risk restratification system for hypertension, hyperlipidemia, and hyperglycemia based on abdominal MRI combined with phenotype, characterized in that, include: The data acquisition unit is used to acquire the DICOM sequence of the same abdominal MRI examination of the subject to be processed, the imaging report associated with the DICOM sequence, the clinical characteristics of hypertension, hyperlipidemia, and hyperglycemia, and the original risk level. The report structuring unit is used to extract the structure from the image report and generate a report fat phenotype vector and a report confidence level. The image phenotype extraction unit is used to determine a first target region for abdominal fat measurement and a second target region for liver fat measurement from the DICOM sequence, and generate DICOM fat phenotype vectors and sub-phenotype image quality confidence scores, respectively. The calibration fusion unit is used to form a protocol-specific calibration mapping based on the image report and DICOM paired data, map the reported fat phenotype vector and the DICOM fat phenotype vector to a unified latent fat phenotype space, and perform gated fusion based on the report confidence, the sub-phenotype image quality confidence, the missing mask and the conflict marker to obtain the joint phenotype vector and the fusion confidence. The residual generation unit is used to determine the conditional expected value of at least one fat phenotype in the joint phenotype vector under the conditions of body mass index, age and sex based on reference object data, and to calculate the body mass index residual risk characteristics based on the deviation between the actual value and the conditional expected value. The risk restratification unit is used to input the three high clinical features (hypertension, hyperlipidemia, and hyperglycemia) that do not include the image fat phenotype into the baseline risk model to obtain the baseline risk probability, input the joint phenotype vector, the body mass index residual risk feature, and the fusion confidence into the image increment model to obtain the image risk increment, obtain the risk adjustment amount based on the baseline risk probability and the image risk increment, and output the risk level adjustment result relative to the original risk level (upward, maintained, or downward) based on the risk adjustment amount and gating rules.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the three-risk restratification method based on abdominal MRI combined phenotype as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the three-risk restratification method based on abdominal MRI combined phenotype as described in any one of claims 1 to 7.

Images