CN-122025036-A - Multi-mode data fusion whole hospital blood glucose risk assessment method and system

CN122025036ACN 122025036 ACN122025036 ACN 122025036ACN-122025036-A

Abstract

S001, in the initial stage of medical image loading, obtain the multi-scale edge intensity of the image to be processed, frequency disturbance range and local structure self-similarity difference, withdraw the sensitive characteristic of the unusual region, form the sensitive characteristic set of the positionable artifact; S002, inputting the artifact sensitive feature set into a structural integrity judging mechanism, comparing the structural integrity judging mechanism with a preset medical image template, identifying the shielding degree, deformation type and structural distortion range of an abnormal region, generating an image quality defect map and labeling structural shielding. The invention builds a closed-loop mechanism of model optimization perceived by image anomaly, effectively avoids the influence of image artifacts and other interferences on blood sugar risk assessment, realizes characteristic stripping, weight regulation and credible reasoning, improves model stability and clinical adaptability, and has self-evolution capability.

Inventors

ZHENG CHAO
SUN QINGNAN
HE MINGYUAN
XIANG PENG
ZHANG YIKAI
HE BIHANG
YAO CHANG

Assignees

浙江大学医学院附属第二医院
浙江和仁科技股份有限公司
智联嘉医(杭州)数字技术有限公司

Dates

Publication Date: 20260512
Application Date: 20251230

Claims (9)

1. The whole hospital blood glucose risk assessment method based on multi-mode data fusion is characterized by comprising the following steps of: s001, in the initial stage of medical image loading, acquiring multi-scale edge intensity, frequency disturbance amplitude and local structure self-similarity difference of an image to be processed, extracting sensitive features of an abnormal region, and forming a positionable artifact sensitive feature set; s002, inputting the artifact sensitive feature set into a structural integrity judging mechanism, comparing the structural integrity judging mechanism with a preset medical image template, identifying the shielding degree, deformation type and structural distortion range of an abnormal region, generating an image quality defect map and labeling structural shielding; S003, extracting edge direction, texture complexity and spectrum distortion parameters of an artifact region based on an image quality defect map, calculating misleading risk scores and generating scoring matrixes, mapping the scoring matrixes to an image coordinate system, and constructing a discriminant feature shielding frame; s004, performing self-adaptive fusion channel regulation and control according to the discriminant feature shielding framework, dynamically adjusting the weight of an abnormal region in the multi-mode feature fusion process, and reducing the participation strength of the abnormal region through a confidence weakening strategy to generate a fusion feature stability index; s005, combining the characteristic stability index and dynamic fluctuation of multi-mode input, calculating a credibility vector, generating an inference credibility interval, adjusting a model judgment threshold and an output boundary according to a credibility state and an interference level, and recording correction intensity and a feedback factor; S006, based on the corrected reasoning result and feedback factor, establishing a fusion memory caching mechanism, archiving an image defect map, artifact characteristics, a weight adjustment result and a trusted vector, performing artifact scene clustering, and constructing an abnormal mode database for guiding loss function bias adjustment and structural parameter optimization in model iterative training to form a model evolution closed loop.
2. The method for whole hospital blood glucose risk assessment based on multi-modal data fusion according to claim 1, wherein step S001 comprises: performing size standardization treatment on the medical image to be treated, and extracting multi-scale edge intensity distribution; Dividing an image into subareas with fixed sizes, and calculating the frequency disturbance amplitude of each subarea; Extracting a plurality of image blocks in an image, and executing structural self-similarity difference analysis; and fusing three indexes of edge strength, frequency disturbance and structural self-similarity, and positioning a sensitive area in an image coordinate system to generate an artifact sensitive characteristic set.
3. The method for whole hospital blood glucose risk assessment based on multi-modal data fusion according to claim 1, wherein step S002 comprises: Matching the artifact sensitive feature set with a preset medical image specification feature template, and identifying a structural abnormality candidate region; calculating the shielding strength index of the candidate region, judging the shielding level and marking the shielding type; Analyzing geometric deformation and structural continuity damage conditions of the candidate region, and identifying deformation types and fracture characteristics; And constructing an image quality defect map in an image coordinate system, and generating structural shielding annotation information.
4. The method for whole hospital blood glucose risk assessment based on multi-modal data fusion according to claim 1, wherein step S003 comprises: extracting edge direction angles of pixels in the artifact region and calculating a direction concentration degree; Dividing the sub-blocks of the artifact region, and calculating a gray level statistic value, an information entropy, a gray level co-occurrence matrix contrast and an energy value to obtain a texture complexity score; performing Fourier transform analysis on spectral features of the artifact region, and calculating the high-frequency energy duty ratio and the direction concentration; and generating a misleading risk scoring matrix by weighting and fusing the scores, and constructing a discriminant feature shielding framework according to the misleading risk scoring matrix.
5. The method for whole hospital blood glucose risk assessment based on multi-modal data fusion according to claim 4, wherein the specific steps of constructing the discriminative feature shielding framework are as follows: positioning the spatial range of each artifact region in an original image coordinate system according to the misleading risk scoring matrix; marking the artifact area with the score higher than the set threshold value as a high risk area, and generating a binary mask image consistent with the image size; in the mask image, pixels corresponding to the high-risk artifact region are assigned zero, and pixels not corresponding to the high-risk region are assigned one; And taking the mask image as an input constraint condition in the subsequent feature extraction, fusion and risk modeling processes, and shielding the interference features of the high-risk region.
6. The method for whole hospital blood glucose risk assessment based on multi-modal data fusion according to claim 1, wherein step S004 comprises: Dividing the image mode input into a trusted region and an interference region according to the feature shielding mask image, and respectively extracting gray histogram features, local contrast features and texture direction consistency features of each type of region; applying a confidence level weakening strategy to the interference area, setting a scaling factor according to the characteristic value of the interference area, and adjusting the response value intensity of the interference area in the channel characteristic diagram; integrating all the scaled characteristic responses into a channel response mapping matrix, and carrying out weighted fusion according to a preset credibility coefficient in a multi-mode fusion process; And calculating a fusion characteristic stability index according to the deviation of the fusion value and the original response value, and generating a fusion characteristic stability thermodynamic diagram in a two-dimensional image form.
7. The method for whole hospital blood glucose risk assessment based on multi-modal data fusion according to claim 1, wherein step S005 comprises: Extracting the fusion characteristic stability score of each pixel point in the image mode, and calculating a unified normalized credibility vector by combining the structural inspection data, the continuous blood glucose monitoring data and the variation amplitude of vital sign parameters; constructing an inference reliability index and a reliability interval distribution diagram based on the reliability vector, comparing the average reliability of the target judgment area with the boundary of the confidence interval, and judging whether the inference result is stable and reliable; And dynamically adjusting a judgment threshold and an output boundary of the blood glucose risk assessment model according to the overall credibility mean value and the interference area coverage ratio of the image, and recording correction strength and credible feedback factors generated in the adjustment process.
8. The method for whole hospital blood glucose risk assessment based on multi-modal data fusion according to claim 1, wherein step S006 comprises: The image quality defect map, the artifact sensitive feature set, the channel weight adjustment result, the credibility evaluation vector and the reasoning result corresponding to the current input image are structurally archived and stored in a cache management database; Extracting correction intensity, artifact area occupation ratio, region quantity and density of an archive sample, generating feature vectors, performing density clustering, and constructing image abnormal input scene classification; and according to the image abnormal scene classification result, the loss function sample weight and the feature extraction path weight are adjusted in the model training stage, so that closed-loop evolution optimization of the model structure is realized.
9. A whole hospital blood glucose risk assessment system for multi-modal data fusion, which is used for realizing the whole hospital blood glucose risk assessment method for multi-modal data fusion according to any one of the claims 1-8, and is characterized by comprising an image anomaly characteristic extraction module, a structural integrity analysis and defect labeling module, an interference characteristic risk modeling and shielding module, a fusion channel regulation and stability assessment module, an inference reliability construction and dynamic judgment module and a memory cache archiving and model evolution optimization module: The image abnormal feature extraction module is used for acquiring the multi-scale edge intensity, the frequency disturbance amplitude and the local structure self-similarity difference of the image to be processed in the initial stage of medical image loading, extracting the sensitive features of an abnormal region and forming a positionable artifact sensitive feature set; the structural integrity analysis and defect labeling module inputs the artifact sensitive feature set into a structural integrity judging mechanism, compares the artifact sensitive feature set with a preset medical image template, identifies the shielding degree, deformation type and structural distortion range of an abnormal region, generates an image quality defect map and labels structural shielding; the interference feature risk modeling and shielding module is used for extracting edge direction, texture complexity and spectrum distortion parameters of an artifact region based on the image quality defect map, calculating misleading risk scores, generating scoring matrixes, mapping the scoring matrixes to an image coordinate system and constructing a discriminant feature shielding frame; The fusion channel regulation and stability evaluation module executes self-adaptive fusion channel regulation according to the discriminant feature shielding framework, dynamically adjusts the weight of an abnormal region in the multi-mode feature fusion process, reduces the participation intensity of the abnormal region through a confidence weakening strategy, and generates a fusion feature stability index; the inference reliability construction and dynamic judgment module is used for combining and fusing characteristic stability indexes and dynamic fluctuation of multi-mode input, calculating a reliability vector, generating an inference reliability interval, adjusting a model judgment threshold and an output boundary according to a reliability state and an interference level, and recording correction strength and feedback factors; and the memory cache archiving and model evolution optimization module establishes a fusion memory cache mechanism based on the corrected reasoning result and feedback factor, archives the image defect map, the artifact characteristics, the weight adjustment result and the credible vector, performs artifact scene clustering, and constructs an abnormal mode database for guiding loss function bias adjustment and structural parameter optimization in model iterative training to form a model evolution closed loop.

Description

Multi-mode data fusion whole hospital blood glucose risk assessment method and system Technical Field The invention relates to the technical field of medical information intelligent risk assessment, in particular to a whole hospital blood glucose risk assessment method and system based on multi-mode data fusion. Background The whole hospital blood glucose risk assessment with multi-mode data fusion refers to the process of systematically predicting and dynamically assessing the blood glucose abnormality risk of all hospitalized or outpatient patients in the whole hospital range. The method comprehensively utilizes various medical data sources (namely 'multi-mode data') including but not limited to structured electronic medical records, laboratory test indexes (such as fasting blood glucose and glycosylated hemoglobin), continuous blood glucose monitoring data, drug use records, vital signs, image data, doctor diagnosis and treatment behaviors and individual characteristics (such as age, gender, past medical history and the like) of patients, and establishes a unified data analysis model by fusing association relations and interaction characteristics among different data. By means of an artificial intelligence algorithm and a deep learning technology, heterogeneous integration and risk weight modeling are carried out on different modal information, so that patient groups with hyperglycemia, hypoglycemia or abnormal blood sugar fluctuation can be identified in real time, and hierarchical classification management, personalized intervention and early warning reminding for a whole hospital are realized. The assessment method not only improves the sensitivity and accuracy of blood glucose risk identification, but also is beneficial to optimizing the blood glucose management flow in a hospital, reduces the incidence rate of diabetes related complications, and has important clinical value and intelligent medical popularization significance. The prior art has the following defects: In the whole hospital blood glucose risk assessment process of actually developing multi-mode data fusion, an image mode (such as abdomen CT, fundus images, lower limb vascular ultrasound and the like) is used as an important information source for assisting in judging typical complications of diabetes (such as diabetic retinopathy, diabetic nephropathy, fatty liver, atherosclerosis and the like), and the reliability of image quality directly influences the judging accuracy of a multi-mode fusion model. However, in the actual image acquisition and processing process, image anomalies such as motion blur, device artifact, occlusion area or signal-to-noise reduction exist in the image due to factors such as autonomous patient movement, acquisition device interference, limitation of scanning conditions or complexity of physiological structures. These abnormal disturbances not only cause local structural information loss or distortion, but also may introduce spurious features that are highly similar to the features of the real lesions, causing the model to deviate in feature extraction and risk classification. For example, a region of occlusion of eyelashes in a fundus image may be mistakenly identified as a bleeding spot, and a metal implant-induced stringing artifact in abdominal CT may be mistakenly identified as a pancreatic lesion shadow, resulting in an overestimate or underestimate of blood glucose risk. Because most fusion models lack explicit recognition and error correction mechanisms for abnormal fluctuation of image quality, the abnormal characteristics are extremely easy to lead to misleading of model training and reasoning processes, and finally influence the clinical reliability and decision security of blood glucose risk assessment results. The above information disclosed in the background section is only for enhancement of understanding of the background of the disclosure and therefore it may include information that does not form the prior art that is already known to a person of ordinary skill in the art. Disclosure of Invention The invention aims to provide a whole hospital blood glucose risk assessment method and system based on multi-mode data fusion, so as to solve the problems in the background technology. In order to achieve the above purpose, the invention provides a multi-mode data fusion whole hospital blood glucose risk assessment method, which comprises the following steps: s001, in the initial stage of medical image loading, acquiring multi-scale edge intensity, frequency disturbance amplitude and local structure self-similarity difference of an image to be processed, extracting sensitive features of an abnormal region, and forming a positionable artifact sensitive feature set; s002, inputting the artifact sensitive feature set into a structural integrity judging mechanism, comparing the structural integrity judging mechanism with a preset medical image template, identifying the shielding degree, deformation typ