Search

CN-120783164-B - Multi-mode imaging evaluation method suitable for atrial flutter cardiogenic stroke

CN120783164BCN 120783164 BCN120783164 BCN 120783164BCN-120783164-B

Abstract

The invention discloses a multi-mode imaging evaluation method suitable for atrial flutter cardiac stroke, which comprises the following steps of collecting and preprocessing multi-mode image data of a heart and a brain, registering the preprocessed data in a multi-mode image mode, respectively extracting and fusing features of the registered multi-mode image data of the heart and the brain, aligning and fusing cross-mode features of the image data of the heart and the brain, constructing a cardiovascular and cerebrovascular integrated evaluation interaction model based on a graph neural network, and predicting end-to-end cardiac stroke risk, wherein the multi-mode imaging evaluation method is to combine the multi-mode heart image data and the multi-mode brain image data to establish a cardiovascular and cerebrovascular integrated evaluation frame, and the image features and the clinical data are fused by using an artificial intelligent method to realize end-to-end stroke risk prediction and evaluation. The invention comprehensively utilizes artificial intelligence and medical technology, and provides a more accurate cardiac stroke assessment method.

Inventors

  • LI YAN
  • GUO JUN
  • LI MEILING
  • GAO SIQI

Assignees

  • 暨南大学附属第一医院（广州华侨医院）

Dates

Publication Date
20260508
Application Date
20250606

Claims (8)

  1. 1. A multi-modal imaging assessment method suitable for atrial flutter cardiogenic stroke, comprising the steps of: step 1, heart and brain multi-mode image data acquisition and preprocessing; step 2, carrying out multi-mode image registration on the preprocessed data; step 3, respectively carrying out feature extraction and fusion on the registered heart and brain multi-mode image data; Step 4, performing cross-modal feature alignment and fusion on the heart and brain image data; step 5, constructing a cardiovascular and cerebrovascular integrated evaluation interaction model based on the graph neural network; Step 6, predicting the risk of end-to-end heart-derived stroke; the heart and brain multi-mode image data acquisition and preprocessing method comprises the following steps: Step 101, reading heart multi-mode image data, including TEE, MRI, CT data and brain multi-mode image data, including DWI-MRI, CTP, MRA data, and carrying out normalization processing on each mode image; step 102, denoising and filtering the data subjected to normalization processing in different modes respectively; the preprocessing data is subjected to multi-mode image registration, and the method comprises the following steps of: step 201, selecting soft tissue high contrast and high resolution cardiac MRI as reference image TEE and cardiac CT are respectively used as images to be registered And Registering the heart multi-modality image data to the same coordinate space, comprising the steps of: for reference images using affine transformation And images to be registered Performing primary registration; Using a similarity measure: , Representing an image Is used as a reference to the entropy of (a), Represents the joint entropy of the data and, Representing the image to be registered after affine transformation, and optimizing similarity measurement by adopting gradient descent: Wherein In order for the rate of learning to be high, Represent the first Affine transformation matrix after sub-optimization, searching for optimal transformation matrix : Obtaining registered heart multi-modal data: After registration, get , After registration, get ; Step 202, selecting brain MRA with high spatial resolution as reference image DWI-MRI, CTP are respectively used as images to be registered And Registering the brain multi-mode image data to the same coordinate space to obtain registered brain multi-mode data: After registration, get , After registration, get ; Step 203, using a symmetry sensing mechanism, optimizing the image registration loss by using the symmetry of the brain image, and improving the registration accuracy.
  2. 2. The method for multi-modal imaging assessment of atrial flutter heart stroke of claim 1, wherein the optimizing the image registration loss using symmetry of brain images using a symmetry sensing mechanism comprises the steps of: Step 20301, for brain images after registration Generating mirror image thereof by horizontal overturn Then, the symmetry loss is calculated To measure the difference between the original image and the mirror image, which is smaller The symmetry is higher; Step 20302, enhancing symmetry of the image by computing a cross-attention mechanism, first mapping image data to a query matrix Key matrix Sum matrix : , wherein, Respectively obtaining mapping matrixes by training; step 20303, calculating an attention matrix A weighted relationship of the respective positions is obtained, wherein, As a scaling factor, for controlling the stability of the calculation; step 20304, further optimizing the registration result, the registration loss function is: , wherein, For structural similarity measurement, measure images And The degree of similarity between the two, For mutual information measurement, the dependency relationship between two images is measured, Is a weight super parameter for controlling the relative importance of two loss terms.
  3. 3. The method for multi-modality imaging evaluation for atrial flutter cardiac stroke of claim 2, wherein the feature extraction and fusion of the registered multi-modality image data of the heart and brain, respectively, comprises the steps of: Step 301, optimizing a vascular reconstruction structure by combining maximum posterior estimation, improving vascular reconstruction precision and improving feature extraction precision; Step 302, extracting the registered heart multi-mode image data features by using a depth feature extraction network to obtain feature vectors with the same dimension 、 And ; Step 303, performing feature fusion on the heart multi-mode data features by adopting a transducer structure to obtain heart multi-mode image data fusion features The heart disease prediction method comprises the steps of including characteristic information of heart anatomy, left auricle thrombus and myocardial fibrosis, and being used for subsequent heart-derived stroke risk prediction; Step 304, extracting the registered brain multi-mode image data image features, carrying out feature extraction and analysis on CTP data by PerfGAT, and then fusing the brain multi-mode image data features to obtain brain multi-mode image data fusion features Contains characteristic information of brain lesions, vascular stenosis and blood supply modes.
  4. 4. A multi-modal imaging assessment method for atrial flutter heart stroke as claimed in claim 3, wherein said optimizing the vascular reconstruction structure in combination with maximum a posteriori estimation comprises the steps of: step 30101, performing initial vessel segmentation on the image after registration of the heart and brain image data; step 30102, maximum a posteriori estimating MAP multiple iterative optimization vessel, ensuring the vessel structure smooth; step 30103, enhancing the vascular details by using a GAN-based super-resolution network, enhancing the vascular contrast by adopting a histogram matching method, and completing the final vascular reconstruction and optimization.
  5. 5. A multi-modal imaging assessment method for atrial flutter heart stroke as claimed in claim 3, wherein said feature extraction and analysis of CTP data using PerfGAT comprises the steps of: Step 30401, obtaining features of the registered CTP data preliminarily extracted by the convolutional neural network ; Step 30402, constructing a graph structure including spatiotemporal dependencies , wherein, Is a collection of nodes of the graph, each node representing a pixel in the CTP image, Is a set of edges of the graph representing spatial relationships between pixels and temporal relationships, i.e., time series data; step 30403, by the formula: For a pair of Processing to calculate a feature update for each node, wherein Represent the first Layer node Is used for the feature vector of (a), Is a node Including spatial neighbors and temporal neighbors, Representing the attention coefficient, measuring the node Opposite node Depending on the spatio-temporal relationship between nodes, Representing a matrix of weights that can be learned, Representing an activation function, through the process, the characteristics of each node can be updated, and space-time dependency information is captured; And step 30404, obtaining CTP characteristics fused with spatial and temporal characteristics after passing through a plurality of GNN layers, and using the CTP characteristics for predicting cerebral ischemia areas and subsequent characteristic fusion and analysis.
  6. 6. A multi-modal imaging assessment method for atrial flutter cardiogenic stroke as recited in claim 3, wherein said cross-modal feature alignment and fusion of heart and brain image data comprises the steps of: step 401, fusion features of cardiac multi-modality image data from the same patient using a contrast learning mechanism Fusion features with brain multi-modality image data Alignment is performed using a cardiac and cerebral feature alignment loss function: Wherein And Is the first The heart and brain multi-modal image data fusion features of individual samples, Is the cosine similarity calculation and the like, The temperature parameter and the loss function maximize the similarity of the positive sample pair and minimize the similarity of the negative sample pair; Step 402, performing feature extraction on clinical data of atrial fibrillation, blood pressure, heart rate, D-dimer, NT-proBNP and complications of a patient by using a multi-layer perceptron structure to obtain clinical data features of the patient ; Step 403, performing cross-modal fusion on the heart and brain multi-modal image data fusion characteristics and the clinical data characteristics, and obtaining a fused characteristic representation by adopting a cross-modal transducer method The relationship between heart, brain and clinical features is captured.
  7. 7. The multi-modal imaging evaluation method suitable for atrial flutter cardiac stroke of claim 6, wherein the establishing of the cardiovascular and cerebrovascular integrated evaluation interaction model based on the graph neural network comprises the following steps: step 501, constructing a cardiovascular and cerebrovascular interaction map Wherein the node , Represent the first The nodes of the individual heart are connected, Represent the first The brain nodes comprise three kinds of heart nodes including left atrium, left auricle and aorta, and two kinds of brain nodes including middle cerebral artery and anterior cerebral artery, and sides Including the heart internal connection edge connecting left atrium, auricle, aorta: Internal connecting edges of brain connecting main cerebral arteries: and connecting the left auricle with a cerebral blood supply area and simulating a heart-brain interaction edge of a thrombus propagation path: ; step 502, the graph neural network calculates the heart brain interaction characteristics, the characteristic vector of each node is transmitted and updated through neighborhood information, and finally the heart brain interaction characteristics are obtained as follows: 。
  8. 8. The method of claim 7, wherein the predicting the risk of end-to-end heart stroke comprises the steps of: step 601, predictive task definition, predicting cardiac stroke risk for a given patient , wherein, Representing input multi-modality image data and clinical features of a patient, A risk of stroke label is indicated, Representing a high risk of being able to take place, Representing a low risk; step 602, to Sending the probability of risk in stroke into a classification network and finally outputting the probability of risk in stroke: , wherein, Is the learnable weight of the classification layer, and the loss function adopts cross entropy: Wherein Is a true risk tag for stroke, The risk probability of model prediction; Step 603, optimizing parameters in the predicted model of the risk of stroke by using the inverse Monte Carlo method, so that the predicted result of the risk of stroke better accords with the distribution of the observed data, and the goal is to solve the model parameters Posterior distribution of (c): , wherein, A set of training data is represented and, Is a likelihood function that is a function of the likelihood, Is a priori distribution, find new model parameters So that the predicted probability of stroke As much as possible fit the real label Wherein, in the distribution of (c), wherein, Wherein In order to normalize the weights, the weights are, , wherein, Is in the sample The likelihood of the lower one is that, Is the prior probability of being a priori, It is the proposed distribution probability that, Is from Sample in (b) to obtain A sample of the parameters is taken, Representing the total number of parameter samples; Step 604, predicting result analysis, high risk area: suggesting that a cardiac stroke may occur, risk area: Further inspection is required.

Description

Multi-mode imaging evaluation method suitable for atrial flutter cardiogenic stroke Technical Field The invention belongs to the crossing field of medicine and artificial intelligence science, and particularly relates to a multi-mode imaging evaluation method suitable for atrial flutter heart stroke. Background Atrial fibrillation (Atrial Fibrillation, AF) is one of the most common cardiac arrhythmias in clinic, with a significant increase in incidence with age. Atrial fibrillation patients are prone to blood flow retention due to atrial contraction dysfunction, and thrombus is formed at the left auricle and other parts. After the thrombus has fallen off and entered the systemic circulation, it reaches the brain via the aortic and carotid systems, which can cause a cardiac stroke (Cardioembolic Stroke, CES). Clinically, atrial fibrillation has proven to be one of the major risk factors for cardiac stroke. Compared with other types of ischemic strokes, atrial fibrillation-related cardiac strokes often have higher disability rate and mortality rate and higher recurrence risk, so that the accurate evaluation of the stroke risk of atrial fibrillation patients has important clinical value for prevention and treatment of atrial fibrillation patients. Traditionally, stroke risk assessment for patients with atrial fibrillation has relied primarily on clinical scoring systems (e.g., CHA2DS 2 -VASc scoring) and single modality imaging examinations, however these methods have significant shortcomings in accuracy and personalized risk prediction. Firstly, a single image mode is difficult to comprehensively reflect complex pathological changes of a cardiovascular system and a cerebrovascular system of a patient, key characteristics such as left auricle thrombosis, myocardial fibrosis and cerebral blood supply abnormality cannot be fully captured, secondly, the traditional registration method is difficult to realize accurate alignment of multi-mode images due to larger resolution, contrast and signal-to-noise ratio differences among different image modes, so that insufficient information fusion is caused, in addition, the traditional method often depends on rules formulated by expert experience when processing a heart-brain interaction relationship, lacks data-driven automatic modeling capability, and has a large improvement space in the aspects of accuracy, individuation evaluation and early warning due to lack of deep mining of multi-mode image characteristics of the heart and the brain and effective fusion of clinical data. Therefore, a comprehensive evaluation method based on deep learning, multi-modal data fusion and graph neural network is needed, not only can accurately register and extract features of heart and brain images, but also can realize cross-modal information alignment and interactive modeling, thereby improving the prediction accuracy of cardiac stroke risks of patients suffering from atrial fibrillation and providing more reliable basis for clinical treatment. Disclosure of Invention In order to solve the problems, the invention provides a multi-mode imaging AI evaluation method suitable for atrial flutter cardiac stroke, which comprehensively utilizes a plurality of image data such as TEE (transesophageal echocardiography), cardiac CT (computed tomography), cardiac MRI (magnetic resonance imaging), brain DWI-MRI (diffusion weighted magnetic resonance imaging), CTP (CT perfusion imaging) and MRA (magnetic resonance vascular imaging), adopts advanced image preprocessing, non-rigid registration, multi-mode feature extraction and fusion technology and introduces GNN (graphic neural network) to construct a cardiovascular and cerebrovascular integrated evaluation interaction model. Through the end-to-end deep learning framework and the reverse Monte Carlo, maximum posterior estimation and other optimization technologies, the invention can realize the accurate prediction of the cardiac stroke risk of patients with atrial fibrillation on the basis of fully extracting and fusing the cross-mode images and clinical data characteristics, and provides a more personalized and prospective risk assessment method for clinic. The invention discloses a multi-mode imaging evaluation method suitable for atrial flutter cardiac stroke, which adopts the technical scheme that the method comprises the following steps: step 1, heart and brain multi-mode image data acquisition and preprocessing; step 2, carrying out multi-mode image registration on the preprocessed data; step 3, respectively carrying out feature extraction and fusion on the registered heart and brain multi-mode image data; Step 4, performing cross-modal feature alignment and fusion on the heart and brain image data; step 5, constructing a cardiovascular and cerebrovascular integrated evaluation interaction model based on the graph neural network; Step 6, predicting the risk of end-to-end heart-derived stroke; The multi-mode imaging evaluation method is to combi