Search

CN-122000058-A - Disease prediction method based on medical data and multi-mode information fusion

CN122000058ACN 122000058 ACN122000058 ACN 122000058ACN-122000058-A

Abstract

The invention discloses a disease prediction method based on medical data and multi-mode information fusion, belonging to the technical field of medical data processing; the method comprises the steps of constructing a coronary artery CT image data set, constructing a SEGRESNET model, obtaining a high-precision vessel probability map by utilizing the SEGRESNET model, extracting peripheral adipose tissue of coronary artery based on the vessel probability map, outputting coronary vessel ROI and peripheral adipose ROI, extracting image histology characteristics of the coronary vessel ROI and the peripheral adipose ROI, constructing an integrated prediction model, and training by using comprehensive feature vectors of the labeling data set and lesion labels of the comprehensive feature vectors. According to the invention, a vascular probability map is obtained through SEGRESNET models, and a stable vascular mask and a central line are obtained by combining thresholding and morphological processing, so that the accurate positioning and extraction of the coronary artery and the surrounding fat region are realized by constructing the surrounding fat of the coronary artery, the ROI deviation and error transmission caused by tiny blood vessels, artifacts and unclear boundaries are reduced, and the stability and consistency of PCAT region extraction and subsequent image histology characteristics are improved.

Inventors

  • WANG JIANI
  • SU LI

Assignees

  • 苏州医朵云健康股份有限公司

Dates

Publication Date
20260508
Application Date
20260126

Claims (10)

  1. 1. A disease prediction method based on medical data and multi-mode information fusion is characterized by comprising the following steps: Step S1, constructing a coronary artery CT image data set, forming a marked data set and an unmarked data set, and preprocessing the marked data set; s2, constructing SEGRESNET a model, and obtaining a high-precision vessel probability map by using the SEGRESNET model; Step S3, based on the vessel probability map, extracting the peripheral adipose tissue of the coronary artery, and outputting the coronary vessel ROI and the peripheral adipose ROI; S4, extracting image histology characteristics of the coronary vessel ROI and the pericoronary fat ROI, splicing clinical characteristics of the patient with the image histology characteristics, and constructing a comprehensive characteristic vector; and S5, constructing an integrated prediction model, and training by using the comprehensive feature vectors of the labeling data set and the lesion labels of the labeling data set.
  2. 2. The disease prediction method based on medical data and multi-mode information fusion of claim 1, wherein the SEGRESNET model comprises an encoder, a decoder and a Dice loss optimization module, the encoder adopts a four-layer structure and comprises four coding blocks, namely a first coding block, a second coding block, a third coding block and a fourth coding block, wherein the four coding blocks are connected in series, and the core operation of each coding block is convolution and downsampling operation; The encoder converts the input high-resolution CT image into a bottleneck characteristic map with low resolution and high channel number through hierarchical convolution and downsampling operation; The decoder adopts a four-layer structure and comprises four decoding blocks, wherein the four decoding blocks are respectively formed by connecting a first decoding block, a second decoding block, a third decoding block and a fourth decoding block in series, and the core operation of the decoding blocks is up-sampling, jump connection with a characteristic diagram output by the encoder and convolution processing; The decoder gradually upsamples the abstract features extracted by the encoder and performs jump connection with the feature images of the corresponding layers of the encoder so as to recover space details, and finally outputs a vascular probability image with the same size as the input image; And the Dice loss optimization module calculates the difference between the predicted vascular probability map and the golden label data, and iteratively optimizes SEGRESNET models.
  3. 3. The disease prediction method based on medical data and multi-modal information fusion of claim 2, wherein a SEGRESNET model is constructed in the step S2, and a SEGRESNET model is utilized to obtain a high-precision vascular probability map, specifically: step S2-1 converting the dataset into five-dimensional tensor data Inputting to an encoder; Step S2-2, the data is subjected to convolution and downsampling operations by an encoder, specifically: the convolution of the first encoded block is connected to the residual: ; Wherein, the The output characteristic diagram of the first coding block is represented, and the result is obtained after convolution and residual connection; Representing a three-dimensional convolution operation; Representing a convolution kernel weight tensor; a bias vector representing a first encoded block; Representing a modified linear element activation function, the formula is ; Representing input tensor data; maximum pooled downsampling of the first encoded block: ; Wherein, the The characteristic diagram is output after pooling is represented; representing a three-dimensional maximum pooling operation; Representing the pooled kernel size; Representing a step size; Representing the number of downsampled slices; The convolution of the second encoded block is connected to the residual: ; Wherein, the Representing a second encoded block output profile; Representing convolution kernel weights; A bias vector representing a second encoded block; Maximum pooled downsampling of the second encoded block: ; the convolution of the third coding block is connected with the residual: ; Maximum pooled downsampling for the third encoded block: ; maximum pooled downsampling for the fourth encoded block: ; maximum pooled downsampling for the fourth encoded block: ; Wherein, the Representing a bottleneck layer feature map; Step S2-3, inputting the bottleneck layer feature map to a decoder to execute up-sampling and jump connection operation, specifically: Upsampling of the first decoding block: ; Wherein, the The feature map after up sampling; Skip concatenation of first decoded block, upsampling result Feature map with third coding block Splicing along the channel dimension: ; Wherein, the The characteristic diagram after splicing is [ ] represents splicing operation along the channel dimension; The convolution processing of the first decoding block, namely carrying out 3x3x3 convolution on the spliced feature images, integrating information and reducing dimension, wherein the expression is as follows: ; Wherein, the An output feature map for the first decoding block; , Convolution weights and offset parameters for the first decoding block; Upsampling of the second decoding block The up-sampling is performed such that, ; Skip connection of second decoding block to be Feature map with encoder second layer Splicing the components to be spliced, ; Convolution processing of the second decoding block: wherein, the method comprises the steps of, , Convolution weights and offset parameters for the second decoding block; Upsampling of the third decoding block: ; skip connection of third decoding block to be Feature map with first coding block Splicing the components to be spliced, ; Convolution processing of the third decoding block: wherein, the method comprises the steps of, , Convolution weights and offset parameters for the third decoding block; Upsampling of the fourth decoding block restores the feature map resolution to the original input size, ; Convolution processing of the fourth decoding block, which reduces the number of channels to 1 using a 1x1x1 convolution, generates a single activation value for each pixel, Wherein, the method comprises the steps of, Is the convolved output tensor; , the final convolution weight and bias parameters; The final output is passed through a Sigmoid function, mapping the activation value of each pixel to a probability of belonging to a vessel: wherein, the method comprises the steps of, Representing a vessel probability map; and S2-4, calculating the difference between the blood vessel probability map and the gold standard, and optimizing.
  4. 4. The method for predicting disease based on medical data and multimodal information fusion of claim 3, wherein the difference between the vascular probability map and the golden standard is calculated and optimized in step S2-4, specifically: For each sample in the batch Calculating predictions Standard with gold Is a degree of overlap of: ; Wherein the method comprises the steps of Representing the sum of the matrix elements, Representation is achieved by element-wise multiplication; Average loss of all samples: Loss value By back propagation algorithm, the gradient of all the learnable parameters relative to the model is calculated ; ; Iteratively updating model parameters according to the calculated gradient by using an SGD optimization algorithm, and performing batch processing on CT images of any single sample Obtaining a corresponding vascular probability map ; Vascular probability map Each pixel value of (1) characterizes its corresponding point location in three-dimensional space as belonging to a particular vessel Is a probability of (2).
  5. 5. The method for predicting a disease based on medical data and multimodal information fusion of claim 1, wherein the step S3 extracts the pericoronary adipose tissue based on the vessel probability map and outputs the coronary vessel ROI and the pericoronary adipose ROI specifically comprises: step S3-1, a vascular probability map is subjected to threshold segmentation and morphological operation to obtain a binary mask, wherein the binary mask is specifically: Setting a fixed binarization threshold value, converting a continuous probability map into a discrete binary mask, and clearly distinguishing a blood vessel foreground from an image background, wherein the expression is as follows: ; Wherein, the Representing the value of the binarization threshold value, Representing an indication function; a probability map of a blood vessel is represented, Representing an initial binary segmentation mask; Adopting morphological operation to process an initial binary segmentation mask, eliminating small noise points and smoothing vessel boundaries, wherein the expression is as follows: ; Wherein, the Representing a neighborhood defining corrosion and expansion operations; Indicating a corrosion operation; representing an expansion operation; representing a final coronary vessel segmentation mask; Step S3-2, extracting the fat ROI around the coronary artery, which specifically comprises the following steps: Mask segmentation from three-dimensional coronary vessels The extracted skeleton is a central line, and the expression is: wherein, the method comprises the steps of, Representing a refinement-based skeletonizing algorithm; Representing the extracted centerline, also a binary mask, where 1 represents the centerline pixel; each point on the central line is taken as a sphere center, a three-dimensional space is constructed in a specific radius, and the three-dimensional space is a coronary surrounding area, and the expression is as follows: ; Wherein, the Representation points Sum point Euclidean distance between them; representing the radius of expansion; A three-dimensional binary mask representing the perivenous space; In the surrounding space And (3) screening out adipose tissues according to the original gray value of the CT image, wherein the expression is as follows: ; Wherein, the Representation points In the original CT image CT value; a range of adipose tissue CT values recognized in medical imaging; A ROI binary mask representing the finally extracted pericoronary adipose tissue.
  6. 6. The method for predicting diseases based on medical data and multimodal information fusion of claim 1, wherein the step S4 extracts image histology characteristics of coronary vessel ROI and pericoronary fat ROI, and splices clinical characteristics of patients with the image histology characteristics to construct comprehensive characteristic vectors, specifically: s4-1, extracting image histology characteristics of the coronary vessel ROI, specifically defining an image histology characteristic extraction function The image histology feature extraction function outputs feature vectors containing multiple types of features, and the specific calculation flow is as follows: defining a voxel set and gray values, and defining a spatial range and a numerical object of feature calculation: Defining a voxel set, and determining the spatial range of feature analysis: ROI mask Is an image of the original CT A three-dimensional binary matrix with the same size, wherein a voxel coordinate point with the value of 1 in the matrix belongs to the ROI, and the value of 0 is background, and the voxel coordinates marked with 1 form a voxel set , ; Wherein, the The three-dimensional coordinate vector represents the spatial position of a certain voxel in the image; Representing a three-dimensional space formed by all integer coordinates, namely the whole CT image; Is a mask In coordinates of The value at which the value is to be calculated, Representing a collection The total number of medium voxels; defining gray values, and defining numerical objects of feature analysis: For collections Each voxel of (a) Gray value of it Directly from the original CT image At the same coordinate Numerical values at the points, expressed as: ; Wherein, the Is a voxel CT value at the location; Step S4-1-2, calculating first-order statistical features, and describing global distribution of voxel intensity in the ROI: The first order statistical features include averages representing CT values within the ROI Measuring degree of dispersion of CT value Measuring degree of deviation of distribution asymmetry Kurtosis measuring how steep a distribution is relative to normal ; ; ; ; ; Step S4-1-3, calculating morphological characteristics and describing three-dimensional geometric characteristics of the ROI: The morphological characteristics include volume, surface area, sphericity, volume Wherein, the method comprises the steps of, A physical volume that is a single voxel; Surface area Wherein, the method comprises the steps of, Is a set of surface voxels of the ROI, Is a voxel Surface area contribution of (c); Sphericity degree Wherein, the method comprises the steps of, To measure the extent to which the ROI is close to a sphere, the value range is [0,1]; S4-1-4, calculating texture features by using a gray level co-occurrence matrix GLCM; and S4-2, constructing a comprehensive feature vector to perform multi-source feature fusion.
  7. 7. The method for predicting disease based on medical data and multimodal information fusion of claim 6, wherein the step S4-1-4 uses a gray level co-occurrence matrix GLCM to calculate texture features, specifically: The gray level co-occurrence matrix GLCM is constructed, which comprises the following steps: image gray level quantization, namely setting a gray level number Mapping the original CT value range of voxels within the ROI to the interval linearly or non-linearly Integers within such that each voxel has a discrete gray level ; Defining spatial relationships by displacement vectors The definition of the term "a" or "an" is, Representing the spatial distance between pairs of voxels, Representing the angle; Statistical symbiotic frequencies, find for each voxel within the ROI the displacement relative to it as Is to the position in the GLCM matrix 1 Is added to the count of (2); for a given displacement vector The GLCM generated is one Matrix of matrix elements The definition is as follows: ; Wherein, the Representing the number of voxel pairs meeting the conditions in brackets; Representing ROI Three-dimensional coordinates of two voxels within; Representing voxels With respect to voxels Must be exactly equal to the defined displacement vector ; Representing voxels Is of the gray scale of Voxel(s) Is of the gray scale of ; Matrix normalization-matrix is performed Divided by the sum of all its elements, the expression is: After normalization, the method comprises the steps of, At this time Interpreted as the probability of occurrence of voxel pairs; Calculating texture indexes, wherein the texture indexes comprise contrast ratio for measuring the change degree of a local image, correlation for measuring linear dependency relationship among pixels, energy for measuring uniformity of the image and homogeneity for measuring similarity of local gray scales, and the contrast ratio is as follows: ; Correlation: wherein, the method comprises the steps of, Is that Line mean and standard deviation of (2); , Is the feature total dimension; For blood vessels Is not equal to the sample number of each sample of (1) Feature extraction is performed on the two ROIs respectively: coronary vessel ROI feature vector: wherein, the method comprises the steps of, A vector dimension representing a coronary vessel; pericoronal fat ROI feature vector: wherein, the method comprises the steps of, Representing the vector dimension of pericoronal fat.
  8. 8. The disease prediction method based on medical data and multi-modal information fusion of claim 6, wherein the constructing of the comprehensive feature vector in step S4-2 performs multi-source feature fusion, specifically: The same sample and the same blood vessel Is spliced with clinical feature vectors to form a comprehensive feature vector ; Wherein, the Representing vector stitching; is the total dimension of the whole feature vector; representing the clinical feature vector of the sample, Is the number of clinical features.
  9. 9. The method for predicting disease based on medical data and multimodal information fusion of claim 1, wherein the step S5 is to construct an integrated prediction model, the integrated prediction model comprises a plurality of base models, and the integrated prediction model is constructed by training the plurality of base models and assigning voting weights according to cross-validation performance of the base models, specifically: step S5-1, evaluating the performance of the base model by adopting a cross-validation mode: Each base model quantifies performance metrics by AUC values, each base model , Sequentially corresponding to a vector classifier SVC, a logistic regression LR and a gradient lifting tree GB; The vector classifier SVC, the logistic regression LR and the gradient lifting tree GB quantify performance indexes through AUC values, and the expression is that the AUC values are calculated: ; Wherein, the In order to indicate the function, Representing the predicted probability of the model for a certain positive sample, Representing the predictive probability of a model for a negative sample, when the condition is When the function value is established, the function value is 1, otherwise, the function value is 0; Representing the number of positive samples in the validation set; representing the number of negative samples in the validation set; for verification set Calculation of Probability of model, the expression is: ; S5-2, calculating the weight of the weighted voting based on the cross-validation AUC value of the base model, and carrying out linear normalization; And S5-3, constructing a weighted voting integrated prediction function by using the trained base model.
  10. 10. The method for predicting disease based on medical data and multimodal information fusion of claim 1, wherein the step S5-2 calculates weights of weighted votes based on cross-validated AUC values of a base model, and performs linear normalization, specifically: wherein, the method comprises the steps of, Is a model Weights of (2); As the value of the AUC, ; Is the first The model is at the first The AUC value of the fold is satisfied ; S5-3, constructing a weighted voting integrated prediction function by using the trained base model, wherein the weighted voting integrated prediction function specifically comprises the following steps: Using the complete training set Training each base model, and expressing as follows: ; feature vector for new sample The predictive probability of the integrated model is the weighted voting result: wherein, the method comprises the steps of, Representing the output probability value range 0,1, In order to normalize the weights, the weights are, Probabilities are predicted for the base model.

Description

Disease prediction method based on medical data and multi-mode information fusion Technical Field The invention belongs to the technical field of medical data processing, and particularly relates to a disease prediction method based on medical data and multi-mode information fusion. Background Coronary atherosclerotic disease is one of the leading causes of cardiovascular events. With the popularity of coronary CT vascular imaging (Coronary CT Angiography, CCTA) in clinic, automated risk assessment and disease prediction based on CCTA is an important direction for intelligent analysis of medical images. Compared with the traditional evaluation mode which only depends on the degree of lumen stenosis, more and more researches show that the image phenotypes such as the blood vessel wall, plaque morphology, and the related change of the inflammation of the adipose tissue around the blood vessel have correlation with the future adverse events, so that the construction of a prediction model which can fuse multi-source information (image, image histology and clinical data) has important significance. However, the coronary structure is small, the walking is complex, and motion artifacts exist, which lead to insufficient coronary segmentation, definition of a region of interest (Region of Interest, ROI) and feature extraction stability, thereby affecting the downstream generalization ability of the disease prediction model. In particular, accurate extraction of pericoronary fat (Peri-coronary Adipose Tissue, PCAT) regions requires reliable vessel localization and consistent spatial definition, otherwise it is prone to contamination of the myocardium, blood or other tissue, causing deviations in the imaging histology. Existing coronary analysis procedures typically first perform coronary segmentation or centerline extraction. The traditional method mostly adopts means such as enhanced filtering, region growing, graph searching, morphological processing and the like based on rules, but is sensitive to noise, artifacts and anatomical variation. In recent years, deep learning segmentation models (such as U-Net and its variants, a residual network combined with a segmentation network of encoder-decoder structure, etc.) have been shown to be prominent in three-dimensional medical image segmentation, enabling the output of vessel masks or probability maps, and optimizing segmentation accuracy by a loss function. Although deep learning improves segmentation, small branches of the coronary, artifacts caused by proximal calcification, and domain offsets from different scan protocols can still lead to segmentation instability. In addition, the binary segmentation result is directly output for subsequent processing, and boundary uncertainty is difficult to express, so that the consistency of subsequent ROI construction is affected. From the above conclusion, the prior art has the defects that the PCAT ROI and the image histology feature are unstable due to the coronary artery positioning and the ROI construction error transmission, and the PCAT ROI and the image histology feature are unstable due to the coronary artery positioning and the ROI construction error transmission; Therefore, a prediction method is needed to solve the problem of obtaining reliable vessel localization in coronary CCTA, so as to stably extract coronary vessel ROI and pericoronary fat ROI. Disclosure of Invention The invention aims to provide a disease prediction method based on medical data and multi-mode information fusion, so as to solve the problems in the background technology. The invention aims to realize the disease prediction method based on medical data and multi-mode information fusion, which is characterized by comprising the following steps: Step S1, constructing a coronary artery CT image data set, forming a marked data set and an unmarked data set, and preprocessing the marked data set; s2, constructing SEGRESNET a model, and obtaining a high-precision vessel probability map by using the SEGRESNET model; Step S3, based on the vessel probability map, extracting the peripheral adipose tissue of the coronary artery, and outputting the coronary vessel ROI and the peripheral adipose ROI; S4, extracting image histology characteristics of the coronary vessel ROI and the pericoronary fat ROI, splicing clinical characteristics of the patient with the image histology characteristics, and constructing a comprehensive characteristic vector; and S5, constructing an integrated prediction model, and training by using the comprehensive feature vectors of the labeling data set and the lesion labels of the labeling data set. Preferably, the SEGRESNET model comprises an encoder, a decoder and a Dice loss optimization module, wherein the encoder adopts a four-layer structure and comprises four coding blocks, namely a first coding block, a second coding block, a third coding block and a fourth coding block, wherein the four coding blocks are connected in serie