A method and system for predicting post-liver transplantation complications based on habitat characteristics

Abstract

The application provides a kind of based on habitat feature's liver transplantation postoperative complication prediction method and system, it is related to liver transplantation postoperative complication prediction field, method includes: obtaining multimodal medical image data and clinical index data and pre-processing and segmentation to carry out multiple scale Laplace-Gaussian space transformation and generate multiple scale image sequence;The region of interest under each scale image sequence is unsupervised clustering to divide and obtain multiple image habitats;Respectively extract the image habitat of each image habitat to construct multiple scale habitat feature set;Multiple scale habitat feature set is reduced in dimension to screen image habitat features, and the image habitat features after screening are fused with clinical index data through multimodal gate fusion network to generate comprehensive risk feature vector;Comprehensive risk feature vector is input into deep survival prediction model and is optimized to output the prediction result of liver transplantation postoperative complication.The application realizes the dynamic prediction of the probability and time of postoperative complication.

Description

Technical Field

[0001] This invention relates to the field of liver transplantation complication prediction technology, and in particular to a method and system for predicting liver transplantation complications based on habitat characteristics. Background Technology

[0002] Liver transplantation (LT), as the ultimate treatment for end-stage liver disease, is characterized by high morbidity and mortality rates of postoperative complications (such as ischemic biliary tract disease, vascular embolism, and early tumor recurrence). Accurate preoperative prediction of postoperative complication risks is crucial for optimizing donor-recipient matching and developing perioperative management strategies.

[0003] Existing forecasting technologies mainly suffer from two major types of bottlenecks: First, at the feature mining level, traditional radiomics methods are typically based on a "holistic" assumption, extracting global texture features from the entire region of interest. However, malignant tumors or diseased liver tissue exhibit high spatial heterogeneity, often containing different physiological "habitats" such as necrotic areas, hypoxic areas, and high-angiogenic areas. Global feature analysis inevitably smooths out these local microenvironmental differences that have significant prognostic value. Furthermore, existing studies are mostly based on single-scale raw images, neglecting the multi-scale texture information exhibited by lesions at different frequency domains and spatial resolutions, resulting in insufficient feature representation capabilities.

[0004] Secondly, at the model construction level, existing methods mostly model it as a simple binary classification problem (i.e., predicting "whether" complications will occur), ignoring the "time" dimension of complication occurrence. Clinically, the speed of complication onset is more critical for the formulation of follow-up strategies. Although the traditional Cox proportional hazards regression model can handle survival data, its strong assumption of "linear correlation between log hazard and covariates" makes it difficult to fit the complex nonlinear interactions between high-dimensional radiomic features, resulting in limited predictive accuracy.

[0005] Therefore, there is an urgent need for a predictive method that can deeply deconstruct the multi-scale heterogeneity within lesions and effectively model nonlinear time-dependent risks. Summary of the Invention Therefore, the purpose of this invention is to provide a method and system for predicting post-liver transplant complications based on habitat characteristics, which can be used to dynamically predict the probability and timing of post-liver transplant complications.

[0006] This invention provides a method for predicting post-liver transplant complications based on habitat characteristics, comprising: Acquire abdominal multimodal medical imaging data and clinical indicator data of the patient to be predicted, and preprocess the medical imaging data and segment the region of interest; perform multi-scale Laplacian-Gaussian spatial transformation on the segmented region of interest image to generate a multi-scale image sequence; Based on voxel grayscale distribution, unsupervised clustering of regions of interest in each scale image sequence is performed using a Gaussian mixture model to divide multiple image habitats; the image omics features of each image habitat are extracted and the differences and topological features between habitats are calculated to construct a multi-scale habitat feature set; The multi-scale habitat feature set is reduced in dimensionality using a sparse-constrained feature selection algorithm to filter image habitat features. The filtered image habitat features and the clinical indicator data are then combined through a multimodal gating fusion network to generate a comprehensive risk feature vector. The comprehensive risk feature vector is input into a deep survival prediction model based on a risk function, and the model parameters of the deep survival prediction model are optimized by maximizing the partial likelihood function. The optimized deep survival prediction model outputs the prediction results of post-liver transplantation complications, including cumulative risk probability and risk stratification results.

[0007] The aforementioned habitat-based method for predicting post-liver transplant complications introduces multi-scale Laplacian-Gaussian spatial transformation and Gaussian mixture models for clustering, dividing lesions into imaging habitats with different physiological significance. This replaces existing methods that rely heavily on single-scale raw images, improving feature representation and thus predictive efficacy. Secondly, a deep survival prediction model based on a risk function is constructed, specifically employing the Cox proportional hazards model. This model uses multi-layer nonlinear activation functions to fit the risk ratio, overcoming the limitations of the linear assumption. By combining partial likelihood functions and the Breslow estimator, this method not only predicts whether complications will occur but also outputs the survival curve indicating when complications occur with the highest probability, providing quantitative evidence for doctors to develop personalized post-operative follow-up schedules. Furthermore, this method uses a gated fusion network, which adaptively learns the weights of imaging and clinical features based on each patient's specific situation, thereby improving the model's robustness across different patient groups.

[0008] In addition, the method for predicting post-liver transplantation complications based on habitat characteristics according to the present invention may also have the following additional technical features: Furthermore, the steps of performing unsupervised clustering of the regions of interest in each scale of the image sequence based on voxel grayscale distribution and using a Gaussian mixture model to divide the image into multiple image habitats include: A Gaussian mixture model was used to fit the voxel gray distribution, and the model parameters of the Gaussian mixture model were iteratively optimized by the expectation-maximization algorithm until the log-likelihood function converged, so as to delineate the image habitat masks with different physiological and pathological significance. The probability density function of the Gaussian mixture model is defined as follows: ; In the formula, K For habitat quantity; For the first k The mixing coefficient of each habitat, and satisfying ; It follows a Gaussian distribution; For the first k The mean of each habitat; For the first k The variance of each habitat; For the region of interest, the first i Gray value of an individual element.

[0009] Furthermore, habitat differences include the Bach distance between habitats, and the formula for calculating the habitat feature set is: ; In the formula, F total Represents a set of habitat features; Indicates the first s Global image omics features extracted at various scales; Indicates the first k Local image omics features extracted from the local environment; S represents the scale number; K Indicates habitat quantity; ⊕ indicates feature splicing operation; H a and H b These represent two different habitat areas; D B ( H a , H b ) represents the Barthel distance between habitats; where, D B ( H a , H b The formula for calculating ) is: ; In the formula, m This represents the mean of characteristics within the corresponding habitat area; s 2 This represents the variance of characteristics within the corresponding habitat region.

[0010] Furthermore, the formula for calculating the comprehensive risk feature vector is as follows: ; In the formula, Z Represents the comprehensive risk feature vector; X img Indicates the habitat characteristics of the filtered images; X cli This represents clinical indicator data; Represents an image habitat feature encoder; This represents the encoder for clinical indicator data; This represents the gating weight, where the calculation formula is: ; In the formula, Sigmoid It is the Sigmoid activation function. W g This represents the weights of the gated fusion network; b g [;] indicates the bias of the gated fusion network; [;] indicates the tensor splicing operation.

[0011] Furthermore, in the step of inputting the comprehensive risk feature vector into the deep survival prediction model based on the risk function, the deep survival prediction model based on the risk function adopts the Cox proportional hazards model, and the risk function of the Cox proportional hazards model is: ; In the formula, Z This is a comprehensive risk feature vector; h 0( t () is the benchmark risk function; h ( t | Z ) represents the risk rate of an individual with a comprehensive risk characteristic vector at time t; Indicates the risk ratio; i represents the parameters to be learned in the fully connected layer; T represents the transpose operation.

[0012] Furthermore, in the step of optimizing the model parameters of the deep survival prediction model by maximizing the partial likelihood function, the negative log-partial likelihood function is used as the loss function for model training. The formula for the loss function is: ; In the formula, L ( i () represents the loss function; E i Indicates an event indicator; R ( ti ) indicates time t i A group of patients who are constantly at risk; The L2 regularization term is used to prevent overfitting. Z i Indicates the patient i The comprehensive risk feature vector; Z j Indicates the patient j The comprehensive risk feature vector; i represents the parameters to be learned in the fully connected layer; T represents the transpose operation.

[0013] Furthermore, the method also includes: outputting an individualized complication-free survival probability curve based on the prediction result, wherein the survival probability is calculated using the following formula: ; In the formula, S 0( t ) is the baseline survival function; x For specific patients; t For time; S(t| x ) indicates a specific patient x The probability of survival at time t; h 0( u ) indicates the integration time u The baseline risk function at that time; Z This is a comprehensive risk feature vector; Z x Indicates a specific patient x The comprehensive risk feature vector; i The parameters to be learned in the fully connected layer are represented by T; T represents the transpose operation. This is the risk ratio.

[0014] Another aspect of the present invention provides a liver transplant complication prediction system based on habitat characteristics, the system comprising: The acquisition module is used to acquire abdominal multimodal medical image data and clinical indicator data of the patient to be predicted, and to preprocess the medical image data and segment the region of interest; and to perform multi-scale Laplacian-Gaussian spatial transformation on the segmented region of interest image to generate a multi-scale image sequence. The clustering module is used to perform unsupervised clustering of regions of interest in each scale image sequence based on voxel grayscale distribution and using Gaussian mixture model to divide them into multiple image habitats; the image omics features of each image habitat are extracted and the differences and topological features between habitats are calculated to construct a multi-scale habitat feature set; The screening module is used to reduce the dimensionality of the multi-scale habitat feature set by using a sparse-constrained feature selection algorithm to screen image habitat features, and to generate a comprehensive risk feature vector by combining the screened image habitat features with the clinical indicator data through a multimodal gating fusion network. The prediction module is used to input the comprehensive risk feature vector into the deep survival prediction model based on the risk function, and optimize the model parameters of the deep survival prediction model by maximizing the partial likelihood function, so as to output the prediction results of post-liver transplantation complications through the optimized deep survival prediction model. The prediction results include cumulative risk probability and risk stratification results.

[0015] In another aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for predicting post-liver transplant complications based on habitat characteristics.

[0016] In another aspect, the present invention provides a data processing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described method for predicting post-liver transplant complications based on habitat characteristics. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method for predicting post-liver transplant complications based on habitat characteristics in the first embodiment of the present invention; Figure 2 This is a system block diagram of the liver transplant complication prediction system based on habitat characteristics in the second embodiment of the present invention; The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0018] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0020] To facilitate understanding of the present invention, several embodiments are given below. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of the present invention will be more thorough and complete.

[0021] Example 1 Please see Figure 1 The image shows a method for predicting post-liver transplantation complications based on habitat characteristics in the first embodiment of the present invention, the method comprising steps S101 to S106: S101. Obtain abdominal multimodal medical imaging data and clinical indicator data of the patient to be predicted, and preprocess the medical imaging data and segment the region of interest.

[0022] Preoperative multimodal medical images of the abdomen of the patient to be predicted are acquired, specifically including enhanced CT or MRI images (covering the arterial phase, portal venous phase, and delayed phase) and clinical indicator data. Due to differences in equipment and scanning parameters leading to variations in physical scale and grayscale distribution, the images undergo the following preprocessing steps to ensure the robustness and cross-patient comparability of the subsequently extracted multiscale radiomics features: (1) Offset field correction (for MRI data): Because MRI imaging is highly susceptible to the inhomogeneities of radio frequency coils and magnetic fields, resulting in low-frequency intensity shifts (i.e., bias fields), which severely interfere with subsequent habitat clustering and feature extraction, this invention employs the N4ITK (N4 Bias Field Correction in Insight Toolkit) algorithm for correction. Specifically, assuming the observed image consists of the true image, the bias field, and additive noise, the mathematical model corresponding to the N4ITK algorithm is as follows: v ( x )= u ( x ) b ( x )+ n ( x In the formula, v ( x () represents the observed image; u ( x () represents a true image without offset; b ( x ) indicates the low-frequency bias field; n ( x ) represents independent additive noise.

[0023] Since this low-frequency bias field manifests as a multiplicative disturbance (i.e., multiplicative offset) in the above model, to simplify the calculation, it is transformed into an additive model through a logarithmic transformation: log v( x )≈log u ( x )+ log b ( x ).

[0024] The N4ITK algorithm fits the smoothed logarithmically transformed low-frequency bias field log using B-splines. b ( x The image is deblurred through iterative deconvolution, with the optimization objective being to minimize the histogram entropy of the intensity distribution of the corrected image, thereby restoring the true grayscale distribution of the tissue.

[0025] (2) Isotropic space resampling: Because raw medical images acquired clinically often have uneven slice thickness (e.g., Z-axis slice thickness, often with a resolution of 5 mm, while the XY plane resolution is 1 mm), this anisotropy can lead to distortion in the calculation of three-dimensional spatial features. Therefore, this application employs a cubic spline interpolation algorithm to uniformly resample all modalities of the image to an isotropic resolution (e.g., 1 mm × 1 mm × 1 mm) to eliminate spatial geometric distortion and ensure that the three-dimensional texture features possess spatial rotation invariance.

[0026] (3) Z-score grayscale normalization: To eliminate systematic differences in absolute grayscale values ​​between images from different patients, Z-score grayscale normalization was performed on the resampled images. The calculation formula is as follows: ; In the formula, I ( x () represents the original voxel grayscale value; I norm ( x () represents the standardized grayscale value; m ROI and s ROI These represent the mean and standard deviation of the grayscale values ​​in the region of interest (ROI), specifically defined as the liver parenchyma. By strictly limiting the statistical range to the target area, interference from background air or external non-tissue regions on the statistical distribution is effectively avoided.

[0027] S102. Perform multi-scale Laplacian-Gaussian spatial transformation on the segmented region of interest image to generate a multi-scale image sequence.

[0028] In order to capture texture information at different spatial scales, this invention uses the Laplacian of Gaussian (LoG) operator to perform multi-scale convolution on the image.

[0029] First, perform a multi-scale LoG transformation. Specifically: Define a 3D Gaussian smoothing kernel G ( x,y,z ; s s The calculation formula is as follows: ; In the formula, s s represents the scale parameter of the Laplacian-Gaussian spatial transformation; (x,y,z) represents the spatial coordinates of a voxel in a 3D image.

[0030] LoG operator Laplace transform defined by Gaussian kernel: ; In the formula, Represents the LoG operator; G ( x,y,z ; s s ) represents a three-dimensional Gaussian smoothing kernel; Transformed image I σs ( x,y,z )for: ; In the formula, I norm The standardized grayscale value; G ( ﹒ ; s s ) indicates that the scale parameter is s s Gaussian smoothing kernel.

[0031] In this embodiment, the scaling parameter of the Laplace-Gaussian spatial transformation is set. s s ∈{1.0,3.0,5.0}, in mm, corresponding to fine texture, medium texture, and coarse texture, respectively. Fine texture reflects tissue microstructure; coarse texture reflects macroscopic heterogeneity; and medium texture reflects subregional heterogeneity of lesions and local boundary transition characteristics, where subregions are the habitats.

[0032] S103. Based on voxel grayscale distribution and using Gaussian mixture model, unsupervised clustering is performed on the region of interest under each scale image sequence to divide it into multiple image habitats.

[0033] In this embodiment, habitat clustering is performed based on the Gaussian Mixture Model (GMM) and the Expectation-Maximization (EM) algorithm.

[0034] A Gaussian mixture model was used to fit the voxel grayscale distribution, and the model parameters were iteratively optimized using the expectation-maximization algorithm until the log-likelihood function converged, in order to delineate image habitat masks with different physiological and pathological significance. Specifically: For each scale s From the Region of Interest (ROI) image, extract the grayscale value set V of all voxels, specifically: V = { v 1, v 2,…, v i}, where, i This represents the total number of voxels within the region of interest. Assume the ROI is composed of... K The habitat consists of several subregions (i.e., habitats) with different gray-level distributions, whose distributions follow a Gaussian mixture model; where the probability density function of the Gaussian mixture model is defined as: ; In the formula, K For habitat quantity; For the first k The mixing coefficient of each habitat, and satisfying ; It follows a Gaussian distribution; For the first k The mean of each habitat; For the first k The variance of each habitat; For the region of interest, the first i Gray value of an individual element.

[0035] S104. Extract the image omics features of each image habitat and calculate the differences and topological features between habitats to construct a multi-scale habitat feature set.

[0036] Based on the generated habitat mask, two main features are extracted: texture features within the habitat and differences between habitats. Specifically: Second-order statistical features are extracted using the Gray Level Co-occurrence Matrix (GLCM). Define P( m , n |d,ω) represents the gray-level pair (d,ω) at distance d and direction ω. m ,n The joint probability of the occurrence of ) is given by , where n = m +1, then the formulas for calculating energy, contrast, and inverse difference moment in GLCM are as follows: Energy is used to measure texture uniformity, and the formula for calculating energy is: ; Contrast ratio is used to measure local changes in matrix values. The formula for calculating contrast ratio is: ; The inverse difference moment is used to measure the local homogeneity of texture. The formula for calculating the inverse difference moment is: ; In the above calculation formula, f ASM Indicates energy; f Con Indicates contrast; f IDM Indicates the inverse difference moment; m Indicates distance as d grayscale pairs ( m , n The grayscale value of ) n Indicates direction as oh grayscale pairs ( m , n The grayscale value of N; g P represents the total number of grayscale values ​​after quantization; m , n ) represents the corresponding gray-level pair at a specific distance d and direction ω. m , n The joint probability of occurrence.

[0037] In addition, grayscale travel matrix features, such as short-travel advantage and long-travel advantage, are extracted to capture coarse and fine textures.

[0038] Furthermore, habitat difference characteristics include Bach distance between habitats to quantify distributional differences between different habitats. The formula for calculating the habitat feature set is as follows: ; In the formula, F total Represents a set of habitat features; Indicates the first s Global image omics features extracted at each scale, where parameter s corresponds to the scale parameter in step S102. s s ; Indicates the first kLocal image omics features extracted from the local environment; S represents the scale number; K Indicates habitat quantity; ⊕ indicates feature splicing operation; H a and H b These represent two different habitat areas; D B ( H a , H b ) represents the Barthel distance between habitats; where, D B ( H a , H b The formula for calculating ) is: ; In the formula, m This represents the mean of characteristics within the corresponding habitat area; s 2 This represents the variance of characteristics within the corresponding habitat region.

[0039] S105. A sparse-constrained feature selection algorithm is used to reduce the dimensionality of the multi-scale habitat feature set to screen image habitat features. The screened image habitat features and clinical indicator data are then combined through a multimodal gating fusion network to generate a comprehensive risk feature vector.

[0040] As a concrete example, the extracted feature dimensions may reach thousands of dimensions. To prevent overfitting, the LASSO (Least Absolute Shrinkage and Selection Operator) algorithm with L1 regularization is used to sparsely select high-dimensional features. The selected image habitat features and clinical indicator data are then input into a multimodal gated fusion network. This gated fusion network dynamically adjusts the contribution of different modalities by learning gate weights. Specifically: The formula for calculating the comprehensive risk feature vector is: ; In the formula, Z Represents the comprehensive risk feature vector; X img Indicates the habitat characteristics of the filtered images; X cli This represents clinical indicator data; Represents an image habitat feature encoder; This represents the encoder for clinical indicator data; This represents the gating weight, where the calculation formula is: ; In the formula, Sigmoid It is the Sigmoid activation function. W g This represents the weights of the gated fusion network; b g [;] indicates the bias of the gated fusion network; [;] indicates the tensor splicing operation.

[0041] S106. Input the comprehensive risk feature vector into the deep survival prediction model based on the risk function, and optimize the model parameters of the deep survival prediction model by maximizing the partial likelihood function, so as to output the prediction results of post-liver transplantation complications through the optimized deep survival prediction model.

[0042] As a concrete example, in the step of inputting the comprehensive risk feature vector into the risk function-based deep survival prediction model, the risk function-based deep survival prediction model adopts the Cox proportional hazards model, and the risk function of the Cox proportional hazards model is: ; In the formula, Z This is a comprehensive risk feature vector; h 0( t () is the benchmark risk function; h ( t | Z ) represents the risk rate of an individual with a comprehensive risk characteristic vector at time t; Indicates the risk ratio; i represents the parameters to be learned in the fully connected layer; T represents the transpose operation.

[0043] Furthermore, in the step of optimizing the model parameters of the deep survival prediction model by maximizing the partial likelihood function, the negative log-partial likelihood function is used as the loss function for model training; the training objective of the model is to maximize the partial likelihood function; finally, the baseline risk function is calculated by combining the Breslow estimator, and the individualized survival probability curve is output.

[0044] The formula for the loss function is: ; In the formula, L ( i () represents the loss function; E i Indicates an event indicator; R ( t i ) indicates time t i A group of patients who are constantly at risk; The L2 regularization term is used to prevent overfitting.Z i Indicates the patient i The comprehensive risk feature vector; Z j Indicates the patient j The comprehensive risk feature vector; i represents the parameters to be learned in the fully connected layer; T represents the transpose operation.

[0045] Specifically, the prediction results include cumulative risk probability and risk stratification results. In some optional embodiments of this application, the method for predicting post-liver transplantation complications based on habitat characteristics further includes: outputting an individualized complication-free survival probability curve based on the prediction results, wherein the survival probability is calculated using the following formula: ; In the formula, S 0( t ) is the baseline survival function; x For specific patients; t For time; S(t| x ) indicates a specific patient x The probability of survival at time t; h 0( u ) indicates the integration time u The baseline risk function at that time; Z This is a comprehensive risk feature vector; Z x Indicates a specific patient x The comprehensive risk feature vector; i The parameters to be learned in the fully connected layer are represented by T; T represents the transpose operation. This is the risk ratio.

[0046] In summary, the habitat-based complication prediction method for liver transplantation described in the above embodiments of this invention introduces multi-scale Laplacian-Gaussian spatial transformation and Gaussian mixture model for clustering, dividing lesions into imaging habitats with different physiological significance. This replaces the existing approach that relies on single-scale raw images, improving feature representation and thus prediction efficiency. Secondly, a deep survival prediction model based on a risk function is constructed, specifically using the Cox proportional hazards model. This model utilizes multi-layer nonlinear activation functions to fit the risk ratio, overcoming the limitations of the linear assumption. By combining partial likelihood functions and the Breslow estimator, this application can not only predict "whether" complications will occur but also output the survival curve with the highest probability of occurrence, providing quantitative evidence for doctors to develop personalized postoperative follow-up schedules. Furthermore, this application employs a gated fusion network, which can adaptively learn the weights of imaging and clinical features according to each patient's specific situation, thereby improving the robustness of the model across different patient groups.

[0047] Example 2 Please see Figure 2 The image shows a habitat-based liver transplantation complication prediction system according to a second embodiment of the present invention. The system includes: The acquisition module is used to acquire abdominal multimodal medical image data and clinical indicator data of the patient to be predicted, and to preprocess the medical image data and segment the region of interest; and to perform multi-scale Laplacian-Gaussian spatial transformation on the segmented region of interest image to generate a multi-scale image sequence. The clustering module is used to perform unsupervised clustering of regions of interest in each scale image sequence based on voxel grayscale distribution and using Gaussian mixture model to divide them into multiple image habitats; the image omics features of each image habitat are extracted and the differences and topological features between habitats are calculated to construct a multi-scale habitat feature set; The screening module is used to reduce the dimensionality of the multi-scale habitat feature set by using a sparse-constrained feature selection algorithm to screen image habitat features, and to generate a comprehensive risk feature vector by combining the screened image habitat features with the clinical indicator data through a multimodal gating fusion network. The prediction module is used to input the comprehensive risk feature vector into the deep survival prediction model based on the risk function, and optimize the model parameters of the deep survival prediction model by maximizing the partial likelihood function, so as to output the prediction results of post-liver transplantation complications through the optimized deep survival prediction model. The prediction results include cumulative risk probability and risk stratification results.

[0048] In summary, the habitat-based liver transplant complication prediction system in the above embodiments of the present invention, by introducing multi-scale Laplacian-Gaussian spatial transformation and Gaussian mixture model for clustering, divides lesions into imaging habitats with different physiological significance, replacing the existing technology's approach of mostly studying original images at a single scale, thus improving feature representation ability and predictive efficacy. Secondly, it constructs a deep survival prediction model based on a risk function, specifically using the Cox proportional hazards model, and employs multi-layer nonlinear activation functions to fit the risk ratio, breaking through the limitations of the linear assumption. By combining partial likelihood functions and Breslow estimators, this application can not only predict "whether" complications will occur, but also output the survival curve with the highest probability of "when" complications will occur, providing a quantitative basis for doctors to formulate personalized postoperative follow-up schedules. Furthermore, this application uses a gated fusion network, which can adaptively learn the weights of imaging features and clinical features according to the specific situation of each patient, thereby improving the robustness of the model in different patient groups.

[0049] Furthermore, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the methods described above.

[0050] Furthermore, embodiments of the present invention also propose a data processing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the methods described above.

[0051] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0052] More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0053] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0054] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0055] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for predicting post-liver transplant complications based on habitat characteristics, characterized in that, include: Acquire abdominal multimodal medical imaging data and clinical indicator data of the patient to be predicted, and preprocess the medical imaging data and segment the region of interest; Perform multi-scale Laplacian-Gaussian spatial transformation on the segmented region of interest image to generate a multi-scale image sequence; Based on voxel grayscale distribution, unsupervised clustering of regions of interest in each scale image sequence is performed using a Gaussian mixture model to divide multiple image habitats; the image omics features of each image habitat are extracted and the differences and topological features between habitats are calculated to construct a multi-scale habitat feature set; The multi-scale habitat feature set is reduced in dimensionality using a sparse-constrained feature selection algorithm to filter image habitat features. The filtered image habitat features and the clinical indicator data are then combined through a multimodal gating fusion network to generate a comprehensive risk feature vector. The comprehensive risk feature vector is input into a deep survival prediction model based on a risk function, and the model parameters of the deep survival prediction model are optimized by maximizing the partial likelihood function. The optimized deep survival prediction model outputs the prediction results of post-liver transplantation complications, including cumulative risk probability and risk stratification results.

2. The method for predicting post-liver transplantation complications based on habitat characteristics according to claim 1, characterized in that, The steps for dividing multiple image habitats by unsupervised clustering of regions of interest in each scale of image sequence based on voxel grayscale distribution and using Gaussian mixture model include: A Gaussian mixture model was used to fit the voxel gray distribution, and the model parameters of the Gaussian mixture model were iteratively optimized by the expectation-maximization algorithm until the log-likelihood function converged, so as to delineate the image habitat masks with different physiological and pathological significance. The probability density function of the Gaussian mixture model is defined as follows: ； In the formula, K For habitat quantity; For the first k The mixing coefficient of each habitat, and satisfying ; It follows a Gaussian distribution; For the first k The mean of each habitat; For the first k The variance of each habitat; For the region of interest, the first i Gray value of an individual element.

3. The method for predicting post-liver transplantation complications based on habitat characteristics according to claim 1, characterized in that, Habitat differences include Bartholin's distance between habitats. The formula for calculating the habitat feature set is: ； In the formula, F total Represents a set of habitat features; Indicates the first s Global image omics features extracted at various scales; Indicates the first k Local image omics features extracted from the local environment; S represents the scale number; K Indicates habitat quantity; ⊕ indicates feature splicing operation; H a and H b These represent two different habitat areas; D B ( H a , H b () indicates the Barthel distance between habitats; in, D B ( H a , H b The formula for calculating ) is: ； In the formula, μ This represents the mean of characteristics within the corresponding habitat area; σ 2 This represents the variance of characteristics within the corresponding habitat region.

4. The method for predicting post-liver transplantation complications based on habitat characteristics according to claim 1, characterized in that, The formula for calculating the comprehensive risk feature vector is: ； In the formula, Z Represents the comprehensive risk feature vector; i>X img Indicates the habitat characteristics of the filtered images; X cli This represents clinical indicator data; Represents an image habitat feature encoder; This represents the encoder for clinical indicator data; This represents the gating weight, where the calculation formula is: ； In the formula, Sigmoid It is the Sigmoid activation function. W g This represents the weights of the gated fusion network; b g [;] indicates the bias of the gated fusion network; [;] indicates the tensor splicing operation.

5. The method for predicting post-liver transplantation complications based on habitat characteristics according to claim 1, characterized in that, In the step of inputting the comprehensive risk feature vector into the risk function-based deep survival prediction model, the risk function-based deep survival prediction model adopts the Cox proportional hazards model, and the risk function of the Cox proportional hazards model is: ； In the formula, Z This is a comprehensive risk feature vector; i>h0( t () is the benchmark risk function; h ( t | Z ) represents the risk rate of an individual with a comprehensive risk characteristic vector at time t; Indicates the risk ratio; θ represents the parameters to be learned in the fully connected layer; T represents the transpose operation.

6. The method for predicting post-liver transplantation complications based on habitat characteristics according to claim 1, characterized in that, In the step of optimizing the model parameters of the deep survival prediction model by maximizing the partial likelihood function, the negative log-partial likelihood function is used as the loss function for model training. The formula for the loss function is: ； In the formula, L ( θ () represents the loss function; E i Indicates an event indicator; R ( t i ) indicates time t i A group of patients who are constantly at risk; The L2 regularization term is used to prevent overfitting. Z i Indicates the patient i The comprehensive risk feature vector; Z j Indicates the patient j The comprehensive risk feature vector; θ represents the parameters to be learned in the fully connected layer; T represents the transpose operation.

7. The method for predicting post-liver transplantation complications based on habitat characteristics according to claim 6, characterized in that, The method further includes: outputting an individualized complication-free survival probability curve based on the prediction results, wherein the survival probability is calculated using the following formula: ； In the formula, S 0( t ) is the baseline survival function; x For specific patients; t For time; S(t| x ) indicates a specific patient x The probability of survival at time t; h 0( u ) indicates the integration time u The baseline risk function at that time; Z This is a comprehensive risk feature vector; Z x Indicates a specific patient x The comprehensive risk feature vector; θ The parameters to be learned in the fully connected layer are represented by T; T represents the transpose operation. This is the risk ratio.

8. A system for predicting post-liver transplant complications based on habitat characteristics, characterized in that, The system includes: The acquisition module is used to acquire abdominal multimodal medical image data and clinical indicator data of the patient to be predicted, and to preprocess the medical image data and segment the region of interest; and to perform multi-scale Laplacian-Gaussian spatial transformation on the segmented region of interest image to generate a multi-scale image sequence. The clustering module is used to perform unsupervised clustering of regions of interest in each scale image sequence based on voxel grayscale distribution and using Gaussian mixture model to divide them into multiple image habitats; the image omics features of each image habitat are extracted and the differences and topological features between habitats are calculated to construct a multi-scale habitat feature set; The screening module is used to reduce the dimensionality of the multi-scale habitat feature set by using a sparse-constrained feature selection algorithm to screen image habitat features, and to generate a comprehensive risk feature vector by combining the screened image habitat features with the clinical indicator data through a multimodal gating fusion network. The prediction module is used to input the comprehensive risk feature vector into the deep survival prediction model based on the risk function, and optimize the model parameters of the deep survival prediction model by maximizing the partial likelihood function, so as to output the prediction results of post-liver transplantation complications through the optimized deep survival prediction model. The prediction results include cumulative risk probability and risk stratification results.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the habitat-based method for predicting post-liver transplant complications as described in any one of claims 1-7.

10. A data processing apparatus, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method for predicting post-liver transplant complications based on habitat characteristics as described in any one of claims 1-7.

Images