Deep radiomics model construction method for vascular disease risk auxiliary judgment

Abstract

The invention provides a deep radiomics model construction method for vascular disease risk auxiliary judgment. The method comprises the following steps: constructing a double-flow branch parallel processing architecture comprising an image flow branch and a feature flow branch; performing deep learning multi-model screening in the image stream branch, screening a deep learning model with optimal performance, and performing multi-level feature extraction on the input three-dimensional medical image through the optimal deep learning model; a space attention mechanism is introduced into an image stream branch; the feature flow branch performs sparse selection on radiomics features and extracts quantitative features significantly associated with vascular disease risks; splicing the quantitative features of the feature flow and the image flow features after spatial attention enhancement by adopting an adaptive weight distribution strategy, and then carrying out weighted fusion; and a multi-task output layer is constructed, a vascular lesion classification task and a lifetime prediction regression task are executed at the same time, and a weighted combination optimization model of classification loss and regression loss is utilized.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence algorithm-assisted medical diagnosis, and particularly to a method for constructing a deep radiomics model for assisting in the judgment of vascular disease risk. Background Art

[0002] Ischemic stroke (AIS) is one of the main causes with relatively high mortality and disability rates worldwide. According to statistics, the mortality rate of AIS accounts for about 1 / 3 of the global mortality rate, and the incidence has been gradually increasing in recent years. Among them, head and neck atherosclerotic diseases are considered to be one of the common causes of AIS in Asian populations. A large number of studies have shown that the progression of head and neck atherosclerotic plaques will increase the risk of cerebrovascular diseases, imposing a great burden on the lives and psychology of patients. At present, the imaging techniques for detecting head and carotid artery plaques mainly include ultrasound, computed tomography angiography (CTA), high-resolution vessel wall imaging, etc. Conventional carotid ultrasound is sensitive to the identification of carotid artery plaques and the dynamic changes of blood flow, but it cannot observe the relevant conditions of intracranial arteries. Head and neck CTA can effectively identify plaques, including the analysis of the degree of lumen stenosis and calcification components, but the evaluation of non-stenotic plaques or low-density plaques is not accurate enough, and the diagnostic effect on vulnerable plaques is limited. High-resolution magnetic resonance vessel wall imaging (HR-VWI) can display the lumen and vessel wall conditions at the same time, and has great advantages in evaluating the lumen or in the detection and analysis of plaque components and morphology. It has high spatial resolution, good tissue contrast, and is safe and radiation-free. Most previous studies have discussed the advantages of HR-VWI technology in evaluating plaque vulnerability compared with other technologies and its value in judging the causes of stroke, but most of them require experienced doctors to analyze the morphology and components of plaques, which are limited by experience and professionalism; in addition, the conventional morphological and some component characteristics of plaques are difficult to quickly and accurately evaluate the nature of plaques, and it is also difficult to accurately predict the occurrence or recurrence risk of AIS.

[0003] In recent years, with the development and application of artificial intelligence (AI), including machine learning (ML) and deep learning (DL), the application of automatically segmenting plaques through algorithms, analyzing learning rules with existing data, extracting a large number of high-dimensional features, and constructing diagnostic and prediction models to improve the speed and accuracy of the evaluation of plaques and subsequent cerebrovascular events has gradually increased. However, the existing methods for vascular disease risk assessment mainly rely on traditional radiomics or single deep learning models. Traditional radiomics is usually based on manually designed features (such as texture, shape, etc.), but the feature extraction process depends on expert experience, it is difficult to capture the deep spatial features in three-dimensional medical images, and the feature redundancy is high, resulting in insufficient generalization ability of the model. Although the method based on a single deep learning model can automatically extract features, its network architecture is fixed, lacking the screening and optimization of the performance of multiple models, and the prediction accuracy may be affected due to improper model selection. In addition, most of the existing methods adopt single-task learning (such as only classification or regression), and it is difficult to simultaneously meet the dual needs of identifying vascular lesion types and predicting survival period in clinical practice. Summary of the Invention

[0004] In order to overcome the deficiencies in the prior art, efficiently and accurately extract features of key value for disease prediction from a large amount of HR-VWI data, and construct a prediction model with excellent performance, the present invention provides a method for constructing a deep radiomics model for assisting in the judgment of vascular disease risk, which significantly improves the prediction accuracy and generalization ability of the model while retaining the rich information of HR-VWI data.

[0005] The technical solution adopted by the present invention to solve its technical problems is as follows:

[0006] A method for constructing a deep radiomics model for assisting in the judgment of vascular disease risk, the steps include:

[0007] Construct a dual-stream branch parallel processing architecture including an image stream branch and a feature stream branch;

[0008] Perform deep learning multi-model screening within the image stream branch, screen the deep learning model with the best performance, and perform multi-level feature extraction on the input three-dimensional medical image through the best deep learning model;

[0009] Introduce a spatial attention mechanism in the image stream branch to enhance the feature expression of the extracted features;

[0010] The feature stream branch performs sparse selection on the radiomics features of the input three-dimensional medical image, and extracts quantitative features significantly associated with vascular disease risk;

[0011] An adaptive weight allocation strategy is adopted. After splicing the quantization features of the feature stream and the image stream features enhanced by spatial attention, weighted fusion is performed.

[0012] A multi-task output layer is constructed to simultaneously perform the vascular lesion classification task and the survival period prediction regression task, and the weighted combination of the classification loss and the regression loss is used to optimize the model.

[0013] Furthermore, the deep learning multi-model screening is specifically as follows:

[0014] Based on multiple existing network architectures, multiple deep learning models are respectively constructed to extract deep image features from HR-VWI.

[0015] The three-dimensional medical images are input to train each deep learning model. After each deep learning model is trained, the prediction performance of each deep learning model is evaluated.

[0016] According to the evaluation results, the deep learning model with the best performance is selected.

[0017] Furthermore, the image stream branch uses the three-dimensional convolutional neural network 3D CNN with the best performance to perform multi-level feature extraction on the input three-dimensional medical images, and captures spatial hierarchical features through convolutional layers, pooling layers, and residual blocks; the three-dimensional convolution operation process of the image stream branch is specifically as follows:

[0018] Input three-dimensional medical image volume where W, H, and D are the width, height, and depth of the input three-dimensional medical image volume respectively, and C is the number of channels;

[0019] Through the l-th layer of three-dimensional convolutional kernel Perform convolution operation, and the output feature map is calculated as: Z (l) = X (l-1) * K (l) + b (l) ;

[0020] where, k w , k h , k d are the width, height, and depth of the convolutional kernel respectively, C in is the number of input channels, C out is the number of output channels; Z (l) represents the output of the l-th layer, X (l-1) represents the output of the l-1-th layer, * represents the three-dimensional convolution operation, and b (l) represents the bias of the l-th layer;

[0021] Generate the activation output through the parametric rectified linear unit PReLU activation function:

[0022] A (l)= max(0, Z (l) ) + a (l) · min(0, Z (l) )

[0023] Where A (l) represents the activation output of the l-th layer, and a (l) represents the learnable parameter.

[0024] Furthermore, the specific operation process of the spatial attention mechanism is as follows:

[0025] Channel compression: Perform global average pooling on the activation output A (l) of the l-th layer of the image stream branch in the channel dimension to generate a channel-compressed feature map M c :

[0026]

[0027] where (i, j, k) represents the index in the spatial dimension;

[0028] Spatial weight calculation: Perform a 3×3×3 convolution operation on the channel-compressed feature map M c and generate a spatial attention weight W s through the Sigmoid function:

[0029] W s = σ(Conv 3×3×3 (M c ))

[0030] where: σ is the sigmoid function, and Conv 3×3×3 represents the use of a 3×3×3 convolution operation;

[0031] Feature enhancement: Multiply the original activation output A (l) element-wise with the spatial attention weight W s to obtain a feature map enhanced by spatial attention

[0032]

[0033] where ⊙ represents the element-wise product.

[0034] Furthermore, the operation process of the feature stream branch is specifically as follows:

[0035] Based on the LASSO algorithm, perform sparse selection on the radiomics features and calculate the optimized feature selection coefficient

[0036]

[0037] Where: β represents the original feature selection coefficient; N represents the number of samples; y represents the target variable; F represents the feature matrix; is the L2 norm, representing the Euclidean length of the vector; λ represents the regularization parameter; ∥.∥ 1 is the L1 norm, representing sparsity;

[0038] Quantitative features significantly associated with the risk of vascular diseases are screened according to the threshold ∈, and the set of selected feature indices is obtained

[0039]

[0040] The screened features are encoded to generate the encoded low-dimensional feature vector h f :

[0041] h f = ReLU(W f F S + b f )

[0042] Where: W f represents the weight of feature encoding; b f represents the bias of feature encoding; F S is the screened feature matrix, and ReLU is the rectified linear unit activation function.

[0043] Furthermore, after splicing the features of the feature stream and the image stream features enhanced by spatial attention, weighted fusion is performed, and the fusion weight α is calculated by the following formula:

[0044] α = σ(W α [h i ; h f )

[0045] Where: σ is the sigmoid function; W α is the learnable parameter matrix; [h i ; h f represents splicing two feature vectors together, where h f represents the low-dimensional feature vector after encoding of the feature stream, and h i represents the image stream feature vector enhanced by spatial attention;

[0046] The weighted fusion is expressed as:

[0047] h fusion = α · h i + (1 - α) · h f

[0048] Where h fusion is the feature output after weighted fusion.

[0049] Further, for the optimization model using the weighted combination of classification loss and regression loss, the classification loss is defined as:

[0050]

[0051] Where: represents the classification loss; Q represents the total number of categories; y c represents the true label; p c represents the predicted probability;

[0052] The regression loss is defined as:

[0053]

[0054] Where: represents the regression loss; y represents the true value; represents the predicted value; δ represents the threshold;

[0055] The weighted combination of classification loss and regression loss is defined as:

[0056]

[0057] Where: represents the total loss obtained by the weighted combination of classification loss and regression loss; λ 1 、λ 2 、λ 3 are the weight coefficients of the loss, controlling the influence of different losses in the total loss; ∥W∥ 2 is the L2 regularization term.

[0058] The beneficial effects of the present invention include:

[0059] Two-stream branch parallel processing is adopted. Among them, the image stream branch constructs a deep learning model based on multiple network architectures, and screens the optimal model through performance evaluation, which can ensure the robustness and reliability of feature extraction. The selected optimal 3D CNN is combined with a spatial attention mechanism to automatically learn the multi-level spatial features of three-dimensional images, which can enhance the focusing ability on key regions; the feature stream branch screens the quantitative features significantly associated with vascular diseases through the LASSO algorithm and encodes them into low-dimensional features, which can reduce the interference of redundant information. That is, the image stream branch focuses on spatial features and can capture visual information such as morphology and texture in the image; while the feature stream branch focuses on quantitative features and can provide statistical information directly related to disease risk. The two-stream branches extract information from the perspectives of images and quantitative features respectively, ensuring that the model can make full use of different types of data features, and at the same time utilize the spatial information and quantitative features of the images to make up for the limitations of a single branch and improve the comprehensiveness and diversity of feature expression. Introducing a spatial attention mechanism in the image stream branch can dynamically adjust the model's attention to key regions in the image, enhance the recognition ability of lesion regions, and combined with the quantitative features of the feature stream branch, the model can more accurately locate and identify regions related to disease risk, improving the accuracy of prediction. Introducing a weight allocation strategy to dynamically fuse the outputs of the image stream and the feature stream, optimizing the adaptability of feature combinations, and enhancing the expression ability of the model. A classification and regression joint output layer is constructed, and the vascular lesion classification and survival period prediction tasks are synchronously optimized through a weighted loss function, improving the clinical applicability of the model.

[0060] In summary, through two-stream feature extraction, spatial attention enhancement and multi-task collaborative optimization, the present invention significantly improves the accuracy of vascular disease risk assessment. Its adaptive fusion strategy and multi-task design enhance the robustness of the model, can meet the needs of clinical diagnosis and prognostic analysis at the same time, and provide efficient and reliable technical support for precision medicine. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 is a schematic diagram of the overall structure of the deep learning radiomics model of the present invention;

[0062] Figure 2 is the overall flow chart of the method for constructing the deep radiomics model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0063] The technical solutions of the present invention will be described clearly and completely below with reference to the accompanying drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0064] In addition, the technical features involved in different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0065] Example 1: This example provides an efficient and robust deep learning radiomics model for medical image analysis, which realizes accurate lesion classification and prognosis prediction, solves the problems of feature alignment and interaction in multi-modal data fusion, and improves the generalization ability of the model on small-sample medical data.

[0066] The deep learning radiomics model adopts a two-stream architecture, including parallel processing of an image stream (3D CNN) and a feature stream (radiomics). The spatial attention mechanism is used to enhance the feature response in the lesion area, and dynamic feature fusion adaptively weights and fuses multi-modal features. Multi-task learning jointly optimizes classification and regression tasks.

[0067] The specific construction steps are as follows:

[0068] (1) Data preprocessing:

[0069] The three-dimensional medical images are standardized, and the formula is:

[0070]

[0071] where represents the pixel value of the standardized image; x represents the original pixel value; μ represents the mean of the original data; represents the standard deviation of the original data; through standardization, the data is adjusted to a distribution with a mean of 0 and a standard deviation of 1 to improve the stability and convergence speed of model training.

[0072] The pre-computed radiomics features are subjected to Min-Max Scaling to scale the features to a specific range (usually 0 to 1), and the formula is:

[0073]

[0074] where x′ is the normalized feature value; x min is the minimum value in the feature; x max is the maximum value in the feature. Through Min-Max Scaling, the features are kept on the same scale, which helps the model to learn better.

[0075] (2) Construct a two-stream branch parallel processing architecture including an image stream branch and a feature stream branch, and introduce a spatial attention mechanism in the image stream branch; automatically learn the spatial hierarchical features of the input three-dimensional medical images through the image stream branch, and provide quantitative features significantly related to the risk of vascular diseases through the feature stream branch;

[0076] Specifically: perform deep learning multi-model screening within the image stream branch: construct multiple deep learning models based on multiple existing network architectures respectively to extract deep-level image features from HR-VWI; input three-dimensional medical images to train each deep learning model, and after the training of each deep learning model is completed, evaluate the prediction performance of each deep learning model; screen the deep learning model with the best performance according to the evaluation results, and perform multi-level feature extraction on the input three-dimensional medical images through the optimal deep learning model;

[0077] The image stream branch uses a three-dimensional convolutional neural network 3D CNN to perform multi-level feature extraction on the input three-dimensional medical images, and captures spatial hierarchical features through convolutional layers, pooling layers, and residual blocks;

[0078] The image stream branch adopts a hierarchical structure:

[0079]

[0080] The CNN layer uses He normal initialization, and the fully connected layer uses Xavier uniform initialization;

[0081] The specific process of three-dimensional convolution operation in the image stream branch is as follows:

[0082] Input the three-dimensional medical image volume where W, H, and D are the width, height, and depth of the input three-dimensional medical image volume respectively, and C is the number of channels;

[0083] Perform convolution operation through the l-th layer of three-dimensional convolution kernel The output feature map is calculated as: Z (l) = X (l-1) * K (l) + b (l) ;

[0084] where k w , k h , k d are the width, height, and depth of the convolution kernel respectively, C in is the number of input channels, C out is the number of output channels; Z (l) represents the output of the l-th layer, X (l-1) represents the output of the (l - 1)-th layer, * represents three-dimensional convolution operation, and b (l) represents the bias of the l-th layer;

[0085] Generate activation output through the parametric rectified linear unit PReLU activation function:

[0086] A (l) = max(0, Z (l) ) + a (l) · min(0, Z(l) )

[0087] Among them, A (l) represents the activation output of the l-th layer, and a (l) represents the learnable parameter.

[0088] The specific operation process of the spatial attention mechanism is as follows:

[0089] Channel compression: Perform global average pooling on the activation output A (l) of the l-th layer of the image stream branch in the channel dimension to generate a channel-compressed feature map M c :

[0090]

[0091] where (i, j, k) represents the index in the spatial dimension;

[0092] Spatial weight calculation: Perform a 3×3×3 convolution operation on the channel-compressed feature map M c and generate a spatial attention weight W s through the Sigmoid function:

[0093] W s = σ(Conv 3×3×3 (M c ))

[0094] where: σ is the sigmoid function, and Conv 3×3×3 represents the use of a 3×3×3 convolution operation;

[0095] Feature enhancement: Multiply the original activation output A (l) element-wise with the spatial attention weight W s to obtain a feature map enhanced by spatial attention

[0096]

[0097] where ⊙ represents the element-wise product.

[0098] The operation process of the feature stream branch is specifically as follows:

[0099] Based on the LASSO algorithm, perform sparse selection on the radiomics features and calculate the optimized feature selection coefficient

[0100]

[0101] where: β represents the original feature selection coefficient; N represents the number of samples; y represents the target variable; F represents the feature matrix; is the L2 norm, representing the Euclidean length of the vector; λ represents the regularization parameter; ∥.∥ 1 is the L1 norm, representing sparsity;

[0102] Screen the quantitative features significantly associated with the risk of vascular diseases according to the threshold ∈, and obtain the set of selected feature indices

[0103]

[0104] Encode the screened features to generate the encoded low-dimensional feature vector h f :

[0105] h f = ReLU(W f F S + b f )

[0106] where: W f represents the weight of feature encoding; b f represents the bias of feature encoding; F S is the screened feature matrix, and ReLU is the rectified linear unit activation function.

[0107] (3) Adopt an adaptive weight allocation strategy. After splicing the features of the feature stream and the image stream features enhanced by spatial attention, perform weighted fusion;

[0108] The fusion weight α is calculated by the following formula:

[0109] α = σ(W α [h i ; h f )

[0110] where: σ is the sigmoid function; W α is the learnable parameter matrix; [h i ; h f means splicing two feature vectors together, where h f represents the low-dimensional feature vector after encoding of the feature stream, and h i represents the image stream feature vector enhanced by spatial attention;

[0111] The weighted fusion is expressed as:

[0112] h fusion = α · h i + (1 - α) · h f

[0113] where h fusion is the feature output after weighted fusion.

[0114] (4) Construct a multi-task output layer to simultaneously perform the vascular lesion classification task and the survival period prediction regression task, and optimize the model using a weighted combination of the classification loss and the regression loss.

[0115] The classification loss is defined as:

[0116]

[0117] Where: represents the classification loss; Q represents the total number of classes; y c represents the true label; p c represents the predicted probability;

[0118] The regression loss is defined as:

[0119]

[0120] Where: represents the regression loss; y represents the true value; represents the predicted value; δ represents the threshold;

[0121] The weighted combination of the classification loss and the regression loss is defined as:

[0122]

[0123] Where: represents the total loss obtained by the weighted combination of the classification loss and the regression loss; λ 1 、λ 2 、λ 3 are the weight coefficients of the loss, controlling the influence of different losses on the total loss; ∥W∥ 2 is the L2 regularization term.

[0124] (5) Model regularization and optimization:

[0125] Adopt a Dropout layer to randomly mask neurons, and combine L2 weight decay to constrain the model complexity;

[0126] Use an adaptive learning rate optimizer to update the parameters, and prevent overfitting through an early stopping strategy.

[0127] Optimizer configuration:

[0128] The optimizer used is AdamW, an adaptive learning rate optimizer, and weight decay is combined for regularization. The update rule of AdamW is as follows:

[0129]

[0130] Where: θ t+1 represents the updated value of the model parameters in the (t + 1)-th iteration; θt denotes the current parameter values of the model in the \(t\)-th iteration; \(\eta\) is the learning rate, which controls the step size of the model parameter update; is the first moment estimate after bias correction, which reflects the average value of the gradients and is usually bias-corrected after multiple updates; is the first moment estimate after bias correction, which reflects the average value of the squared gradients and is similar to the momentum concept of RMSProp; \(\epsilon\) is a small constant to avoid division by zero; \(\lambda\) is the weight decay coefficient, which is used to perform L2 regularization on the parameters to prevent overfitting.

[0131] Regularization strategy:

[0132] Dropout rate: Set to 0.3, applicable to fully connected layers. During training, 30% of the neurons will be randomly dropped in each round to reduce the model's dependence on the training data.

[0133] Weight decay: Set to 1e-4. An L2 regularization term is added to the loss function to constrain the model's parameters, making them smoother and preventing overfitting.

[0134] Early Stopping: When the loss on the validation set does not decrease for 10 consecutive epochs, terminate the training, which helps to avoid overfitting of the model on the training set.

[0135] Example 2: Performing deep learning multi-model screening within the imaging stream branch as described in Example 1, specifically:

[0136] Deep learning model construction: Based on advanced network architectures such as ResNet3D18, ResNet3D50, ResNext3D18, and ResNext3D50, multiple deep learning models are constructed respectively. These models automatically extract deep image features from HR-VWI data through multi-layer convolution and pooling operations.

[0137] Optimal model selection: After the training of each deep learning model is completed, the ROC curve is used to evaluate the prediction performance of each model, and metrics such as AUC, sensitivity, specificity, and accuracy are calculated to select the best discriminative model for the identification of culprit plaques in carotid artery atherosclerosis. To optimize computing resources and improve training efficiency, in this embodiment, ROI sub-images are uniformly used for model training, and the size of the ROI sub-images is standardized to 64×64×64 before training. In addition, 3D image enhancement techniques such as random flipping and random cropping are applied to the training data. After evaluating the deep learning modeling experiment, the best-performing deep learning model is selected, and multi-level feature extraction is performed on the input three-dimensional medical images through this optimal deep learning model. Three mainstream machine learning algorithms, namely logistic regression (LR), support vector machine (SVM), and random forest (RF), are used to train the model. The ROC curve is used to evaluate the prediction performance of each model, and metrics such as AUC, sensitivity, specificity, PPV, NPV, and accuracy are calculated to select the best discriminative model for the identification of culprit plaques in carotid artery atherosclerosis based on deep learning radiomics.

[0138] Regarding the dataset used in the processes of model training, evaluation, etc. in the above implementation: Patients with carotid artery stenosis > 30% detected by CTA or DSA due to cerebrovascular diseases in the First Clinical Hospital of China Three Gorges University from September 2020 to June 2024 were retrospectively collected, and HR-VWI examinations were performed on these patients within two weeks. This study complied with the Declaration of Helsinki and was approved by the Medical Ethics Committee of Yichang Central People's Hospital, and the informed consent of the subjects was waived. (Approval number: 2023-167-01)

[0139] All patients were divided into a symptomatic group and an asymptomatic group according to whether clinical symptoms occurred within 2 weeks before the MRI examination and / or whether the head MRI showed acute / subacute cerebral infarction. Clinical symptoms included classic transient ischemic attack (TIA) and anterior and posterior circulation ischemic strokes. Typical TIA was defined as abnormal focal neurological deficits lasting less than 24 hours, and complete ischemic stroke was characterized by sudden onset of focal neurological deficits lasting 24 hours. Clinical characteristics were recorded, including gender, age, and risk factors for atherosclerosis (hypertension, diabetes).

[0140] Inclusion criteria: (1) For the symptomatic group, the patient has indeed suffered a cerebrovascular event and there are diffusion-restricted lesions on the DWI sequence; (2) MRA or CTA shows stenosis of the anterior and posterior circulation arteries. Exclusion criteria: (1) Non-atherosclerotic intracranial arterial diseases, including vasculitis, moyamoya disease, and dissection; (2) Chronic ischemic stroke / transient ischemic attack (TIA) (>12 weeks); (3) Suspected cardioembolic stroke; (4) Known coagulation disorders; (5) Clinical MRI contraindications, such as patients with pacemakers, certain metal implants, or severe claustrophobia. To ensure the accuracy and generalization ability of the model, the dataset is divided into a training set (70%), a test set (20%), and a validation set (10%).

[0141] Obviously, the above embodiments are merely examples for clear illustration and not limitations on the implementation manners. For those of ordinary skill in the art, other different forms of changes or variations can be made based on the above description. It is not necessary and impossible to list all the implementation manners here. And the obvious changes or variations derived therefrom are still within the protection scope of the present invention.

Claims

1. A method for constructing a deep radiomics model for assisting in the risk assessment of vascular disease, with characteristic steps include: Construct a dual-stream branch parallel processing architecture including an image stream branch and a feature stream branch; Perform deep learning multi-model screening in the image stream branch to select the best performing deep learning model, and use the best deep learning model to perform multi-level feature extraction on the input three-dimensional medical image; Introducing spatial attention mechanism in the image stream branch to enhance the extracted feature expression; The feature stream branch performs sparse selection on the radiomic features of the input three-dimensional medical images to extract quantitative features that are significantly associated with the risk of vascular diseases. Adopting the adaptive weight allocation strategy, the quantitative features of the feature stream are spliced ​​with the image stream features after spatial attention enhancement, and then weighted fusion is performed; A multi-task output layer was constructed to simultaneously perform the vascular lesion classification task and the survival prediction regression task, and the model was optimized using a weighted combination of classification loss and regression loss.

2. The method for constructing a deep radiomics model for auxiliary judgment of vascular disease risk according to claim 1, characterized in that: The deep learning multi-model screening is specifically as follows: Based on multiple existing network architectures, multiple deep learning models are constructed to extract deep image features from HR-VWI respectively; Input three-dimensional medical images to train each deep learning model, and after the training of each deep learning model is completed, evaluate the prediction performance of each deep learning model; Filter the best performing deep learning model based on the evaluation results.

3. The method for constructing a deep radiomics model for auxiliary judgment of vascular disease risk according to claim 1, characterized in that: The image stream branch uses the best performing 3D convolutional neural network 3D CNN to perform multi-level feature extraction on the input 3D medical image, and captures spatial hierarchical features through convolutional layers, pooling layers and residual blocks; the 3D convolution operation process of the image stream branch is specifically as follows: Input 3D medical image volume Where W, H, and D are the width, height, and depth of the input 3D medical image volume, respectively, and C is the number of channels; Through the lth layer of three-dimensional convolution kernel After convolution operation, the output feature map is calculated as: Z (l) =X (l-1) *K (l) +b (l) ; Among them, k w , k h , k d are the width, height, and depth of the convolution kernel, respectively, and C in is the number of input channels, C out is the number of output channels; Z (l) represents the output of layer l, X (l-1) represents the output of the l-1th layer, * represents the three-dimensional convolution operation, b (l) represents the bias of the lth layer; The activation output is generated by the parameterized rectified linear unit PReLU activation function: AND (l) =max(0,Z (l) )+a (l) min(0,Z (l) ) Among them, A (l) represents the activation output of layer l, a (l) represents a learnable parameter.

4. The method for constructing a deep radiomics model for auxiliary judgment of vascular disease risk according to claim 3, characterized in that: The specific operation process of the spatial attention mechanism is as follows: Channel compression: The activation output A of the lth layer of the image stream branch (l) Perform global average pooling in the channel dimension to generate a channel compression feature map M c : Where (i, j, k) represents the index in the spatial dimension; Spatial weight calculation: feature map M after channel compression c Perform a 3×3×3 convolution operation and generate the spatial attention weight W through the Sigmoid function s : W s =σ(Conv 3×3×3 (M c )) Where: σ is the sigmoid function, Conv 3×3×3 Indicates the use of 3×3×3 convolution operation; Feature enhancement: The original activation output A (l) and the spatial attention weight W s Multiply element by element to get the feature map after spatial attention enhancement where ⊙ denotes the element-wise product.

5. The method for constructing a deep radiomics model for auxiliary judgment of vascular disease risk according to claim 4, characterized in that: The operation process of the feature flow branch is specifically as follows: Sparse selection of radiomics features based on LASSO algorithm and calculation of optimized feature selection coefficients Where: β represents the original feature selection coefficient; N represents the number of samples; y represents the target variable; F represents the feature matrix; is the L2 norm, indicating the Euclidean length of the vector; λ is the regularization parameter; ∥.∥1 is the L1 norm, indicating sparsity; According to the threshold ∈, the quantitative features significantly associated with the risk of vascular disease are screened and the selected feature index set is obtained. Encode the filtered features to generate the encoded low-dimensional feature vector h f : h f =ReLU(W f F S +b f ) Where: W f represents the weight of feature encoding; b f represents the bias of feature encoding; F S is the filtered feature matrix, and ReLU is the rectified linear unit activation function.

6. The method for constructing a deep radiomics model for auxiliary judgment of vascular disease risk according to claim 5, characterized in that: The features of the feature stream are spliced ​​with the features of the image stream after spatial attention enhancement, and then weighted fusion is performed, and the fusion weight α is calculated by the following formula: α=σ(W α [h i ;h f ]) Where: σ is the sigmoid function; W α is the learnable parameter matrix; [h i ;h f ] means concatenating two feature vectors together, where h f represents the low-dimensional feature vector after feature stream encoding, h i Represents the image stream feature vector after spatial attention enhancement; The weighted fusion is expressed as: h fusion =a·h i +(1-a)·h f where h fusion It is the feature output after weighted fusion.

7. The method for constructing a deep radiomics model for auxiliary judgment of vascular disease risk according to any one of claims 2 to 6, characterized in that: The weighted combination optimization model using classification loss and regression loss, where the classification loss is defined as: in: represents the classification loss; Q represents the total number of categories; y c represents the true label; p c represents the predicted probability; The regression loss is defined as: in: represents regression loss; y represents the true value; represents the predicted value; δ represents the threshold; The weighted combination of classification loss and regression loss is defined as: in: It represents the total loss obtained by weighted combination of classification loss and regression loss; λ1, λ2, λ3 are the weight coefficients of loss, which control the influence of different losses on the total loss; ∥W∥2 is the L2 regularization term.

Images