A method for predicting prognosis of acute ischemic stroke based on CTA image

Abstract

The application discloses an acute ischemic stroke prognosis prediction method based on CTA images, relates to the technical field of medical image processing, and fuses a three-dimensional visual transformer and a meta-learning strategy to train a target prognosis prediction model based on the three-dimensional visual transformer and matched with a target data acquisition condition.The target prognosis prediction model can effectively model a spatial dependence relationship in a three-dimensional CTA image across regions and layers, significantly improves the recognition ability of complex cerebral vascular structures and ischemic regions compared with a traditional convolution network which only depends on a local receptive field, and further more, the introduction of the meta-learning strategy can improve the model generalization ability under small samples and multiple data acquisition conditions, so that the target prognosis prediction model directly takes the CTA image to be predicted as input to realize automatic prognosis prediction, does not need manual segmentation or feature extraction, avoids the dependence on manual experience in the traditional method, and improves the standardization degree and repeatability of the model application.

Description

Technical Field

[0001] This application relates to the field of medical image processing technology, and in particular to a method for predicting the prognosis of acute ischemic stroke based on CTA images. Background Technology

[0002] Acute ischemic stroke (AIS) is one of the leading causes of death and disability. Its pathogenesis primarily involves cerebral artery occlusion leading to ischemic necrosis of brain tissue. Computed Tomography Angiography (CTA) is a routine preoperative examination for AIS. This routine preoperative imaging examination can rapidly obtain information on cerebral vascular structure and thrombus-related data, including key imaging features such as thrombus length, density, and permeability. These features have been shown to be closely related to AIS prognosis. Extracting and analyzing cerebral vascular structure and thrombus-related information from CTA images is of great value in predicting AIS prognosis and guiding clinical decision-making. Currently, AIS prognosis prediction based on CTA images mainly includes the following technical approaches: (1) Traditional statistical modeling methods: represented by Logistic regression, which models images by manually extracting image features and clinical variables. This method has a certain degree of interpretability, but it relies on manual feature engineering, is difficult to process high-dimensional and complex image data, and has problems such as strong subjectivity and poor repeatability, making it difficult to achieve standardized application. (2) Machine learning methods: Random forest, support vector machine (SVM), XGBoost and other algorithms can model some nonlinear relationships, but still rely on manually designed image features. They have limited ability to model the spatial structure of high-dimensional image data, and also have problems of strong subjectivity and poor repeatability, making it difficult to achieve standardized application.

[0003] (3) Deep learning method: Features are extracted from CTA images using convolutional neural networks (CNN) and convolutional structures in 3D CNN to achieve AIS prognostic classification prediction. This method can achieve automatic feature extraction and reduce manual intervention. However, CNN-type models are limited by local convolutional structures and rely on local receptive fields, making it difficult to capture long-range dependencies across regions and layers in CTA images. In addition, deep learning models usually require a large amount of labeled data for training, while medical image data is expensive to acquire and difficult to label, leading to the model being prone to overfitting. Summary of the Invention

[0004] This application addresses the aforementioned problems and technical needs by proposing a method for predicting the prognosis of acute ischemic stroke based on CTA imaging. The technical solution of this application is as follows: A method for predicting the prognosis of acute ischemic stroke based on CTA images, comprising: CTA images of patients with acute ischemic stroke (AIS) were obtained and the prognostic results were labeled to obtain sample CTA images with prognostic results labels, thus constructing a sample dataset. Based on the prognostic result labels and the data acquisition conditions of the sample CTA images, multiple subtasks are constructed from the sample dataset. Each subtask includes multiple sample CTA images, and each subtask is divided into a support set and a query set. The sample CTA images are input into a prognostic prediction model based on a 3D vision transformer. The prognostic prediction model adaptively divides the sample CTA images into multiple voxel blocks according to anatomical structure and resolution. After linear projection of each voxel block, a global classification label CLS token is concatenated and a positional encoding is added to obtain an embedding sequence. Multiple Transformer coding layers are used to model the relationship between different voxel blocks based on the embedding sequence using a self-attention mechanism. The vector corresponding to the CLS token output by the last Transformer coding layer is input into the classification head to obtain the predicted prognostic result. Using the predicted prognostic values ​​and labels of sample CTA images in the support set and query set of each subtask, the prognostic prediction model is iteratively trained according to the meta-learning mechanism to obtain the initial model parameters of the prognostic prediction model. Using sample CTA images with prognostic labels obtained according to the target data acquisition conditions, a prognostic prediction model with initial model parameters is trained to obtain a target prognostic prediction model. The CTA images of the patients to be predicted obtained according to the target data acquisition conditions are then input into the target prognostic prediction model to obtain the predicted prognostic values.

[0005] Further technical solutions include: methods for predicting the prognosis of acute ischemic stroke also include: After linear projection of each voxel block, multiple different global classification markers (CLS tokens) are concatenated and positional encoding is added to obtain the embedding sequence. The multiple different global classification markers (CLS tokens) correspond to different anatomical regions. A self-attention mechanism is employed in each Transformer encoding layer, using the attention mask corresponding to each global classification label CLS token to calculate the self-attention weight between the current global classification label CLS token and the vector within the corresponding anatomical region. The vectors corresponding to each global classification token (CLS token) output from the last Transformer encoding layer are fused and then input into the classification head.

[0006] A further technical solution is that this method for predicting the prognosis of acute ischemic stroke also includes: Generate CLS basic markers; encode the clinical information and physiological parameters of the AIS patients to which the sample CTA images belong as auxiliary features, and fuse and map the auxiliary features and CLS basic markers to obtain the global classification marker CLS token.

[0007] A further technical solution is that this method for predicting the prognosis of acute ischemic stroke also includes: After inputting the CTA images of the patient to be predicted into the target prognosis prediction model, the intermediate layer feature map of the target prognosis prediction model is extracted and combined with the weight information in the classification head to calculate the importance of each spatial location in the CTA images to be predicted and generate an activation distribution map. The activation distribution map is then overlaid with the CTA images to be predicted to form a visual heat map, which visually demonstrates the degree of attention the target prognosis prediction model pays to different spatial locations.

[0008] Its further technical solution is to iteratively train the prognostic prediction model according to the meta-learning mechanism, including: Initialize the initial model parameters of the prognostic prediction model; Multiple subtasks are randomly selected in each iteration; For each subtask extracted in the current iteration, the predicted prognostic values ​​and labels of the sample CTA images in the support set of the current subtask are used to perform gradient descent updates based on the current initial model parameters. During gradient descent updates, parameters related to key anatomical structures are given higher update weights to obtain specific temporary parameters for the current subtask. The loss function of the current subtask is calculated based on the predicted prognostic values ​​and labels of the sample CTA images in the query set of the current subtask, using the specific temporary parameters of the current subtask. The loss function includes cross-entropy loss, feature consistency loss, and regularization constraints. The meta-loss is obtained by combining the loss functions of all sub-tasks extracted in the current iteration. The gradient of the meta-loss with respect to the current initial model parameters is calculated, and the initial model parameters of the prognostic prediction model are updated to enter the next iteration until the iteration termination condition is reached.

[0009] The further technical solution involves constructing the sample dataset, which also includes: At least one of the following data augmentation operations is performed on the acquired sample CTA images: spatial transformation, intensity transformation, and perturbation transformation for vascular regions. The perturbation transformation for vascular regions performed on the sample CTA images includes at least one of the following: anisotropic deformation and elastic deformation to simulate vascular anatomical differences, perturbation of vascular enhancement degree and time window shift based on contrast agent dynamics, selective intensity attenuation of vascular regions to simulate vascular occlusion or low perfusion, imaging noise modeling and resolution / reconstruction kernel perturbation combined with Poisson and Gaussian noise.

[0010] A further technical solution involves adaptively dividing the sample CTA images into multiple voxel blocks based on anatomical structure and resolution, including: Adaptive triaxial segmentation is performed based on the spatial resolution of the sample CTA images in the three axes and the anatomical regions they belong to. The higher the spatial resolution of the sample CTA images in each axis, the smaller the segmentation scale of the sample CTA images in the current axis. The segmentation scale of the regions where key anatomical structures are located in the sample CTA images is smaller than the segmentation scale of the background regions.

[0011] The further technical solution is that each subtask constructed from the sample dataset includes multiple sample CTA images with various different prognostic result labels obtained under the same data acquisition conditions. The data acquisition conditions include the scanning equipment used, scanning parameters, and patient population characteristics of the AIS patients in the sample.

[0012] The further technical solution is that the shallow Transformer coding layer in the prognostic prediction model adopts a local window self-attention mechanism, while the deep Transformer coding layer adopts a cross-window or global self-attention mechanism; the prognostic prediction model adopts a voxel block merging strategy and a channel compression strategy, and uses sparse attention or attention masks based on key anatomical structures for attention calculation.

[0013] A further technical solution for obtaining sample CTA images with prognostic labels includes: Acquire initial CTA images acquired under different data acquisition conditions and stored in DICOM format, and annotate the prognostic results of AIS patients in the sample to which each initial CTA image belongs; Data preprocessing is performed on the initial CTA image with prognostic result labels to obtain sample CTA images with prognostic result labels. The data preprocessing includes: reconstructing the grayscale value of the initial CTA image based on DICOM header information and then cropping the image; normalizing the initial CTA image using the Z-score normalization method; resampling the initial CTA image to a predetermined spatial resolution using trilinear interpolation; and aligning the initial CTA image to a predetermined reference template space using a registration method.

[0014] The beneficial technical effects of this application are: This application discloses a method for predicting the prognosis of acute ischemic stroke based on CTA images. This method integrates a 3D visual transformer and a meta-learning strategy to train a target prognosis prediction model based on the 3D visual transformer and matched with the target data acquisition conditions. This target prognosis prediction model can effectively model the spatial dependencies across regions and layers in 3D CTA images. Compared with the traditional convolutional network method that only relies on local receptive fields, it significantly improves the ability to identify complex cerebral vascular structures and ischemic regions. In addition, the introduction of the meta-learning strategy can improve the model's generalization ability under small sample and multiple data acquisition conditions. The obtained target prognosis prediction model can directly use the CTA image to be predicted as input to achieve automated prognosis prediction without manual segmentation or feature extraction. This avoids the dependence on human experience in traditional methods and improves the standardization and reproducibility of model application.

[0015] This method introduces a meta-learning strategy, enabling the prognostic prediction model to learn initial model parameters with good transferability under limited training data. When faced with data from different data acquisition conditions, only minor fine-tuning is needed to adapt to obtain the target prognostic prediction model, significantly reducing the model's dependence on large-scale labeled data and allowing it to adapt to different data distributions. Through the synergistic effect of standardized preprocessing procedures and the meta-learning mechanism, the method effectively mitigates the data distribution bias caused by different scanning devices, scanning parameters, and population characteristics, improving the model's stability in real clinical environments. Maintaining the stability of vascular structure representation during cross-task learning further enhances the model's generalization performance in small CTA samples and cross-domain scenarios.

[0016] The prognostic prediction model, based on the classic 3D vision transformer, has been optimized in terms of voxel segmentation and CLS token construction to suit the characteristics of CTA images. This further enhances the model's ability to represent cerebral vascular structures and lesion areas, thereby improving the accuracy and rationality of the model's predictions. Furthermore, to address the high computational complexity and limited data scale of CTA 3D data, a lightweight and efficient modeling mechanism is introduced to improve the model's training efficiency and inference speed in small-sample CTA scenarios.

[0017] To address the common problem of traditional deep learning models struggling to explain their predictive basis and thus hindering clinical application, this method introduces a visualization analysis strategy. This strategy visually displays the key imaging regions the model focuses on during prediction, enabling physicians to understand the model's decision-making rationale and thereby improving the interpretability and clinical value of the prognostic prediction model. The method has a clear workflow, is easily integrated with existing medical information systems (such as PACS systems), and can achieve automated analysis and prognostic assessment of patient imaging data, making it highly valuable for widespread application. Attached Figure Description

[0018] Figure 1 This is a flowchart of a method for predicting the prognosis of acute ischemic stroke according to an embodiment of this application.

[0019] Figure 2 This is a visualization of a cross-sectional image on which a CTA image to be predicted is compared with a heatmap obtained using a target prognostic prediction model.

[0020] Figure 3 This is a visualization of a cross-sectional image on which a CTA image to be predicted is compared with a heatmap obtained using a target prognostic prediction model. Detailed Implementation

[0021] The specific embodiments of this application will be further described below with reference to the accompanying drawings.

[0022] This application discloses a method for predicting the prognosis of acute ischemic stroke based on CTA images. This method uses a prognostic prediction model based on a 3D visual transformer to analyze CTA images and introduces a meta-learning strategy to enhance the generalization ability of the prognostic prediction model under conditions of small samples and multiple data acquisitions. Please refer to... Figure 1 The flowchart shown illustrates that the method includes the following steps: Step 1: Obtain CTA images of sample AIS patients and label the prognostic results to obtain sample CTA images with prognostic result labels, thus constructing a sample dataset.

[0023] The CTA images obtained in this step typically come from multiple different medical centers, such as... Figure 1 Taking data from two medical centers as an example, during the actual data acquisition phase, initial CTA images stored in DICOM format were obtained from the medical centers' image archiving and communication systems, and the data acquisition conditions for each initial CTA image were determined. Each data acquisition condition consisted of preoperative examination data of the sample AIS patients, and the prognostic results were annotated in conjunction with clinical follow-up results to obtain initial CTA images with prognostic labels.

[0024] Initial CTA images with prognostic labels were preprocessed to obtain sample CTA images with prognostic labels. The preprocessing included: firstly, grayscale reconstruction of the initial CTA images based on DICOM header information, followed by image cropping according to clinically common window width and level settings to preserve key grayscale ranges related to cerebral vessels and lesions. To eliminate intensity distribution differences between different initial CTA images, Z-score normalization was used to normalize the initial CTA images to ensure uniform statistical distribution characteristics. Furthermore, considering the differences in voxel size between CTA images acquired by different scanning devices, trilinear interpolation was used to resample the initial CTA images to a predetermined spatial resolution, achieving a uniform spatial resolution. Additionally, to reduce spatial offset caused by patient positioning and scanning differences, registration methods were used to align the initial CTA images to a predetermined reference template space, thereby achieving spatial consistency. Registration methods used included rigid registration, non-rigid registration, and deep learning registration methods.

[0025] Even when data from different medical centers is combined, the number of directly acquired CTA images remains relatively limited, exhibiting a small sample size. Therefore, to enhance the generalization ability of prognostic prediction models in complex clinical settings, data augmentation operations are performed on the acquired CTA images to expand the sample dataset. These operations include at least one of the following: spatial transformation, intensity transformation, and perturbation transformation targeting vascular regions. The spatial transformations performed include at least one of rotation, translation, scaling, and mirroring.

[0026] The intensity transformation performed includes at least one of window width and window level perturbation, gamma correction, noise superposition, and blurring.

[0027] To address the characteristics of CTA images, the perturbation transformations performed on vascular regions include at least one of the following: anisotropic deformation and elastic deformation to simulate vascular anatomy differences; vascular enhancement perturbation and time window shift based on contrast agent kinetics; selective intensity attenuation of vascular regions to simulate vascular occlusion or low perfusion; and imaging noise modeling and resolution / reconstruction kernel perturbation combining Poisson and Gaussian noise. This more realistically reflects the changes in CTA images under different scanning conditions and pathological states.

[0028] In addition, to address the class imbalance problem in prognostic classification, random oversampling is performed on the CTA images corresponding to prognostic result labels that do not meet the required number in the sample dataset to expand the minority class samples, thereby improving the model's ability to identify patients with poor prognosis.

[0029] Step 2: The sample CTA images in the constructed sample dataset are from different medical centers, and the data acquisition conditions of different medical centers often differ. To address the model generalization problem caused by the limited number of sample CTA images and the differences in data acquisition conditions across multiple centers, this application introduces model meta-learning into a deep learning framework to optimize model parameters.

[0030] Therefore, we first construct multiple sub-tasks from the sample dataset, such as Figure 1 Taking the construction of 10 sub-tasks as an example, the sample CTA images in each sub-task are divided into a support set and a query set. The support set and the query set each contain multiple sample CTA images. The support set is used to quickly update the model parameters in the current task, while the query set is used to evaluate the performance of the updated model.

[0031] To address the characteristics of CTA image data, during the subtask construction phase, subtasks are not only divided based on prognostic outcome labels, but also multi-source task distributions are constructed by combining the data acquisition conditions of the sample CTA images. Each subtask thus constructed includes multiple sample CTA images with various prognostic outcome labels acquired under the same data acquisition conditions, to more realistically simulate the heterogeneity of clinical data. The data acquisition conditions for each CTA image include the scanning equipment used to acquire the CTA image, the set scanning parameters, and the patient population characteristics of the AIS patient to which the CTA image belongs. The set scanning parameters include at least one of contrast agent type and contrast agent dosage, and the patient population characteristics include at least one of patient lesion type, patient age, and patient gender. In another embodiment, the perturbation transformation fusion meta-learning task construction process for vascular regions executed during the data augmentation phase is also used as a perturbation source for different task distributions to improve the model's rapid adaptability and robustness in small-sample CTA scenarios.

[0032] Step 3: During the iterative training of the prognostic prediction model according to the meta-learning mechanism, the sample CTA images are input into the prognostic prediction model based on the 3D vision transformer to obtain the predicted prognostic results.

[0033] Please refer to Figure 1The prognostic prediction model in this application first performs adaptive triaxial segmentation of the sample CTA image based on anatomical structure and resolution to obtain multiple voxel blocks. The three axes include the axial, sagittal, and coronal directions. That is, in this application, the sample CTA image is not simply segmented isotropically and uniformly, but rather, considering the differences in anatomical structure and resolution imbalances of the CTA image in the axial, sagittal, and coronal planes, a structure-aware three-dimensional patch segmentation strategy is adopted. Specifically, adaptive triaxial segmentation is performed based on the spatial resolution of the sample CTA image in each of the three axes and the anatomical region to which it belongs. The higher the spatial resolution of the sample CTA image in each axis, the smaller the segmentation scale of the sample CTA image in the current axis direction; the lower the spatial resolution of the sample CTA image in each axis direction, the larger the segmentation scale of the sample CTA image in the current axis direction, thereby maintaining the effective representation of spatial information. On the other hand, by enhancing vascular response or using priors (such as vascular region probability distribution or attention guidance) to adaptively segment different regions, the segmentation scale of key anatomical structures in the sample CTA images is smaller than that of the background regions. This allows for refined modeling of key anatomical structures and compressed representation of non-critical background regions, ultimately improving the model's ability to represent cerebral vascular structures and lesion regions. The key anatomical structures referred to in this application are cerebral vascular structures and lesion regions.

[0034] Then, after linear projection of each voxel block, the global classification token (CLS token) is concatenated and positional encoding is added to obtain the embedding sequence. For example... Figure 1 In the diagram, the symbol * represents the global classification token (CLS), and 0-9 represent positional encodings. The obtained embedding sequences are sequentially input into multiple Transformer encoding layers. Each Transformer encoding layer uses a self-attention mechanism to encode the embedding sequences, thereby modeling the relationships between different voxel blocks based on the embedding sequences. Then, the vector corresponding to the CLS token output from the last Transformer encoding layer is input into the classification head to obtain the predicted prognostic value.

[0035] The core of the self-attention mechanism is to calculate a weighted sum of a query, key, and value. The embedding sequence includes vectors corresponding to voxel blocks or global classification tokens (CLS tokens), and any two vectors... and Self-attention weights between The calculation formula is:

[0036] in, It is a vector The query vector, It is a vector The key vector, It is a vector The key vector is obtained by calculating the self-attention weights and then summing the relationships between different vectors using a weighted average. Global feature representation As a vector output, the global feature representation The weighted sum of all vectors can be expressed as:

[0037] in, It is a vector The value vector.

[0038] In another embodiment based on this structure, unlike the single CLS token design, the prognostic prediction model linearly projects each voxel block and concatenates multiple different global classification marker CLS tokens, adding positional encoding to obtain an embedding sequence. These multiple different global classification marker CLS tokens correspond to different anatomical regions. Typically, these multiple different global classification marker CLS tokens correspond to different vascular regions and background regions. A self-attention mechanism is employed in each Transformer encoding layer, using the attention mask corresponding to each global classification marker CLS token to calculate the self-attention weight between the current global classification marker CLS token and the vector within the corresponding anatomical region. Then, the vectors corresponding to each global classification marker CLS token output from the last Transformer encoding layer are fused and input into the classification head. This embodiment uses multi-granularity CLS tokens, which can further enhance the prognostic prediction model's global modeling capability for key vascular structures.

[0039] In another embodiment, for each global classification marker (CLS token), after generating the basic CLS marker using conventional methods, it is not directly used as the CLS token. Instead, the clinical information and physiological parameters of the AIS patient to which the sample CTA image belongs are further encoded as auxiliary features. Then, the auxiliary features and the basic CLS marker are fused and mapped to obtain the global classification marker (CLS token). This approach can achieve joint representation of image features and clinical information, further improving the accuracy and rationality of model predictions.

[0040] On the other hand, a dynamic weighting strategy based on attention weights is introduced in the Transformer encoding process, so that the contribution of different voxel blocks to the global classification label CLS token can reflect their importance in prognostic judgment, thereby improving the interpretability of the model.

[0041] Compared to traditional convolutional neural networks, the prognostic prediction model structure used in this application has advantages in modeling long-range dependencies. However, CTA images with three-dimensional data structures present challenges in terms of high computational complexity and limited data size. Therefore, in another embodiment, the shallow Transformer encoding layer of this prognostic prediction model employs a local window self-attention mechanism, while the deep Transformer encoding layer employs a cross-window or global self-attention mechanism. This hierarchical Transformer structure uses local window attention in the shallow layers to reduce computational overhead, and gradually introduces cross-window or global attention in the deep layers to achieve long-range dependency modeling, thus reducing computational overhead while ensuring modeling effectiveness. Furthermore, this prognostic prediction model employs a voxel block merging strategy and a channel compression strategy to reduce feature dimensions while preserving key vascular structure information. Specific implementation methods can refer to existing approaches. In addition, sparse attention or attention masks based on key anatomical structures are used for attention computation, focusing attention computation on blood vessels and suspected lesion areas, thereby reducing redundant computation. Furthermore, the model size can be further compressed through parameter sharing or low-rank decomposition to improve training efficiency and inference speed in small-sample CTA scenarios.

[0042] In practical applications, an improved Transformer model based on an attention mechanism or a fusion structure of convolution and Transformer can also be used to extract features from 3D images.

[0043] Step 4: Using the predicted prognostic values ​​and labels of the sample CTA images in the support set and query set of each subtask, the prognostic prediction model is iteratively trained according to the meta-learning mechanism to obtain the initial model parameters of the prognostic prediction model.

[0044] The process of iteratively training the prognostic prediction model according to the meta-learning mechanism is as follows: (1) Initialize the initial model parameters of the prognostic prediction model. .

[0045] (2) Randomly select multiple subtasks in each iteration.

[0046] (3) For each subtask extracted in the current iteration: Using the predicted prognostic values ​​and prognostic labels of sample CTA images from the support set of the current subtask in the current initial model parameters Based on this, gradient descent is performed to update the parameters, obtaining the specific temporary parameters for the current subtask. .

[0047] Using the predicted prognostic values ​​and prognostic labels of sample CTA images in the query set of the current subtask, and applying specific provisional parameters of the current subtask. Based on this, the loss function for the current subtask is calculated. For classification tasks, this loss function is generally the cross-entropy loss.

[0048] (4) The meta-loss is obtained by combining the loss functions of all sub-tasks extracted in the current iteration. The gradient of the meta-loss with respect to the current initial model parameters is calculated and the initial model parameters of the prognostic prediction model are updated. Proceed to the next iteration until the iteration termination condition is met.

[0049] By iterating this process repeatedly across multiple subtasks, the prognostic prediction model learns a set of initial model parameters with good generalization ability. .

[0050] In the above process, during the inner loop update, the support set of each subtask is used to update the current initial model parameters. When performing gradient descent updates based on the existing parameters, parameters related to key anatomical structures are assigned higher update weights to obtain specific temporary parameters for the current subtask. This aims to improve the ability of prognostic prediction models to quickly adapt to key anatomical structures.

[0051] Furthermore, during the meta-update phase, the query set of the current subtask is used to apply specific temporary parameters of the current subtask. When calculating the loss function for the current subtask, in addition to calculating the cross-entropy loss, feature consistency loss and regularization constraints are also introduced to ensure that the prognostic prediction model maintains the stability of vascular structure representation during cross-task learning, thereby further enhancing its generalization performance in CTA small sample and cross-domain scenarios.

[0052] In addition to the classic Model-Agnostic Meta-Learning (MAML) strategy, this application can also employ transfer learning, domain adaptation, or other meta-learning methods to improve the model's generalization ability under different data distributions.

[0053] After obtaining the initial model parameters of the prognostic prediction model through iterative training in steps 1-4 above, the model application stage can begin. When faced with new data acquisition conditions or CTA image data from different centers, it is only necessary to obtain a small number of sample CTA images with prognostic result labels according to the target data acquisition conditions. The specific method is similar to step 1 and will not be repeated here. Figure 1Taking the acquisition of CTA images with prognostic labels from medical center 3 according to the target data acquisition conditions as an example, these samples are then used to train a prognostic prediction model with initial model parameters. A small number of gradient updates are performed based on the initial model parameters to fine-tune the model parameters. The optimization objective is to minimize the classification loss, thereby obtaining a target prognostic prediction model that is adapted to the current target data acquisition conditions.

[0054] The target prognostic prediction model can then be applied to real-world clinical scenarios with target data acquisition conditions to automatically analyze and predict the prognosis of new patient CTA image data. Specifically, the target CTA images of the patient to be predicted are acquired according to the target data acquisition conditions, and then input into the target prognostic prediction model to obtain the predicted prognostic outcome value for the patient.

[0055] To improve the interpretability of the model in clinical applications, after inputting the CTA images of the patient to be predicted into the target prognostic prediction model, the intermediate layer feature maps of the target prognostic prediction model are extracted and combined with the weight information in the classification head to calculate the importance of each spatial location in the CTA images to be predicted, generating an activation distribution map. Then, the activation distribution map is overlaid with the CTA images to be predicted to form a visual heatmap, visually demonstrating the degree of attention the target prognostic prediction model pays to different spatial locations. In one embodiment, the above visualization method is based on Class Activation Mapping (CAM), which highlights key anatomical regions related to prognostic prediction by weighted fusion of feature responses. Other visualization techniques, such as gradient-based visualization methods or feature importance analysis methods, can also be used to interpret the model's decision-making basis. Through the above visualization operations, the decision-making basis of the target prognostic prediction model can be intuitively presented, enabling clinicians to combine image interpretation results with the output of the target prognostic prediction model to assist in judgment, thereby improving the interpretability and clinical application value of the target prognostic prediction model.

[0056] For example in Figure 2 In this example, the patient's CTA image showed occlusion of the right internal carotid artery C7 segment (as indicated by the arrow), while the right middle cerebral artery trunk and branches were not visualized. The prognostic prediction value output by the target prognostic prediction model indicated a good prognosis, and a visual heatmap was generated. Figure 2 A visualization heatmap of a cross-sectional image is shown. The closer the color is to red, the higher the weight assigned by the model. It can be seen that the high-weight areas in the visualization heatmap are located in the blood supply area of ​​the right middle cerebral artery, mainly concentrated in the infarct ischemic area.

[0057] In another example, the CTA image of the patient to be predicted showed occlusion of the C7 segment of the left internal carotid artery (as indicated by the arrow), and the main trunk and branches of the left middle cerebral artery were not visualized. The prognostic prediction value output by the target prognostic prediction model indicated a poor prognosis and a visual heatmap was output. Figure 3 A visualization heatmap of a cross-sectional image is shown. The closer the color is to red, the higher the weight assigned by the model. It can be seen that the high-weight areas in the visualization heatmap are located in the blood supply area of ​​the left middle cerebral artery, mainly concentrated in the infarct ischemic area.

[0058] Depend on Figure 2 and Figure 3 The examples show that this target prognostic prediction model has certain biological and clinical relevance in distinguishing different prognostic categories.

[0059] The above descriptions are merely preferred embodiments of this application, and this application is not limited to the above embodiments. It is understood that other improvements and variations that can be directly derived or conceived by those skilled in the art without departing from the spirit and concept of this application should be considered to be included within the protection scope of this application.

Claims

1. A method for predicting the prognosis of acute ischemic stroke based on CTA images, characterized in that, The prognostic prediction method for acute ischemic stroke includes: CTA images of patients with acute ischemic stroke (AIS) were obtained and the prognostic results were labeled to obtain sample CTA images with prognostic results labels, thus constructing a sample dataset. Based on the prognostic result labels and the data acquisition conditions of the sample CTA images, multiple subtasks are constructed from the sample dataset. Each subtask includes multiple sample CTA images, and each subtask is divided into a support set and a query set. The sample CTA image is input into the prognostic prediction model based on a 3D vision transformer. The prognostic prediction model performs adaptive triaxial partitioning of the sample CTA image according to anatomical structure and resolution to obtain multiple voxel blocks. After linear projection of each voxel block, a global classification label CLS token is concatenated and a positional encoding is added to obtain an embedding sequence. Multiple Transformer coding layers are used to model the relationship between different voxel blocks based on the embedding sequence using a self-attention mechanism. The vector corresponding to the CLS token output by the last Transformer coding layer is input into the classification head to obtain the predicted prognostic result. Using the predicted prognostic values ​​and prognostic labels of sample CTA images in the support set and query set of each subtask, the prognostic prediction model is iteratively trained according to the meta-learning mechanism to obtain the initial model parameters of the prognostic prediction model. Using sample CTA images with prognostic labels obtained according to the target data acquisition conditions, a prognostic prediction model with initial model parameters is trained to obtain a target prognostic prediction model. The CTA images of the patient to be predicted obtained according to the target data acquisition conditions are then input into the target prognostic prediction model to obtain the predicted prognostic value.

2. The method for predicting the prognosis of acute ischemic stroke according to claim 1, characterized in that, The method for predicting the prognosis of acute ischemic stroke also includes: After linear projection of each voxel block, multiple different global classification markers (CLS tokens) are concatenated and positional encoding is added to obtain the embedding sequence. The multiple different global classification markers (CLS tokens) correspond to different anatomical regions. A self-attention mechanism is employed in each Transformer encoding layer, using the attention mask corresponding to each global classification label CLS token to calculate the self-attention weight between the current global classification label CLS token and the vector within the corresponding anatomical region. The vectors corresponding to each global classification token (CLS token) output from the last Transformer encoding layer are fused and then input into the classification head.

3. The method for predicting the prognosis of acute ischemic stroke according to claim 1, characterized in that, The method for predicting the prognosis of acute ischemic stroke also includes: Generate CLS basic markers; encode the clinical information and physiological parameters of the AIS patients to which the sample CTA images belong as auxiliary features, and fuse and map the auxiliary features and CLS basic markers to obtain the global classification marker CLS token.

4. The method for predicting the prognosis of acute ischemic stroke according to claim 1, characterized in that, The method for predicting the prognosis of acute ischemic stroke also includes: After inputting the CTA images of the patient to be predicted into the target prognosis prediction model, the intermediate layer feature map of the target prognosis prediction model is extracted and combined with the weight information in the classification head to calculate the importance of each spatial location in the CTA images to be predicted and generate an activation distribution map. The activation distribution map is superimposed on the CTA images to be predicted to form a visual heat map to visually demonstrate the degree of attention the target prognosis prediction model pays to different spatial locations.

5. The method for predicting the prognosis of acute ischemic stroke according to claim 1, characterized in that, Iterative training of the prognosis prediction model according to the meta-learning mechanism includes: Initialize the initial model parameters of the prognosis prediction model; Multiple subtasks are randomly selected in each iteration; For each subtask extracted in the current iteration, the predicted prognostic values ​​and labels of the sample CTA images in the support set of the current subtask are used to perform gradient descent updates based on the current initial model parameters. During gradient descent updates, parameters related to key anatomical structures are given higher update weights to obtain specific temporary parameters for the current subtask. The loss function of the current subtask is calculated based on the predicted prognostic values ​​and labels of the sample CTA images in the query set of the current subtask, using the specific temporary parameters of the current subtask. The loss function includes cross-entropy loss, feature consistency loss, and regularization constraints. The meta-loss is obtained by combining the loss functions of all sub-tasks extracted in the current iteration. The gradient of the meta-loss with respect to the current initial model parameters is calculated, and the initial model parameters of the prognostic prediction model are updated to enter the next iteration until the iteration termination condition is reached.

6. The method for predicting the prognosis of acute ischemic stroke according to claim 1, characterized in that, The constructed sample dataset also includes: At least one of the following data augmentation operations is performed on the acquired sample CTA images: spatial transformation, intensity transformation, and perturbation transformation for vascular regions. The perturbation transformation for vascular regions performed on the sample CTA images includes at least one of the following: anisotropic deformation and elastic deformation to simulate vascular anatomical differences, perturbation of vascular enhancement degree and time window shift based on contrast agent dynamics, selective intensity attenuation of vascular regions to simulate vascular occlusion or low perfusion, imaging noise modeling and resolution / reconstruction kernel perturbation combined with Poisson and Gaussian noise.

7. The method for predicting the prognosis of acute ischemic stroke according to claim 1, characterized in that, Based on anatomical structure and resolution, adaptive triaxial segmentation of the sample CTA images yielded multiple voxel blocks, including: Adaptive triaxial segmentation is performed based on the spatial resolution of the sample CTA images in the three axes and the anatomical regions they belong to. The higher the spatial resolution of the sample CTA images in each axis, the smaller the segmentation scale of the sample CTA images in the current axis. The segmentation scale of the regions where key anatomical structures are located in the sample CTA images is smaller than the segmentation scale of the background regions.

8. The method for predicting the prognosis of acute ischemic stroke according to claim 1, characterized in that, Each subtask constructed from the sample dataset consists of multiple sample CTA images with various prognostic labels, acquired under the same data acquisition conditions, including the scanning equipment used, scanning parameters, and patient population characteristics of the AIS patients in the sample.

9. The method for predicting the prognosis of acute ischemic stroke according to claim 1, characterized in that, The prognostic prediction model employs a local window self-attention mechanism in the shallow Transformer coding layer and a cross-window or global self-attention mechanism in the deep Transformer coding layer. The prognostic prediction model uses a voxel block merging strategy and a channel compression strategy, and uses sparse attention or attention masks based on key anatomical structures for attention calculation.

10. The method for predicting the prognosis of acute ischemic stroke according to claim 1, characterized in that, Obtaining CTA images with prognostic labels also includes: Acquire initial CTA images acquired under different data acquisition conditions and stored in DICOM format, and annotate the prognostic results of AIS patients in the sample to which each initial CTA image belongs; Data preprocessing is performed on the initial CTA image with prognostic result labels to obtain sample CTA images with prognostic result labels. The data preprocessing includes: reconstructing the grayscale value of the initial CTA image based on DICOM header information and then cropping the image; normalizing the initial CTA image using the Z-score normalization method; resampling the initial CTA image to a predetermined spatial resolution using trilinear interpolation; and aligning the initial CTA image to a predetermined reference template space using a registration method.

Images