Method and system for predicting revascularization grade from cerebral perfusion parameters

Abstract

The application discloses a brain perfusion parameter prediction blood vessel recanalization grade method and system, and belongs to the technical field of medical image processing. The method comprises the following steps: acquiring a preoperative CTP parameter map and a CTA maximum density projection reconstruction map of an acute large blood vessel occlusive stroke patient and a dynamic scanning sequence of CTP scanning; extracting functional perfusion features from a CTP channel and morphological structure features from a CTA channel through a multi-modal feature fusion network; extracting a collateral compensation vector based on a CTP dynamic sequence through a pre-constructed collateral circulation embedding module; modeling six discrete grades into an ordered category by using an mTICI hierarchical ordered regression prediction head of a cumulative link model; and modulating a prediction output by the collateral compensation vector through an attention gate mechanism, and outputting an mTICI grade prediction probability distribution. The application realizes preoperative prospective prediction of an mTICI grade after mechanical thrombectomy.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method and system for predicting vascular recanalization levels using cerebral perfusion parameters. Background Technology

[0002] Acute large vessel occlusive stroke is a leading cause of death and disability in adults, and mechanical thrombectomy has become the standard treatment for large vessel occlusion. Clinically, the modified Thrombolysis in Cerebral Infarction (mTICI) scale is used to assess the degree of recanalization after mechanical thrombectomy, with mTICI grades 2b to 3 representing successful recanalization and closely associated with good clinical outcomes. However, a key problem exists in current clinical practice: the potential degree of recanalization after mechanical thrombectomy cannot be predicted preoperatively. Clinical assessment of the mTICI grade is only retrospectively performed via digital subtraction angiography after thrombectomy. This makes it difficult to accurately select the most likely beneficiary patient group preoperatively and to rationally allocate limited interventional medical resources.

[0003] Chinese patent application CN113256748A discloses an image processing method and device based on cerebral vascular CT images. This method uses three-channel image data from NCCT, CTA, and CTA+ as input to a deep neural network, outputting segmentation labels for ischemic and infarct areas. This technical solution trains the neural network on CTP perfusion data and uses the scanning results from a conventional CT scanner to obtain imaging tissue window parameters, providing a new approach for assessing tissue status in emergency stroke. However, this technical solution mainly addresses the static assessment of tissue ischemia, and its output is a spatial segmentation map of the ischemic and infarct areas, which is a retrospective description of tissue status. This solution fails to establish a predictive correlation between preoperative imaging features and postoperative vascular recanalization outcomes, and cannot achieve a prospective assessment of mTICI grading after mechanical thrombectomy. Furthermore, this solution uses mTICI only as a screening criterion for successful recanalization rather than a predictive target, ignoring the ordered characteristics between different mTICI grades and the crucial influence of collateral circulation on recanalization prognosis.

[0004] In recent years, research on stroke prognosis prediction based on deep learning has made some progress. Some studies have used CT perfusion imaging to predict the final infarct volume, but these methods all treat mTICI as a known input variable rather than a prediction target. Other studies have explored using CT angiography to predict the first-pass effect of thrombectomy, but they use a simple binary classification method, failing to fully utilize the ordered nature of mTICI grading and neglecting the modulating effect of collateral circulation status on recanalization prognosis. Overall, current technologies lack prospective preoperative prediction methods for mTICI grading, especially lacking a comprehensive prediction framework that combines multi-parameter perfusion imaging with collateral circulation function assessment. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for predicting vascular recanalization levels using cerebral perfusion parameters. This method is based on multi-parameter maps of preoperative CT perfusion imaging and CT angiography data. Through multimodal feature fusion and ordered regression prediction architecture, it achieves a prospective assessment of mTICI grading after mechanical thrombectomy, providing decision support for clinical patient screening and interventional resource allocation.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for predicting vascular recanalization grade using cerebral perfusion parameters includes the following steps:

[0008] S1. Obtain preoperative CTP parameter maps and CTA maximum intensity projection reconstruction maps for patients with acute large vessel occlusive stroke. The CTP parameter maps include CBF parameter maps, CBV parameter maps, MTT parameter maps, and Tmax parameter maps, and obtain the dynamic scanning sequence of CTP scans.

[0009] S2. Input the CTP parameter map and the CTA maximum density projection reconstruction map into a pre-trained multimodal feature fusion network, where the CTP channel extracts functional perfusion features and the CTA channel extracts morphological features. The functional perfusion features and morphological features are then processed by the feature fusion layer to generate a fused feature vector.

[0010] S3. Based on CTP scanning, extract the time-concentration curve from the dynamic scanning sequence, and extract the delayed distribution characteristics of the contrast agent arrival in the affected hemisphere from the dynamic changes of the time-concentration curve through the collateral circulation embedding module, and encode and generate the collateral compensation vector.

[0011] S4. Input the fused feature vector into the mTICI hierarchical ordered regression prediction head, use the cumulative linking model to model the six discrete levels of mTICI 0 to 3 as ordered categories, and calculate the ordered transition probability between adjacent levels.

[0012] S5. By using an attention gating mechanism, the collateral compensation vector is modulated to the output of the ordered regression prediction head, thereby strengthening the influence weight of collateral circulation on recanalization prognosis and outputting the prediction probability distribution of each level of mTICI.

[0013] This invention also provides a system for predicting vascular recanalization levels using cerebral perfusion parameters, comprising:

[0014] The data acquisition module is configured to acquire preoperative CTP parameter maps and CTA maximum intensity projection reconstruction maps of patients with acute large vessel occlusive stroke. The CTP parameter maps include CBF parameter maps, CBV parameter maps, MTT parameter maps, and Tmax parameter maps, and to acquire dynamic scanning sequences of CTP scans.

[0015] The multimodal feature fusion module is configured to extract and fuse CTP parameter maps and CTA maximum density projection reconstruction maps through a multimodal feature fusion network. The CTP channel extraction function injects features, and the CTA channel extracts morphological and structural features to generate a fused feature vector.

[0016] The collateral circulation embedding module is configured to extract time-concentration curves based on the dynamic scanning sequence of the CTP scan, extract the delayed distribution characteristics of contrast agent arrival in the affected hemisphere from the dynamic changes of the time-concentration curves, and encode and generate a collateral compensation vector.

[0017] The ordered regression prediction module is configured to perform ordered regression analysis on the fused feature vector using the mTICI hierarchical ordered regression prediction head, and to model mTICI levels 0 to 3 as ordered categories using a cumulative linking model.

[0018] The gated modulation output module is configured to modulate the output of the ordered regression prediction module by using the side compensation vector through an attention gating mechanism, and output the prediction probability distribution of each level of mTICI.

[0019] The beneficial effects of this invention are as follows:

[0020] First, this invention is the first to achieve prospective prediction of mTICI grading based on preoperative images, breaking through the limitation of existing technologies that can only retrospectively assess the degree of vascular recanalization after surgery, and providing a quantitative decision-making basis for clinical preoperative patient screening and interventional resource allocation.

[0021] Second, this invention uses a five-channel multimodal fusion architecture to organically combine CTP functional perfusion features with CTA morphological and structural features, fully mining the prognostic information contained in preoperative images, and has a more comprehensive feature expression capability compared with single-modal methods.

[0022] Third, this invention innovatively adopts an ordered regression prediction head and a cumulative link model to model the mTICI classification as an ordered category rather than an independent category, making full use of the ordered relationship between the levels. Compared with traditional multi-classification methods, it has more predictive characteristics that are in line with clinical practice.

[0023] Fourth, this invention introduces a collateral circulation embedding module to extract collateral compensation features from the dynamic sequence of time-concentration curves, and modulates the final prediction output through an attention gating mechanism, so that the prediction model can fully reflect the key impact of collateral circulation on recanalization prognosis.

[0024] Fifth, this invention outputs the predicted probability distribution for each level rather than a single classification result, providing richer confidence information for clinical decision-making and making it easier for doctors to assess the reliability and uncertainty of the prediction based on the probability distribution. Attached Figure Description

[0025] Figure 1 This is a flowchart of the method for predicting vascular recanalization level using cerebral perfusion parameters according to the present invention.

[0026] Figure 2 This is a diagram of the architecture of the cerebral perfusion parameter prediction system for vascular recanalization levels according to the present invention. Detailed Implementation

[0027] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of protection of the present invention.

[0028] This invention provides a method for predicting vascular recanalization levels using cerebral perfusion parameters. The core idea of ​​this method is to utilize multi-parameter maps of preoperative CT perfusion imaging and CT angiography data, and through a deep learning-based multimodal feature fusion and ordered regression prediction architecture, to achieve a prospective assessment of the mTICI grade after mechanical thrombectomy. Figure 1 As shown, the method includes the following steps:

[0029] Step S1: Multimodal image data acquisition and preprocessing. This step acquires preoperative CTP parameter maps and CTA maximum intensity projection reconstruction maps for patients with acute large vessel occlusive stroke. Specifically, the CTP parameter maps are derived from the post-processing results of the patient's preoperative CT perfusion scan, including four parameter maps: cerebral blood flow (CBF), cerebral blood volume (CBV), mean transit time (MTT), and time to maximum (Tmax). Simultaneously, the raw data of the dynamic scan sequence during the CTP scan are acquired.

[0030] In one embodiment of the present invention, CTP scanning uses a 320-slice CT scanner for whole-brain perfusion imaging. Scanning parameters are set as follows: tube voltage 80kV, tube current 150mAs, scan slice thickness 5mm, reconstruction slice thickness 0.625mm, and contrast agent iohexol injection solution 350mgI / mL, injection rate 5mL / s, injection dose 50mL. After scanning, four parameter maps (CBF, CBV, MTT, and Tmax) are generated using a dedicated post-processing workstation. In a preferred embodiment of the present invention, the matrix size of the CTP parameter maps is uniformly resampled to 256×256 pixels, and the number of layers is typically 16 to 24 layers depending on the coverage area.

[0031] The CTA maximum intensity projection reconstruction image is derived from the post-processing results of the patient's preoperative CT angiography scan. Preferably, the CTA scan uses the same CT equipment as CTP, with scanning parameters set to tube voltage of 120kV, tube current of 280mAs, and a slice thickness of 0.625mm for reconstruction. Contrast agent injection is triggered by contrast agent tracking technology. After the scan, the axial thin-slice image is reconstructed in three dimensions using the Maximum Intensity Projection (MIP) algorithm to generate coronal, sagittal, and axial MIP images. In a preferred embodiment of the present invention, the CTA-MIP image is also resampled to 256×256 pixels to match the size of the CTP parameter map.

[0032] In the multimodal image data preprocessing stage, this invention performs spatial registration and intensity normalization on CTP parameter maps and CTA-MIP images. Spatial registration employs a rigid body registration algorithm to register the CTA-MIP image to the spatial coordinate system of the CTP parameter map, with registration accuracy controlled within 1 mm. For intensity normalization, the CBF parameter map is normalized using the median of the healthy side of the whole brain, with values ​​standardized to between 0 and 2.5; the CBV parameter map is normalized to between 0 and 1.5 using a similar method; the MTT parameter map is normalized to the range of 0 to 12 s; the Tmax parameter map is normalized to the range of 0 to 20 s; and the CTA-MIP image is normalized to the range of 0 to 1 after adjusting the window width and window level.

[0033] After preprocessing, four CTP parameter maps and one CTA-MIP image are stacked into a five-channel three-dimensional tensor with dimensions of 5×256×256×N, where N is the number of layers. In one embodiment of the present invention, N is set to 20 layers, covering the main brain parenchyma regions from the basal ganglia to the centrum semioval. This five-channel tensor serves as the input data for the subsequent multimodal feature fusion network.

[0034] Step S2: Multimodal feature fusion and extraction. In this step, the five-channel multimodal image data is input into a pre-trained multimodal feature fusion network. Functional perfusion features are extracted from the CTP channel and morphological structural features are extracted from the CTA channel. A fused feature vector is generated through the feature fusion layer.

[0035] The multimodal feature fusion network employs a dual-branch encoder architecture. The CTP branch encoder takes a four-channel CTP parameter map (CBF, CBV, MTT, Tmax) as input and uses a three-dimensional residual convolutional neural network for feature extraction. In a preferred embodiment of the invention, the CTP encoder comprises four residual blocks, each containing two 3×3×3 convolutional layers with channel numbers of 64, 128, 256, and 512 respectively. Each residual block is followed by a 2×2×2 max-pooling layer for downsampling. The output of the CTP encoder is a 512-dimensional function-injected feature vector.

[0036] The functional perfusion features extracted by the CTP branch encoder include, but are not limited to, the following categories: the first category is volumetric features, including the ischemic core volume and penumbra volume and their ratio; the second category is asymmetric features, including perfusion parameter differences between the affected and healthy sides; and the third category is temporal features, including delay parameters reflecting the collateral circulation status.

[0037] Specifically, the volume of the ischemic core The volume of the brain region defined as having a brain fat flow (CBF) less than 30% of the median CBF of the healthy side is calculated using the following formula: ,in: Let be the volume of the i-th voxel, in mL; The cerebral blood flow value of the i-th voxel is expressed in mL / 100g / min. The median CBF value of the healthy hemisphere; This is an indicator function; it takes a value of 1 when the condition is met, and 0 otherwise. In embodiments of the present invention, the typical range of ischemic core volume is 0 to 150 mL. A value exceeding 70 mL usually indicates a poor prognosis.

[0038] Half-dark band volume Defined as the volume of the region where Tmax is greater than 6s but CBF is greater than 30% of the healthy side median, the calculation formula is as follows: ,in: The peak time value for the i-th voxel is given in seconds. The penumbra volume reflects the extent of tissue with perfusion abnormalities but without irreversible damage, and is a key indicator for assessing the potential benefit of thrombectomy.

[0039] Core-half-shadow band volume ratio The calculation is as follows: ,in: This is a dimensionless ratio, ranging from 0 to 1. A smaller value (e.g., less than 0.4) indicates the presence of a larger salvageable tissue area and a higher potential for thrombectomy benefit.

[0040] Infusion asymmetry index The formula reflecting the degree of perfusion difference between the affected and healthy sides is as follows: ,in: The average perfusion parameter value for the affected hemisphere; This represents the average perfusion parameter value for the healthy hemisphere. The perfusion parameter P can be either CBF or CBV. The value ranges from -1 to 1, with positive values ​​indicating reduced perfusion on the affected side and larger values ​​indicating more pronounced asymmetry. In one embodiment of the invention, when calculated based on CBF... A value greater than 0.3 indicates significant perfusion asymmetry.

[0041] The CTA branch encoder takes a single-channel CTA-MIP image as input and also employs a three-dimensional residual convolutional network architecture. In one embodiment of the present invention, the network structure of the CTA encoder is similar to that of the CTP encoder, but the number of input channels is 1, and the output is a 512-dimensional morphological feature vector.

[0042] The morphological and structural features extracted by the CTA branch encoder include, but are not limited to, the following categories: the first category is occlusion features, including the location of the occlusion site and the assessment of the degree of occlusion; the second category is thrombosis features, including thrombus load length and thrombus density; and the third category is vascular morphology features, including proximal vascular tortuosity and diameter variation.

[0043] In a preferred embodiment of the present invention, the thrombus burden length The thrombus burden length is calculated by measuring the length of the occluded segment along the vessel centerline, in mm, with a typical range of 5 to 30 mm. Thrombus burden length is a significant factor influencing the difficulty of thrombectomy and the success rate of recanalization. When the thrombus exceeds 15mm, the difficulty of thrombectomy increases significantly. Proximal vessel tortuosity. It is obtained by calculating the ratio of the length of the blood vessel's course to the straight-line distance; it is dimensionless, and its typical value ranges from 1.0 to 2.5. A value greater than 1.8 indicates significant vascular tortuosity, which may increase the difficulty of delivering interventional devices.

[0044] The feature fusion layer fuses the functionally infused feature vector of the CTP branch and the morphological structural feature vector of the CTA branch. In one embodiment of the invention, feature fusion is implemented using a concatenated fully connected approach. Specifically, two 512-dimensional feature vectors are concatenated into a 1024-dimensional vector, which is then mapped to a 256-dimensional fused feature vector through a two-layer fully connected network (with a hidden layer dimension of 512). The fully connected layers employ ReLU activation and Dropout regularization (dropout rate 0.3).

[0045] In another embodiment of the invention, feature fusion employs an attention-weighted fusion mechanism, which calculates the mutual attention weights of two feature vectors through a cross-modal attention module to achieve adaptive feature fusion. Compared to simple concatenation, this method can better capture complementary information between modalities.

[0046] Step S3: Collateral circulation function embedding. This step extracts the time-concentration curve based on the dynamic scanning sequence of the CTP scan, and extracts the delayed distribution characteristics of the contrast agent arrival in the affected hemisphere from the dynamic changes of the time-concentration curve through the collateral circulation embedding module 3, and encodes and generates a collateral compensation vector.

[0047] CTP scans consist of dynamic scan sequences containing multiple temporal phases of raw CT images, typically with a scan duration of 45 to 60 seconds and a temporal resolution of 1 to 2 seconds, comprising 25 to 45 temporal phases. Time-density curves can be extracted from each region of interest (ROI). The time-concentration curve reflects the concentration change of the contrast agent in the region over time. The basic shape of the time-concentration curve includes the baseline segment, the rising segment, the peak segment, and the falling segment, and its morphological characteristics are closely related to the local cerebral blood flow perfusion status.

[0048] The core idea of ​​the collateral circulation embedded module 3 is to quantify the compensatory capacity of collateral circulation by analyzing the arrival delay distribution of time-concentration curves of different regions in the affected hemisphere relative to the corresponding regions in the healthy hemisphere. In patients with good collateral circulation, although the contrast agent cannot directly reach the ischemic area through the occluded vessel, it can be supplied compensatorily through collateral pathways such as leptomeningeal collaterals or the Circle of Willis. This is manifested as a relatively delayed arrival time of the contrast agent in the affected region, but still a certain degree of enhancement can be achieved.

[0049] In one embodiment of the present invention, the collateral compensation delay time Defined as the difference between the peak arrival time of contrast agent in the ischemic region of the affected hemisphere and the peak arrival time in the corresponding region of the healthy hemisphere, the formula is as follows: ,in: The peak arrival time of the time-concentration curve for the ischemic region on the affected side is in seconds. The peak arrival time of the time-concentration curve for the mirror region of the healthy side is given in seconds. This represents the delay time for collateral compensation, measured in seconds (s), with a typical range of 0 to 15 seconds. A shorter duration (e.g., less than 4 seconds) indicates good collateral circulation compensation, allowing the contrast agent to reach the ischemic area quickly via collateral pathways; when... A larger value (e.g., greater than 8s) indicates poor collateral circulation compensation.

[0050] Furthermore, this invention divides the affected hemisphere into multiple anatomical regions and calculates the contrast agent arrival delay for each region, forming a delay distribution vector. In a preferred embodiment of this invention, the affected hemisphere is divided into 10 regions based on the 10-region template of the Alberta Stroke Project Early CT Score (ASPECTS), including the caudate nucleus (C), lentiform nucleus (L), internal capsule (IC), insula (I), and the M1-M6 regions of the middle cerebral artery, totaling 10 regions. Time-concentration curves are extracted for each region, and the arrival delay relative to the healthy side is calculated, forming a 10-dimensional delay distribution vector.

[0051] The collateral circulation embedding module 3 uses a one-dimensional convolutional neural network to encode the features of the time-concentration curve sequence. In one embodiment of the invention, the network input is the time-concentration curves (dimension 10×T, where T is the number of time phases) of 10 regions on the affected side. Through three layers of one-dimensional convolution (kernel sizes of 5, 3, and 3, and channel numbers of 32, 64, and 128 respectively) and global average pooling, a 128-dimensional collateral compensation vector is output. This vector encodes the compensatory characteristics of collateral circulation in different regions of the affected hemisphere, including information such as delayed arrival of contrast agent, attenuation of peak concentration, and abnormal curve morphology.

[0052] Preferably, the side-branch recurrent embedding module 3 also introduces a temporal attention mechanism, assigning different attention weights to different phases of the time-concentration curve, focusing on information-rich phases such as the rising and peak segments. Through learnable temporal attention weights, the network can adaptively focus on the most discriminative temporal features.

[0053] Step S4: mTICI hierarchical ordered regression prediction. In this step, the fused feature vectors are input into the mTICI hierarchical ordered regression prediction head. The cumulative linking model is used to model the six discrete levels of mTICI from level 0 to level 3 as ordered categories, and the ordered transition probability between adjacent levels is calculated.

[0054] The mTICI grading system is a standard scale used clinically to assess the degree of recanalization after mechanical thrombectomy. It includes the following six grades: mTICI 0 (no perfusion), mTICI 1 (perfusion of distal perforators only), mTICI 2a (partial perfusion, less than 50%), mTICI 2b (mostly perfusion, greater than 50%), mTICI 2c (almost complete perfusion), and mTICI 3 (complete perfusion). These six grades have a clear order of importance: 0 < 1 < 2a < 2b < 2c < 3, and a higher grade indicates a better degree of recanalization.

[0055] Traditional multi-classification methods treat the six levels as independent categories, ignoring the ordered relationships between them. This invention innovatively employs an ordered regression prediction head, leveraging the ordered nature of the levels through a cumulative linking model to improve the rationality and accuracy of predictions.

[0056] The basic idea of ​​the cumulative linking model is to transform the ordered classification problem into a series of binary classification problems, where each binary classification problem determines whether the response variable is less than or equal to a certain level. Let... Let be an mTICI-level random variable, taking values ​​of Corresponding numerical code Cumulative probability Let mTICI represent the probability that the level is less than or equal to k, satisfying the following constraints: The cumulative linking model establishes the following relationship between the cumulative probability and the fused feature vector through a logistic linking function: ,in: For logical functions, ; Let be the threshold parameter for the k-th level, which is a learnable parameter and satisfies ... Orderly constraints; For the regression coefficient vector, the dimension and the fusion feature vector same; This is the 256-dimensional fused feature vector output from step S2.

[0057] Transforming the above equation yields an explicit expression for the cumulative probability: ,in: For the Sigmoid function, This maps the linear output to a probability range of 0 to 1.

[0058] The predicted probability for each level can be calculated by the difference of the cumulative probabilities: The agreement In one embodiment of the invention, the network implementation of the ordered regression prediction head includes a linear layer (256-dimensional input, 1-dimensional output) for computation. and 5 learnable threshold parameters. To ensure that the threshold parameter satisfies the ordering constraint, cumulative softplus parameterization is used: ,in These are unconstrained parameters.

[0059] The ordered regression predictor head is trained using a negative log-likelihood loss function: ,in: This represents the number of training samples; The true mTICI rank of the i-th sample; Let be the fused feature vector of the i-th sample. This loss function encourages the model to predict a high probability consistent with the true level.

[0060] Step S5: Attention Gating Modulation and Probability Distribution Output. This step modulates the output of the ordered regression prediction head by the side compensation vector through the attention gating mechanism, strengthens the influence weight of side circulation on recanalization prognosis, and outputs the prediction probability distribution of each level of mTICI.

[0061] Collateral circulation status is a key factor affecting the effectiveness of mechanical thrombectomy and patient prognosis. Good collateral circulation can maintain the survival of ischemic tissue during main vessel occlusion and promote perfusion recovery after recanalization. Therefore, this invention introduces an attention gating mechanism to dynamically modulate mTICI grading prediction based on collateral circulation characteristics.

[0062] The core idea of ​​attention gating is to generate a scalar gating weight based on the side support compensation vector, which adaptively scales the linear output of ordered regression. Specifically, the gating weight... The calculation is as follows: ,in: This is the 128-dimensional lateral compensation vector output in step S3; It is a 128-dimensional learnable weight vector; It is a learnable bias scalar; For the Sigmoid function, ensure .

[0063] Gating weights The physical meaning is the modulation strength of the side circulation on the prediction result. When the side circulation characteristics indicate good compensation, The trend towards higher values ​​strengthens the prediction's tendency to shift towards higher mTICI levels; when collateral circulation characteristics indicate poor compensation, The values ​​tend to be lower, making the predictions more conservative.

[0064] By incorporating gating weights into the cumulative link model, the modulated cumulative probability is calculated as follows: ,in: The regression coefficient vector is 128-dimensional, and the influence of the lateral compensation vector on the cumulative probability is directly introduced. right Scaling and The direct contribution of the side-branch loop features can influence the final prediction through two pathways.

[0065] The final output of the mTICI prediction probability distribution for each level is a six-dimensional vector: Each component corresponds to a prediction probability of mTICI levels 0, 1, 2a, 2b, 2c, and 3, respectively, satisfying the following conditions: .

[0066] In one embodiment of the present invention, the output of the prediction result includes, in addition to the probability distribution vector, the following derived indices:

[0067] (1) Maximum probability prediction level , as a single classification prediction result;

[0068] (2) Probability of successful re-pass This corresponds to the cumulative probability of mTICI level 2b-3;

[0069] (3) For predicting uncertainty indicators, the entropy of the probability distribution is used. In quantification, a higher entropy value indicates greater uncertainty.

[0070] The method of this invention demonstrated good predictive performance in a multicenter retrospective validation. Using 200 patients with acute large vessel occlusive stroke as the validation set, the concordance rate between the predicted mTICI and the actual post-thrombectomy DSA assessment of mTICI reached 82%. The predictive sensitivity for successful recanalization of mTICI grades 2b to 3 reached 89%, the specificity reached 76%, and the area under the ROC curve (AUC) reached 0.87. These results indicate that the method of this invention can provide reliable preoperative mTICI grading prediction for clinical practice.

[0071] like Figure 2 As shown, the present invention also provides a system for predicting vascular recanalization levels using cerebral perfusion parameters. This system is used to execute the method for predicting vascular recanalization levels using cerebral perfusion parameters in the above-described method embodiments. The system includes a data acquisition module 1, a multimodal feature fusion module 2, a collateral circulation embedding module 3, an ordered regression prediction module 4, and a gated modulation output module 5.

[0072] Data acquisition module 1 is configured to acquire preoperative CTP parameter maps and CTA maximum intensity projection reconstruction maps for patients with acute large vessel occlusive stroke. The CTP parameter maps include CBF parameter maps, CBV parameter maps, MTT parameter maps, and Tmax parameter maps, derived from the post-processing results of CT perfusion scans. The CTA maximum intensity projection reconstruction maps are derived from the post-processing results of CT angiography scans. Data acquisition module 1 also includes a preprocessing submodule, which performs preprocessing operations such as spatial registration and intensity normalization. As described in the method embodiment, the preprocessed multimodal image data is stacked into a five-channel three-dimensional tensor as input for subsequent processing.

[0073] The multimodal feature fusion module 2 is configured to extract and fuse features from the CTP parameter map and the CTA maximum density projection reconstruction map through a multimodal feature fusion network. This module comprises two branches: a CTP encoder and a CTA encoder. The CTP encoder extracts functional perfusion features from the four-channel CTP parameter map, including ischemic core volume, penumbra volume, core-penumbra volume ratio, and perfusion asymmetry index. The CTA encoder extracts morphological features from the CTA-MIP image, including occlusion site, thrombus load length, and proximal vessel tortuosity. The features from the two branches are merged through a feature fusion layer to generate a fused feature vector. As described in the method embodiment, the fused feature vector has a dimension of 256.

[0074] The collateral circulation embedding module 3 is configured to extract time-concentration curves based on dynamic scanning sequences from CTP scans. It extracts the distribution characteristics of the contrast agent arrival delay in the affected hemisphere from the dynamic changes of the time-concentration curves and encodes them to generate collateral compensation vectors. This module uses a one-dimensional convolutional neural network to encode the features of the time-concentration curves of multiple anatomical regions on the affected side, outputting collateral compensation vectors. As described in the method embodiment, the collateral compensation vectors have a dimension of 128, encoding the compensatory capacity characteristics of collateral circulation.

[0075] The ordered regression prediction module 4 is configured to perform ordered regression analysis on the fused feature vectors using the mTICI hierarchical ordered regression prediction head. This module employs a cumulative linking model to model mTICI levels 0 to 3 as ordered categories, including a linear layer and a threshold parameter. The linear layer calculates a linear combination of the fused feature vectors, and the threshold parameter defines the decision boundary between adjacent levels. As described in the method embodiment, the threshold parameter is parameterized using cumulative softplus to ensure ordered constraints.

[0076] The gated modulation output module 5 is configured to modulate the output of the ordered regression prediction module using an attention gating mechanism, outputting the predicted probability distributions for each level of mTICI. This module calculates the gating weights based on the side-support compensation vectors, adaptively scaling the linear output of the ordered regression to allow the side-support cyclic features to dynamically influence the prediction results. The final output is a six-dimensional probability distribution vector, corresponding to the predicted probabilities of the six levels of mTICI.

[0077] In one embodiment of the present invention, the system further includes a model training module configured to perform end-to-end joint training of the multimodal feature fusion network, the side-branch recurrent embedding module, the ordered regression prediction head, and the attention gating mechanism using labeled data. Training employs a negative log-likelihood loss function, the optimizer uses the Adam algorithm, the initial learning rate is set to 0.001, the training epochs are 100, and the batch size is 16.

[0078] In another embodiment of the present invention, the system further includes a result visualization module, configured to display the predicted probability distribution in the form of a chart, and to overlay key areas such as the ischemic core and penumbra on the original image, providing clinicians with intuitive decision support information.

[0079] This invention presents a method and system for predicting vascular recanalization levels based on cerebral perfusion parameters. Through the synergistic effects of multimodal feature fusion, collateral circulation function embedding, ordered regression prediction, and attention gating modulation, it achieves prospective prediction of mTICI grading after mechanical thrombectomy. Compared to existing technologies, this invention fully utilizes prognostic information contained in preoperative CTP and CTA images, introduces an ordered regression model to capture the ordered characteristics of mTICI grading, and enhances the predictive ability for recanalization outcomes through collateral circulation embedding. This invention provides a new quantitative tool for clinical decision-making in acute large vessel occlusive stroke, helping to accurately screen the patient population most likely to benefit preoperatively and optimize the allocation of interventional medical resources.

[0080] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for predicting vascular recanalization level using cerebral perfusion parameters, characterized in that, Includes the following steps: S1. Obtain preoperative CTP parameter maps and CTA maximum intensity projection reconstruction maps for patients with acute large vessel occlusive stroke. The CTP parameter maps include CBF parameter maps, CBV parameter maps, MTT parameter maps, and Tmax parameter maps, and obtain dynamic scanning sequences of CTP scans. S2. Input the CTP parameter map and the CTA maximum density projection reconstruction map into a pre-trained multimodal feature fusion network, wherein the CTP channel extracts functional perfusion features and the CTA channel extracts morphological structural features, and the functional perfusion features and the morphological structural features are processed by a feature fusion layer to generate a fused feature vector. S3. Based on the dynamic scanning sequence of the CTP scan, extract the time-concentration curve, and extract the delayed distribution characteristics of the contrast agent arrival in the affected hemisphere from the dynamic changes of the time-concentration curve through a pre-constructed collateral circulation embedding module, and encode and generate a collateral compensation vector. S4. Input the fused feature vector into the mTICI hierarchical ordered regression prediction head, use the cumulative linking model to model the six discrete levels of mTICI 0 to 3 as ordered categories, and calculate the ordered transition probability between adjacent levels. S5. The side compensation vector is modulated to the output of the ordered regression prediction head through the attention gating mechanism to strengthen the influence weight of the side circulation on the recanalization prognosis and output the prediction probability distribution of each level of mTICI.

2. The method for predicting vascular recanalization level using cerebral perfusion parameters according to claim 1, characterized in that, In step S2, the functional perfusion characteristics include ischemic core volume, penumbra volume, core-penumbra volume ratio, and perfusion asymmetry index; the morphological and structural characteristics include occlusion site, thrombus load length, and proximal vessel tortuosity.

3. The method for predicting vascular recanalization level using cerebral perfusion parameters according to claim 1, characterized in that, In step S3, the collateral circulation embedding module divides the affected hemisphere into multiple anatomical regions, extracts the time-concentration curve of each region, calculates the contrast agent arrival delay relative to the corresponding region of the healthy side, forms a delay distribution vector, and generates the collateral compensation vector through a one-dimensional convolutional neural network encoding.

4. The method for predicting vascular recanalization level using cerebral perfusion parameters according to claim 3, characterized in that, The multiple anatomical regions are defined based on the ASPECTS scoring template, including the caudate nucleus, lentiform nucleus, internal capsule, insula, and the M1 to M6 regions of the middle cerebral artery.

5. The method for predicting vascular recanalization level using cerebral perfusion parameters according to claim 1, characterized in that, In step S4, the cumulative linking model establishes a relationship between the cumulative probability and the fused feature vector through a logical linking function. The cumulative probability represents the probability that the mTICI level is less than or equal to a specific level, and the predicted probability of each level is obtained by calculating the difference between the cumulative probabilities.

6. The method for predicting vascular recanalization level using cerebral perfusion parameters according to claim 5, characterized in that, The cumulative linking model includes multiple threshold parameters, which are parameterized using cumulative softplus to ensure ordered constraints. The ordered regression prediction head is trained using a negative log-likelihood loss function.

7. The method for predicting vascular recanalization level using cerebral perfusion parameters according to claim 1, characterized in that, In step S5, the attention gating mechanism generates gating weights based on the side compensation vector. The gating weights are constrained between 0 and 1 by the Sigmoid function and are used to adaptively scale the linear output of the ordered regression.

8. The method for predicting vascular recanalization level using cerebral perfusion parameters according to claim 1, characterized in that, In step S2, the multimodal feature fusion network adopts a dual-branch encoder architecture. The CTP branch encoder takes a four-channel CTP parameter map as input, and the CTA branch encoder takes a single-channel CTA maximum density projection reconstruction map as input. Both branches adopt a three-dimensional residual convolutional neural network.

9. The method for predicting vascular recanalization level using cerebral perfusion parameters according to claim 1, characterized in that, Step S1 also includes a preprocessing step for the CTP parameter map and the CTA maximum density projection reconstruction map. The preprocessing includes spatial registration and intensity normalization. After preprocessing, the four CTP parameter maps and one CTA maximum density projection reconstruction map are stacked into a five-channel three-dimensional tensor.

10. A system for predicting vascular recanalization levels using cerebral perfusion parameters, used to implement the method for predicting vascular recanalization levels using cerebral perfusion parameters according to any one of claims 1-9, characterized in that, include: The data acquisition module is configured to acquire preoperative CTP parameter maps and CTA maximum intensity projection reconstruction maps of patients with acute large vessel occlusive stroke. The CTP parameter maps include CBF parameter maps, CBV parameter maps, MTT parameter maps, and Tmax parameter maps, and to acquire dynamic scanning sequences of CTP scans. The multimodal feature fusion module is configured to extract and fuse the CTP parameter map and the CTA maximum density projection reconstruction map through a multimodal feature fusion network, wherein the CTP channel extraction function injects features, the CTA channel extracts morphological and structural features, and generates a fused feature vector; The collateral circulation embedding module is configured to extract time-concentration curves based on the dynamic scanning sequence of the CTP scan, extract the delayed distribution characteristics of contrast agent arrival in the affected hemisphere from the dynamic changes of the time-concentration curves, and encode and generate a collateral compensation vector. The ordered regression prediction module is configured to perform ordered regression analysis on the fused feature vector using the mTICI hierarchical ordered regression prediction head, and to model mTICI levels 0 to 3 as ordered categories using a cumulative linking model. The gated modulation output module is configured to modulate the output of the ordered regression prediction module by the side compensation vector through an attention gating mechanism, and output the prediction probability distribution of each level of mTICI.

Images