Acute cerebral apoplexy thrombolysis aid decision-making method based on DWI-FLAIR mismatching

Abstract

The invention provides an acute cerebral apoplexy thrombolysis aid decision-making method based on DWI-FLAIR mismatching. The acute cerebral apoplexy thrombolysis aid decision-making method comprises the following steps: acquiring a diffusion weighted imaging sequence image, a liquid attenuation inversion recovery sequence image and associated clinical data of an acute cerebral apoplexy patient; obtaining a focus core area based on the diffusion weighted imaging sequence image and the liquid attenuation inversion recovery sequence image; a multi-dimensional mismatching feature set is quantitatively extracted in the focus core area and the half-dark band prediction area, and the multi-dimensional mismatching feature set comprises an area mismatching intensity ratio, a texture heterogeneity parameter of a mismatching area and a spatial topological relation parameter of the mismatching area and a preset functional brain area template; based on the multi-dimensional mismatching feature set, calculating a mismatching credibility index used for quantitatively evaluating DWI-FLAIR mismatching saliency and stability; and fusing the mismatching credibility index and the associated clinical data, and calculating an individualized comprehensive decision index reflecting the expected thrombolysis net benefit. According to the scheme, the objectization and standardization level of acute cerebral apoplexy thrombolysis decision is improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent medical technology, specifically relating to a decision-making aid method for thrombolysis in acute stroke based on DWI-FLAIR mismatch. Background Technology

[0002] In the treatment of acute ischemic stroke (AIS), intravenous thrombolysis is a crucial means of restoring blood flow and salvaging the ischemic penumbra. However, its application is strictly limited by the time window and carries risks such as intracranial hemorrhage. Therefore, accurate screening of patients who may benefit from thrombolysis is essential. In recent years, the concept of the tissue window (i.e., based on the state of ischemic tissue as shown on imaging rather than simply clock time) has gained attention. Among these, DWI-FLAIR mismatch, as a potential imaging biomarker, has been studied for identifying patients with unknown onset time or those on the edge of the time window who may still have salvageable brain tissue. However, in clinical practice, existing decision-making models based on DWI-FLAIR mismatch suffer from subjectivity and inconsistency. Current clinical judgment mainly relies on the visual comparison of signal changes in high-signal areas on DWI and corresponding areas on FLAIR sequences by neuroradiologists or stroke physicians. This method is highly dependent on personal experience, and different assessors (or even the same assessor at different times) may reach inconsistent conclusions. It lacks objective and reproducible quantitative standards, affecting the homogenization and standardization of diagnosis and treatment. Summary of the Invention

[0003] In view of the above-mentioned defects or deficiencies in the prior art, a method for assisting decision-making on thrombolysis in acute stroke based on DWI-FLAIR mismatch is provided, including the following steps: Acquire diffusion-weighted imaging sequences, fluid attenuation inversion recovery sequences, and associated clinical data from patients with acute stroke; Based on the diffusion-weighted imaging sequence and the fluid attenuation inversion recovery sequence, the core area of ​​the acute ischemic lesion is obtained. In the predicted area of ​​the lesion core region and penumbra, a multidimensional mismatch feature set is quantitatively extracted. The multidimensional mismatch feature set includes: region mismatch intensity ratio, texture heterogeneity parameter of mismatch region, and spatial topological relationship parameter between mismatch region and preset functional brain region template. Based on the multidimensional mismatch feature set, a mismatch confidence index is calculated to quantitatively evaluate the significance and stability of DWI-FLAIR mismatch. By integrating the mismatch confidence index with the associated clinical data, an individualized comprehensive decision-making index reflecting the expected net benefit of thrombolysis is calculated.

[0004] According to the technical solution provided in this application, the calculation of the mismatch confidence index used to quantitatively evaluate the significance and stability of DWI-FLAIR mismatch includes the following steps: Based on a pre-defined expected physiological evolution model, the onset time to imaging examination in the associated clinical data is input, and the theoretical curve of the expected FLAIR sequence signal intensity in the core area of ​​the lesion and the predicted penumbra area is calculated. In the liquid attenuation inversion recovery sequence image, the actual measured curve of the FLAIR signal intensity distribution in the corresponding region is measured; The fitting deviation between the actual measured curve and the theoretical curve is calculated, and the fitting deviation is used to quantify the degree of difference between the actual imaging performance and the expected physiological evolution model based on the onset time. The fitting deviation is corrected by combining the spatial topological relationship parameters in the multidimensional mismatch feature set to generate the mismatch confidence index.

[0005] According to the technical solution provided in this application, the step of correcting the fitting deviation by combining it with the spatial topological relationship parameters in the multidimensional mismatch feature set to generate the mismatch confidence index includes the following steps: Based on the numerical range of the fitting deviation and the degree of overlap between the mismatched regions in the spatial topological relationship parameters and the templates of key functional brain regions, the mismatched regions are mapped onto the clinical benefit-risk stratification map. The hierarchical map includes two dimensions: the first dimension represents the physiological significance of the mismatch, which is determined by the degree of fit deviation; the second dimension represents the functional importance of the brain regions involved, which is determined by the degree of overlap. Based on the quadrant or region where the mismatched region is located in the hierarchical map, assign it a clinical decision tendency category and the corresponding standard confidence interval; Using the clinical decision-making tendency category as the core, and labeling the certainty of the category according to the standard confidence interval, the mismatch confidence index, which combines qualitative conclusions and quantitative uncertainty assessments, is comprehensively generated.

[0006] According to the technical solution provided in this application, the step of assigning a clinical decision tendency category and a corresponding standard confidence interval to the mismatched region based on the quadrant or region where the hierarchical map is located includes the following steps: The system queries a clinical recommendation protocol knowledge base, which contains pre-stored structured clinical recommendation protocols associated with each of the aforementioned clinical decision-making tendency categories. The specific coordinate data corresponding to the mismatched region in the hierarchical map is used as an index key and input into the clinical recommendation protocol knowledge base to retrieve and output the target clinical recommendation protocol that matches it. Based on the key decision factor entries defined in the target clinical recommendation scheme, reverse query the specific values ​​corresponding to the current patient's associated clinical data and the multidimensional mismatch feature set; The specific values ​​obtained from the query are compared item by item with the preset level thresholds in the key decision factor entries, and the standard confidence interval is calculated based on the proportion of the number of items that meet the preset level thresholds to the total number of items.

[0007] According to the technical solution provided in this application, the structured clinical recommendation scheme includes: the strength of the thrombolytic therapy recommendation for this category, a list of recommended supplementary assessment items, and key risk warning information.

[0008] According to the technical solution provided in this application, the acquisition of diffusion-weighted imaging sequences, fluid attenuation inversion recovery sequences, and associated clinical data of patients with acute stroke includes the following steps: Based on the patient's unique identifier, the raw data, including diffusion-weighted imaging sequences and fluid attenuation inversion recovery sequences, collected within the acute stroke onset time window are retrieved from the image archiving and communication system. The raw data is post-processed to generate spatially aligned diffusion-weighted imaging sequence images and liquid attenuation inversion recovery images, which serve as the core image pairs for subsequent analysis. Based on the patient's unique identifier, structured clinical data directly related to this acute stroke event is extracted from the hospital information system to form the associated clinical data. The associated clinical data includes: time from onset to imaging examination, stroke neurological deficit severity score, and information on contraindications for thrombolytic therapy.

[0009] According to the technical solution provided in this application, the method further includes the following steps: Using the liquid attenuation inversion recovery sequence image as the reference coordinate space, the contour map of the lesion core area and the probability map of the mismatched area are aligned to this space to generate a spatially synchronized image base layer, core area contour layer and mismatched area layer. The core region contour layer and the mismatched region layer are rendered by applying color coding mapping rules and transparency gradient algorithms respectively, and then blended onto the image base layer in a semi-transparent overlay manner to generate an initial visualization map. Within the mismatched region layer, the point of maximum fitting deviation is located based on the spatial distribution map of the fitting deviation. Based on the spatial relationship with the preset functional brain region template, the centroid of the mismatched region is calculated and defined as the functional overlap center point; The point with the largest fitting deviation and the center point of functional overlap are highlighted on the initial visualization map to form a visualization verification map.

[0010] According to the technical solution provided in this application, after forming the visual verification map, the following steps are also included: For each pixel in the visualization verification map, establish its index relationship with the original image database and the multidimensional mismatch feature set, and create a spatial-feature mapping table; When a click event is captured on any of the aforementioned key feature points or any mismatched region pixels, a data query service is triggered; the data query service queries the spatial-feature mapping table based on the coordinates of the click location and performs the following operations: From the original image database, extract the original image slices corresponding to the position of the coordinate point on the diffusion-weighted imaging sequence image and the liquid attenuation inversion recovery sequence image, and display them side by side in an independent comparison window; Extract all original feature values ​​of the coordinate point or its region and the preset reference threshold used in the calculation process from the multidimensional mismatch feature set and related calculation model, and display them in a structured table format.

[0011] According to the technical solution provided in this application, obtaining the core area of ​​an acute ischemic lesion based on the diffusion-weighted imaging sequence and the fluid attenuation inversion recovery sequence includes the following steps: The diffusion-weighted imaging sequence image is preprocessed to calculate and generate an apparent diffusion coefficient map. Based on a preset absolute apparent diffusion coefficient threshold, initial segmentation is performed on the apparent diffusion coefficient map to obtain the core area of ​​the first candidate lesion. Based on the liquid attenuation inversion recovery sequence image, local signal intensity features and texture features of each voxel are extracted within the spatial range of the core area of ​​the first candidate lesion to construct a local feature map of the FLAIR sequence. By combining the quantified values ​​of the apparent diffusion coefficient map with the local feature map of the FLAIR sequence, multimodal collaborative analysis is performed to reclassify the histopathophysiological state of voxels in the core region of the first candidate lesion. Based on the reclassification results, voxels classified as meeting the characteristics of potentially reversible ischemic tissue are excluded from the first candidate lesion core area; the connected regions formed by the remaining voxels are identified as the core area of ​​acute ischemic lesions.

[0012] According to the technical solution provided in this application, the multimodal collaborative analysis includes the following steps: Based on the tissue recoverability discrimination rule, each voxel in the core area of ​​the first candidate lesion is analyzed in parallel. The input to the tissue recoverability discrimination rule is the corresponding feature value of the same voxel on the apparent diffusion coefficient map and the local feature map of the FLAIR sequence, and the output is the classification result of whether the voxel belongs to irreversible infarct core tissue or potentially recoverable ischemic tissue. The tissue recoverability discrimination rule is configured as follows: for voxels whose apparent diffusion coefficient value is lower than a preset threshold, but whose corresponding FLAIR local features do not show a typical acute infarction evolution pattern, they are identified as potentially recoverable ischemic tissue.

[0013] Compared with the prior art, the beneficial effects of this application are as follows: I. Enhancing the objectivity and standardization of decision-making: By quantitatively extracting multidimensional mismatch feature sets (such as intensity ratio, texture heterogeneity, and spatial topological relationships), the originally subjective image interpretation is transformed into objective and measurable data indicators, which greatly reduces the evaluation differences among different physicians. This is conducive to establishing unified image evaluation standards within hospitals and even regional stroke networks, and promotes the homogenization of medical quality.

[0014] II. Achieving Intelligent Fusion and Dimensionality Reduction of Multi-Source Information: This invention not only processes imaging information but also actively correlates and fuses key clinical data. By calculating mismatch credibility indices and individualized comprehensive decision-making indices, the system simulates and assists in completing the most complex multi-factor trade-off process in clinical decision-making. It condenses scattered imaging and clinical information into quantitative indicators that can directly assist in judgment, reducing the cognitive load on emergency physicians under time pressure and lowering the risk of decision-making errors due to information overload or omission. Third, it supports individualized and precise clinical decision-making: The final generated individualized comprehensive decision index is a quantitative result that integrates patient-specific imaging characteristics and clinical conditions. It can more precisely reflect the net benefit that a specific patient is expected to obtain from thrombolytic therapy, providing physicians with a more personalized decision-making reference anchor that goes beyond simple time windows or qualitative mismatches, and helps to achieve the goal of precision medicine that treats the individual.

[0015] IV. Optimizing Clinical Pathways and Improving Treatment Efficiency: This provides a structured decision support tool for stroke green channels. Standardized analysis processes and clear quantitative outputs can accelerate communication efficiency among multidisciplinary teams (such as neurology, radiology, and emergency medicine), allowing discussions to focus on decision-making logic based on the same set of objective evidence. This may shorten decision-making time and ensure that eligible patients receive appropriate treatment more quickly. Attached Figure Description

[0016] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1A flowchart illustrating the steps of the thrombolysis-assisted decision-making method for acute stroke based on DWI-FLAIR mismatch provided in this application embodiment. Detailed Implementation

[0017] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] As mentioned in the background section, this application proposes a decision-making aid method for thrombolysis in acute stroke based on DWI-FLAIR mismatch, such as... Figure 1 As shown, it includes the following steps: S1. Acquire diffusion-weighted imaging sequences, fluid attenuation inversion recovery sequences, and associated clinical data of patients with acute stroke; Specifically, based on the patient's emergency room number or hospital number (i.e., unique identifier), the system automatically retrieves and downloads brain MRI scan data collected after the stroke but before the thrombolysis decision from the Picture Archiving and Communication System (PACS) via the DICOM protocol. From this data, the system separates the raw image data of diffusion-weighted imaging (DWI) sequences and fluid attenuation inversion recovery (FLAIR) sequences. Simultaneously, through HL7 or a database interface, the system extracts structured clinical data closely related to the stroke event from the Hospital Information System (HIS) or stroke registry database. This data typically includes: time from onset to imaging examination accurate to the minute; the National Institutes of Health Stroke Scale (NIHSS) score (a typical representative of stroke neurological deficit severity scoring) assessed and recorded by the attending physician; the patient's age; past medical history (e.g., hypertension, diabetes); current medication history (especially anticoagulant use); and a list of contraindications for intravenous thrombolysis (e.g., recent surgery, active bleeding).

[0020] S2. Based on the diffusion-weighted imaging sequence and the liquid attenuation inversion recovery sequence, the core area of ​​the acute ischemic lesion is obtained. In practice, the DWI sequence data is first post-processed to calculate and generate an apparent diffusion coefficient (ADC) map. Then, a preset ADC value threshold is applied (e.g., a commonly used threshold is 620 × 10⁻⁶). -6 mm 2The ADC image was binarized and segmented to initially identify brain tissue regions showing significantly restricted water molecule diffusion; these regions were defined as the core area of ​​the candidate lesion. Simultaneously, the FLAIR sequence image and the ADC image were rigidly or non-linearly registered to ensure they were in the same anatomical coordinate system.

[0021] S3. In the predicted area of ​​the lesion core region and penumbra, a multidimensional mismatch feature set is quantitatively extracted. The multidimensional mismatch feature set includes: region mismatch intensity ratio, texture heterogeneity parameter of mismatch region, and spatial topological relationship parameter between mismatch region and preset functional brain region template. Specifically, the penumbra prediction region is typically defined as a ring-shaped region extending outwards from the core area of ​​the candidate lesion by a certain distance (e.g., 15-20 mm) or a region predicted by a more complex perfusion model. Within this region, the system automatically calculates the following three types of features, which together constitute a multidimensional mismatch feature set: Region mismatch intensity ratio: This parameter quantifies the degree of mismatch between DWI and FLAIR signals. In practice, the average signal intensity ratio of regions below a threshold on the ADC image (i.e., the core region) to the corresponding region on the FLAIR image can be calculated; or more precisely, the proportion of voxel volume within the penumbra prediction region where the ADC value is slightly decreased but the FLAIR signal is not significantly increased (i.e., the signal intensity ratio is less than 1.15 relative to the contralateral normal brain tissue) to the total predicted region volume can be calculated.

[0022] Texture heterogeneity parameter of mismatched regions: This parameter is used to evaluate the signal uniformity of mismatched regions on FLAIR images, reflecting the homogeneity of tissue changes. In practice, algorithms such as Gray-Level Co-occurrence Matrix (GLCM) can be applied to mismatched regions of FLAIR images to extract texture features such as contrast, entropy, and uniformity. A region with high heterogeneity (high entropy, low uniformity) may indicate inconsistent evolutionary stages of ischemic injury or contamination with other signals.

[0023] Spatial topological relationship parameters between mismatched regions and preset functional brain region templates: This parameter assesses the importance of brain functions affected by the mismatched regions. During implementation, a standard brain atlas template (such as the AAL template) needs to be pre-imported. This template has already segmented the cerebral cortex and subcortical structures into brain regions associated with specific functions (such as motor, language, and sensory functions). The system calculates the overlap volume and overlap ratio between the mismatched regions and key functional brain regions (such as the primary motor cortex and Broca's area) through spatial mapping, or calculates the Euclidean distance from the centroid of the mismatched region to the nearest key functional brain region.

[0024] S4. Based on the multidimensional mismatch feature set, calculate the mismatch confidence index used to quantitatively evaluate the significance and stability of DWI-FLAIR mismatch; Specifically, this step aims to synthesize multiple features into an easily understandable score. In implementation, a rule-based scoring system or a lightweight machine learning model (such as logistic regression or support vector machine) can be built. This model takes the aforementioned extracted multidimensional mismatch feature set as input and outputs a scalar value between 0 and 1 or between 0 and 100, which is the mismatch confidence index. A higher index value indicates that the system is more confident in the existence of a significant and stable DWI-FLAIR mismatch.

[0025] S5. Integrate the mismatch confidence index with the associated clinical data to calculate an individualized comprehensive decision index that reflects the expected net benefit of thrombolysis.

[0026] Specifically, in implementation, a decision fusion model needs to be constructed. This model uses the mismatch confidence index calculated in the previous step and key variables extracted from clinical data (such as time from onset to imaging examination, NIHSS score, and age) as inputs. An individualized comprehensive decision index is calculated through weighted summation, evidence-based fusion, or multi-parameter logistic regression models. This index is typically a continuous or graded value, designed to intuitively reflect the net balance between the expected benefits (such as the probability of neurological function improvement) and risks (such as the probability of symptomatic intracranial hemorrhage) of intravenous thrombolysis for a specific patient. For example, an index above a certain threshold corresponds to recommending thrombolysis, below another threshold corresponds to not recommending it, and in between corresponds to requiring careful evaluation.

[0027] The technical advantage of this implementation lies in transforming the complex decision-making process, which originally relied on the subjective experience of physicians and the integration of fragmented information, into a standardized, automated, and quantitative analysis workflow. Its technical principle is to objectively quantify mismatch features using radiomics methods, and then simulate clinical trade-off logic through data fusion models. This firstly solves the problem of inconsistent interpretations caused by differences in individual experience in clinical assessment, improving the homogeneity of diagnosis and treatment. Secondly, through automated processing and multi-source information integration, it greatly reduces the cognitive load on emergency room physicians under time pressure, lowering the risk of decision-making errors due to information omissions or misjudgments. Finally, by outputting an individualized comprehensive decision index that integrates all key imaging and clinical information of the patient, it provides physicians with a more targeted decision-making reference anchor point that transcends a single time window or qualitative description, promoting the advancement of acute stroke thrombolysis decision-making towards precision medicine.

[0028] In a preferred embodiment, the calculation of the mismatch confidence index used to quantify the significance and stability of the DWI-FLAIR mismatch includes the following steps: Based on a pre-defined expected physiological evolution model, the onset time to imaging examination in the associated clinical data is input, and the theoretical curve of the expected FLAIR sequence signal intensity in the core area of ​​the lesion and the predicted penumbra area is calculated. In the liquid attenuation inversion recovery sequence image, the actual measured curve of the FLAIR signal intensity distribution in the corresponding region is measured; The fitting deviation between the actual measured curve and the theoretical curve is calculated, and the fitting deviation is used to quantify the degree of difference between the actual imaging performance and the expected physiological evolution model based on the onset time. The fitting deviation is corrected by combining the spatial topological relationship parameters in the multidimensional mismatch feature set to generate the mismatch confidence index.

[0029] When implementing a model based on a pre-defined, anticipated physiological evolution, and inputting the time from onset to imaging examination to calculate the expected theoretical curve of FLAIR signal intensity, the first step is to establish or adopt a mathematical model describing the change in signal intensity of brain tissue on the FLAIR sequence over time after acute cerebral ischemia. This model can be obtained by fitting a large amount of retrospective clinical data or constructed based on the pathophysiological processes of post-ischemic cytotoxic edema and vasogenic edema. The core input of the model is the "time from onset to imaging examination" (T). For each voxel or sub-region within the lesion core area and the predicted penumbra, the model outputs an expected FLAIR signal intensity value or a signal intensity curve changing over time, based on time T. For example, the model may reveal that the expected increase in FLAIR signal in the ischemic core area is small within 3 hours after onset; while between 3 and 6 hours, the signal is expected to increase significantly. In this way, the system generates one or a set of theoretical curves corresponding to the specific examination time points of the patient.

[0030] In liquid attenuation inversion recovery sequence imaging, the actual measurement curve of the FLAIR signal intensity distribution in the corresponding region is measured on the registered real FLAIR image of the patient, in the same lesion core area and penumbra prediction area, to measure the actual average signal intensity value or extract the spatial distribution histogram of the signal intensity to form the actual measurement curve.

[0031] Calculating the deviation between the actual measured curve and the theoretical curve is the core step in quantifying mismatch. In practice, various methods can be used to calculate the deviation. For example, the root mean square error (RMSE) and mean absolute error (MAE) of the two curves at corresponding time points (or spatial locations) can be calculated, or the lag exponent of the actual measured curve relative to the theoretical curve can be calculated. If the actual FLAIR signal intensity is significantly lower than the intensity predicted by the model based on the onset time (i.e., a large negative deviation or low similarity), it strongly suggests a DWI-FLAIR mismatch, indicating tissue ischemia without the expected significant edema signal.

[0032] The step of correcting the fit deviation by combining it with spatial topological parameters from a multidimensional mismatch feature set to generate a mismatch confidence index is an optimization of the preliminary quantification results. In practice, the correction logic can be reflected in the following: even with the same fit deviation, if the mismatched region is entirely located in a non-critical functional area (such as the prefrontal white matter), its clinical significance may be low; however, if it highly overlaps with the hand motor function area or the language center, its clinical significance and decision-making value are extremely high. Therefore, the system can use the fit deviation as a base score, multiplied by a weighting coefficient determined by spatial topological parameters (such as the overlap ratio with key brain regions). (For example, the higher the overlap ratio, the larger the weighting coefficient, and the more likely the final index is to support the existence of a meaningful mismatch), thereby generating a final mismatch confidence index that better aligns with clinical decision-making logic.

[0033] In a preferred embodiment, the step of correcting the fitting deviation by combining it with the spatial topological relationship parameters in the multidimensional mismatch feature set to generate the mismatch confidence index includes the following steps: Based on the numerical range of the fitting deviation and the degree of overlap between the mismatched regions in the spatial topological relationship parameters and the templates of key functional brain regions, the mismatched regions are mapped onto the clinical benefit-risk stratification map. The hierarchical map includes two dimensions: the first dimension represents the physiological significance of the mismatch, which is determined by the degree of fit deviation; the second dimension represents the functional importance of the brain regions involved, which is determined by the degree of overlap. Based on the quadrant or region where the mismatched region is located in the hierarchical map, assign it a clinical decision tendency category and the corresponding standard confidence interval; Using the clinical decision-making tendency category as the core, and labeling the certainty of the category according to the standard confidence interval, the mismatch confidence index, which combines qualitative conclusions and quantitative uncertainty assessments, is comprehensively generated.

[0034] When implementing the steps of mapping mismatched regions to the clinical benefit-risk stratification map based on the numerical range of the fit deviation and the degree of overlap between the mismatched regions and the key functional brain region templates in the spatial topological relationship parameters, the first step is to define this two-dimensional map. The horizontal axis (first dimension) represents the physiological significance of the mismatch, determined by the fit deviation. For example, the fit deviation can be divided into three intervals: highly significant (deviation > threshold A), moderately significant (threshold B < deviation ≤ threshold A), and mildly or insignificant (deviation ≤ threshold B). The vertical axis (second dimension) represents the functional importance of the brain regions involved, determined by the proportion of overlap between the mismatched region and the key functional brain region template. For example, it can be divided into high-functionality regions (overlap ratio > threshold X), medium-functionality regions (threshold Y < overlap ratio ≤ threshold X), and low-functionality regions (overlap ratio ≤ threshold Y). This forms a 3×3 nine-square grid map, where each square represents a specific combination pattern of physiological significance and functional importance.

[0035] Assigning clinical decision bias categories and corresponding standard confidence intervals to mismatched regions within the hierarchical map is crucial for bridging quantitative analysis and clinical recommendations. In practice, each cell in the map needs to be pre-defined with its corresponding clinical decision bias category. For example, mismatches located in highly significant cells within high-functioning regions might be assigned a category strongly recommending thrombolysis; while mismatches located in mild or insignificant cells within low-functioning regions would be assigned a category with weak evidence of mismatch, suggesting limited benefit from thrombolysis. Each category is associated with a standard confidence interval. This interval can be derived from historical data on the probability that patients classified as belonging to that category ultimately demonstrated a true mismatch (or a good prognosis after thrombolysis). For example, a strong recommendation category might correspond to a confidence interval of [85%, 95%], indicating that the system considers the judgment correct with 85%-95% confidence.

[0036] The process of generating a mismatch confidence index, which combines qualitative conclusions with quantitative uncertainty assessments, is the final information encapsulation step. Centered on clinical decision-making propensity categories and labeling the certainty of each category according to standard confidence intervals, the system comprehensively encapsulates this information. In implementation, the system output is no longer a single number, but a structured information body. For example, the output might be: "Clinical decision-making propensity: Strongly recommend reassessing the time window and actively considering thrombolysis. Confidence level: 90% (interval: 85%-95%)". Here, 90% serves as a scalar representation of the mismatch confidence index, while the underlying categories and intervals provide richer semantics.

[0037] In a preferred embodiment, assigning a clinical decision tendency category and corresponding standard confidence interval to the mismatched region based on its quadrant or region in the hierarchical map includes the following steps: The system queries a clinical recommendation protocol knowledge base, which contains pre-stored structured clinical recommendation protocols associated with each of the aforementioned clinical decision-making tendency categories. The specific coordinate data corresponding to the mismatched region in the hierarchical map is used as an index key and input into the clinical recommendation protocol knowledge base to retrieve and output the target clinical recommendation protocol that matches it. Based on the key decision factor entries defined in the target clinical recommendation scheme, reverse query the specific values ​​corresponding to the current patient's associated clinical data and the multidimensional mismatch feature set; The specific values ​​obtained from the query are compared item by item with the preset level thresholds in the key decision factor entries, and the standard confidence interval is calculated based on the proportion of the number of items that meet the preset level thresholds to the total number of items.

[0038] When implementing a clinical recommendation protocol knowledge base, using the specific coordinates of the mismatched region in the hierarchical map as the index key, and retrieving and outputting the matching target clinical recommendation protocol steps, the first step is to build and maintain a clinical recommendation protocol knowledge base. This knowledge base is essentially a database, where each record corresponds to a specific region (or a group of adjacent regions) in the hierarchical map. Each record contains not only a clinical decision-making tendency category (e.g., recommending thrombolysis but requiring careful clinical evaluation), but also a richer set of structured clinical recommendation protocols. These protocols may include: specific thrombolytic drug recommendations for that category (e.g., standard dose of alteplase), a list of contraindications requiring immediate verification, recommended supplementary examinations to be completed before thrombolysis (e.g., immediate blood pressure recheck, blood tests for coagulation function), and key risk points to be discussed during informed consent interviews with the patient's family. The system uses the coordinates of the mismatched region in the map, similar to a geodesic search... Figure 1 In this way, the pre-stored structured scheme can be quickly retrieved.

[0039] Based on the key decision factors defined in the target clinical recommendation, the system reverse-engineers the specific numerical steps corresponding to the current patient's associated clinical data and the multidimensional mismatch feature set, embodying the idea of ​​dynamic validation. In each recommendation in the knowledge base, in addition to the suggested content, several key decision factors on which this recommendation is based are predefined. For example, for a strong recommendation category, its key factors might include: the deviation from the target must be greater than a certain value, the NIHSS score must be within a certain range, and the onset time must be less than a certain value. After retrieving the target recommendation, the system does not directly adopt it, but automatically searches for the corresponding actual values ​​in the patient's specific data based on these factor entries.

[0040] The core of generating personalized confidence intervals lies in comparing the specific numerical values ​​obtained from the query with preset level thresholds for each of the key decision factor items, and calculating the standard confidence interval based on the proportion of items that meet the preset level thresholds out of the total number of items. During implementation, the system performs the comparison item by item. For example, if a factor item requires an NIHSS score ≥ 6, and the patient actually scores 8, then this item meets the requirement; if it requires "onset time < 4.5 hours," and the patient actually experienced 3 hours, this item also meets the requirement. Assuming there are a total of 5 key factor items, and the patient meets 4 of them, the compliance rate is 80%. The system can use this rate as a basis to scale the "baseline confidence interval" (e.g., [80%, 90%]) associated with the recommendation. A higher compliance rate will result in a narrower and higher lower limit for the final "standard confidence interval" (e.g., if 100% compliance is achieved, the interval becomes [85%, 95%]), indicating sufficient evidence and greater confidence. Conversely, a low compliance rate will result in a wider interval and a lower lower limit (e.g., if 60% compliance is achieved, the interval becomes [70%, 85%]), indicating a gap in the evidence and prompting the system to indicate increased uncertainty in the conclusion.

[0041] In a preferred embodiment, the structured clinical recommendation scheme includes: the strength of the thrombolytic therapy recommendation for this category, a list of recommended supplementary assessment items, and key risk warning information.

[0042] When implementing the structured clinical recommendation scheme, including the strength of thrombolytic therapy recommendations for this category, the list of recommended supplementary assessment items, and key risk warning information, a standardized output template needs to be designed for each clinical decision tendency category in the knowledge base (such as "strong recommendation", "cautious recommendation", "insufficient evidence", "not recommended").

[0043] First, the "strength of recommendation for thrombolytic therapy for this category" is not a simple "yes" or "no," but a graded statement containing operational semantics. For example: Strong recommendation: This can be stated as "There is clear and significant imaging and clinical evidence supporting intravenous thrombolysis. It is recommended to initiate the thrombolysis procedure immediately, unless there are clear absolute contraindications." Cautious recommendation: This can be stated as "There is potential evidence of thrombolysis benefit, but it is accompanied by uncertainty or relative risk. It is recommended to make a decision after fully communicating the risks, taking into account the patient's specific condition and the family's wishes." Insufficient evidence: This can be stated as "Current imaging and clinical evidence is insufficient to clearly support or oppose thrombolysis. It is recommended to prioritize completing the following supplementary assessments or seek multidisciplinary consultation." Not recommended: This can be stated as "Current evidence suggests that the likelihood of benefit from thrombolysis is low, or the risks significantly outweigh the benefits. Routine intravenous thrombolysis is not recommended." This gradient-based intensity description is more effective than binary judgments in reflecting the complexity of clinical decision-making.

[0044] Secondly, the "Recommended Supplemental Assessment Items List" consists of actionable items proposed to address current decision-making uncertainties or to optimize decision-making. It is a highly scenario-based list. For example, for a patient categorized as "cautiously recommended" but with an onset time close to the 4.5-hour upper limit, the list might include: "1. Immediately recheck blood pressure to ensure it is below 185 / 110 mmHg; 2. Urgently draw blood to check platelet count, prothrombin time (PT) / international normalized ratio (INR), and blood glucose; 3. Attempt to contact family members to further confirm the precise onset time." For a patient with "insufficient evidence" but suspected large vessel occlusion, the list might directly recommend: "Immediately complete a head and neck CTA to assess the condition of the large vessels and prepare for possible endovascular thrombectomy." This list seamlessly integrates the systematic analytical conclusions into the hospital's emergency clinical pathway.

[0045] Finally, the "Key Risk Warning Information" section highlights key points that need to be specifically communicated to the treatment team regarding this type of recommendation. Unlike a general list of contraindications, it identifies the most critical risks based on current patient data. For example, for a patient over 80 years old with a high NIHSS score and a "Strong Recommendation" rating, the system might warn: "Note: This patient is elderly and has experienced a severe stroke. The risk of symptomatic intracranial hemorrhage (sICH) after thrombolysis is higher than average (reference risk approximately X%). Enhanced monitoring of neurological function and vital signs is necessary before and after treatment." Alternatively, for a patient with a history of mild cerebral microbleeds and a "Caution Recommendation" rating, the message might be: "The patient's history suggests the presence of cerebral microbleeds. While not an absolute contraindication, this potential risk should be emphasized in the informed consent process." In a preferred embodiment, acquiring diffusion-weighted imaging sequences, fluid attenuation inversion recovery sequences, and associated clinical data of patients with acute stroke includes the following steps: Based on the patient's unique identifier, the raw data, including diffusion-weighted imaging sequences and fluid attenuation inversion recovery sequences, collected within the acute stroke onset time window are retrieved from the image archiving and communication system. The raw data is post-processed to generate spatially aligned diffusion-weighted imaging sequence images and liquid attenuation inversion recovery images, which serve as the core image pairs for subsequent analysis. Based on the patient's unique identifier, structured clinical data directly related to this acute stroke event is extracted from the hospital information system to form the associated clinical data. The associated clinical data includes: time from onset to imaging examination, stroke neurological deficit severity score, and information on contraindications for thrombolytic therapy.

[0046] When retrieving raw data containing diffusion-weighted imaging (DWI) and fluid-attenuated inversion recovery (FLAIR) sequences acquired within the acute stroke onset time window from the Picture Archiving and Communication System (PACS) based on the patient's unique identifier, the system first needs to receive a trigger signal from the emergency triage or stroke green channel system via a hospital information integration platform (such as ESB Enterprise Service Bus) or a direct database interface. This signal carries the patient's unique identifier (such as hospital number or visit number). Using this identifier as an index, the system sends a DICOM query request to the PACS server. The query criteria must be explicitly limited to the time window following the acute stroke onset (e.g., the past 24 hours) and require specific sequence descriptions to ensure that only raw DICOM data of matching DWI and FLAIR sequences are retrieved. This process achieves automated integration with clinical workflows, avoiding delays and errors associated with manual image searching and uploading.

[0047] Post-processing the raw data to generate spatially aligned diffusion-weighted imaging sequences and fluid attenuation inversion recovery images serves as the core image pair for subsequent analysis and is a prerequisite for accurate quantitative analysis. The acquired raw DWI data typically contains images with multiple diffusion-sensitive gradient directions (b-values). The first step in post-processing is to calculate and generate an apparent diffusion coefficient (ADC) map, a quantitative map that eliminates the T2 transmission effect and truly reflects the degree of water molecule diffusion restriction. For FLAIR sequences, post-processing may include removing significant field inhomogeneities. The most critical step is spatial alignment (registration). Because slight patient movement may occur during scanning, rigid or non-rigid image registration algorithms (such as those based on mutual information or normalized cross-correlation) must be used. The ADC map or a DWI image with a specific b-value is treated as a floating image and registered with the FLAIR image (reference image). The same transformation matrix is ​​applied to all DWI / ADC data, ensuring that the anatomical structures of the two sequences correspond precisely at the sub-pixel level, forming a "core image pair" that can be compared at the voxel level.

[0048] Based on the patient's unique identifier, the system extracts structured clinical data directly related to the acute stroke event from the Hospital Information System (HIS), forming a linking clinical data step, which is crucial for obtaining the decision context. The system then utilizes the patient's unique identifier again to access the HIS, Emergency Electronic Medical Record (EMR), or stroke registry database via a data interface. The extraction must be structured, meaning the data exists in a clearly defined field format for easy program parsing. For example, the time from onset to imaging examination needs to be calculated from the onset time and MRI completion time timestamp fields; the "stroke neurological deficit severity score" needs to be extracted from the stroke scale's structured entry form, including the NIHSS total score and scores for each sub-item; and "thrombolytic therapy-related contraindication information" needs to be automatically retrieved and logically judged from multiple structured data sources such as medical records, diagnostic codes, surgical records, and laboratory results (e.g., "recent major surgery history," "INR>1.7"), and then summarized into a list.

[0049] In a preferred embodiment, the method further includes the following steps: Using the liquid attenuation inversion recovery sequence image as the reference coordinate space, the contour map of the lesion core area and the probability map of the mismatched area are aligned to this space to generate a spatially synchronized image base layer, core area contour layer and mismatched area layer. The core region contour layer and the mismatched region layer are rendered by applying color coding mapping rules and transparency gradient algorithms respectively, and then blended onto the image base layer in a semi-transparent overlay manner to generate an initial visualization map. Within the mismatched region layer, the point of maximum fitting deviation is located based on the spatial distribution map of the fitting deviation. Based on the spatial relationship with the preset functional brain region template, the centroid of the mismatched region is calculated and defined as the functional overlap center point; The point with the largest fitting deviation and the center point of functional overlap are highlighted on the initial visualization map to form a visualization verification map.

[0050] In the specific implementation of the steps involving aligning the contour map of the lesion core area with the probability map of the mismatched region to the reference coordinate space using the liquid attenuation inversion recovery sequence image, and generating spatially synchronized image base layer, core area contour layer, and mismatched region layer, the FLAIR image is first determined as the display background (base layer) because it provides a clear view of the brain anatomy. Then, the binarized mask image of the lesion core area and the probability map (value range 0-1) representing each voxel belonging to the mismatched region, generated by an algorithm (e.g., based on a fitted deviation distribution), are used as two independent layers. Using the already completed image registration results, or performing a new round of precise registration, it is ensured that these two layers are completely aligned with the FLAIR base layer in three-dimensional space. At this point, the system memory maintains three spatially corresponding data layers.

[0051] The initial visualization process involves rendering the core area outline layer and the mismatch area layer using color-coded mapping rules and transparency gradient algorithms, respectively. These layers are then blended onto the base image layer with a semi-transparent overlay to generate the initial visualization map. This process is both an art and a science of visualization. To avoid visual clutter, distinct colors that align with medical intuition are designed for different layers. For example, bright red or magenta is often used to outline the core area, suggesting warnings and irreversible damage; green or blue gradients are used to fill the mismatch area layer, with higher probability voxels displaying more saturated and brighter colors, suggesting vitality and potential salvageability. The most crucial element is the transparency gradient algorithm. For instance, areas with a mismatch probability below 0.3 can be set to be completely transparent (not displayed), with opacity increasing linearly between 0.3 and 0.7, and areas with a probability above 0.7 displayed with higher opacity. This way, the final image overlaid on a FLAIR background will only clearly show the highly probable mismatch areas, while blurring or hiding low-probability areas, thus highlighting the visual focus. The core area is typically overlaid with a semi-transparent solid color or outline to clearly define its location and extent.

[0052] Within the mismatched region layer, the system locates the point of maximum fitting deviation based on the spatial distribution map of the fitting deviation. Based on the spatial relationship with the preset functional brain region template, it calculates the centroid of the mismatched region and defines it as the functional overlap center point. This step provides an "anchor point" for interactive verification. In the rendered 3D volumetric data of the mismatched region, the system searches for the coordinates of the single voxel with the highest fitting deviation value and marks it as the "point of maximum fitting deviation," which is usually the location of the most significant physiological mismatch. Simultaneously, the system calculates the geometric centroid coordinates of the entire 3D volume of the mismatched region. Then, this centroid coordinate is mapped onto the preset functional brain region template. If the centroid falls within a key functional area, it is directly defined as the "functional overlap center point." If the centroid is outside a functional area, the system calculates the centroids of the sub-regions overlapping with each key functional area and selects the centroid corresponding to the sub-region with the maximum overlap as the center point. The spatial coordinates of these two points are calculated and stored.

[0053] The point of maximum fitting deviation and the center point of functional overlap are highlighted on the initial visualization atlas, forming the visualization verification atlas step, completing the final annotation. The system annotates these two points on the fused image at their corresponding 3D coordinate positions using striking and unique icons (such as stars, crosses, and triangles), possibly supplemented with brief text labels. At this point, a "visual verification atlas" integrating the original anatomical background, the core extent of the lesion, probabilistic mismatch areas, and key feature point annotations is generated and ready for display.

[0054] This implementation significantly lowers the barrier for doctors to understand and verify the results of complex algorithms. Its technical principle lies in utilizing information visualization technology to map multidimensional, abstract digital information (probability, goodness of fit) into color, brightness, transparency, and spatial location attributes that the human visual system is sensitive to. Through layered overlay and intelligent rendering, multiple information such as lesion location, extent, mismatch significance gradient, and key sites are condensed into a single view. Marking the point of maximum goodness of fit guides doctors to focus on the location most likely to have a true pathophysiological mismatch; marking the center point of functional overlap directly highlights the clinical value of the mismatched area. This atlas allows doctors to form an intuitive impression of the overall results and key details of the system analysis within seconds, providing a clear navigation map for subsequent in-depth interactive verification. It is a key human-computer interface for building a reliable and usable clinical support system.

[0055] In a preferred embodiment, after forming the visual verification map, the method further includes the following steps: For each pixel in the visualization verification map, establish its index relationship with the original image database and the multidimensional mismatch feature set, and create a spatial-feature mapping table; When a click event is captured on any of the aforementioned key feature points or any mismatched region pixels, a data query service is triggered; the data query service queries the spatial-feature mapping table based on the coordinates of the click location and performs the following operations: From the original image database, extract the original image slices corresponding to the position of the coordinate point on the diffusion-weighted imaging sequence image and the liquid attenuation inversion recovery sequence image, and display them side by side in an independent comparison window; Extract all original feature values ​​of the coordinate point or its region and the preset reference threshold used in the calculation process from the multidimensional mismatch feature set and related calculation model, and display them in a structured table format.

[0056] Specifically, in the implementation of visual verification of each pixel in the atlas, establishing its index relationship with the original image database and the multidimensional mismatch feature set, and creating a spatial-feature mapping table, a core data structure needs to be built in the background. When the visual atlas is generated, the system creates a table in memory or a temporary database. Each record in the table corresponds to each pixel in the atlas on the screen (or corresponds to the three-dimensional voxel behind it). Each record contains at least: spatial coordinate index: the three-dimensional coordinates (x, y, z) of the point in the original DWI sequence, ADC map, or FLAIR sequence image; feature value index: a pointer or direct stored value pointing to the relevant calculation result of the voxel in the "multidimensional mismatch feature set", such as the ADC value, FLAIR signal intensity, fit deviation, local texture feature value, etc. of the point; metadata: the partition to which the point belongs (such as belonging to the core area, the mismatch area, or neither). This "spatial-feature mapping table" is the bridge connecting visual presentation and underlying data evidence.

[0057] When a click event is captured on any key feature point or any pixel in a mismatched region, a data query service step is triggered, involving front-end interaction monitoring. The system deploys an event listener on the interactive interface displaying the visual verification atlas. When a doctor clicks on the atlas using a mouse or touch to identify the marked point of maximum fit deviation, the center point of functional overlap, or directly clicks any pixel displayed as a mismatched region (green / blue), the listener captures the screen coordinates of this click event.

[0058] The data query service, based on the coordinates of the clicked location, queries the spatial-feature mapping table and performs the following steps, forming the core of the interactive response. First, the service reverse-maps the screen coordinates back to 3D voxel coordinates. Then, using these coordinates as the key, it quickly queries the spatial-feature mapping table to retrieve all index information for that point. Next, the service executes two key operations in parallel: Extracting and displaying raw image slices side-by-side: Based on the retrieved 3D coordinates, the service accurately extracts slice images of the corresponding coordinate point in the transverse, sagittal, and coronal planes from the raw image database on both the original DWI image (e.g., an image with b=1000 s / mm²) and the original FLAIR image. These images are displayed side-by-side in a pop-up independent window, allowing doctors to directly see the "raw material" analyzed by the algorithm and confirm the original image representation of the location in both modalities.

[0059] Extracting and displaying structured feature data: Simultaneously, based on the feature value index, the service retrieves all relevant raw feature values ​​for that point (or a small region centered on it) from the computation cache or database. For example, "ADC value: 520×10..." - 6 mm 2 " / s", "FLAIR signal intensity (relative to the contralateral side): 1.05", "Fit deviation: -2.3", "Local entropy: 0.8". Furthermore, the preset reference thresholds used by the system in its calculations are listed, such as "ADC threshold: 620", "FLAIR signal ratio threshold: 1.15". All this information is presented in a clearly structured table next to the original image. The doctor can immediately see why the system considers this a mismatch: because the ADC value (520) is below the threshold (620), while the FLAIR signal ratio (1.05) does not exceed the threshold (1.15), and the fit deviation is negative.

[0060] The technical principle of this implementation lies in using a pre-built spatial-feature mapping table data structure to achieve millisecond-level reverse tracing from visual interaction to massive amounts of underlying data and intermediate calculation results. The benefits are: First, doctors can click on any area of ​​interest or suspicion at any time to instantly access all original evidence and calculation details, presenting the basis for the algorithm's judgment transparently and completely. This transparency is the foundation for building trust with doctors. Second, it greatly enhances the value of teaching, quality control, and multidisciplinary discussions. During case discussions, the basis for specific conclusions can be displayed at any time by clicking; during quality control retrospective analysis, the entire analytical chain can be reproduced. Third, it optimizes doctors' cognitive efficiency. Doctors no longer need to switch back and forth between system reports, original PACS images, and multiple parameter tables, or manually search for corresponding relationships; all related information is presented collaboratively with a single click, significantly reducing the mental burden of information integration and making human-computer collaborative decision-making smooth and efficient.

[0061] In a preferred embodiment, obtaining the core area of ​​an acute ischemic lesion based on the diffusion-weighted imaging sequence and the fluid attenuation inversion recovery sequence includes the following steps: The diffusion-weighted imaging sequence image is preprocessed to calculate and generate an apparent diffusion coefficient map. Based on a preset absolute apparent diffusion coefficient threshold, initial segmentation is performed on the apparent diffusion coefficient map to obtain the core area of ​​the first candidate lesion. Based on the liquid attenuation inversion recovery sequence image, local signal intensity features and texture features of each voxel are extracted within the spatial range of the core area of ​​the first candidate lesion to construct a local feature map of the FLAIR sequence. By combining the quantified values ​​of the apparent diffusion coefficient map with the local feature map of the FLAIR sequence, multimodal collaborative analysis is performed to reclassify the histopathophysiological state of voxels in the core region of the first candidate lesion. Based on the reclassification results, voxels classified as meeting the characteristics of potentially reversible ischemic tissue are excluded from the first candidate lesion core area; the connected regions formed by the remaining voxels are identified as the core area of ​​acute ischemic lesions.

[0062] In the specific implementation of preprocessing the diffusion-weighted imaging sequence image, calculating and generating an apparent diffusion coefficient map, and performing initial segmentation on the apparent diffusion coefficient map based on a preset absolute apparent diffusion coefficient threshold to obtain the core area of ​​the first candidate lesion, the system first processes the original DWI sequence data. Using at least two different b values ​​(e.g., b=0 and b=1000 s / mm)... 2 The apparent diffusion coefficient (ADC) value of each voxel is calculated using a single-exponential model to generate an ADC map, which quantitatively reflects the degree of water molecule diffusion restriction. Subsequently, an absolute ADC threshold validated in the literature or calibrated using our own data (e.g., a commonly used range of 550-620 × 10⁻⁶) is applied. -6 The whole-brain ADC map was binarized and segmented using a speed of mm² / s. All voxels with ADC values ​​below this threshold were marked as abnormal, and the connected regions formed by these voxels were defined as the core region of the first candidate lesion. This step quickly identified the extent of brain tissue where water molecule diffusion was significantly restricted.

[0063] Based on the fluid attenuation inversion recovery sequence image, the step of extracting local signal intensity and texture features of each voxel within the spatial range of the core region of the first candidate lesion to construct a local feature map of the FLAIR sequence is crucial for introducing second modality information. The system does not simply observe the signal intensity of the FLAIR image, but performs refined quantitative feature extraction. First, on the FLAIR image already registered with the ADC image space, the analysis area is limited to the aforementioned core region of the first candidate lesion. For each voxel within this region, its local signal intensity features are calculated, such as: the voxel's own FLAIR signal intensity value, the ratio of this value to the signal intensity of the corresponding voxel in its contralateral mirror brain region, or the mean and standard deviation of the signal intensity within a small neighborhood (e.g., 3×3×3 voxels) centered on the voxel. Simultaneously, "texture features" are extracted, reflecting the pattern of the local structure of the image. A commonly used method is to calculate the gray-level co-occurrence matrix (GLCM) and extract parameters such as contrast, energy (uniformity), and entropy (randomness). These parameters can capture subtle changes and spatial distribution patterns of tissue edema. By combining these intensities of each voxel with texture feature values, a multidimensional "FLAIR sequence local feature map" is constructed for the candidate core region.

[0064] Combining the quantified values ​​of the apparent diffusion coefficient map with the local feature map of the FLAIR sequence, multimodal collaborative analysis is performed to reclassify the histopathological state of voxels within the core region of the first candidate lesion. This collaborative analysis is not simply a parallel presentation of two pieces of information, but rather aims to discover the correlation patterns between DWI and FLAIR features to distinguish different pathological states. In practice, the ADC value of each voxel can be concatenated with its corresponding FLAIR local feature vector (such as signal ratio, local homogeneity) to form a richer fused feature vector. Then, a pre-trained classification model (such as random forest, support vector machine, or a simple rule engine) is used to analyze this fused feature vector. The training objective of this model is to learn to distinguish between two tissue types: one is a pattern with low ADC values ​​and FLAIR features exhibiting typical acute cytotoxic edema followed by vasogenic edema (i.e., significantly increased signal and homogenized texture); the other is a pattern with low ADC values ​​but FLAIR features exhibiting "mismatch" or "no significant evolution" (i.e., insignificant signal increase and still high texture heterogeneity).

[0065] Based on the reclassification results, voxels classified as potentially reversible ischemic tissue are excluded from the first candidate lesion core area. The connected regions formed by the remaining voxels are then identified as the acute ischemic lesion core area, ultimately achieving precise purification of the core area. The classification model predicts for each voxel within the first candidate lesion core area whether it belongs to irreversible infarct core tissue or potentially reversible ischemic tissue. The system removes voxels classified as the latter from the initial segmentation results. Finally, the corrected connected regions composed of the remaining voxels consistently identified as infarct cores are officially identified as the "acute ischemic lesion core area" for all subsequent analyses. This area theoretically more closely approximates the actual extent of irreversible tissue damage.

[0066] In a preferred embodiment, the multimodal collaborative analysis includes the following steps: Based on the tissue recoverability discrimination rule, each voxel in the core area of ​​the first candidate lesion is analyzed in parallel. The input to the tissue recoverability discrimination rule is the corresponding feature value of the same voxel on the apparent diffusion coefficient map and the local feature map of the FLAIR sequence, and the output is the classification result of whether the voxel belongs to irreversible infarct core tissue or potentially recoverable ischemic tissue. The tissue recoverability discrimination rule is configured as follows: for voxels whose apparent diffusion coefficient value is lower than a preset threshold, but whose corresponding FLAIR local features do not show a typical acute infarction evolution pattern, they are identified as potentially recoverable ischemic tissue.

[0067] When implementing the parallel analysis step for each voxel within the core region of the first candidate lesion based on the tissue recoverability discrimination rule, it is necessary to first define or train the discrimination rule. This rule can exist in various forms, but its essence is a function or model F, which takes two types of input features of a single voxel as variables: one type is the quantized value from the apparent diffusion coefficient map, i.e., the ADC value of the voxel; the other type is the corresponding feature value from the local feature map of the FLAIR sequence, such as the ratio of the voxel's signal intensity on FLAIR to that of the contralateral normal brain region (FLAIR ratio), and parameters reflecting its local texture uniformity (such as the energy value based on the gray-level co-occurrence matrix). For thousands of voxels within the candidate core region, the system can call this rule F in parallel, independently calculating and judging each voxel, and outputting a classification label.

[0068] The input to the tissue recoverability discrimination rule is the corresponding feature value of the same voxel on the apparent diffusion coefficient map and the local feature map of the FLAIR sequence. The output is the classification result of whether the voxel belongs to irreversible infarct core tissue or potentially recoverable ischemic tissue, clarifying the data flow of the rule. In implementation, this rule can be a very concise set of "if-then" logic. For example, a simplified rule based on clinical experience could be: if (ADC value < threshold T_adc) and (FLAIR ratio > threshold T_flair), then it is determined as "irreversible infarct core tissue"; if (ADC value < threshold T_adc) and (FLAIR ratio ≤ threshold T_flair), then it is determined as "potentially recoverable ischemic tissue". Here, the threshold T_flair (such as 1.15 or 1.2) is used to define the boundary of "typical acute infarction evolution pattern", and a FLAIR signal ratio higher than this threshold is considered to indicate the presence of definite edema. More complex rules can be incorporated into texture features. For example, even if the FLAIRratio is slightly above the threshold, if the local texture heterogeneity is high (high entropy), it may be predisposed to be an organization that is still in the early stages of evolution and has potential for recovery.

[0069] The tissue recoverability discrimination rule is configured as follows: for voxels with an apparent diffusion coefficient value below a preset threshold, but whose corresponding FLAIR local features do not exhibit a typical acute infarction evolution pattern, they are classified as potentially recoverable ischemic tissue. This directly addresses a core clinical issue: how to correct the overjudgment tendency of the single ADC threshold method. The preset threshold here refers to the ADC threshold (e.g., 620 × 10⁻⁶). -6 mm 2 If a typical acute infarction evolution pattern is not observed, it needs to be defined and quantified using local FLAIR features, such as the previously mentioned FLAIR signal ratio not exceeding a specific threshold, or exhibiting a heterogeneous state in combination with texture features. The rule is explicitly configured to: when this combination of features of "limited diffusion but no edema signal" is captured, a judgment is made that favors the tissue's potential for recovery. This configuration principle enables the rule to identify the special and important situation of "mismatch existing within the core area".

[0070] The technical principle of this implementation lies in transforming the complex problem of tissue state discrimination into a logical judgment based on a clear combination of imaging biomarkers. By setting the condition that "ADC value is lower than a preset threshold," the rule first confirms the presence of severe ischemic damage in the tissue. Based on this, by examining the second condition that "FLAIR local features exhibit a typical acute infarction evolution pattern," the rule identifies the histological stage of the injury: tissue that has progressed to the stage of significant vasogenic edema (high FLAIR signal) is considered more likely to have suffered irreversible damage; while tissue in the early stage of predominantly cytotoxic edema, before significant vasogenic edema has formed (with little change in FLAIR signal), is considered ischemic tissue with the potential for reversal. This step-by-step discrimination logic not only makes the entire analysis process highly interpretable (doctors can clearly understand the basis for each interpretation made by the system), but also achieves a more specific definition of the infarct core by focusing on the key phenomenon of ADC-FLAIR performance separation. This rule can be trained and optimized based on large-scale clinical imaging-prognostic pairing data to make its discrimination boundary more precise. Ultimately, it ensures that the lesion core area output by the system is a purer "real" infarct core that excludes most "false positive" core tissues, laying the most accurate anatomical foundation for subsequent thrombolysis benefit assessment.

[0071] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A decision-making aid method for thrombolysis in acute stroke based on DWI-FLAIR mismatch, characterized in that, Includes the following steps: Acquire diffusion-weighted imaging sequences, fluid attenuation inversion recovery sequences, and associated clinical data from patients with acute stroke; Based on the diffusion-weighted imaging sequence and the fluid attenuation inversion recovery sequence, the core area of ​​the acute ischemic lesion is obtained. In the predicted area of ​​the lesion core region and penumbra, a multidimensional mismatch feature set is quantitatively extracted. The multidimensional mismatch feature set includes: region mismatch intensity ratio, texture heterogeneity parameter of mismatch region, and spatial topological relationship parameter between mismatch region and preset functional brain region template. Based on the multidimensional mismatch feature set, a mismatch confidence index is calculated to quantitatively evaluate the significance and stability of DWI-FLAIR mismatch. By integrating the mismatch confidence index with the associated clinical data, an individualized comprehensive decision-making index reflecting the expected net benefit of thrombolysis is calculated.

2. The method for assisting decision-making in thrombolysis for acute stroke due to DWI-FLAIR mismatch according to claim 1, characterized in that, The calculation, used to quantify the confidence index of the DWI-FLAIR mismatch significance and stability, includes the following steps: Based on a pre-defined expected physiological evolution model, the onset time to imaging examination in the associated clinical data is input, and the theoretical curve of the expected FLAIR sequence signal intensity in the core area of ​​the lesion and the predicted penumbra area is calculated. In the liquid attenuation inversion recovery sequence image, the actual measured curve of the FLAIR signal intensity distribution in the corresponding region is measured; The fitting deviation between the actual measured curve and the theoretical curve is calculated, and the fitting deviation is used to quantify the degree of difference between the actual imaging performance and the expected physiological evolution model based on the onset time. The fitting deviation is corrected by combining the spatial topological relationship parameters in the multidimensional mismatch feature set to generate the mismatch confidence index.

3. The method for assisting decision-making in thrombolysis for acute stroke due to DWI-FLAIR mismatch according to claim 2, characterized in that, The step of correcting the fitting deviation by combining it with the spatial topological relationship parameters in the multidimensional mismatch feature set to generate the mismatch confidence index includes the following steps: Based on the numerical range of the fitting deviation and the degree of overlap between the mismatched regions in the spatial topological relationship parameters and the templates of key functional brain regions, the mismatched regions are mapped onto the clinical benefit-risk stratification map. The hierarchical map includes two dimensions: the first dimension represents the physiological significance of the mismatch, which is determined by the degree of fit deviation; the second dimension represents the functional importance of the brain regions involved, which is determined by the degree of overlap. Based on the quadrant or region where the mismatched region is located in the hierarchical map, assign it a clinical decision tendency category and the corresponding standard confidence interval; Using the clinical decision-making tendency category as the core, and labeling the certainty of the category according to the standard confidence interval, the mismatch confidence index, which combines qualitative conclusions and quantitative uncertainty assessments, is comprehensively generated.

4. The method for assisting decision-making in thrombolysis for acute stroke due to DWI-FLAIR mismatch according to claim 3, characterized in that, Assigning clinical decision tendency categories and corresponding standard confidence intervals to the mismatched regions based on their location in the hierarchical map includes the following steps: The system queries a clinical recommendation protocol knowledge base, which contains pre-stored structured clinical recommendation protocols associated with each of the aforementioned clinical decision-making tendency categories. The specific coordinate data corresponding to the mismatched region in the hierarchical map is used as an index key and input into the clinical recommendation protocol knowledge base to retrieve and output the target clinical recommendation protocol that matches it. Based on the key decision factor entries defined in the target clinical recommendation scheme, reverse query the specific values ​​corresponding to the current patient's associated clinical data and the multidimensional mismatch feature set; The specific values ​​obtained from the query are compared item by item with the preset level thresholds in the key decision factor entries, and the standard confidence interval is calculated based on the proportion of the number of items that meet the preset level thresholds to the total number of items.

5. The method for assisting decision-making in thrombolysis for acute stroke due to DWI-FLAIR mismatch according to claim 4, characterized in that, The structured clinical recommendation protocol includes: the strength of the thrombolytic therapy recommendation for this category, a list of recommended supplementary assessment items, and key risk warning information.

6. The method for assisting decision-making in thrombolysis for acute stroke due to DWI-FLAIR mismatch according to claim 1, characterized in that, The acquisition of diffusion-weighted imaging sequences, fluid attenuation inversion recovery sequences, and associated clinical data of patients with acute stroke includes the following steps: Based on the patient's unique identifier, the raw data, including diffusion-weighted imaging sequences and fluid attenuation inversion recovery sequences, collected within the acute stroke onset time window are retrieved from the image archiving and communication system. The raw data is post-processed to generate spatially aligned diffusion-weighted imaging sequence images and liquid attenuation inversion recovery images, which serve as the core image pairs for subsequent analysis. Based on the patient's unique identifier, structured clinical data directly related to this acute stroke event is extracted from the hospital information system to form the associated clinical data. The associated clinical data includes: time from onset to imaging examination, stroke neurological deficit severity score, and information on contraindications for thrombolytic therapy.

7. The method for assisting decision-making in thrombolysis for acute stroke due to DWI-FLAIR mismatch according to claim 1, characterized in that, The method further includes the following steps: Using the liquid attenuation inversion recovery sequence image as the reference coordinate space, the contour map of the lesion core area and the probability map of the mismatched area are aligned to this space to generate a spatially synchronized image base layer, core area contour layer and mismatched area layer. The core region contour layer and the mismatched region layer are rendered by applying color coding mapping rules and transparency gradient algorithms respectively, and then blended onto the image base layer in a semi-transparent overlay manner to generate an initial visualization map. Within the mismatched region layer, the point of maximum fitting deviation is located based on the spatial distribution map of the fitting deviation. Based on the spatial relationship with the preset functional brain region template, the centroid of the mismatched region is calculated and defined as the functional overlap center point; The point with the largest fitting deviation and the center point of functional overlap are highlighted on the initial visualization map to form a visualization verification map.

8. The method for assisting decision-making in thrombolysis for acute stroke due to DWI-FLAIR mismatch according to claim 7, characterized in that, After forming the visual verification map, the following steps are also included: For each pixel in the visualization verification map, establish its index relationship with the original image database and the multidimensional mismatch feature set, and create a spatial-feature mapping table; When a click event is captured on any of the aforementioned key feature points or any mismatched region pixels, a data query service is triggered; the data query service queries the spatial-feature mapping table based on the coordinates of the click location and performs the following operations: From the original image database, extract the original image slices corresponding to the position of the coordinate point on the diffusion-weighted imaging sequence image and the liquid attenuation inversion recovery sequence image, and display them side by side in an independent comparison window; Extract all original feature values ​​of the coordinate point or its region and the preset reference threshold used in the calculation process from the multidimensional mismatch feature set and related calculation model, and display them in a structured table format.

9. The method for assisting decision-making in thrombolysis for acute stroke due to DWI-FLAIR mismatch according to claim 1, characterized in that, The process of obtaining the core area of ​​an acute ischemic lesion based on the diffusion-weighted imaging sequence and the fluid attenuation inversion recovery sequence includes the following steps: The diffusion-weighted imaging sequence image is preprocessed to calculate and generate an apparent diffusion coefficient map. Based on a preset absolute apparent diffusion coefficient threshold, initial segmentation is performed on the apparent diffusion coefficient map to obtain the core area of ​​the first candidate lesion. Based on the liquid attenuation inversion recovery sequence image, local signal intensity features and texture features of each voxel are extracted within the spatial range of the core area of ​​the first candidate lesion to construct a local feature map of the FLAIR sequence. By combining the quantified values ​​of the apparent diffusion coefficient map with the local feature map of the FLAIR sequence, multimodal collaborative analysis is performed to reclassify the histopathophysiological state of voxels in the core region of the first candidate lesion. Based on the reclassification results, voxels classified as meeting the characteristics of potentially reversible ischemic tissue are excluded from the first candidate lesion core area; the connected regions formed by the remaining voxels are identified as the core area of ​​acute ischemic lesions.

10. The method for assisting decision-making in thrombolysis for acute stroke due to DWI-FLAIR mismatch according to claim 1, characterized in that, The multimodal collaborative analysis includes the following steps: Based on the tissue recoverability discrimination rule, each voxel in the core area of ​​the first candidate lesion is analyzed in parallel. The input to the tissue recoverability discrimination rule is the corresponding feature value of the same voxel on the apparent diffusion coefficient map and the local feature map of the FLAIR sequence, and the output is the classification result of whether the voxel belongs to irreversible infarct core tissue or potentially recoverable ischemic tissue. The tissue recoverability discrimination rule is configured as follows: for voxels whose apparent diffusion coefficient value is lower than a preset threshold, but whose corresponding FLAIR local features do not show a typical acute infarction evolution pattern, they are identified as potentially recoverable ischemic tissue.

Images