Postoperative stroke analysis method, system and device based on multi-modal image analysis

Abstract

The application discloses a postoperative stroke analysis method based on multi-modal image analysis, comprising the following steps: inputting medical data, including the patient's neurological function assessment, intraoperative parameters, preoperative and postoperative DWI images; labeling the medical data with whether postoperative stroke after vascular intervention surgery occurs or not, and forming a data set by combining the medical data with the label; constructing a deep learning model; training the deep learning model by using the data set to obtain a risk prediction model for predicting whether postoperative stroke after vascular intervention surgery occurs or not; inputting the medical data of the patient into the risk prediction model to output the prediction result of the risk of postoperative stroke after vascular intervention surgery. The application also provides a postoperative stroke analysis system and device. The method provided by the application can realize super-early risk prediction, brain region positioning and individualized rehabilitation guidance of overt stroke after vascular intervention surgery.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of medical intelligent assistance, and particularly relates to a postoperative stroke analysis method, system and device based on multi-modal image analysis. BACKGROUND

[0002] The current clinic is faced with three core problems: active early warning of occult stroke is missing: the existing diagnosis relies on postoperative DWI manual reading, which requires joint analysis of radiology and neurology and takes time, and the positioning error rate is as high as 20%-35% (especially for basal ganglia and subcortical small infarction); the mechanism of symptom transformation is unknown: there is a lack of prediction model based on multi-modal data (image + clinical), which cannot distinguish the transformation risk of asymptomatic infarction to symptomatic stroke. The timeliness of rehabilitation intervention is insufficient: the clinic usually starts intervention after the patient shows overt symptoms (median delay time > 24 hours), missing the 6-hour golden treatment window after stroke.

[0003] Patent CN118737436A discloses a brain stroke diagnosis and treatment intelligent decision system. Although the system uses AI image analysis technology, it has the problems of single data type and mainly relying on image data. At the same time, it is aimed at general brain stroke patients and fails to integrate the perioperative multi-modal data specific to transcatheter aortic valve implantation, which is crucial for prediction. Therefore, the system cannot capture the unique risk factors of brain stroke after transcatheter aortic valve implantation, and naturally cannot solve the above-mentioned "symptom transformation prediction" problem.

[0004] Patent CN119400394A discloses a thrombolysis prediction method based on deep learning and multi-modal fusion. Although the method integrates image and clinical features, its application target is to judge whether it is suitable for "thrombolytic treatment". More importantly, its model construction relies on static and single-time-point image data, which belongs to a kind of "static evaluation" model. It cannot handle and analyze the dynamic evolution characteristics of DWI images from preoperative to postoperative during the perioperative period of transcatheter aortic valve implantation, and ignores the key variable of intraoperative operation. Therefore, the method cannot realize "early active early warning" of overt stroke.

[0005] In summary, the existing technologies either have too general application scenarios or have too single and static data dimensions, and all of them cannot effectively solve the series of clinical problems such as risk prediction of occult cerebral infarction to overt stroke transformation, responsibility lesion positioning and targeted rehabilitation guidance after transcatheter aortic valve implantation. SUMMARY

[0006] The application aims to provide a postoperative explicit stroke analysis method, system and device based on multi-modal image analysis, which can effectively reduce the time and cost of multi-department cooperation, shorten the time window from diagnosis to stroke intervention, and provide comprehensive and reliable reference for super-early risk prediction, brain region positioning and individualized rehabilitation guidance of explicit stroke.

[0007] In order to achieve the first object of the application, the following scheme is provided: a postoperative explicit stroke analysis method based on multi-modal image analysis after transaortic valve cardiovascular intervention, comprising the following steps:

[0008] Input medical data, including patient's neurological function assessment, intraoperative parameters, preoperative and postoperative DWI images;

[0009] Label the medical data with whether postoperative stroke after vascular intervention surgery occurs, and form a data set with the medical data and the label;

[0010] Construct a deep learning model, including a multi-source data integration module, an image feature extraction module, a feature fusion module and a prediction module;

[0011] The multi-source data integration module includes an embedding layer and a fully connected layer, the embedding layer is used to map the classification variables in the input neurological function assessment and intraoperative parameters into dense vectors, the fully connected layer is used to standardize the continuous variables in the input neurological function assessment and intraoperative parameters to output feature vectors, and the multi-modal feature weights are calculated based on the dense vectors and the feature vectors to obtain corresponding structured data feature vectors;

[0012] The image feature extraction module is used to pre-process the DWI images by an offset correction formula and calculate the corresponding spatio-temporal differences, and a threshold segmentation algorithm is used to identify new infarct lesions in the DWI images and extract corresponding imageomic features, and the corresponding spatio-temporal-image features are constructed based on the calculated spatio-temporal differences and the extracted imageomic features;

[0013] The feature fusion module performs multi-modal feature fusion on the structured data feature vectors and the spatio-temporal-image features by adaptive weight gating to output corresponding fusion feature vectors;

[0014] The prediction module predicts according to the fusion feature vector to output a prediction result;

[0015] The data set is used to train the deep learning model to obtain a risk prediction model for predicting whether postoperative stroke after vascular intervention surgery occurs;

[0016] Inputting medical data of a patient into a risk prediction model to output a prediction result of a risk of postoperative stroke after a vascular interventional surgery, the prediction result including whether the postoperative stroke after the vascular interventional surgery occurs and a lesion location.

[0017] The present application realizes the super-early risk prediction of overt stroke, brain region positioning and individualized rehabilitation guidance by integrating preoperative neurological function scores, preoperative and postoperative DWI images and intraoperative parameter factors.

[0018] Specifically, the vascular interventional surgery refers to transcatheter structural heart intervention surgery, and the transcatheter structural heart intervention surgery includes transcatheter aortic valve replacement (TAVR) and transcatheter mitral valve edge-to-edge suturing surgery.

[0019] Specifically, the DWI image needs to be processed by brain template registration before input.

[0020] Specifically, the input DWI image is registered by using the MNI152 template, and the registration accuracy error is controlled within ±1 mm.

[0021] Specifically, the DWI image includes preoperative and postoperative head DWI images within 24 hours, and the image parameters are as follows: DICOM format, layer thickness ≤3 mm, b value = 1000 s / mm2, and matrix size ≥256×256.

[0022] Specifically, the neurological function assessment includes a modified Rankin scale (mRS) score (0-6 points) for reflecting the preoperative neurological function state and a Mini-Mental State Examination (MMSE) score (0-30 points) for assessing cognitive function, wherein an mRS score ≥2 points is used as a baseline dysfunction classification variable, and an MMSE score ≤24 points is significantly related to the risk of postoperative stroke.

[0023] Specifically, the intraoperative parameters include anesthesia type, whether a second valve is implanted, artificial valve type, perivalvular leakage, sheath tube placement time, intraoperative bleeding volume, left coronary sinus implantation depth and non-coronary sinus implantation depth, wherein perivalvular leakage ≥3 levels is used as an independent risk factor.

[0024] Specifically, the image feature extraction module adopts an improved 3D ResNet-50 network structure, which includes:

[0025] (1) DWI image preprocessing algorithm:

[0026] Offset correction formula: ;

[0027] wherein, is the signal intensity without diffusion weighting, b is the diffusion sensitivity coefficient, ADC is the apparent diffusion coefficient, and S is the diffusion weighted signal intensity;

[0028] (2) Spatiotemporal feature extraction algorithm:

[0029] Let the preoperative image be and the postoperative image be The spatiotemporal difference calculation formula is:

[0030] ;

[0031] wherein 24 hours;

[0032] (3) New infarct lesion identification algorithm:

[0033] Threshold segmentation combined with morphological processing is adopted:

[0034] ;

[0035] wherein ( is the background mean value, is the standard deviation), =27mm³.

[0036] Specifically, the image features include quantitative parameters obtained by comparing preoperative and postoperative head DWI: the total number of new infarct lesions in each brain region (TLN), the volume of each new infarct lesion (ILV), and the total volume of new infarct lesions (TLV), wherein TLV>500mm³ is set as a high-risk threshold for stroke, and imageomic features, which include shape feature sphericity , texture feature contrast , i and j represent the gray values (gray levels) of two adjacent pixels in the image.

[0037] For example, for an 8-bit grayscale image, the value range of i and j is 0-255 (corresponding to 256 gray levels), is an element in the gray level co-occurrence matrix (GLCM), representing the probability of adjacent occurrence of pixels with gray value i and pixels with gray value j under a certain direction and distance, GLCM is the core matrix of texture analysis, which quantifies image texture by statistical relationship of pixel pairs.

[0038] Specifically, the brain regions are divided and constructed based on the Harvard-Oxford brain atlas, including basal ganglia region, cortical region and cerebellum region, and the corresponding symptom types are movement disorder, aphasia and ataxia, respectively.

[0039] Specifically, the multi-source data integration module adopts a hierarchical coding strategy, including:

[0040] (1) Classification variable embedding algorithm:

[0041] For classification variables such as valve type, a learnable embedding is used:

[0042] ;

[0043] Wherein is the embedding dimension, which is set to 64;

[0044] (2) Continuous variable standardization formula:

[0045] , wherein is the mean value of the training set, is the standard deviation;

[0046] (3) Multi-modal feature weight calculation:

[0047] ;

[0048] Wherein, represents feature splicing, MLP is a multi-layer perception machine, and Softmax is a key nonlinear normalization function, which converts the original score (logits) output by the MLP into a probability distribution, ensures that the sum of the weights of all modal features is 1, and highlights the most important features;

[0049] (4) Structured data feature vector generation:

[0050] ;

[0051] Specifically, the attention mechanism includes an improved multi-head cross-attention mechanism and an asymmetric multi-modal attention mechanism:

[0052] (1) Multi-head attention calculation formula:

[0053] ;

[0054] Wherein ;

[0055] ;

[0056] (2) Position encoding fusion:

[0057] For the temporal DWI data, introduce position encoding:

[0058] ;

[0059] ;

[0060] (3) Adaptive weight gating:

[0061] ;

[0062] where is a sigmoid function, is a learnable weight matrix, is a clinical feature vector; is an image feature vector; is a bias term, used to adjust the threshold of the gating activation, enhancing the flexibility of the model.

[0063] (4) Asymmetric attention score:

[0064] ; where Gate comes from the output of (3) adaptive weight gating; Element-wise multiplication (Hadamard product), used to dynamically modulate the attention weights through gating; is a key nonlinear normalization function, which converts the original scores (logits) output by the MLP into a probability distribution, ensuring that the sum of the weights of all modal features is 1, while highlighting the most important features; Q is the query matrix (Query), which comes from the feature representation of the current modality; K is the key matrix (Key), which comes from the feature representation of other modalities; symbol T represents the transpose operation on matrix K; represents the dimension of the key vector, used to scale the dot product result (to prevent gradient vanishing).

[0065] where is an inter-modal prior weight matrix, is an adjustment intensity parameter.

[0066] Specifically, the prediction module adopts a hierarchical prediction strategy, including:

[0067] (1) Risk probability calculation formula:

[0068] ;

[0069] where is a fusion feature vector, is a sigmoid activation function; is a learnable weight matrix; is a bias term;

[0070] (2) Brain region localization algorithm:

[0071] Probabilistic mapping based on Harvard-Oxford atlas:

[0072] ;

[0073] where i {basal ganglia, cortex, cerebellum};

[0074] (3) Symptom prediction model:

[0075] Multi-label classification is used:

[0076] ;

[0077] where j {motor disorder, aphasia, ataxia}; is the fusion feature vector; is the learnable weight vector for the jth symptom; is the bias term for the jth symptom;

[0078] (4) Confidence calculation:

[0079] where H is the entropy function:

[0080] ;

[0081] Specifically, the deep learning model is trained using a composite loss function:

[0082] (1) Overall loss function:

[0083] ;

[0084] where , , , ;

[0085] (2) Focal Loss specific formula:

[0086] ;

[0087] where , set , ;

[0088] (3) Region localization loss:

[0089] ;

[0090] wherein , ;

[0091] (4) Regularization term:

[0092] wherein , .

[0093] Specifically, the model training adopts an adaptive learning strategy, including:

[0094] (1) Learning rate scheduling algorithm:

[0095] ;

[0096] wherein , is the maximum number of training epochs;

[0097] (2) Data augmentation strategy:

[0098] Image enhancement: random rotation , random scaling ;

[0099] Noise injection: wherein ;

[0100] (3) Early stopping condition:

[0101] When the consecutive epochs satisfy , stop training.

[0102] Specifically, the model performance index quantification standard is:

[0103] Classification performance: , sensitivity , specificity , ;

[0104] Positioning accuracy: Dice coefficient , Hausdorff distance ; timeliness: single instance inference time 5 seconds, system response time 60 seconds.

[0105] In order to achieve the second object of the application, the following technical scheme is provided: a postoperative stroke analysis system for executing the steps of the postoperative stroke analysis method based on multi-modal image analysis described above, including an input unit, a data analysis unit and an auxiliary unit;

[0106] The input unit is configured to collect medical data of a patient.

[0107] The data analysis unit is configured to analyze the collected medical data to output a prediction result, that is, to run the risk prediction model, extract a structured data feature vector and a space-time-image feature from the collected medical data, perform feature fusion, and predict a risk of postoperative cerebral stroke after a transvascular interventional operation.

[0108] The assistance unit is configured to generate corresponding medication and rehabilitation strategy references according to the output prediction result.

[0109] Specifically, the analysis system adopts a distributed architecture.

[0110] In order to achieve the third object of the present application, the following technical scheme is provided: a postoperative cerebral stroke analysis device for performing the steps of the postoperative cerebral stroke analysis method based on multi-modal image analysis.

[0111] Compared with the prior art, the present application has the following beneficial effects:

[0112] Integrated dimension is novel: for the first time, key operation parameters (such as artificial valve type and implantation depth) during TAVR are combined with preoperative and postoperative dynamic DWI image features to break through the limitations of traditional models relying only on preoperative or static data.

[0113] Precise prediction and positioning: through deep learning models and refined feature engineering, high-precision prediction of explicit stroke risk (target AUC 0.85, sensitivity 0.80) is achieved, and anatomic localization of the responsible lesion based on the Harvard-Oxford atlas standard is also achieved.

[0114] Clinical decision-making loop: the system not only provides risk values, but also maps lesion locations and expected neurological deficits (such as motor disorders and aphasia) to directly generate targeted rehabilitation recommendations (such as robot-assisted training and AI speech rehabilitation), achieving a clinical decision-making loop from prediction to intervention.

[0115] Efficient and practical deployment: the system uses lightweight deployment optimization (such as INT8 quantization) to seamlessly integrate into existing hospital information systems, achieving an automatic and fast (single instance prediction <5 seconds) analysis process with strong clinical practicality. BRIEF DESCRIPTION OF DRAWINGS

[0116] Figure 1 The overall flowchart of the postoperative cerebral stroke analysis method based on multi-modal image analysis provided by the present embodiment is shown in the figure.

[0117] Figure 2 A deep learning model architecture schematic diagram is provided for the embodiment;

[0118] Figure 3 A multi-head cross-attention mechanism structure schematic diagram is provided for the embodiment;

[0119] Figure 4 A system distributed architecture schematic diagram is provided for the embodiment;

[0120] Figure 5 A personalized treatment recommendation system interface output by the postoperative stroke analysis system of the vascular intervention surgery is provided for the embodiment. DETAILED DESCRIPTION

[0121] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the following will combine the accompanying drawings in the embodiments of the present application to make a clear and complete description of the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. The components of the embodiments of the present application described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0122] As shown in the figure, a postoperative stroke analysis method based on multi-modal image analysis is provided for the embodiment, and the specific steps are as follows: Figure 1

[0123] Step S1: Medical data input and preprocessing

[0124] The input medical data includes:

[0125] DWI image data: preoperative and postoperative head DWI image within 24 hours, DICOM format, layer thickness 3mm, b value = 1000 s / mm²

[0126] Neurological function assessment: mRS score (0-6) and MMSE score (0-30)

[0127] Intraoperative parameters: including 8 key parameters such as type of anesthesia, type of valve, and perivalvular leakage level.

[0128] Preprocessing of DWI image:

[0129] Offset correction: ;

[0130] ​MNI152 template registration: registration accuracy error is controlled within ±1mm.

[0131] Step S2: data labeling and dataset construction

[0132] The label was annotated by whether symptomatic cerebral stroke occurred within 30 days as the gold standard, and the annotation was confirmed by 3 neuroradiologists with a consistency kappa value >0.85.

[0133] Dataset construction: training set 70% (700 cases), validation set 15% (150 cases), and test set 15% (150 cases).

[0134] Step S3: deep learning model construction

[0135] As shown in Figure 2 , a deep learning model containing four core modules was constructed:

[0136] (1) Multi-source data integration module:

[0137] Classification variable embedding:

[0138] .

[0139] Continuous variable standardization:

[0140] .

[0141] Feature weight calculation:

[0142] .

[0143] Integration vector generation

[0144] ;

[0145] (2) Image feature extraction module:

[0146] Calculation of spatial and temporal differences;

[0147] ;

[0148] Identification of new infarction lesions:

[0149] ;

[0150] Extraction of radiomics features:

[0151] Sphericity ;

[0152] Contrast ;

[0153] (3) Feature fusion module:

[0154] As Figure 3 shown, the improved multi-head cross attention mechanism is adopted:

[0155] Multi-head attention:

[0156] ;

[0157] .

[0158] Position encoding:

[0159] ;

[0160] .

[0161] Adaptive gating:

[0162] .

[0163] Asymmetric attention:

[0164] .

[0165] (4) Prediction module:

[0166] Risk probability prediction:

[0167] .

[0168] Brain region localization

[0169] ;

[0170] Symptom prediction:

[0171] .

[0172] Confidence calculation:

[0173] .

[0174] Steps S4-S7: Each module implements the corresponding function according to the above algorithm, and the specific parameter settings are:

[0175] Learning rate: , batch size: , training rounds: maximum 200 rounds, early stop .

[0176] Step S8: Model training process. A composite loss function is used:

[0177] ;

[0178] :

[0179] .

[0180] Region localization loss:

[0181] .

[0182] Regularization:

[0183] .

[0184] Using adaptive learning rate schedule:

[0185] .

[0186] Step S9: Model application, i.e. inputting new patient data into the trained model, and the output includes: 30-day symptomatic stroke occurrence probability (0-1), responsible brain region localization and confidence, corresponding symptom type prediction, and individualized treatment recommendations.

[0187] The embodiment also provides a postoperative stroke analysis system for executing the steps of the postoperative stroke analysis method based on multi-modal image analysis described above.

[0188] As shown in Figure 4 , the system architecture provided by the embodiment includes:

[0189] (1) Input unit:

[0190] DICOM image interface: support for multiple manufacturers' equipment, transmission rate ≥ 100MB / s;

[0191] HIS system interface: HL7 FHIR standard, real-time data synchronization;

[0192] Intraoperative parameter input: mobile APP, supporting offline input.

[0193] (2) Data analysis unit: a distributed processing architecture is adopted, which is specifically as follows:

[0194] Data Pipeline: DICOM -> pre-processing -> feature extraction -> model inference -> result output.

[0195] Load balancing:

[0196] ;

[0197] Performance indicators:

[0198] Singleton processing delay < 60 seconds;

[0199] Concurrent processing capacity ≥ 50 cases / minute;

[0200] GPU inference time < 5 seconds.

[0201] (3) Security mechanism:

[0202] Data transmission: AES-256 encryption;

[0203] Access control: multi-level permissions based on RBAC;

[0204] Audit log: MD5 verification, storage period ≥ 3 years.

[0205] (4) Auxiliary unit, individualized treatment strategy generation:

[0206] Medication guidance:

[0207] :

[0208] :

[0209] :

[0210] return "Suggest replacing with monoclonal antibody treatment (clopidogrel 75mg)";

[0211] else:

[0212] return "Suggest starting dual antiplatelet therapy (aspirin 100mg + clopidogrel 75mg)";

[0213] else:

[0214] return "Routine anticoagulant therapy".

[0215] Rehabilitation plan:

[0216] :

[0217] plans = {"basal ganglia-motor disorders": "upper limb robot-assisted training 30 minutes / day x 6 weeks",

[0218] "language area-aphasia": "AI speech rehabilitation system 20 minutes / day",

[0219] "cerebellum-ataxia": "balance function training 45 minutes / day + VR technology"}

[0220] return plans.get(f"{{ }-{ }", "standard rehabilitation program").

[0221] To better illustrate the effectiveness of the technical solution provided in this embodiment, a specific explanation is given based on clinical validation cases.

[0222] Patient basic information: Age: 78 years old, male, underlying disease: severe aortic stenosis.

[0223] The input data is as follows:

[0224] Preoperative data: DWI imaging: no acute infarction lesions; mRS score: 1 point; MMSE score: 26 points.

[0225] Intraoperative parameters: Anesthesia type: local anesthesia; Valve type: Venus A; Second valve: no; Paravalvular leak: grade 3; Sheath duration: 145 min; Blood loss: 200 ml; Left coronary sinus depth: 5.1 mm; Noncoronary sinus depth: 3.7 mm.

[0226] Model prediction results:

[0227] System output

[0228] {" ": 0.82,"confidence": 0.91,

[0229] " {"basal ganglia": 0.75,"cortex": 0.20,"cerebellum": 0.05},

[0230] " {"Motor impairment": 0.78,"Aphasia": 0.15,"Ataxia":0.07}.

[0231] like Figure 5 As shown, the individualized treatment recommendations output by the post-vascular interventional stroke analysis system are as follows:

[0232] Medication guidance: It is recommended to start dual antiplatelet therapy (aspirin 100mg + clopidogrel 75mg).

[0233] Rehabilitation plan: Start upper limb robot-assisted motor training the next day (30 minutes / day x 6 weeks).

[0234] Monitoring strategy: Perform DWI again 24 hours after surgery and closely observe changes in upper limb muscle strength.

[0235] Clinical validation results:

[0236] The patient developed mild weakness in the left limbs 26 hours after the surgery.

[0237] DWI follow-up showed a new small infarct lesion (volume 580 mm³) in the right basal ganglia region.

[0238] Prediction accuracy: The model's predictions are highly consistent with actual outcomes.

[0239] Therapeutic effect: After early intervention, patient's 3-month follow-up mRS score is 1.

[0240] Performance evaluation:

[0241] Model performance indicators:

[0242] AUC: 0.87 (95% CI: 0.83-0.91).

[0243] Sensitivity: 0.83.

[0244] Specificity: 0.78.

[0245] F1-score: 0.80.

[0246] Dice coefficient: 0.89.

[0247] System performance indicators:

[0248] Average response time: 45 seconds.

[0249] Concurrent processing capacity: 65 cases / minute.

[0250] System availability: 99.7%.

[0251] Prediction accuracy: 83.5%.

[0252] The embodiment also provides a postoperative stroke analysis device for executing the steps of the postoperative stroke analysis method based on multi-modal image analysis provided by the above-mentioned embodiments.

[0253] The scheme provided by the present application is different from the prior art in that:

[0254] Static intraoperative parameter integration: for the first time, key operation parameters (such as artificial valve type, sheath tube placement duration, etc.) during TAVR are systematically combined with imaging genomics modeling, breaking the limitation of traditional preoperative data;

[0255] Precise mapping of brain regions and symptoms: based on the responsibility area positioning of the Harvard-Oxford brain atlas, targeted rehabilitation is guided.

[0256] Strong clinical practicability: the risk prediction model adopts lightweight deployment (single instance prediction <5 seconds), which is suitable for existing hospital information systems.

[0257] In addition, the terms "upper", "lower", "inner", "outer", "front", "back" are used only for descriptive purposes and not to indicate or imply relative importance. Unless otherwise specifically stated, the relative steps, numerical expressions, and numerical values of the components and steps set forth in these embodiments do not limit the scope of the present application.

[0258] Of course, the above only is the specific embodiment of the present application, and is not to limit the scope of the present application, any equivalent changes or modifications made in accordance with the application described in the scope of the application, should be included in the scope of the application.

[0259] Finally, it should be noted that the above-described embodiments, only for the specific embodiments of the present application, in order to illustrate the technical solutions of the present application, and not to limit, the scope of protection of the present application is not limited to this, although the foregoing detailed description of the present application, those skilled in the art should understand that any familiar with the technical field of the technical personnel in the technical range of the present application, it still can be modified or easily thought of changes to the technical solution recorded in the foregoing examples, or the equivalent replacement of some of the technical features; and these modifications, changes or replacement, and the corresponding technical solution of the spirit and scope of the present application embodiment technical solution, all should be covered in the scope of protection of the present application. Therefore, the scope of protection of the present application should be said to the scope of protection of the claims.

Claims

1. A method for postoperative stroke analysis based on multimodal image analysis, comprising the following steps: Input medical data, including the patient's neurological function assessment, intraoperative parameters, and preoperative and postoperative DWI images, including preoperative and postoperative DWI images of the head within 24 hours after surgery. Medical data are labeled based on whether postoperative stroke occurred after transvascular interventional surgery, and the medical data and labels are combined to form a dataset. Construct a deep learning model, including a multi-source data integration module, an image feature extraction module, a feature fusion module, and a prediction module; The multi-source data integration module includes an embedding layer and a fully connected layer. The embedding layer is used to map categorical variables in the input neurofunctional assessment and intraoperative parameters into dense vectors. The fully connected layer is used to standardize continuous variables in the input neurofunctional assessment and intraoperative parameters to output feature vectors. Multimodal feature weights are calculated based on the dense vectors and feature vectors to obtain the corresponding structured data feature vectors. The image feature extraction module is used to preprocess DWI images and calculate corresponding spatiotemporal differences. Simultaneously, a threshold segmentation algorithm is used to identify newly developed infarcts in the DWI images and extract corresponding radiomics features. Based on the calculated spatiotemporal differences and the extracted radiomics features, corresponding spatiotemporal-image features are constructed. Let the preoperative image be... Postoperative images are The formula for calculating spatiotemporal differences is: ;in The feature fusion module, through adaptive weight gating, performs multimodal feature fusion of structured data feature vectors and spatiotemporal-image features to output the corresponding fused feature vector; The prediction module makes a prediction based on the fused feature vector and outputs the prediction result. The dataset was used to train a deep learning model to obtain a risk prediction model for predicting whether stroke will occur after transvascular interventional surgery. The patient's medical data is input into the risk prediction model to output a prediction of the risk of stroke after transvascular interventional surgery.

2. The postoperative stroke analysis method based on multimodal image analysis according to claim 1, characterized in that, Before input, the DWI images need to be registered using a brain template. The MNI152 template is used to register the input DWI images, and the registration accuracy error is controlled within ±1mm.

3. The postoperative stroke analysis method based on multimodal image analysis according to claim 1, characterized in that, The neurological function assessment includes a modified Rankin Scale score to reflect preoperative neurological function and a Mini-Mental State Scale score to assess cognitive function.

4. The postoperative stroke analysis method based on multimodal image analysis according to claim 1, characterized in that, The intraoperative parameters include the type of anesthesia, whether a second valve is implanted, the type of artificial valve, paravalvular leak, duration of sheath placement, intraoperative blood loss, depth of implantation in the left coronary sinus, and depth of implantation in the non-coronary sinus.

5. The postoperative stroke analysis method based on multimodal image analysis according to claim 1, characterized in that, The imaging features include quantitative parameters obtained through preoperative and postoperative DWI comparisons of the head, as well as radiomics features: The quantitative parameters include the total number of newly developed infarcts in each brain region, the volume of each newly developed infarct, and the total volume of newly developed infarcts. The image omics features include shape feature sphericity and texture feature contrast.

6. The postoperative stroke analysis method based on multimodal image analysis according to claim 5, characterized in that, The brain regions are constructed based on the Harvard-Oxford brain atlas and include the basal ganglia, cortex, and cerebellum.

7. The postoperative stroke analysis method based on multimodal image analysis according to claim 1, characterized in that, The model training employs an adaptive learning strategy, including: a learning rate scheduling algorithm, data augmentation strategies including rotation and noise injection; and an early stopping condition of continuous operation. Each epoch satisfies Training should be stopped at this time.

8. A postoperative stroke analysis system, characterized in that, The steps for performing the postoperative stroke analysis method based on multimodal image analysis as described in any one of claims 1 to 7 include an input unit, a data analysis unit, and an auxiliary unit. The input unit is used to collect the patient's medical data; The data analysis unit is used to analyze the collected medical data to output prediction results; The auxiliary unit is used to generate corresponding medication and rehabilitation strategy references based on the output prediction results.

9. A postoperative stroke analysis device, characterized in that, The steps for performing the postoperative stroke analysis method based on multimodal image analysis as described in any one of claims 1 to 7.

Images