Prediction system for acute ischemic stroke infarction core area expansion risk

Abstract

The invention relates to the technical field of artificial intelligence, in particular to an acute ischemic stroke infarction core area expansion risk prediction system, which comprises an acquisition module for acquiring head MRI-DWI images and clinical information of a patient; the first prediction module is used for predicting the expansion risk of an infarction core region according to the clinical information and the head MRI-DWI image; and the second prediction module is used for marking an expanded area of the infarction core area according to the head MRI-DWI image. In order to solve the problem that an image processing scheme in the prior art is low in diagnosis efficiency improvement, two models are introduced to effectively process multi-modal information composed of a head MRI-DWI image of a patient and clinical information, the possibility of further progression of the condition of the patient according to the current multi-modal information is predicted, and the probability of further progression of the condition of the patient according to the current multi-modal information is predicted. When the infarction core area is expanded, the range is possibly expanded on the image, and the prediction range and the quantitative evaluation result of the infarction core area are visually displayed on the primary diagnosis MRI-DWI sequence, so that a doctor can make further diagnosis conveniently.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a prediction system for the risk of expansion of the infarct core area in acute ischemic stroke. Background Art

[0002] Acute ischemic stroke (AIS) is a common cardiovascular and cerebrovascular disease with high morbidity, mortality, and disability. Magnetic resonance diffusion imaging (MRI-DWI) has a higher sensitivity than non-contrast enhanced computed tomography (CT). The sensitivity of early diagnosis of AIS is 88% to 100%, and the specificity is 95% to 100%. It has become the preferred examination method in clinical practice. In clinical practice, after standardized treatment according to relevant guidelines, most AIS patients improve, but some patients still have worsening disease progression, and the typical imaging manifestation is the expansion of the infarct core area on MRI-DWI images.

[0003] At present, the judgment of the disease progression of newly diagnosed AIS patients mainly relies on the National Institute of Health Stroke Scale (NIHSS) score and neuroimaging examination (CT, MRI) at 24h, 48h, and 72h after admission, combined with laboratory indicators (plasma brain natriuretic peptide, blood glucose, etc.). Although the NIHSS score is a widely used tool, it may be difficult to evaluate patients who cannot move due to coma, muscle relaxation or other reasons; and there is a certain lag in the judgment of the disease progression by NIHSS score, imaging review and laboratory examination; and there may be differences in the subjective prediction and interpretation of the disease progression by medical staff. The above reasons make it difficult to predict the disease progression of newly diagnosed AIS patients, resulting in the accuracy of the evaluation results being affected.

[0004] Recently, artificial intelligence technology has begun to be applied to the diagnosis and treatment of acute ischemic stroke (AIS). Artificial intelligence aims to develop decision support tools. Therefore, it is necessary to combine imaging data with other patient characteristics to improve the diagnostic and predictive capabilities of decision support models. However, current research on acute ischemic stroke usually only focuses on automatic segmentation and image processing of MRI images, and does not involve research on the prediction of disease progression. Summary of the invention

[0005] In view of the above problems existing in the prior art, a system for predicting the risk of expansion of the infarct core area in acute ischemic stroke is now provided.

[0006] The specific technical solutions are as follows:

[0007] A prediction system for the risk of expansion of the infarct core area in acute ischemic stroke, comprising:

[0008] An acquisition module, which acquires head MRI-DWI images and clinical information from the patient;

[0009] A first prediction module, the first prediction module is connected to the acquisition module, and the first prediction module predicts the risk of expansion of the infarct core area according to the clinical information and the head MRI-DWI image;

[0010] A second prediction module, wherein the second prediction module is connected to the acquisition module, and the second prediction module marks the enlarged area of ​​the infarct core area according to the head MRI-DWI image.

[0011] On the other hand, the acquisition module includes:

[0012] An image acquisition module, wherein the image acquisition module is connected to an external MRI device to acquire the head MRI-DWI image output by the MRI device;

[0013] A medical information collection module is connected to the hospital information system, and the medical information collection module collects the clinical information.

[0014] On the other hand, the first prediction module uses a first prediction model to predict the expansion risk, and the first prediction module includes:

[0015] An image extraction module, wherein the image extraction module extracts image features from the head MRI-DWI image;

[0016] A multidimensional data extraction module, wherein the multidimensional data extraction module extracts multimodal features from the clinical information;

[0017] A fusion prediction module is connected to the image extraction module and the multi-dimensional data extraction module respectively, and the fusion prediction module predicts the expansion risk according to the image features and the multi-modal features.

[0018] On the other hand, the image extraction module comprises:

[0019] A convolution module, wherein the convolution module extracts first image features at different levels from the head MRI-DWI image through a series of convolution layers;

[0020] A plurality of first preprocessing modules, wherein the first preprocessing modules are respectively connected to the convolution layers at corresponding levels in the convolution modules to respectively obtain the first image features;

[0021] The first preprocessing module flattens and linearizes the hierarchical image features to obtain second image features;

[0022] A first convolution layer, wherein the first convolution layers are respectively connected to the first preprocessing modules, and the image fusion convolution layer performs convolution processing on all the second image features to form second image features;

[0023] A first linear layer, wherein the first linear layer is connected to the first convolutional layer, and the first linear layer performs linearization processing on the second image feature to obtain the image feature.

[0024] On the other hand, the multidimensional data extraction module includes:

[0025] A plurality of table insertion modules, wherein the table insertion modules sequentially embed the clinical information into a table and establish a corresponding relationship between input and output to form an input table;

[0026] An encoding layer, the encoding layer is connected to the table insertion module, and the encoding layer performs one-hot encoding on the input table to form encoding information;

[0027] An embedding layer, the embedding layer is connected to the encoding layer, and the embedding layer generates an embedding vector according to the encoding information;

[0028] A second linear layer, wherein the second linear layer is connected to the embedding layer, and the second linear layer generates the multimodal feature according to the embedding vector.

[0029] On the other hand, the fusion prediction module includes:

[0030] A second convolutional layer, wherein the second convolutional layer receives the image feature and the multimodal feature respectively, and the second convolutional layer concatenates the image feature and the multimodal feature to extract a first fusion feature;

[0031] A third linear layer, the third linear layer is connected to the second convolutional layer, and the third linear layer linearly processes the first fused features to obtain second fused features;

[0032] A Softmax function module, wherein the Softmax function module is connected to the third linear layer, and the Softmax function module performs normalization mapping on each element in the second fusion feature to obtain a third fusion feature;

[0033] A third convolution layer, the third convolution layer is connected to the Softmax function module, and the third convolution layer extracts a fourth fusion feature after splicing the third fusion feature and the multimodal feature;

[0034] A fourth linear layer, wherein the fourth linear layer is connected to the Softmax function module and the third convolutional layer, and the fourth linear layer extracts a fifth fusion feature after splicing the third fusion feature and the image feature;

[0035] The fourth linear layer generates the expansion risk according to the fifth fused feature and the fourth fused feature.

[0036] On the other hand, the second prediction module comprises:

[0037] A pre-segmentation module, wherein the pre-segmentation module segments the head MRI-DWI image through a coding and decoding layer to obtain an infarct core area labeled image;

[0038] A residual network, wherein the residual network is connected to the pre-segmentation module, and the residual network extracts the annotated image of the infarct core area to obtain a multi-scale image patch;

[0039] A multi-scale MAP model, the multi-scale MAP model is connected to the residual network, and the multi-scale MAP model generates multi-scale features according to the multi-scale image patch and neighborhood pixels associated with the multi-scale image patch;

[0040] An estimation module, wherein the estimation module is connected to the multi-scale MAP model, and the estimation module predicts the expanded area according to the multi-scale features.

[0041] On the other hand, the pre-segmentation module comprises:

[0042] A plurality of sequentially connected convolutional layers, each of which extracts features from an input image using a convolution kernel to form a downsampled image;

[0043] The input image of the first convolutional layer is the head MRI-DWI image;

[0044] The input image of the convolutional layer of the next level is the downsampled image of the previous level;

[0045] A plurality of skip structures corresponding to the convolutional layer, wherein the skip structures collect the input image of the corresponding convolutional layer as original image features;

[0046] A plurality of upsampling layers, each of which upsamples the input image and the original image features at a corresponding level to obtain an upsampled image;

[0047] The input image of the upsampling layer of the first level is the downsampling image of the convolutional layer of the last level;

[0048] The input image of the upsampling layer of the next level is the upsampling image of the previous level;

[0049] The upsampling layer at the last level outputs the annotated image of the infarct core area.

[0050] The above technical solution has the following advantages or beneficial effects:

[0051] In order to solve the problem that the image processing schemes in the prior art can only achieve image segmentation when processing acute ischemic stroke and have low improvement in diagnostic efficiency, this scheme introduces two models to effectively process the multimodal information composed of the patient's head MRI-DWI images and clinical information. The scheme predicts the possibility of further progression of the patient's condition based on the current multimodal information, as well as the possible expansion range on the image when the core area of ​​the infarction expands. The predicted range and quantitative evaluation results of the core area of ​​the infarction are intuitively displayed on the initial diagnosis MRI-DWI sequence, which is convenient for doctors to make further diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The embodiments of the present invention will be described more fully with reference to the attached drawings, which are provided for illustration and description only and are not intended to limit the scope of the present invention.

[0053] Figure 1 It is an overall schematic diagram of an embodiment of the present invention;

[0054] Figure 2 This is a schematic diagram of a collection module in an embodiment of the present invention;

[0055] Figure 3 This is a schematic diagram of a first prediction model in an embodiment of the present invention;

[0056] Figure 4 This is a schematic diagram of a pre-extraction module in an embodiment of the present invention;

[0057] Figure 5 This is a schematic diagram of a multidimensional data extraction module in an embodiment of the present invention;

[0058] Figure 6 This is a schematic diagram of a fusion prediction module in an embodiment of the present invention;

[0059] Figure 7 This is a schematic diagram of a second prediction module in an embodiment of the present invention;

[0060] Figure 8 Schematic diagram of a pre-segmentation module in an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0063] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0064] The present invention comprises:

[0065] A prediction system for the risk of expansion of the infarct core area in acute ischemic stroke, such as Figure 1 As shown, including:

[0066] Acquisition module 1, acquisition module 1 acquires head MRI-DWI images and clinical information of the patient;

[0067] A first prediction module 2, the first prediction module 2 is connected to the acquisition module 1, and the first prediction module 2 predicts the risk of expansion of the infarct core area according to clinical information and head MRI-DWI images;

[0068] The second prediction module 3 is connected to the acquisition module 2, and the second prediction module 3 marks the enlarged area of ​​the infarct core area according to the head MRI-DWI image.

[0069] Specifically, the image processing scheme in the prior art can only achieve image segmentation when processing acute ischemic stroke, and the traditional manual judgment method has lag and low improvement in diagnostic efficiency. In this embodiment, the first prediction module 2 and the second prediction module 3 are introduced to effectively process the multimodal information composed of the patient's head MRI-DWI image and clinical information. Among them, the first prediction module 2 predicts the possibility of further deterioration of the patient's disease according to the current multimodal information, and the second prediction module 3 predicts the possible range on the image when the infarction core area expands, and intuitively displays the predicted range and quantitative evaluation results of the infarction core area on the initial diagnosis MRI-DWI sequence, which is convenient for doctors to make further diagnosis.

[0070] In the implementation process, the above prediction system is mainly configured as a software embodiment in a computer system, such as an in-hospital electronic information system, and is used to connect other in-hospital electronic information systems, such as MRI examination equipment and electronic medical record systems, through a software interface to obtain head MRI-DWI images and clinical information. The first prediction module 2 and the second prediction module 3 are configured with a pre-trained artificial intelligence model, which can predict the progression of acute ischemic stroke (Acute Ischemic Stroke, AIS) based on the input multimodal information, as well as the possible infarct core area when expansion occurs.

[0071] In one embodiment, Figure 2 As shown, the acquisition module 1 includes:

[0072] An image acquisition module 11, the image acquisition module 11 is connected to an external MRI device to obtain a head MRI-DWI image output by the MRI device;

[0073] The medical information collection module 12 is connected to the hospital information system, and the medical information collection module 12 collects clinical information.

[0074] Specifically, in order to achieve better data collection effects and facilitate subsequent predictions, in this solution, an image acquisition module 11 is provided to obtain the head MRI-DWI images output by the MRI device, and a medical information acquisition module 12 is provided to collect clinical information.

[0075] Clinical information includes general information, medical history information, and laboratory tests. General information includes: patient's sex (Sex), age (Age); medical history information includes: smoking history (History of smoking), drinking history (History of alcohol), diabetes history (History of diabetes), myocardial infarction (History of myocardial infarction), coronary atherosclerosis (History of coronary atherosclerosis), atrial fibrillation (History of atrial fibrillation), hypertension (History of hypertension) and stroke (History of stroke); laboratory tests include: prothrombin time (Prothrombin time, PT), fibrinogen (Fibrinogen), D-dimer (D-dimer), serum troponin I (Serum troponin I), blood sugar (Blood sugar), blood lipids (Blood lipids), plasma brain natriuretic peptide (Brain natriuretic peptide, BNP).

[0076] The head MRI-DWI image is an image obtained based on Magnetic Resonance Imaging-Diffuse Weighted Imaging (MRI-DWI). In order to achieve better recognition accuracy, the image acquisition module 11 also includes a pre-processing module for performing basic processing on the image before starting processing.

[0077] The processing flow of the preprocessing module includes: merging the head MRI-DWI image sequences in DCM format into NII files, and then removing the skull of these NII images using FSL software (Version: 6.0.6.5). The open source public MRI T1WI average brain image is selected as the standard brain template (http: / / www.bmap.ucla.edu / portfolio / atlases / ICBM_Template / ), and the MRI-DWI images in all data sets are aligned to the standard brain template. The alignment process is as follows: First, the sample and template orientation are turned to be consistent, and then 4 rounds of alignment are performed, the first two rounds are rigid transformations, and the last two rounds are affine transformations. After each round of alignment, the Dice Coefficient (DC) is calculated, and the alignment process will be terminated early when DC>0.94.

[0078] In one embodiment, the first prediction module uses the first prediction model to predict the enlarged risk, such as Figure 3 As shown, the first prediction model includes:

[0079] An image extraction module 21, the image extraction module 21 extracts image features from the head MRI-DWI image;

[0080] A multi-dimensional data extraction module 22, which extracts multi-modal features from clinical information;

[0081] The fusion prediction module 23 is connected to the image extraction module 21 and the multi-dimensional data extraction module 22 respectively. The fusion prediction module 23 predicts the expanded risk according to the image features and the multi-modal features.

[0082] Specifically, in order to achieve better recognition effects, in this embodiment, an image extraction module 21 is provided to extract image features from the head MRI-DWI images. At the same time, for a large amount of clinical information, the clinical information is vectorized by the multidimensional data extraction module 22 to extract multimodal features. Finally, the fusion prediction module 23 fuses the image features and the multimodal features to predict the enlarged risk.

[0083] In one embodiment, Figure 4 As shown, the image extraction module 21 includes:

[0084] A convolution module 211 , which extracts first image features at different levels from the head MRI-DWI image through a series of convolution layers;

[0085] A plurality of first preprocessing modules 212, wherein the first preprocessing modules 212 are respectively connected to the convolution layers at corresponding levels in the convolution module 211 to respectively obtain first image features;

[0086] The first preprocessing module 212 flattens and linearizes the hierarchical image features to obtain second image features;

[0087] A first convolution layer 213, the first convolution layer 213 is connected to the first preprocessing module 212 respectively, and the first convolution layer performs convolution processing on all the second image features to form the second image features;

[0088] The first linear layer 214 is connected to the first convolutional layer 213, and the first linear layer 214 performs linearization processing on the second image feature to obtain the image feature.

[0089] Specifically, in order to achieve a better image processing process, in this embodiment, the convolution module 211 is first used to extract the first image features at different levels from the head MRI-DWI image using a convolution kernel through a series of sequentially connected convolution layers, and then the first preprocessing module 212 is respectively connected to the convolution layers at the corresponding levels in the convolution module 211 to respectively obtain the first image features, and flatten and linearize them to obtain linearized image features to input into the first convolution layer 213. The first convolution layer superimposes all the second image features in the Z-axis direction, and then uses a three-dimensional convolution kernel to perform convolution processing to form the second image features. In this step, the fusion of image features in different dimensions is realized. Finally, the second image features are linearized and reduced in dimension by the first linear layer 214 to obtain image features, thereby achieving a better image extraction process.

[0090] In one embodiment, Figure 5 As shown, the multidimensional data extraction module 22 includes:

[0091] A plurality of table inserting modules 221, the table inserting modules 221 sequentially embed clinical information into a table and establish a corresponding relationship between input and output to form an input table;

[0092] The encoding layer 222 is connected to the table inserting module 221, and the encoding layer 222 performs one-hot encoding on the input table to form encoding information;

[0093] An embedding layer 223, the embedding layer 223 is connected to the encoding layer 222, and the embedding layer 223 generates an embedding vector according to the encoding information;

[0094] The second linear layer 224 is connected to the embedding layer 223, and the second linear layer 224 generates multimodal features according to the embedding vector.

[0095] Specifically, considering that the amount of clinical information collected in the electronic medical record system is large, in this embodiment, multiple table insertion modules 221 are set. The table insertion module 221 embeds the clinical information in a table in turn and establishes a corresponding relationship between input and output to form an input table. The table insertion module 221 refers to the IT layer. Interlated Tabular uses a mechanism called "in-place" to correspond input data with output data. This mechanism greatly improves the expressive power of the model and also introduces structured regularization technology to prevent overfitting of the model. Subsequently, the encoding layer 222 performs one-hot encoding on the structured input table to form encoding information, thereby realizing binary representation of multiple dimensions. After preprocessing the tabular data, the embedding layer 223 generates an embedding vector based on the encoding information, and then the second linear layer 224 linearizes the embedding vector to generate multimodal features, thereby realizing the encoding and extraction of clinical information.

[0096] In one embodiment, Figure 6 As shown, the fusion prediction module 23 includes:

[0097] A second convolutional layer 231, the second convolutional layer 231 receives the image features and the multimodal features respectively, and the second convolutional layer concatenates the image features and the multimodal features to extract a first fusion feature;

[0098] A third linear layer 232, the third linear layer 232 is connected to the second convolutional layer 231, and the third linear layer 232 linearly processes the first fused feature to obtain a second fused feature;

[0099] A Softmax function module 233, the Softmax function module 233 is connected to the third linear layer 232, and the Softmax function module 233 performs normalization mapping on each element in the second fusion feature to obtain a third fusion feature;

[0100] A third convolution layer 234, the third convolution layer 234 is connected to the Softmax function module 233, and the third convolution layer 234 extracts the fourth fusion feature after splicing the third fusion feature and the multimodal feature;

[0101] A fourth linear layer 235, the fourth linear layer 235 is connected to the Softmax function module 233 and the third convolutional layer 234, and the fourth linear layer 235 extracts the fifth fusion feature after splicing the third fusion feature and the image feature;

[0102] The fourth linear layer 235 generates an expansion risk according to the fifth fused feature and the fourth fused feature.

[0103] Specifically, in order to achieve effective prediction of the risk of expansion, in this solution, the second convolution layer 231 is first used to receive the image features and the multimodal features respectively, and the second convolution layer extracts the first fusion features after vectorizing and splicing the image features and the multimodal features, and then the third linear layer 232 linearly processes the first fusion features to obtain the second fusion features to avoid model overfitting, and then the Softmax function module 233 normalizes and maps the elements of multiple dimensions in the vectorized second fusion features to obtain the third fusion features, thereby completing the feature fusion process. Subsequently, based on the original input of the multimodal features, the third convolution layer 234 extracts the fourth fusion features through the convolution kernel after splicing the third fusion features and the multimodal features. For the original input of the image features, the fourth linear layer 235 extracts the fifth fusion features after splicing the third fusion features and the image features, and finally the fourth linear layer 235 classifies and generates the expansion risk according to the fifth fusion features and the fourth fusion features, thereby achieving effective prediction of the expansion risk.

[0104] In one embodiment, Figure 7 As shown, the second prediction module 3 includes:

[0105] A pre-segmentation module 31, the pre-segmentation module 31 segments the head MRI-DWI image through the codec layer to obtain an infarct core area labeled image;

[0106] A residual network 32, the residual network 32 is connected to the pre-segmentation module 31, and the residual network 32 extracts the annotated image of the infarct core area to obtain a multi-scale image patch;

[0107] A multi-scale MAP model 33, wherein the multi-scale MAP model 33 is connected to the residual network 32, and the multi-scale MAP model 33 generates multi-scale features according to the multi-scale image patch and the neighborhood pixels associated with the multi-scale image patch;

[0108] The estimation module 34 is connected to the multi-scale MAP model 33, and the estimation module 34 obtains the enlarged area according to the multi-scale feature prediction.

[0109] Specifically, in order to achieve a better prediction effect of the expanded area, in this solution, the head MRI-DWI image is first segmented by the pre-segmentation module 31 to obtain the annotated image of the infarct core area. This step is implemented by the U-shaped model structure of the usual convolution layer and upsampling layer. Subsequently, the annotated image of the infarct core area is extracted by the residual network 32 to obtain a multi-scale image patch. Specifically, the residual network includes a plurality of residual blocks, which are the core components of the network and can help the model learn more complex feature representations. Each residual block includes two or three convolution layers and a skip connection (Shortcut Connection) for identity mapping, which enables the network to increase the depth without adding additional parameters, thereby extracting more features, and then forming symmetric patch classifications of different scales as multi-scale image patches. Based on the multi-scale image patches, a multi-scale MAP model 33 can be established to integrate the classification results of patches of different sizes. In the process of integration, in order to facilitate subsequent predictions, the neighborhood information between patches is also added. The specific process of integration includes: taking the sum of the output results of the residual network 32 as the likelihood probability, and then defining the spatial neighborhood constraint function as the prior probability. Since the neighborhood constraint function considers the spatial dependency between adjacent pixels at each scale, the multi-scale classification results can be effectively combined. Finally, the estimation module 34 predicts the expanded area based on the multi-scale features to define whether the image prediction range of each AIS patient exceeds the MASK range of the initial infarction core area, and thus visualizes the prediction range of the infarction core area 48 hours after the initial diagnosis of the MRI-DWI image, automatically calculates the number of voxels in the mask, and outputs the predicted volume of the infarction core area.

[0110] In one embodiment, Figure 8 As shown, the pre-segmentation module 31 includes:

[0111] A plurality of sequentially connected convolutional layers 311, each convolutional layer 311 extracts features of an input image using a convolution kernel to form a downsampled image;

[0112] The input image of the first convolutional layer 311 is a head MRI-DWI image;

[0113] The input image of the next convolutional layer 311 is the downsampled image of the previous layer;

[0114] A plurality of skip structures 312 corresponding to the convolutional layer 311, the skip structure 312 collecting an input image of the corresponding convolutional layer as an original image feature;

[0115] A plurality of upsampling layers 313 connected in sequence, each upsampling layer 313 upsamples the input image and the original image features at the corresponding level to obtain an upsampled image;

[0116] The input image of the first upsampling layer 313 is the downsampled image of the last convolution layer;

[0117] The input image of the next upsampling layer 313 is the upsampling image of the previous level;

[0118] The last upsampling layer 313 outputs an infarct core area labeled image.

[0119] Specifically, in order to achieve a better segmentation process for the infarct core area, a U-shaped segmentation model is constructed in this embodiment, which mainly includes a downsampling process performed by multiple sequentially connected convolutional layers 311 and an upsampling process performed by multiple sequentially connected upsampling layers 313. The image segmentation is achieved through the downsampling and upsampling processes performed successively, and a better segmentation effect is achieved through the original image features of the jump structure input part.

[0120] As an optional implementation, in the pre-segmentation module 31, multiple sequentially connected convolutional layers 311 and the convolution module 211 of the image extraction module 21 in the first prediction model can be reused, that is, in the process of convolution of multiple sequentially connected convolutional layers 311, the output image features can be directly input into the first preprocessing module 212 on the corresponding level for processing as extracted image features, thereby reducing the scale of the overall model.

[0121] During the training process, MRI-DWI images of AIS patients diagnosed within 24 hours of onset and reexamination MRI-DWI images within 48 hours of admission and corresponding clinical information can be collected; at the same time, the patient's neurological function score (NIHSS score) is collected as annotation information. More than 300 AIS patients who met the inclusion criteria were screened, and during the model construction and verification process, the data were randomly divided into training set (dataset 1), verification set (dataset 2) and test set (dataset 3), with a ratio of 7:2:1.

[0122] For the sample images, the above preprocessing method is used to remove the skull and align them to obtain MRI-DWI images for training. Then the two models are trained separately. Data set 1 is used for training the above deep learning model, and data set 2 is used to select the model with the best performance and terminate the training process. Data set 3 is used to evaluate the generalization ability of the model.

[0123] The above are only preferred embodiments of the present invention, and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the description and illustrations of the present invention should be included in the protection scope of the present invention.

Claims

1. A prediction system for the risk of expansion of the infarct core area in acute ischemic stroke, characterized in that: include: An acquisition module, which acquires head MRI-DWI images and clinical information from the patient; A first prediction module, the first prediction module is connected to the acquisition module, and the first prediction module predicts the risk of expansion of the infarct core area according to the clinical information and the head MRI-DWI image; A second prediction module, wherein the second prediction module is connected to the acquisition module, and the second prediction module marks the enlarged area of ​​the infarct core area according to the head MRI-DWI image.

2. The prediction system according to claim 1, characterized in that The acquisition module comprises: An image acquisition module, wherein the image acquisition module is connected to an external MRI device to acquire the head MRI-DWI image output by the MRI device; A medical information collection module is connected to the hospital information system, and the medical information collection module collects the clinical information.

3. The prediction system according to claim 1, characterized in that The first prediction module uses a first prediction model to predict the expansion risk, and the first prediction model includes: An image extraction module, wherein the image extraction module extracts image features from the head MRI-DWI image; A multidimensional data extraction module, wherein the multidimensional data extraction module extracts multimodal features from the clinical information; A fusion prediction module is connected to the image extraction module and the multi-dimensional data extraction module respectively, and the fusion prediction module predicts the expansion risk according to the image features and the multi-modal features.

4. The prediction system according to claim 3, characterized in that The image extraction module comprises: A convolution module, wherein the convolution module extracts first image features at different levels from the head MRI-DWI image through a series of convolution layers; A plurality of first preprocessing modules, wherein the first preprocessing modules are respectively connected to the convolution layers at corresponding levels in the convolution modules to respectively obtain the first image features; The first preprocessing module flattens and linearizes the hierarchical image features to obtain second image features; a first convolution layer, wherein the first convolution layers are respectively connected to the first preprocessing modules, and the first convolution layer performs convolution processing on all the second image features to form second image features; A first linear layer, wherein the first linear layer is connected to the first convolutional layer, and the first linear layer performs linearization processing on the second image feature to obtain the image feature.

5. The prediction system according to claim 3, characterized in that: The multidimensional data extraction module comprises: A plurality of table insertion modules, wherein the table insertion modules sequentially embed the clinical information into a table and establish a corresponding relationship between input and output to form an input table; An encoding layer, the encoding layer is connected to the table insertion module, and the encoding layer performs one-hot encoding on the input table to form encoding information; An embedding layer, the embedding layer is connected to the encoding layer, and the embedding layer generates an embedding vector according to the encoding information; A second linear layer, wherein the second linear layer is connected to the embedding layer, and the second linear layer generates the multimodal feature according to the embedding vector.

6. The prediction system according to claim 3, characterized in that: The fusion prediction module includes: A second convolutional layer, wherein the second convolutional layer receives the image feature and the multimodal feature respectively, and the second convolutional layer concatenates the image feature and the multimodal feature to extract a first fusion feature; A third linear layer, the third linear layer is connected to the second convolutional layer, and the third linear layer linearly processes the first fused features to obtain second fused features; A Softmax function module, wherein the Softmax function module is connected to the third linear layer, and the Softmax function module performs normalization mapping on each element in the second fusion feature to obtain a third fusion feature; A third convolution layer, the third convolution layer is connected to the Softmax function module, and the third convolution layer extracts a fourth fusion feature after splicing the third fusion feature and the multimodal feature; A fourth linear layer, wherein the fourth linear layer is connected to the Softmax function module and the third convolutional layer, and the fourth linear layer extracts a fifth fusion feature after splicing the third fusion feature and the image feature; The fourth linear layer generates the expansion risk according to the fifth fused feature and the fourth fused feature.

7. The prediction system according to claim 1, characterized in that The second prediction module comprises: A pre-segmentation module, wherein the pre-segmentation module segments the head MRI-DWI image through a coding and decoding layer to obtain an infarct core area labeled image; A residual network, wherein the residual network is connected to the pre-segmentation module, and the residual network extracts the annotated image of the infarct core area to obtain a multi-scale image patch; A multi-scale MAP model, the multi-scale MAP model is connected to the residual network, and the multi-scale MAP model generates multi-scale features according to the multi-scale image patch and neighborhood pixels associated with the multi-scale image patch; An estimation module, wherein the estimation module is connected to the multi-scale MAP model, and the estimation module predicts the expansion area according to the multi-scale features.

8. The prediction system according to claim 7, characterized in that The pre-segmentation module comprises: A plurality of sequentially connected convolutional layers, each of which extracts features from an input image using a convolution kernel to form a downsampled image; The input image of the first convolutional layer is the head MRI-DWI image; The input image of the convolutional layer of the next level is the downsampled image of the previous level; A plurality of skip structures corresponding to the convolutional layer, wherein the skip structures collect the input image of the corresponding convolutional layer as original image features; A plurality of upsampling layers, each of which upsamples the input image and the original image features at a corresponding level to obtain an upsampled image; The input image of the upsampling layer of the first level is the downsampling image of the convolutional layer of the last level; The input image of the upsampling layer of the next level is the upsampling image of the previous level; The upsampling layer at the last level outputs the annotated image of the infarct core area.

Images