Deep learning-based stroke prognosis prediction method and system

Abstract

The invention discloses a deep learning-based stroke prognosis prediction method and system. The method comprises the steps of obtaining a head magnetic resonance medical image data set of a stroke patient before treatment and a corresponding mRS score of stroke prognosis; segmenting a core infarction focus image from the obtained magnetic resonance medical image by adopting a DWI sequence; establishing a six-score cSVD load scoring standard, scoring the core infarction focus image, and constructing a model data set; 3, constructing a prediction model for predicting the stroke prognosis mRS score, dividing the model data set in the step 3 into a training set, a verification set and a test set, training the prediction model by adopting the training set, verifying the prediction model on the verification set, testing by adopting the test set, screening out the prediction model of which the precision meets the requirement, and determining the stroke prognosis mRS score. And predicting the mRS score of the stroke prognosis of the to-be-evaluated patient by adopting the screened prediction model. According to the method, the prediction capability of the image features obtained by MRI (Magnetic Resonance Imaging) flat scanning of the LAA-AIS patient on prognosis is fully excavated, and a new reference direction is provided for prognosis evaluation of the LAA-AIS patient.

Description

Technical Field

[0001] The present invention belongs to the technical field of clinical medical image prediction, and specifically relates to a method and system for predicting stroke prognosis based on deep learning. Background Art

[0002] Large artery atherosclerosis acute ischemic stroke (LAA-AIS) is one of the most important types of stroke, characterized by severe clinical symptoms and poor prognosis. Therefore, prognostic studies on this type of patients are particularly important. With the advancement of imaging technology, imaging plays an important role in the assessment of vascular stenosis, the degree of brain tissue ischemia, the selection of patient treatment options, and the prediction of prognosis in patients with LAA-AIS.

[0003] Currently, the imaging assessment widely used in the LAA-AIS patient population is multimodal computed tomography (CT) examination sequences, such as one-stop CT angiography (CTA) and CT perfusion (CTP) imaging techniques. These techniques have been widely used in the detection of AIS patients. For AIS patients, CTA has become the preferred method for large vessel stenosis or occlusion, among which multiphase CTA (mCTA) has many advantages over traditional single-phase CTA (sCTA). The mCTA imaging scheme is to collect images of three phases: arterial peak phase, venous peak phase, and late venous phase after peripheral intravenous bolus injection of contrast agent. The mCTA method can optimize the detection of large vessel occlusion, more quickly and intuitively evaluate the status of collateral circulation in LAA-AIS patients, and improve patients' tolerance to exercise and poor hemodynamics, thereby improving the reliability between raters

[110] . CTP imaging can automatically calculate the patient's cerebral blood flow (CBF), cerebral blood volume (CBV), and time to top (Tmax) of contrast agents to all tissues using software. Currently, many multicenter studies, such as DAWN and DEFUSE 3, use RAPID software to automatically process CTP data and screen out AIS patients whose time of onset is beyond the traditional treatment window or whose onset time is unknown based on the calculated parameters for further intravascular treatment [111,78]. However, CT examinations have inherent disadvantages: radiation and contrast agent uptake. With the development of magnetic resonance imaging (MRI) technology, the scanning time of conventional MRI has been significantly shortened. In particular, the scanning time of the diffusion weighted imaging (DWI) sequence, which is most sensitive to acute infarction lesions, has been controlled to less than one minute. In previous studies on the prognosis of AIS, clinical signs were important indicators for predicting prognosis, such as the NIHSS score, but the score is subjective and unstable. At present, the scanning time and technical parameters of magnetic resonance imaging have reached the requirements of emergency assessment, and the simple use of DWI sequences can provide more confidence in clinical diagnosis and treatment. In the analysis of MR DWI sequences, the most widely used method is to use commercial software to extract the area with restricted diffusion in the DWI sequence. This area is used as the core infarct volume. The volume is combined with clinical data scores, and the traditional logistic regression (LR) method is used to establish a prognosis prediction model for AIS patients.However, this method ignores the image and location characteristics of the lesion itself, as well as other related imaging markers of intracranial cerebral small vessel disease, such as white matter hyperintensities and microbleeding foci. These features have a great impact on the prognosis of AIS patients, so the accuracy of predicting the prognosis of AIS patients using existing methods is low. Therefore, it is necessary to develop a more accurate prediction method. Summary of the invention

[0004] The purpose of the present invention is to address the shortcomings of the existing technology and provide a prediction method for stroke prognosis based on deep learning. This method can fully tap the predictive ability of image features obtained from MRI plain scans of LAA-AIS patients for prognosis, improve the prediction accuracy, and provide a new reference direction for the prognosis assessment of LAA-AIS patients.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0006] A method for predicting stroke prognosis based on deep learning, comprising the following steps:

[0007] Step 1: Obtain a head magnetic resonance medical image dataset of stroke patients before treatment and the corresponding mRS score of stroke prognosis;

[0008] Step 2: Segment the core infarct lesion image using diffusion-weighted DWI sequence on the acquired magnetic resonance medical image;

[0009] Step 3, establish a 6-point total chronic lacunar isotonic cerebral small vessel disease cSVD burden scoring standard and score the core infarct lesion image segmented in step 2 according to the scoring standard to obtain a core infarct lesion image labeled with a total cSVD burden score, and construct the core infarct lesion image labeled with a total cSVD burden score and the corresponding mRS score of stroke prognosis as a model data set;

[0010] Step 4: Construct a prediction model for predicting the mRS score of stroke prognosis. Divide the model data set of step 3 into a training set, a validation set, and a test set. Use the training set to train the prediction model and verify it on the validation set. Use the test set to test it. Screen out the prediction model with the required accuracy. Use the screened prediction model to predict the mRS score of the stroke prognosis of the patient to be evaluated.

[0011] Furthermore, step 1 screens out eligible stroke patients from the MR-STARS database. The screening criteria for stroke patients are:

[0012] (1) First acute anterior circulation stroke;

[0013] (2) The time from onset to examination was less than 24 hours, and the patient received a standard magnetic resonance imaging (MRI) examination sequence;

[0014] (3) the presence of stenosis or occlusion of large vessels in the anterior circulation;

[0015] (4) The patients only received routine standardized medical treatment.

[0016] Furthermore, the prognostic mRS score includes 0-6, and the patients are divided into two groups according to the mRS score: good prognosis and poor prognosis. Among them, the mRS score of 0-2 indicates a good prognosis, and the mRS score of 3-6 indicates a poor prognosis.

[0017] Furthermore, in step 2, the apparent diffusion coefficient map ADC of the DWI sequence is extracted from the magnetic resonance medical image, the ADC map is analyzed and the DWI infarct volume is calculated, and the ADC value in the map < 620 μm 2 The area with a diameter of 0.1 mm / s was determined as the core infarct lesion, and the infarct volume was calculated and the core infarct lesion image was extracted.

[0018] Furthermore, the 6-point total cSVD burden scoring standard constructed in step 3 is:

[0019] (1) The presence of intracranial lacunar lesions was scored as 1 point;

[0020] (2) If the number of intracranial microbleeds (CMBs) is 1 to 4, 1 point is given; if the number of CMBs is ≥5, 2 points are given;

[0021] (3) the number of PVS in the intracranial basal ganglia was ≥20, scored 1 point;

[0022] (4) If the sum of the intracranial P-WMH and D-WMH scores is 3-4 points, it is scored as 1 point; if the sum of the intracranial P-WMH and D-WMH scores is 5-6 points, it is scored as 2 points.

[0023] Furthermore, the grading criteria for P-WMH, i.e., paraventricular WMH, are as follows:

[0024] (1) 0 points = normal, no lesions were found on the T2 FLAIR sequence or the lesion diameter was <3 mm;

[0025] (2) 1 point = mild, the lesion is pen-cap or linear;

[0026] (3) 2 points = moderate, the lesion showed a halo sign or a smooth band shape;

[0027] (4) 3 points = severe: the lesion is irregular in shape and extends into the deep white matter; the lesion diameter is >10 mm.

[0028] Furthermore, the grading criteria for D-WMH (deep WMH) are as follows:

[0029] (1) 0 points = normal, no lesions were found in the T2FLAIR sequence or the lesion diameter was <3 mm;

[0030] (2) 1 point = mild, the lesions are dot-shaped or focal-shaped, with a single diameter between 3 and 10 mm;

[0031] (3) 2 points = moderate, the lesions began to fuse, or the diameter of a single lesion was between 10 and 20 mm;

[0032] (4) 3 points = severe, large lesions or fused lesions with a diameter > 20 mm are seen in the deep white matter.

[0033] Furthermore, in step 4, a ResNet 50 neural network is used to construct a prediction model for predicting the mRS score of stroke prognosis. The ResNet 50 neural network includes seven parts, among which the first part is used to initialize the input image, including a convolutional layer, a batch normalization layer, an activation function layer, and a maximum pooling layer; the second to fifth parts are residual block calculations, the sixth part is an average pooling layer for reducing the dimension of the feature vector, and the seventh part is a linear layer for outputting the model prediction results.

[0034] Furthermore, in step 4, the model data set is randomly divided into modeling samples and test sets, and then the modeling samples are divided into training sets and validation sets using the five-fold cross-validation CV method, that is, the modeling data set is divided into five parts of equal size, four of which are used as training sets and one as a validation set each time, and multiple cross-validations are performed. After that, the same five-fold cross-validation method is used in the test set to obtain the sensitivity, specificity, precision, accuracy, and AUC value of the prediction model, and the multiple evaluation results are averaged, and finally the model parameters with the highest accuracy are screened out as the final parameters of the prediction model.

[0035] Another object of the present invention is to provide a system according to the above-mentioned method for predicting stroke prognosis based on deep learning, comprising:

[0036] An image data acquisition module is used to obtain a head magnetic resonance medical image dataset of stroke patients before treatment and an mRS score for stroke prognosis;

[0037] A core infarct lesion image extraction module is used to segment the magnetic resonance medical image obtained by the image data acquisition module into a core infarct lesion image using a DWI sequence;

[0038] The total cSVD load scoring module is used to establish a 6-point total cSVD load scoring standard and score the core infarct lesion images segmented by the core infarct lesion image extraction module according to the scoring standard to obtain a core infarct lesion image dataset with annotated scores;

[0039] The model building and prediction module is used to build a prediction model for the mRS score of stroke prognosis, and divide the data set in the total cSVD load score module into a training set, a validation set, and a test set. The prediction model is trained with the training set and verified on the validation set, and then tested with the test set to screen out the prediction model with the required accuracy. The screened prediction model is used to predict the mRS score of the stroke prognosis of the patient to be evaluated.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention uses diffusion-weighted DWI sequences to segment core infarction lesion images and fuses the two indicators of a 6-point total cSVD load score to establish a prediction model, which fully considers the association between cSVD load and stroke recurrence, secondary intracranial hemorrhage and poor prognosis of LAA-AIS patients, and effectively mines the image features obtained by MRI plain scan of LAA-AIS patients, extracts high-dimensional image features of acute infarction lesions, which can also avoid the drawbacks of evaluating patient prognosis only by infarction volume in the past, thereby greatly improving the prediction accuracy of the prediction model, and at the same time proves that the two indicators of conventional MRI information and a 6-point total cSVD load score can effectively predict the prognosis of AIS patients, providing a basis and reference value for medical workers to carry out further treatment work, thereby reducing the workload of medical workers. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a flowchart of a method for predicting stroke prognosis based on deep learning in an embodiment of the present invention.

[0042] Figure 2 The overall performance of the four models in the embodiments of the present invention in predicting the prognosis of 157 patients is shown in the ROC curve: (a) Model 1, (b) Model 2, (c) Model 3, and (d) Model 4. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present invention will be described clearly and completely below in combination with 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.

[0044] 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.

[0045] The present invention will be further described below in conjunction with specific embodiments, but the present invention is not limited thereto.

[0046] like Figure 1As shown, the embodiment of the present invention discloses a method for predicting stroke prognosis based on deep learning, comprising the following steps:

[0047] Step 1: Obtain a dataset of head magnetic resonance medical images of stroke patients before treatment and mRS scores for stroke prognosis;

[0048] The study population of this embodiment comes from a multicenter database (MR-based Stroke Mechanism and Future Risk Score, MR-STARS) established by multiple medical centers led by the General Hospital of the People's Liberation Army from August 2015 to August 2021, with a clinical trial registration number (Clinicaltrials.gov, NCT02580097). MR-STARS uses AIS patients as research subjects (onset time is less than 24h), and deeply explores the pathogenesis of AIS, the choice of treatment options, and the early prediction of prognosis of AIS patients. The study was approved by the ethics committees of multiple research centers, and all participants or their families signed written informed consent.

[0049] The patients screened in this example were all from eligible LAA-AIS patients in the MR-STARS database, where the screening criteria were: (1) first acute anterior circulation stroke; (2) the time from onset to examination was less than 24 hours, and the patients received standard MRI examination sequences, including: T2WI, T2FLAIR, DWI, SWI, MRA, etc.; (3) MRA clearly found stenosis or occlusion of large vessels in the anterior circulation; (4) the patients only received routine standardized medical treatment; the criteria for excluding patients were: (1) incomplete baseline data; (2) any DWI sequence that could not be manually segmented due to motion artifacts or cerebral hemorrhage; (3) previous infarction and hemorrhagic lesions were found in the MRI sequence; (4) when the patient was diagnosed with hereditary cSVD, such as autosomal dominant arteriopathy with subcortical infarcts and leukoencephalopathy (5) accompanied by systemic diseases, such as renal failure, severe heart failure, severe liver damage, etc.

[0050] The mRS score of the functional prognosis of each LAA-AIS patient at 90 days was obtained from the MR-STARS database. The prognosis mRS score includes 0-6. In this embodiment, the patients were divided into two groups according to the mRS score: a good prognosis group and a poor prognosis group, wherein an mRS score of 0-2 indicates a good prognosis group and an mRS score of 3-6 indicates a poor prognosis group.

[0051] Step 2: Segment the core infarct lesion image using diffusion-weighted DWI sequence on the acquired magnetic resonance medical image;

[0052] The infarct volume of LAA-AIS patients was measured using the apparent diffusion coefficient (ADC) map of the DWI sequence. DWI infarct volume analysis was performed using NeuBrainCARE software (Beta1.2.10.24, iSchemaView). By analyzing the ADC map, it was determined that the ADC value within the map was <620 μm. 2 The area with a size of 1 / 2 is the core infarct area. The core infarct area is extracted and the core infarct volume is calculated. The results automatically output by the software are further analyzed and checked by radiologists. If there are large differences, the lesions can be manually segmented and the segmented lesions can be re-entered into the software to correct the output results.

[0053] Step 3, establish a 6-point total chronic lacunar isotonic cerebral small vessel disease cSVD burden scoring standard and score the core infarct lesion image segmented in step 2 according to the scoring standard to obtain a core infarct lesion image labeled with a total cSVD burden score, and construct the core infarct lesion image labeled with a total cSVD burden score and the corresponding mRS score of stroke prognosis as a model data set;

[0054] This example establishes a 6-point total chronic lacunar cerebral small vessel disease (cSVD) burden scoring standard, which mainly includes four types of clinically common chronic cSVD markers, namely white matter hyperintensity (WMH), perivascular space (PVS), cerebral microbleed (CMB) and lacunar lesions, for evaluation, as follows:

[0055] Ⅰ. WMH: Lesions usually occur in deep white matter areas, such as the bilateral periventricular areas, the radiating corona area, or the semiovale center. The signal changes are dot-like or patchy with high signals in T2WI and T2FLAIR sequences. The WMH scoring method in this embodiment adopts the Fazekas grading method. According to the location of WMH, WMH is divided into two categories: periventricular WMH (periventricular WMH, P-WMH) and deep WMH (deep WMH, D-WMH) for evaluation. The specific grading standards are as follows:

[0056] P-WMH: (1) 0 points = normal, no lesions were found on the T2 FLAIR sequence or the lesion diameter was <3 mm; (2) 1 point = mild, the lesion was pen-cap-like or linear; (3) 2 points = moderate, the lesion was "halo"-like or smooth band-like; (4) 3 points = severe, the lesion was irregular in shape and extended into the deep white matter, and the lesion diameter was >10 mm;

[0057] D-WMH: (1) 0 points = normal, no lesions were found in the T2FLAIR sequence or the lesion diameter was <3 mm; (2) 1 point = mild, the lesions were dot-shaped or focal-shaped, and the diameter of a single lesion was between 3 and 10 mm; 3) 2 points = moderate, the lesions began to fuse, or the diameter of a single lesion was between 10 and 20 mm; (4) 3 points = severe, large lesions or fused lesions were seen in the deep white matter, with a diameter >20 mm.

[0058] Ⅱ. PVS: The signal of PVS is consistent with that of cerebrospinal fluid in MRI sequence, and the size is usually less than 3mm; it is located in the semioval center, basal ganglia, brainstem and hippocampus; the border of PVS is clear and smooth, mostly linear in the semioval area, and mostly round and oval in the basal ganglia area. The shape characteristics of PVS in different parts are related to the distribution of perforating arteries. PVS needs to be differentiated from old intracranial cavities, which are usually larger than 3mm in size and have less smooth borders. High signal rings (proliferating glial cells) are common in T2FLAIR sequences. In this study, the patients included in the study were patients with anterior circulation vascular stenosis, so the central semiovale PVS (CS-PVS) and basal ganglia PVS (BG-PVS) were evaluated separately. PVS was graded according to the 4-point scoring method commonly used in clinical practice: 1 point = 0-10, 2 points = 11-20; 3 points = 21-40, 4 points = >40 or uncountable sieve-like changes. The PVS was counted at the level with the largest number. If the number of bilateral PVS was asymmetric, the number of PVS on the side with the most severe lesions was taken.

[0059] Ⅲ. Chronic lacunar lesions: usually appear as round or oval lesions with a diameter of 3-15 mm. The lesions are mostly located in the basal ganglia, internal capsule, centrum semiovale or brainstem. Lacunar lesions show cerebrospinal fluid-like low signal in T2 FLAIR sequence, surrounded by high signal ring, and low signal in DWI sequence. Lacunar lesions are often associated with old infarction or microbleeding in the area supplied by intracranial perforating arteries. It needs to be differentiated from PVS.

[0060] IV. CMB: CMB appears as a small round or oval signal cavity on the SWI sequence, usually 2-5 mm in diameter, sometimes up to 10 mm. Iron or calcium deposition should be excluded during the SWI evaluation.

[0061] The 6-point total cSVD burden scoring standard constructed according to the above evaluation method is:

[0062] (1) The presence of intracranial lacunar lesions was scored as 1 point;

[0063] (2) 1 to 4 intracranial CMBs, 1 point; 5 or more CMBs, 2 points;

[0064] (3) the number of intracranial BG-PVS was ≥20, scored 1 point;

[0065] (4) If the sum of the intracranial P-WMH and D-WMH scores is 3-4 points, it is scored as 1 point; if the sum of the intracranial P-WMH and D-WMH scores is 5-6 points, it is scored as 2 points.

[0066] Step 4: construct a prediction model for predicting the mRS score of stroke prognosis, divide the model data set of step 3 into a training set, a validation set and a test set, use the training set to train the prediction model and verify it on the validation set, and then use the test set to test it, screen out the prediction model with the accuracy that meets the requirements, and use the screened prediction model to predict the mRS score of the stroke prognosis of the patient to be evaluated;

[0067] This embodiment uses the ResNet 50 neural network to construct a prediction model for predicting the mRS score of stroke prognosis. The core idea of ​​ResNet is to add residual blocks to CNN. When the depth of CNN increases, the loss function of the training set gradually decreases to a saturation value. When the number of network layers continues to increase, the loss function will not always tend to be stable, but will gradually increase, and the gradient will disappear. The introduction of residual blocks can set the weight parameters of these layers to 0, so that the input is directly used as the initial result of the output, thereby skipping this part.

[0068] The ResNet 50 of this embodiment has a depth of 50 layers, and a neural network including seven parts can be abstracted. The first part is mainly used to initialize the input image, including a convolution layer, a batch normalization layer, an activation function layer, and a maximum pooling layer, and finally outputs a 56×56 feature matrix. The second to fifth parts are residual block calculations, including two residual block structures, Res_Block_1 and Res_Block_2, respectively. The second part consists of 1 Res_Block_1 residual block and 2 Res_Block_2 residual blocks, the third part consists of 1 Res_Block_1 residual block and 3 Res_Block_2 residual blocks, the fourth part consists of 1 Res_Block_1 residual block and 5 Res_Block_2 residual blocks, and the fifth part consists of 1 Res Block1 residual block and 2 Res Block 2 residual blocks. The left side of Res Block 1 consists of two 1×1 convolutional layers and one 3×3 convolutional layer. The structure first reduces the input feature vector to 64 dimensions through a 1×1 convolutional layer, then learns the features through a 3×3 convolutional layer, and finally restores the feature vector to 256 dimensions through a 1×1 convolutional layer to fit the residual function; the right side is a 1×1 convolution operation on the input matrix, and the new feature matrix is ​​output after adding both sides. The Res_Block_2 residual block is similar to Res_Block_1, but the right side is an identity mapping of the input feature vector, so that the matrix dimensions of the input and output remain unchanged, which is convenient for multiple residual blocks to be connected in series. The sixth part is followed by an average pooling layer, which is used to reduce the dimension of the feature vector. Finally, in the seventh part, the linear layer outputs the model prediction results.

[0069] In this embodiment, 137 randomly divided patient data are used as modeling samples, and the remaining 20 patients are used as test sets for external cohorts. In order to avoid overfitting of model training due to a small sample size, this embodiment uses a five-fold cross validation (cross validation, CV) method to divide the modeling sample into a training set and a validation set, eliminating the impact on the model caused by the imbalance of a single data set division. This method divides the modeling sample into 5 parts of the same size, 4 of which are used as training sets and 1 as a validation set each time, and cross-validates 5 times. Finally, the same 5-fold cross validation method is used in the external test set to obtain the sensitivity, specificity, precision, accuracy, and average area under the receiver operating characteristic curve (receiver operator curve, ROC) of the prediction model, and the multiple evaluation results are averaged, and finally the model parameters with the highest accuracy are selected as the final parameters of the prediction model. According to the final parameters, a prediction model for predicting the mRS score of the patient's stroke prognosis is constructed, and the prediction model is used to predict the mRS score of the patient's stroke prognosis to be evaluated.

[0070] In order to illustrate the effect of this embodiment, the prediction model (model 4) established in this embodiment is compared with the prediction accuracy of the model established by using the logistic regression LR method to include the infarct core volume index (model 1), the model established by using the LR method to include the two indicators of infarct core volume and initial NIHSS score on admission (model 2), and the prediction model established by using the deep learning DL method based on the image features of the core infarct lesions segmented by the DWI sequence (model 3). The comparison results are shown in Tables 1 and Figure 2 shown.

[0071] Table 1 Results of applying the four models to the entire sample population (157 patients)

[0072]

[0073] The results in Table 1 show that the AUC value of model 1 is 0.77 [0.69-0.84]; the AUC value of model 2 is 0.83 [0.77-0.90], the AUC value of model 3 is 0.84 [0.78-0.90]; the AUC value of model 4 is 0.89 [0.84-0.94]. Figure 2 The results show that model 4 has the best application effect in all populations. Therefore, using this prediction model in the prognosis prediction of stroke patients can greatly improve the accuracy of the prediction, provide a basis and reference value for medical workers to carry out further work, and reduce the workload of medical workers.

[0074] The embodiment of the present invention further provides a system for predicting stroke prognosis based on deep learning, comprising:

[0075] An image data acquisition module is used to obtain a head magnetic resonance medical image dataset of stroke patients before treatment and an mRS score for stroke prognosis;

[0076] A core infarct lesion image extraction module is used to segment the magnetic resonance medical image obtained by the image data acquisition module into a core infarct lesion image using a DWI sequence;

[0077] The total cSVD load scoring module is used to establish a 6-point total cSVD load scoring standard and score the core infarct lesion images segmented by the core infarct lesion image extraction module according to the scoring standard to obtain a core infarct lesion image dataset with annotated scores;

[0078] The model building and prediction module is used to build a prediction model for the mRS score of stroke prognosis, and divide the data set in the total cSVD load score module into a training set, a validation set, and a test set. The prediction model is trained with the training set and verified on the validation set, and then tested with the test set to screen out the prediction model with the required accuracy. The screened prediction model is used to predict the mRS score of the stroke prognosis of the patient to be evaluated.

[0079] 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 contents of the specification of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting stroke prognosis based on deep learning, characterized in that: The steps include: Step 1: Obtain a head magnetic resonance medical image dataset of stroke patients before treatment and the corresponding mRS score of stroke prognosis; Step 2: Segment the core infarct lesion image using diffusion-weighted DWI sequence on the acquired magnetic resonance medical image; Step 3, establish a 6-point total chronic lacunar isotonic cerebral small vessel disease cSVD burden scoring standard and score the core infarct lesion image segmented in step 2 according to the scoring standard to obtain a core infarct lesion image labeled with a total cSVD burden score, and construct the core infarct lesion image labeled with a total cSVD burden score and the corresponding mRS score of stroke prognosis as a model data set; Step 4: Construct a prediction model for predicting the mRS score of stroke prognosis. Divide the model data set of step 3 into a training set, a validation set, and a test set. Use the training set to train the prediction model and verify it on the validation set. Use the test set to test it. Screen out the prediction model with the required accuracy. Use the screened prediction model to predict the mRS score of the stroke prognosis of the patient to be evaluated.

2. The method for predicting stroke prognosis based on deep learning according to claim 1, characterized in that: Step 1: Screen eligible stroke patients from the MR-STARS database. The screening criteria for stroke patients are: (1) First acute anterior circulation stroke; (2) The time from onset to examination was less than 24 hours, and the patient received a standard magnetic resonance imaging (MRI) examination sequence; (3) the presence of stenosis or occlusion of large vessels in the anterior circulation; (4) The patients only received routine standardized medical treatment.

3. The method for predicting stroke prognosis based on deep learning according to claim 1, characterized in that: The prognostic mRS score ranges from 0 to 6. Patients are divided into two groups according to the mRS score: good prognosis and poor prognosis. An mRS score of 0 to 2 indicates a good prognosis, and an mRS score of 3 to 6 indicates a poor prognosis.

4. The method for predicting stroke prognosis based on deep learning according to claim 1, characterized in that: In step 2, the apparent diffusion coefficient map ADC of the DWI sequence is extracted from the magnetic resonance medical image, the ADC map is analyzed and the DWI infarct volume is calculated, and the ADC value <620μm in the map is 2 The area with a diameter of 0.1 mm / s was determined as the core infarct lesion, and the infarct volume was calculated and the core infarct lesion image was extracted.

5. The method for predicting stroke prognosis based on deep learning according to claim 1, characterized in that: The 6-point total cSVD burden score constructed in step 3 is: (1) The presence of intracranial lacunar lesions was scored as 1 point; (2) If the number of intracranial microbleeds (CMBs) is 1 to 4, 1 point is given; if the number of CMBs is ≥5, 2 points are given; (3) the number of PVS in the intracranial basal ganglia was ≥20, scored 1 point; (4) If the sum of the intracranial P-WMH and D-WMH scores is 3-4 points, it is scored as 1 point; if the sum of the intracranial P-WMH and D-WMH scores is 5-6 points, it is scored as 2 points.

6. The method for predicting stroke prognosis based on deep learning according to claim 5, characterized in that: The grading criteria for P-WMH, i.e., paraventricular WMH, are as follows: (1) 0 points = normal, no lesions were found on the T2 FLAIR sequence or the lesion diameter was <3 mm; (2) 1 point = mild, the lesion is pen-cap or linear; (3) 2 points = moderate, the lesion showed a halo sign or a smooth band shape; (4) 3 points = severe: the lesion is irregular in shape and extends into the deep white matter; the lesion diameter is >10 mm.

7. The method for predicting stroke prognosis based on deep learning according to claim 5, characterized in that: The grading criteria for D-WMH (deep WMH) are as follows: (1) 0 points = normal, no lesions were found in the T2FLAIR sequence or the lesion diameter was <3 mm; (2) 1 point = mild, the lesions are dot-shaped or focal-shaped, with a single diameter between 3 and 10 mm; (3) 2 points = moderate, the lesions began to fuse, or the diameter of a single lesion was between 10 and 20 mm; (4) 3 points = severe, large lesions or fused lesions with a diameter > 20 mm are seen in the deep white matter.

8. The method for predicting stroke prognosis based on deep learning according to claim 1, characterized in that: In step 4, the ResNet 50 neural network is used to construct a prediction model for predicting the mRS score of stroke prognosis. The ResNet 50 neural network consists of seven parts, among which the first part is used to initialize the input image, including the convolution layer, batch normalization layer, activation function layer and maximum pooling layer; the second to fifth parts are residual block calculations, the sixth part is the average pooling layer, which is used to reduce the dimension of the feature vector, and the seventh part is the linear layer, which is used to output the model prediction results.

9. The method for predicting stroke prognosis based on deep learning according to claim 1, characterized in that: In step 4, the model data set is randomly divided into modeling samples and test sets, and then the modeling samples are divided into training sets and validation sets using the five-fold cross-validation CV method, that is, the modeling data set is divided into five parts of equal size, four of which are used as training sets and one as validation sets each time, and multiple cross-validations are performed. After that, the same five-fold cross-validation method is used in the test set to obtain the sensitivity, specificity, precision, accuracy, and AUC value of the prediction model, and the multiple evaluation results are averaged, and finally the model parameters with the highest accuracy are screened out as the final parameters of the prediction model.

10. A system for predicting stroke prognosis based on deep learning according to any one of claims 1 to 9, characterized in that: include: An image data acquisition module is used to obtain a head magnetic resonance medical image dataset of stroke patients before treatment and an mRS score for stroke prognosis; A core infarct lesion image extraction module is used to segment the magnetic resonance medical image obtained by the image data acquisition module into a core infarct lesion image using a DWI sequence; The total cSVD load scoring module is used to establish a 6-point total cSVD load scoring standard and score the core infarct lesion images segmented by the core infarct lesion image extraction module according to the scoring standard to obtain a core infarct lesion image dataset with annotated scores; The model building and prediction module is used to build a prediction model for the mRS score of stroke prognosis, and divide the data set in the total cSVD load score module into a training set, a validation set, and a test set. The prediction model is trained with the training set and verified on the validation set, and then tested with the test set to screen out the prediction model with the required accuracy. The screened prediction model is used to predict the mRS score of the stroke prognosis of the patient to be evaluated.

Images