A method, device, terminal and storage medium for establishing a large vessel occlusion identification model

Abstract

The present application discloses a method, device, terminal and storage medium for establishing a large vessel occlusion recognition model. The method includes the following steps: obtaining a standard CT image of the brain and performing angiography to obtain angiography image sets of multiple stages respectively; preprocessing the angiography image sets of multiple stages, and then performing vascular segmentation processing to generate two-dimensional maximum intensity projection images of multiple stages; creating a training set combination to obtain multiple groups of model training sets and performing deep learning model training one by one to obtain multiple deep learning models; selecting a deep learning model with the best evaluation result as a large vessel occlusion recognition model. The present application greatly improves the diagnostic detection rate of LVO, not only optimizing the workflow of radiologists in detecting LVO, but also screening out patients who meet the conditions for thrombectomy.

Description

Technical Field

[0001] The present application relates to a scheme for establishing a large vessel occlusion identification model, specifically, a method, device, terminal and storage medium for establishing a large vessel occlusion identification model based on multi-phase CTA, belonging to the cross-technical field of artificial intelligence and medical analysis. Background Art

[0002] Acute ischemic stroke (AIS) is caused by thrombotic or embolic occlusion of a cerebral artery. Blockage of a proximal cerebral artery causes deep ischemia, leading to a collapse of cellular energetics, followed by cell death within minutes. The infarct core is surrounded by an ischemic penumbra, where neurons will die within hours. Accurate identification of this "at-risk tissue" allows patients to benefit from treatment. In cases of massive ischemic stroke, the extent of damage will increase over the next few days. In extreme cases, the mass effect coupled with tissue damage can lead to increased intracranial pressure and loss of cerebral blood circulation.

[0003] Large vessel occlusion (LVO) stroke is one of the most time-sensitive diagnoses in medicine, and the therapeutic effect of endovascular therapy is closely related to time. Opening the embolized artery can restore blood flow to the tissue. The basic goal of acute cerebral infarction treatment is to limit the extent of the infarction by quickly and permanently recanalizing the blocked artery.

[0004] The main imaging methods for diagnosing AIS are non-contrast computed tomography (NCCT), CT perfusion (CTP), and computed tomography angiography (CTA). The advantages of CT-based methods are that they are faster and more widely used than MRI in emergency situations.

[0005] NCCT is highly sensitive for hemorrhagic lesions, and a key role for NCCT is also to detect ischemic signs in established ischemic lesions. The main finding of NCCT is a subcortical area of ​​low attenuation within the vascular territory. However, the sensitivity of NCCT is relatively low within the first 24 hours, especially within the limited time window of thrombolytic therapy.

[0006] The extent of acute infarction is usually estimated by CT perfusion parameters (mean transit time, cerebral blood volume, and cerebral blood flow). In CTP, cerebral blood flow and brain volume are altered in patients with acute stroke. However, the coverage of the brain by CTP is still limited on many detector scanners, which is generally considered a disadvantage of CTP. Newer 256-slice scanners can provide whole-brain coverage, and this coverage is likely to become widely used in the next few years. For example, CTP implies relatively higher radiation exposure, coarser resolution, and thicker slices compared with CTA.

[0007] The use of CTA allows the visualization of the anatomy of the aortic arch, carotid arteries, and cerebral arteries, as well as the presence of carotid or cerebral artery stenosis or thrombus. Among CT techniques, CTA has several advantages, and it is highly specific and sensitive than NCCT in detecting early irreversible ischemia. In addition, it offers the possibility of evaluating the entire cerebral vascular system, which is not available with CT perfusion.

[0008] Occlusion sites in CTA are determined by radiologists who read medical images, so radiological LVO detection and treatment are subject to variable delays and human expertise. Since penumbra tissue can be converted to infarction at a rate of 10.1 mL / hour, if potential LVO cases are not diagnosed in advance, the time window between the patient's baseline imaging and mechanical thrombectomy may be increased. Therefore, LVO must be detected as soon as possible to reduce the risk of neurological loss in patients with AIS, thereby reducing morbidity.

[0009] In summary, how to propose a solution for establishing a large vessel occlusion identification model based on multi-phase CTA on the basis of the above-mentioned existing technologies has become a common concern of technicians in this field. Summary of the invention

[0010] In view of the above-mentioned defects in the prior art, the purpose of this application is to propose a method, device, terminal and storage medium for establishing a large vessel occlusion identification model, as follows.

[0011] In a first aspect, a method for establishing a large vessel occlusion identification model comprises the following steps:

[0012] Obtain standard CT images of the brain and perform angiography to obtain angiography image sets at multiple stages;

[0013] Preprocessing the angiography image sets of multiple stages, and then performing blood vessel segmentation processing on the preprocessed angiography image sets of each stage to generate two-dimensional maximum intensity projection images of multiple stages;

[0014] A training set combination is created based on two-dimensional maximum intensity projection images of multiple stages to obtain multiple groups of model training sets, and deep learning model training is performed one by one using the multiple groups of model training sets to obtain multiple deep learning models;

[0015] The diagnostic performance of multiple deep learning models was evaluated, and the deep learning model with the best evaluation result was selected as the large vessel occlusion recognition model.

[0016] Preferably, the step of acquiring a standard CT image of the brain and performing angiography to obtain angiography image sets of multiple stages comprises the following steps:

[0017] Obtain standard CT images of the brain;

[0018] Angiography is performed based on the standard CT images to obtain a first-stage angiography image set, a second-stage angiography image set and a third-stage angiography image set, wherein the first-stage angiography image set is acquired from the aortic arch to the apex during the peak arterial phase, the second-stage angiography image set is acquired from the skull base to the apex during the peak venous phase, and the third-stage angiography image set is acquired from the skull base to the apex during the late venous phase.

[0019] Preferably, preprocessing the angiography image sets of multiple stages, and then performing blood vessel segmentation processing on the preprocessed angiography image sets of each stage to generate two-dimensional maximum intensity projection images of multiple stages, comprises the following steps:

[0020] converting the angiography image sets of multiple phases from a medical digital imaging format into an array format, and performing a normalization process to obtain a corresponding pre-processed angiography image set;

[0021] Using a blood vessel segmentation algorithm to perform image segmentation processing on the pre-processed angiography image sets at multiple stages;

[0022] The pre-processed angiography image set after the image segmentation process is subjected to maximum intensity projection generation to obtain a first-stage two-dimensional maximum intensity projection image, a second-stage two-dimensional maximum intensity projection image, and a third-stage two-dimensional maximum intensity projection image.

[0023] Preferably, the step of converting the angiography image sets of multiple stages from a medical digital imaging format to an array format and performing normalization processing to obtain a corresponding pre-processed angiography image set comprises the following steps:

[0024] The first-stage angiography image set, the second-stage angiography image set and the third-stage angiography image set were converted from the medical digital imaging and communication format to the NumPy array format, and isotropic sampling was performed to standardize each voxel to a volume of 1 mm3. Each axial angiography image was adjusted to 500×500 pixels, converted to grayscale and normalized to a value between 0 and 1, to obtain the first-stage preprocessed angiography image set, the second-stage preprocessed angiography image set and the third-stage preprocessed angiography image set, respectively.

[0025] Preferably, the method of combining and creating training sets based on two-dimensional maximum intensity projection images of multiple stages to obtain multiple groups of model training sets, and using the multiple groups of model training sets to perform deep learning model training one by one to obtain multiple deep learning models includes the following steps:

[0026] According to the first-stage two-dimensional maximum intensity projection image, the second-stage two-dimensional maximum intensity projection image and the third-stage two-dimensional maximum intensity projection image, all seven groups of model training sets are created by permutation and combination;

[0027] Use all seven groups of the model training sets to perform deep learning model training one by one. The output of the deep learning model has two output classes, and seven deep learning models corresponding to the seven groups of the model training sets are obtained respectively.

[0028] Preferably, the step of evaluating the diagnostic performance of multiple deep learning models and selecting a deep learning model with the best evaluation result as a large vessel occlusion recognition model comprises the following steps:

[0029] Statistical software was used to evaluate the diagnostic performance of the seven deep learning models on a validation set and a test set, and a deep learning model with the best evaluation result was selected as a large vessel occlusion recognition model, wherein the output form of the large vessel occlusion recognition model was a positive diagnosis result and a negative diagnosis result.

[0030] In a second aspect, a device for establishing a large vessel occlusion identification model includes the following modules:

[0031] A data acquisition module is configured to acquire a standard CT image of the brain and perform angiography to obtain angiography image sets at multiple stages;

[0032] The data preprocessing and blood vessel segmentation module is configured to preprocess the angiography image sets of multiple stages, and then perform blood vessel segmentation processing on the preprocessed angiography image sets of each stage to generate two-dimensional maximum intensity projection images of multiple stages;

[0033] The training set creation and model training module is configured to create training sets based on the two-dimensional maximum intensity projection images of multiple stages to obtain multiple groups of model training sets, and use the multiple groups of model training sets to perform deep learning model training one by one to obtain multiple deep learning models;

[0034] The model selection and output module is configured to evaluate the diagnostic performance of multiple deep learning models and select a deep learning model with the best evaluation result as a large vessel occlusion recognition model.

[0035] In a third aspect, a terminal includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps in the method for establishing a large vessel occlusion identification model as described above when executing the computer program.

[0036] In a fourth aspect, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the method for establishing a large vessel occlusion identification model as described above are implemented.

[0037] The advantages of this application are mainly reflected in the following aspects:

[0038] The method for establishing a large vessel occlusion recognition model based on multi-phase CTA proposed in this application can accurately detect LVO in the examination of AIS patients. The method of this application uses a deep learning model of a convolutional neural network as the basic model, which can effectively classify LVO on CTA images. The model has three stages. By using a combination of different stages, the diagnostic detection rate of LVO is greatly improved. The method of this application not only optimizes the workflow of radiologists in detecting LVO, but also can screen out patients who meet the conditions for thrombectomy, meeting the actual needs of both doctors and patients.

[0039] Corresponding to the above method, the device, terminal and storage medium for establishing a large vessel occlusion recognition model proposed in this application realize the effective recognition of large vessel occlusion with a systematic and standardized processing flow, thereby providing technical support for subsequent surgery or medical procedures. The hardware solution has high adaptability and compatibility and can be effectively applied to actual medical scenarios.

[0040] The specific implementation methods of the present application are further described below in conjunction with the accompanying drawings of the embodiments, so that the technical solution of the present application is easier to understand and grasp. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The drawings constituting a part of this application are used to provide a further understanding of this application, so that other features, purposes and advantages of this application become more obvious. The schematic embodiment drawings and their descriptions of this application are used to explain this application and do not constitute an improper limitation on this application. In the drawings:

[0042] Figure 1 It is a schematic diagram of the general process of the method for establishing a large vessel occlusion identification model in an embodiment of the present application;

[0043] Figure 2 is a schematic diagram of an operation example flow chart of a method for establishing a large vessel occlusion identification model in an embodiment of the present application;

[0044] Figure 3 is an image important area positioning diagram of an operation example in an embodiment of the present application;

[0045] Figure 4 is a prediction map of the occlusion site of the operation example in the embodiment of the present application;

[0046] Figure 5 is a false positive discovery prediction diagram of an operating example in the embodiment of the present application;

[0047] Figure 6 Schematic diagram of the architecture of a device for establishing a large vessel occlusion identification model in an embodiment of the present application. DETAILED DESCRIPTION

[0048] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0049] This application proposes a scheme for establishing a large vessel occlusion recognition model based on multi-phase CTA. The established large vessel occlusion recognition model can accurately detect LVO in the examination of AIS patients and can classify LVO on CTA images. The model has three stages, and the diagnosis detection rate of LVO is improved by using a combination of different stages. The specific scheme is as follows.

[0050] like Figure 1 , Figure 2 As shown, a method for establishing a large vessel occlusion identification model comprises the following steps:

[0051] S1. Obtain a standard CT image of the brain and perform angiography to obtain angiography image sets at multiple stages. This step can be specifically described as the following process.

[0052] S11. Since CT angiography is performed in multiple hospitals using multiple different CT scanners, the exact scanning parameters of different patients vary slightly. The implementation of the method first requires obtaining standard CT images of the brain.

[0053] S12. Perform angiography based on the standard CT image, and summarize the first stage angiography image set, the second stage angiography image set and the third stage angiography image set, wherein the first stage angiography image set is acquired from the aortic arch to the apex during the peak arterial phase, the second stage angiography image set is acquired from the skull base to the apex during the peak venous phase, and the third stage angiography image set is acquired from the skull base to the apex during the late venous phase.

[0054] The standard CT angiography workflow includes image acquisition with slice thicknesses of 0.625 mm to 1 mm, which are reconstructed into maximum intensity projections with 24 mm slice thickness and 4 mm spacing.

[0055] S2. Preprocessing the angiography image sets of multiple stages, and then performing blood vessel segmentation processing on the preprocessed angiography image sets of each stage to generate two-dimensional maximum intensity projection images of multiple stages; this step can be concretized as the following process.

[0056] S21, converting the angiography image sets of multiple stages from a medical digital imaging format into an array format, and performing normalization processing to obtain a corresponding pre-processed angiography image set.

[0057] Furthermore, in this step, the first-stage angiography image set, the second-stage angiography image set and the third-stage angiography image set need to be converted from the medical digital imaging and communication format into the NumPy array format, and isotropic sampling is performed to standardize each voxel to a volume of 1 mm3. Each axial angiography image is adjusted to 500×500 pixels, converted to grayscale and normalized to a value between 0 and 1, and the first-stage preprocessed angiography image set, the second-stage preprocessed angiography image set and the third-stage preprocessed angiography image set are obtained respectively.

[0058] S22, using a blood vessel segmentation algorithm to perform image segmentation processing on the pre-processed angiography image sets of multiple stages.

[0059] Given that the dataset used for deep learning applications in this embodiment is relatively small, a vessel segmentation algorithm is used here to segment the intracranial vasculature (i.e., signal) from potential non-vasculature structures such as brain parenchyma and skull (i.e., noise) to improve the signal-to-noise ratio. The vessel segmentation workflow is to compare the center pixel intensity with its surrounding pixel neighborhood. If the pixel intensity is greater than 50% of the surrounding pixels, it is assumed that the pixel belongs to a vessel. Additional processing of connected components, thresholds, and Sobel filters is performed. In general, this segmentation process takes about 40 seconds for each CTA.

[0060] S23, performing maximum intensity projection on the preprocessed angiography image set after the image segmentation process to obtain a first-stage two-dimensional maximum intensity projection image, a second-stage two-dimensional maximum intensity projection image, and a third-stage two-dimensional maximum intensity projection image.

[0061] In this example, an algorithm was created to select 40 axial skull images from each multiphase CT angiography examination, extending from the skull vertex to the circle of Willis, to include all relevant intracranial vasculature. The final preprocessing step involves generating maximum intensity projections along the axial planes of these 40 slices to generate the final 2D maximum intensity projection image for model training. Figure 3 .

[0062] exist Figure 3 In the figure, A shows an arterial (phase 1) image showing a sudden occlusion of the distal branch of the right middle cerebral artery M1 (arrow) and a lack of distal vessels. B shows a venous (phase 2) image showing a slight asymmetry and mild increase in vessels in the contralateral left hemisphere. C shows a late venous (phase 3) image showing opacity of the ipsilateral right hemisphere vasculature downstream of the occlusion. D shows a heat map showing the most discriminating area (red) for the model that predicted correctly.

[0063] S3. Create training sets based on the two-dimensional maximum intensity projection images of multiple stages to obtain multiple groups of model training sets, and use the multiple groups of model training sets to train deep learning models one by one to obtain multiple deep learning models. This step can be concretized into the following process.

[0064] S31. Create seven groups of model training sets according to the first-stage two-dimensional maximum intensity projection images, the second-stage two-dimensional maximum intensity projection images, and the third-stage two-dimensional maximum intensity projection images in permutations and combinations.

[0065] In this example, the DenseNet-121 architecture was tested to achieve the highest accuracy, and the model was started using pre-trained weights from ImageNet. All models were built using the Keras toolkit in Python with a learning rate of 0.00001, 10 epochs before convergence, and a batch size of 3.

[0066] The details of all seven groups of model training sets are as follows.

[0067] A first model training set includes the two-dimensional maximum intensity projection images of the first stage;

[0068] A second model training set, including the second-stage two-dimensional maximum intensity projection image;

[0069] A third model training set, comprising the two-dimensional maximum intensity projection image of the third stage;

[0070] a fourth model training set, comprising the first-stage two-dimensional maximum intensity projection image and the second-stage two-dimensional maximum intensity projection image;

[0071] a fifth model training set, comprising the second-stage two-dimensional maximum intensity projection image and the third-stage two-dimensional maximum intensity projection image;

[0072] a sixth model training set, comprising the two-dimensional maximum intensity projection image of the first stage and the two-dimensional maximum intensity projection image of the third stage;

[0073] The seventh model training set includes the first-stage two-dimensional maximum intensity projection image, the second-stage two-dimensional maximum intensity projection image, and the third-stage two-dimensional maximum intensity projection image.

[0074] S32. Use all seven groups of the model training sets to perform deep learning model training one by one. The output of the deep learning model has two output classes, and seven deep learning models corresponding to the seven groups of the model training sets are obtained respectively.

[0075] To help view the class distinction of the model, the present application embodiment also uses the gradient weighted class activation mapping (Grad-CAM) method to generate a localization map of the most important areas in the image. Examples of these gradient weighted class activation mapping methods are as follows Figure 4 and Figure 5 shown.

[0076] exist Figure 4In the figure, the top row shows a representative CT slice of delayed venous phase CT angiography. The middle row shows the preprocessed maximum intensity projection images, which serve as the input to the model. The bottom row is an overlay of heat maps, with red areas showing the most discriminating regions. These hot spots are associated with the location of occlusions (respectively: left [L] M2, L M1, right [R] M2, base).

[0077] exist Figure 5 In the figure, A, axial arterial phase CT angiogram; B, axial venous phase CT angiogram, images in A and B show no evidence of large vessel occlusion. Additional noncontrast-enhanced brain CT images (not shown) show no evidence of acute infarction. C, diffusion-weighted MRI obtained 6 hours after presentation, shows an area of ​​restricted diffusion (arrow) consistent with acute infarction in the left paramedian pons. D, heat map overlay shows focal activity projected onto the central brain.

[0078] S4. Evaluate the diagnostic performance of multiple deep learning models, and select the deep learning model with the best evaluation result as the large vessel occlusion recognition model. This step can be concretized into the following process.

[0079] The diagnostic performance of the seven deep learning models was evaluated on the validation set and the test set using statistical software (SAS, version 9.4; SAS, Cary, NC). The results showed that when the additional delay stage was combined with the first stage (i.e., the first and third stages and the first, second and third stages), there was an additive effect, resulting in model performance that was superior to single-phase CT angiography, indicating that the initial phase delay and delayed enhancement were important. The deep learning model with the best evaluation result was selected as the large vessel occlusion recognition model, and the output of the large vessel occlusion recognition model was in the form of a positive diagnosis result and a negative diagnosis result.

[0080] In summary, the above method has the following advantages and beneficial effects:

[0081] 1. Universality. The model established by the method is discriminative across patient demographics, multiple institutions and scanners, and does not exclude additional head and neck vascular lesions. Therefore, the prediction model population of the present invention is heterogeneous. This heterogeneity indicates that the prediction results can be generally applied in clinical settings;

[0082] 2. Efficient,The model developed is a deep learning convolutional neural network for detecting large vessel occlusions in multi-phase CTA,with high area under the receiver operating characteristic curve, especially when using delayed phases.,Both anterior and posterior circulation occlusions were detected.,Our work is an important first step in incorporating deep learning into large vessel occlusion (LVO) classification in emergency situations,and shortens the time to LVO detection, indicating that it can serve as a tool for radiologists and clinicians to prioritize potential LVO patients in emergency situations and become a powerful tool for detecting all potential LVO subtypes suitable for mechanical thrombectomy;

[0083] 3. Accurate, the dataset in the embodiment includes multi-phase CTA images, and we are able to study the role of delayed CTA in the performance of our model through three independent time-resolved data points. Multi-phase CTA has been shown to clearly and accurately depict LVO, while providing a more detailed assessment of the intracranial pial collateral network in the normal hemisphere and the LVO hemisphere;

[0084] 4. High R&D potential. In the model established by the method, the classifier predicted that it was positive for LVO with a probability score close to 100%. Although no evidence of LVO was found in the CTA, the patient's subsequent MRI approximately 6 hours later showed a small acute ischemic infarction in the left paramedian pons, such as Figure 5 As shown, this directly suggests the possibility that the model can detect smaller vasculature infarcts that are not visible to radiologists on CTA.

[0085] On the other hand, Figure 6 As shown, the present application also relates to a device for establishing a large vessel occlusion identification model, comprising the following modules:

[0086] A data acquisition module is configured to acquire a standard CT image of the brain and perform angiography to obtain angiography image sets at multiple stages;

[0087] The data preprocessing and blood vessel segmentation module is configured to preprocess the angiography image sets of multiple stages, and then perform blood vessel segmentation processing on the preprocessed angiography image sets of each stage to generate two-dimensional maximum intensity projection images of multiple stages;

[0088] The training set creation and model training module is configured to create training sets based on the two-dimensional maximum intensity projection images of multiple stages to obtain multiple groups of model training sets, and use the multiple groups of model training sets to perform deep learning model training one by one to obtain multiple deep learning models;

[0089] The model selection and output module is configured to evaluate the diagnostic performance of multiple deep learning models and select a deep learning model with the best evaluation result as a large vessel occlusion recognition model.

[0090] In a possible implementation, the data acquisition module includes:

[0091] a first data acquisition unit configured to acquire a standard CT image of the brain;

[0092] The second data acquisition unit is configured to perform angiography based on the standard CT image, and summarize to obtain a first-stage angiography image set, a second-stage angiography image set, and a third-stage angiography image set, respectively, wherein the first-stage angiography image set is acquired from the aortic arch to the apex during the peak arterial phase, the second-stage angiography image set is acquired from the skull base to the apex during the peak venous phase, and the third-stage angiography image set is acquired from the skull base to the apex during the late venous phase.

[0093] In a possible implementation, the data preprocessing and blood vessel segmentation module includes:

[0094] A data preprocessing unit configured to convert the angiography image sets of multiple stages from a medical digital imaging format into an array format and perform normalization processing to obtain a corresponding preprocessed angiography image set;

[0095] A first blood vessel segmentation unit is configured to perform image segmentation processing on the pre-processed angiography image sets of multiple stages using a blood vessel segmentation algorithm;

[0096] The second blood vessel segmentation unit is configured to perform maximum intensity projection on the preprocessed angiography image set after image segmentation processing to obtain a first-stage two-dimensional maximum intensity projection image, a second-stage two-dimensional maximum intensity projection image, and a third-stage two-dimensional maximum intensity projection image.

[0097] In a possible implementation, the training set creation and model training module includes:

[0098] A training set creation unit is configured to create all seven groups of model training sets according to the first-stage two-dimensional maximum intensity projection image, the second-stage two-dimensional maximum intensity projection image, and the third-stage two-dimensional maximum intensity projection image in permutations and combinations;

[0099] The model training unit is configured to use all seven groups of the model training sets to perform deep learning model training one by one. The output of the deep learning model has two output classes, and seven deep learning models corresponding to the seven groups of the model training sets are obtained respectively.

[0100] In a possible implementation, the model selection and output module includes:

[0101] The model selection and output unit is configured to use statistical software to evaluate the diagnostic performance of the seven deep learning models on a validation set and a test set, and select a deep learning model with the best evaluation result as a large vessel occlusion recognition model, wherein the output form of the large vessel occlusion recognition model is a positive diagnosis result and a negative diagnosis result.

[0102] In another aspect, the present application also relates to a terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps in the above-mentioned commodity intelligent identification method are implemented, for example Figure 1 Alternatively, when the processor executes the computer program, the functions of each module / unit in the above-mentioned device embodiments are implemented, for example Figure 6 The functions of each module / unit shown.

[0103] On the other hand, the present application also relates to a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps in the commodity intelligent identification method as described above are implemented.

[0104] Among them, the readable storage medium can be a computer storage medium or a communication medium. The communication medium includes any medium that facilitates the transmission of a computer program from one place to another. The computer storage medium can be any available medium that can be accessed by a general or special-purpose computer. For example, a readable storage medium is coupled to a processor so that the processor can read information from the readable storage medium and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can be located in an application-specific integrated circuit (ASIC). In addition, the ASIC can be located in a user device. Of course, the processor and the readable storage medium can also exist in a communication device as discrete components. The readable storage medium can be a read-only memory (ROM), a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, and an optical data storage device, etc.

[0105] Corresponding to the above method, the device, terminal and storage medium for establishing a large vessel occlusion recognition model proposed in this application realize the effective recognition of large vessel occlusion with a systematic and standardized processing flow, thereby providing technical support for subsequent surgery or medical procedures. The hardware solution has high adaptability and compatibility and can be effectively applied to actual medical scenarios.

[0106] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and that the present application can be implemented in other specific forms without departing from the spirit and basic features of the present application. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive, and the scope of the present application is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present application.

[0107] Finally, it should be understood that although this specification is described according to implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A method for establishing a large vessel occlusion identification model, characterized in that: The steps include: Obtaining standard CT images of the brain and performing angiography to obtain angiography image sets at multiple stages; Preprocessing the angiography image sets of multiple stages, and then performing blood vessel segmentation processing on the preprocessed angiography image sets of each stage to generate two-dimensional maximum intensity projection images of multiple stages; A training set combination is created based on two-dimensional maximum intensity projection images of multiple stages to obtain multiple groups of model training sets, and deep learning model training is performed one by one using the multiple groups of model training sets to obtain multiple deep learning models; Evaluate the diagnostic performance of multiple deep learning models and select the deep learning model with the best evaluation result as the large vessel occlusion recognition model; The step of acquiring a standard CT image of the brain and performing angiography to obtain angiography image sets of multiple stages comprises the following steps: Obtain standard CT images of the brain; Angiography is performed based on the standard CT images to obtain a first-stage angiography image set, a second-stage angiography image set and a third-stage angiography image set, wherein the first-stage angiography image set is acquired from the aortic arch to the apex during the peak arterial phase, the second-stage angiography image set is acquired from the skull base to the apex during the peak venous phase, and the third-stage angiography image set is acquired from the skull base to the apex during the late venous phase.

2. The method for establishing a large vessel occlusion identification model according to claim 1, characterized in that: Preprocessing the angiography image sets of multiple stages, and then performing blood vessel segmentation processing on the preprocessed angiography image sets of each stage to generate two-dimensional maximum intensity projection images of multiple stages, including the following steps: converting the angiography image sets of multiple stages from a medical digital imaging format into an array format, and performing normalization processing to obtain a corresponding pre-processed angiography image set; Using a blood vessel segmentation algorithm to perform image segmentation processing on the pre-processed angiography image sets of multiple stages; The pre-processed angiography image set after the image segmentation process is subjected to maximum intensity projection generation to obtain a first-stage two-dimensional maximum intensity projection image, a second-stage two-dimensional maximum intensity projection image, and a third-stage two-dimensional maximum intensity projection image.

3. The method for establishing a large vessel occlusion identification model according to claim 2, characterized in that: The method of converting the angiography image sets of multiple stages from a medical digital imaging format into an array format and performing normalization processing to obtain a corresponding pre-processed angiography image set comprises the following steps: The first-stage angiography image set, the second-stage angiography image set and the third-stage angiography image set were converted from the medical digital imaging and communication format to the NumPy array format, and isotropic sampling was performed to standardize each voxel to a volume of 1 mm3. Each axial angiography image was adjusted to 500×500 pixels, converted to grayscale and normalized to a value between 0 and 1, to obtain the first-stage preprocessed angiography image set, the second-stage preprocessed angiography image set and the third-stage preprocessed angiography image set, respectively.

4. The method for establishing a large vessel occlusion identification model according to claim 2, characterized in that: The method of combining and creating training sets based on two-dimensional maximum intensity projection images of multiple stages to obtain multiple groups of model training sets, and using the multiple groups of model training sets to perform deep learning model training one by one to obtain multiple deep learning models includes the following steps: According to the first-stage two-dimensional maximum intensity projection image, the second-stage two-dimensional maximum intensity projection image and the third-stage two-dimensional maximum intensity projection image, all seven groups of model training sets are created by permutation and combination; Use all seven groups of the model training sets to perform deep learning model training one by one. The output of the deep learning model has two output classes, and seven deep learning models corresponding to the seven groups of the model training sets are obtained respectively.

5. The method for establishing a large vessel occlusion identification model according to claim 4, characterized in that: The step of evaluating the diagnostic performance of multiple deep learning models and selecting a deep learning model with the best evaluation result as a large vessel occlusion recognition model comprises the following steps: Statistical software was used to evaluate the diagnostic performance of the seven deep learning models on a validation set and a test set, and a deep learning model with the best evaluation result was selected as a large vessel occlusion recognition model, wherein the output form of the large vessel occlusion recognition model was a positive diagnosis result and a negative diagnosis result.

6. A device for establishing a large vessel occlusion identification model, characterized in that: Includes the following modules: A data acquisition module is configured to acquire a standard CT image of the brain and perform angiography to obtain angiography image sets at multiple stages; The data preprocessing and blood vessel segmentation module is configured to preprocess the angiography image sets of multiple stages, and then perform blood vessel segmentation processing on the preprocessed angiography image sets of each stage to generate two-dimensional maximum intensity projection images of multiple stages; The training set creation and model training module is configured to create training sets based on the two-dimensional maximum intensity projection images of multiple stages to obtain multiple groups of model training sets, and use the multiple groups of model training sets to perform deep learning model training one by one to obtain multiple deep learning models; A model selection and output module is configured to evaluate the diagnostic performance of multiple deep learning models and select a deep learning model with the best evaluation result as a large vessel occlusion recognition model; The step of acquiring a standard CT image of the brain and performing angiography to obtain angiography image sets of multiple stages comprises the following steps: Obtain standard CT images of the brain; Angiography is performed based on the standard CT images to obtain a first-stage angiography image set, a second-stage angiography image set and a third-stage angiography image set, wherein the first-stage angiography image set is acquired from the aortic arch to the apex during the peak arterial phase, the second-stage angiography image set is acquired from the skull base to the apex during the peak venous phase, and the third-stage angiography image set is acquired from the skull base to the apex during the late venous phase.

7. A terminal, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method implements the steps in the method for establishing a large vessel occlusion identification model according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the method for establishing a large vessel occlusion identification model according to any one of claims 1 to 5 are implemented.

Images