Hematoma enlargement risk prediction method, equipment and medium

Abstract

The invention discloses a hematoma enlargement risk prediction method and device and a medium, and relates to the field of smart medical treatment, and the method comprises the steps: inputting a baseline CT image of a target patient into a hematoma segmentation module, and obtaining a baseline hematoma segmentation result and hematoma region segmentation features of the target patient; calculating baseline hematoma radiomics characteristics of the target patient according to the baseline hematoma segmentation result of the target patient; calculating the baseline hematoma volume of the target patient according to the baseline hematoma segmentation result of the target patient; and inputting the baseline hematoma image omics features, the baseline hematoma volume and the hematoma region segmentation features of the target patient into a hematoma feature mixing module to obtain a final hematoma enlargement risk prediction result of the target patient. The risk that hematoma of the cerebral hemorrhage part expands in the future can be predicted only through the baseline CT image of the cerebral hemorrhage patient, and the prediction efficiency and accuracy are improved.

Description

Technical Field

[0001] The present application relates to the field of smart medical care, and in particular to a method, device and medium for predicting the risk of hematoma expansion. Background Art

[0002] With the aging of the population and the progress of urbanization, the prevalence of stroke (commonly known as "stroke") is obvious. Stroke has five characteristics: high morbidity, high disability rate, high mortality rate, high recurrence rate, and high economic burden. Stroke is divided into two types: hemorrhagic stroke and ischemic stroke. Among them, hemorrhagic stroke (cerebral hemorrhage) accounts for about 15% of the stroke population, with a low incidence rate, but it is the most deadly type of stroke. Within 6 hours after the onset of cerebral hemorrhage, a hematoma will form at the bleeding site, and under the action of hemodynamics, it will continue to expand in the next 24 hours, endangering the patient's life and prognosis. Studies have shown that for every 1 ml increase in hematoma volume, the patient's risk of death will increase by about 5%. The currently commonly used criterion for bleeding expansion is that the admission hematoma volume of patients with cerebral hemorrhage has increased by 6 ml or expanded by 33% in the next 24 hours. However, since the mechanism of hematoma expansion is still unclear, it is a major challenge to predict the risk of future hematoma expansion based on the data collected when the patient is admitted to the hospital. Therefore, there is a lack of effective tools to identify patients with high-risk cerebral hemorrhage who may have hematoma expansion in clinical practice, which hinders the development of effective intervention measures for hematoma expansion. Therefore, the development of a risk prediction tool for hematoma expansion that can accurately and quickly identify patients with cerebral hemorrhage is crucial to improving the prognosis of patients with cerebral hemorrhage and promoting the development of effective intervention measures for hematoma expansion.

[0003] At present, the most commonly used detection method in clinical practice is to predict the risk of hematoma expansion of cerebral hemorrhage through the BAT five-level scale score. It combines the imaging markers on the computed tomography (CT) image of cerebral hemorrhage and the time from cerebral hemorrhage to the first CT imaging to comprehensively predict the risk value of hematoma expansion. For the imaging markers on the CT image of cerebral hemorrhage, the mixed sign is recorded as 1 point, and the absence is recorded as 0 point. The low-density sign is recorded as 2 points, and the absence is recorded as 0 point. For the time from cerebral hemorrhage to the first CT imaging, less than 2.5 hours is recorded as 2 points, and not less than 2.5 hours is recorded as 0 points. This scale determines cerebral hemorrhage patients with a total score of not less than 3 points (i.e., 3 to 5 points) as high risk, and other scores (i.e., 0 to 2 points) as low risk. The scale score has been verified in multi-center practice and has been proven to be a tool with good predictive effect.

[0004] The current method for predicting the risk of hematoma expansion in cerebral hemorrhage is mainly for experienced clinicians to manually calculate the final BAT score by manually interpreting the imaging markers on the CT images of cerebral hemorrhage, combined with the time from the onset of cerebral hemorrhage to the first CT imaging. However, since different doctors have different judgments on the same imaging marker, the BAT score will also be affected by subjective judgment and have different score risks. In addition, since the BAT scale score only uses the pre-set imaging marker features on the CT image to judge the risk of hematoma expansion, the CT image is not fully utilized, resulting in a low prediction accuracy of the BAT score for hematoma expansion. Summary of the invention

[0005] The purpose of this application is to provide a method, device and medium for predicting the risk of hematoma expansion, which can predict the risk of future expansion of hematoma in the site of cerebral hemorrhage only through the baseline CT images of patients with cerebral hemorrhage, thereby improving the prediction efficiency and accuracy.

[0006] To achieve the above objectives, this application provides the following solutions.

[0007] In a first aspect, the present application provides a method for predicting the risk of hematoma expansion, comprising:

[0008] Inputting the baseline CT image of the target patient into the hematoma segmentation module to obtain the baseline hematoma segmentation result and hematoma region segmentation features of the target patient; the hematoma region segmentation features include multi-scale hematoma growth depth features, multi-scale hematoma region depth features and initial hematoma expansion risk prediction results;

[0009] Calculating the baseline hematoma radiomics features of the target patient according to the baseline hematoma segmentation result of the target patient; the baseline hematoma radiomics features include: first-order grayscale features, shape features, and grayscale co-occurrence matrix texture features;

[0010] Calculate the target patient's baseline hematoma volume according to the target patient's baseline hematoma segmentation result;

[0011] The target patient's baseline hematoma radiomics features, baseline hematoma volume, and hematoma area segmentation features are input into the hematoma feature mixing module to obtain the final hematoma expansion risk prediction results for the target patient.

[0012] Optionally, the baseline CT image of the target patient is input into the hematoma segmentation module to obtain the baseline hematoma segmentation result and hematoma region segmentation features of the target patient, which also includes:

[0013] Construct a hematoma segmentation network;

[0014] A training set for the hematoma segmentation module was constructed based on the baseline CT images and follow-up CT images of the sample cases;

[0015] The five-fold cross-validation method was used to divide the hematoma segmentation module training set into five pairs of non-overlapping training sets and validation sets, and the stochastic gradient descent algorithm was used to train the hematoma segmentation network to obtain the trained hematoma segmentation network as the hematoma segmentation module.

[0016] Optionally, a hematoma segmentation module training set is constructed based on the baseline CT images and follow-up CT images of the sample cases, including:

[0017] Obtain baseline CT images and follow-up CT images of sample cases, and determine hematoma segmentation labels;

[0018] Calculate the baseline hematoma volume and follow-up hematoma volume of the sample case according to the hematoma segmentation label of the sample case, and determine the hematoma expansion prediction label;

[0019] Perform skull stripping on the baseline CT images and follow-up CT images of the sample cases to obtain the skull stripping baseline CT images and skull stripping follow-up CT images of the sample cases;

[0020] The skull stripping follow-up CT image of the sample case is registered with the skull stripping baseline CT image to obtain the registered follow-up CT image of the sample case, and the registered image deformation field is determined;

[0021] The registered image deformation field is used to register the hematoma segmentation label of the follow-up CT image of the sample case with the hematoma segmentation label of the baseline CT image to obtain the registered hematoma segmentation label of the follow-up CT image of the sample case;

[0022] The skull stripping baseline CT image of the sample case, the hematoma segmentation label of the baseline CT image, the registered follow-up CT image, and the registered hematoma segmentation label of the follow-up CT image are resampled and cropped to obtain the processed baseline CT image, the processed follow-up CT image, and the corresponding hematoma segmentation label of the sample case;

[0023] A hematoma segmentation module training set is constructed based on the processed baseline CT images of the sample cases, the processed baseline CT images and the corresponding hematoma segmentation labels, and the hematoma expansion prediction labels.

[0024] Optionally, the hematoma segmentation module includes a shared encoder, two private segmentation decoders, a hematoma growth prediction head, a hematoma segmentation head and an initial hematoma expansion risk prediction head; the shared encoder, the first private segmentation encoder and the hematoma growth prediction head are connected in sequence to form a hematoma growth prediction subnetwork; the shared encoder, the second private segmentation encoder and the hematoma segmentation head are connected in sequence to form a baseline hematoma segmentation subnetwork; the initial hematoma expansion risk prediction head is respectively connected to the private segmentation decoder of the hematoma growth prediction subnetwork and the private segmentation decoder of the baseline hematoma segmentation subnetwork to form an initial hematoma expansion risk prediction subnetwork;

[0025] The hematoma growth prediction subnetwork is used to predict hematoma growth according to the input baseline CT image, obtain hematoma growth prediction results, and extract multi-scale hematoma growth depth features;

[0026] The baseline hematoma segmentation subnetwork is used to perform baseline hematoma segmentation according to the input baseline CT image, obtain the baseline hematoma segmentation result, and extract the multi-scale hematoma region depth features;

[0027] The initial hematoma expansion risk prediction subnetwork is used to perform initial hematoma expansion risk prediction based on the input multi-scale hematoma growth depth features and multi-scale hematoma area depth features to obtain an initial hematoma expansion risk prediction result.

[0028] Optionally, the shared encoder includes five downsampling modules connected in sequence, each of which is composed of two repeated convolution blocks, each of which includes a 3D convolutional neural network with a convolution kernel size of 3, an instance normalization layer and a leakage corrected linear unit activation layer; the private segmentation decoder includes five upsampling modules connected in sequence, each of which is composed of two repeated transposed convolution blocks, each of which includes a 3D transposed convolutional neural network with a convolution kernel size of 3, an instance normalization layer and a leakage corrected linear unit activation layer; the hematoma growth prediction head and the hematoma segmentation head are both convolutional neural networks with a convolution kernel size of 1; the initial hematoma expansion risk prediction head is a fully connected layer.

[0029] Optionally, the target patient's baseline hematoma radiomics features, baseline hematoma volume, and hematoma region segmentation features are input into a hematoma feature mixing module to obtain a final hematoma expansion risk prediction result for the target patient, which also includes:

[0030] Construct a hybrid network of hematoma features;

[0031] Based on the hematoma segmentation module training set and the hematoma segmentation module, the baseline hematoma segmentation results, hematoma area segmentation features and hematoma expansion prediction labels of the sample cases are determined;

[0032] The baseline hematoma radiomics features of the sample cases were calculated according to the baseline hematoma segmentation results of the sample cases;

[0033] The baseline hematoma volume of the sample case is calculated according to the baseline hematoma segmentation results of the sample case;

[0034] A hematoma feature hybrid module training set was constructed based on the hematoma area segmentation features, baseline hematoma radiomics features, baseline hematoma volume, and hematoma expansion prediction labels of the sample cases.

[0035] The five-fold cross-validation method was used to divide the hematoma feature hybrid module training set into five pairs of non-overlapping training sets and validation sets, and the stochastic gradient descent algorithm was used to train the hematoma feature hybrid network to obtain the trained hematoma feature hybrid network as the hematoma feature hybrid module.

[0036] Optionally, the hematoma feature mixing module is a three-layer fully connected network; an exponential linear unit is provided between the first fully connected layer and the second fully connected layer of the three-layer fully connected network, and an exponential linear unit and a random inactivation layer are provided between the second fully connected layer and the third fully connected layer.

[0037] Optionally, the baseline hematoma radiomics features of the target patient are calculated according to the baseline hematoma segmentation result of the target patient, including:

[0038] Based on the pyradiomics toolkit, the baseline hematoma radiomics features of the target patient were calculated according to the baseline hematoma segmentation results of the target patient.

[0039] In a second aspect, the present application provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method for predicting the risk of hematoma expansion.

[0040] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for predicting the risk of hematoma expansion.

[0041] According to the specific embodiments provided in the present application, the present application discloses the following technical effects: the present application provides a method, device and medium for predicting the risk of hematoma expansion, uses a deep learning method, adopts a two-stage risk prediction, and combines multidimensional hematoma characteristics including baseline hematoma imaging genomics features, baseline hematoma volume, multi-scale hematoma growth depth features, multi-scale hematoma area depth features and initial hematoma expansion risk prediction results to comprehensively predict the risk of hematoma expansion. It can achieve rapid and accurate prediction of the patient's future hematoma expansion risk only through baseline CT imaging data without the help of other clinical information of the patient with cerebral hemorrhage, providing strong support for promoting the verification of existing clinical anti-hematoma expansion treatment methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0043] Figure 1 A flow chart of the method for predicting the risk of hematoma expansion provided in this application.

[0044] Figure 2 Schematic diagram of the network architecture of the hematoma segmentation module provided in this application.

[0045] Figure 3 Schematic diagram of the network architecture of the hematoma feature mixing module provided in this application.

[0046] Figure 4 A diagram of the implementation steps of the method for predicting the risk of hematoma expansion provided in this application. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0048] The present application proposes a method, device and medium for predicting the risk of hematoma expansion. Through a deep learning method, the risk of hematoma expansion is predicted based on CT images. The multi-scale features of the hematoma area and the expansion area on the CT image can be comprehensively extracted to comprehensively judge the risk of hematoma expansion, thereby realizing an objective, automated and high-precision prediction of the future risk of hematoma expansion in patients with cerebral hemorrhage.

[0049] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0050] In an exemplary embodiment, the present application provides a method for predicting the risk of hematoma expansion, which is executed by a computer device, and can be executed by a computer device such as a terminal or a server alone, or by a terminal and a server together. In the embodiment of the present application, Figure 1 As shown, the method includes the following steps 1 to 4.

[0051] Step 1: Input the baseline CT image of the target patient into the hematoma segmentation module to obtain the baseline hematoma segmentation result and hematoma region segmentation features of the target patient. The hematoma region segmentation features include multi-scale hematoma growth depth features, multi-scale hematoma region depth features and initial hematoma expansion risk prediction results.

[0052] Step 2: Calculate the baseline hematoma radiomics features of the target patient based on the baseline hematoma segmentation results of the target patient. The baseline hematoma radiomics features include: first-order grayscale features, shape features, and gray-level co-occurrence matrix texture features;

[0053] Step 3: Calculate the baseline hematoma volume of the target patient according to the baseline hematoma segmentation result of the target patient.

[0054] Step 4: Input the target patient's baseline hematoma radiomics features, baseline hematoma volume, and hematoma area segmentation features into the hematoma feature mixing module to obtain the final hematoma expansion risk prediction result of the target patient.

[0055] In the acute phase after the onset of cerebral hemorrhage (generally 72 hours after the onset), a mass will form at the bleeding site, forming a "hematoma". The CT image formed by the patient's initial scan is called the baseline CT image, and the image formed by the follow-up scan during the acute phase is called the follow-up CT image. The present application aims to automatically complete the risk prediction of whether the hematoma formed in the cerebral hemorrhage area will expand in the future using only the baseline CT image. Specifically, the present application includes two stages, each of which is implemented by a module, namely a hematoma segmentation module and a hematoma feature mixing module. Among them, the hematoma segmentation module can segment the baseline hematoma area, and can comprehensively predict the growth area of ​​the baseline hematoma area and the follow-up hematoma area. At the same time, the hematoma segmentation module also extracts multi-scale hematoma area features and multi-scale hematoma growth features to jointly predict the initial hematoma expansion risk. Subsequently, based on the hematoma segmentation module, the hematoma feature mixing module comprehensively utilizes multi-scale hematoma area features, multi-scale hematoma growth features, initial hematoma expansion risk, and imaging genomics features of hematoma area and hematoma growth area to predict the final hematoma expansion risk. This application has fast analysis speed, strong generalization ability on unknown data, and clinical practicality. The above two modules are explained below.

[0056] First, hematoma segmentation module.

[0057] The method for determining the hematoma segmentation module includes: constructing a hematoma segmentation network; constructing a hematoma segmentation module training set based on baseline CT images and follow-up CT images of sample cases; using a five-fold cross-validation method to divide the hematoma segmentation module training set into five pairs of non-overlapping training sets and validation sets, and using a stochastic gradient descent algorithm to train the hematoma segmentation network, and obtaining the trained hematoma segmentation network as the hematoma segmentation module.

[0058] Among them, a hematoma segmentation module training set is constructed according to the baseline CT image and the follow-up CT image of the sample case, including: obtaining the baseline CT image and the follow-up CT image of the sample case, and determining the hematoma segmentation label; calculating the baseline hematoma volume and the follow-up hematoma volume of the sample case according to the hematoma segmentation label of the sample case, and determining the hematoma expansion prediction label; performing skull stripping on the baseline CT image and the follow-up CT image of the sample case to obtain the skull stripping baseline CT image and the skull stripping follow-up CT image of the sample case; registering the skull stripping follow-up CT image of the sample case with the skull stripping baseline CT image to obtain the registered follow-up CT image of the sample case, and determining the registered image deformation field; The registered image deformation field is used to register the hematoma segmentation label of the follow-up CT image of the sample case with the hematoma segmentation label of the baseline CT image to obtain the registered hematoma segmentation label of the follow-up CT image of the sample case; the skull stripping baseline CT image, the hematoma segmentation label of the baseline CT image, the registered follow-up CT image and the registered hematoma segmentation label of the follow-up CT image of the sample case are resampled and cropped to obtain the processed baseline CT image, the processed follow-up CT image and the corresponding hematoma segmentation label of the sample case; the hematoma segmentation module training set is constructed according to the processed baseline CT image of the sample case, the processed baseline CT image and the corresponding hematoma segmentation label and the hematoma expansion prediction label.

[0059] The hematoma segmentation module uses the high-dimensional feature extraction capability of the convolutional neural network for the baseline hematoma area, and uses the Pytorch deep learning framework and the target segmentation capability based on the U-shaped neural network to complete the accurate segmentation of the baseline hematoma area and the voxel-level prediction of the future hematoma growth area. The specific segmentation and growth prediction process is described as follows.

[0060] 1) Image collection and hematoma delineation: Baseline CT images and follow-up CT images of hematoma are collected, and binary images of hematoma area are manually delineated and reviewed as hematoma segmentation labels.

[0061] 2) Calculate the hematoma expansion prediction label: Use the hematoma segmentation label to calculate the follow-up hematoma volume and the baseline hematoma volume. When the follow-up hematoma volume exceeds the baseline hematoma volume by 6 ml, it is judged as hematoma expansion, otherwise it is judged as no expansion.

[0062] 3) Align and calculate the hematoma growth segmentation label as follows.

[0063] a) Skull stripping: The existing CT-BET tool is used to strip the skull of the baseline CT image and the follow-up CT image to obtain the skull stripping baseline CT image and the skull stripping follow-up CT image.

[0064] b) Registration: The skull stripping follow-up CT image was aligned to the skull stripping baseline CT image using the ANTs tool through two rigid registrations and two affine registrations, and the registered image deformation field was obtained by registering the follow-up CT image to the baseline CT image, removing the images that failed the registration.

[0065] c) Hematoma segmentation label registration: Using the registered image deformation field, the ANTs tool is used to further complete the registration of the hematoma segmentation label of the follow-up CT image to the baseline CT image.

[0066] d) Calculating the hematoma growth segmentation label: The hematoma growth segmentation label is obtained by subtracting the baseline hematoma segmentation label from the follow-up hematoma segmentation label.

[0067] 4) Data preprocessing: The baseline CT images after skull removal and the corresponding hematoma segmentation labels, as well as the follow-up CT images after skull removal and registration and the corresponding registered hematoma segmentation labels, were resampled to unify the voxel resolution to 0.45×0.45×5mm. 3 , and crop the image matrix size to 320×320×24 to obtain the processed baseline CT image, the processed follow-up CT image and the corresponding hematoma segmentation label.

[0068] 5) Dataset construction: Use the processed baseline CT images, the processed follow-up CT images and the corresponding hematoma segmentation labels and hematoma expansion prediction labels to construct the hematoma segmentation module training set and test set.

[0069] 6) Model training: The five-fold cross-validation method was used to divide the hematoma segmentation module training set into five pairs of non-overlapping training sets and validation sets, and the stochastic gradient descent algorithm was used to train the hematoma segmentation network. The network was evaluated on the validation set to adjust the hyperparameters, and finally the hyperparameters with the best effect on the five validation sets were selected. The training was completed on the complete training set to obtain the hematoma segmentation module.

[0070] 7) Model generalization evaluation: The hematoma segmentation module trained on the complete training set is tested and evaluated on another test set to simulate the model's baseline hematoma segmentation error, hematoma growth prediction error, and initial hematoma expansion prediction risk error in real scenarios.

[0071] The overall network architecture of the hematoma segmentation module is described below. This module accepts a baseline CT image as input and outputs hematoma growth prediction results, baseline hematoma segmentation results, and initial hematoma expansion risk prediction results through close cooperation between four components.

[0072] like Figure 2As shown, the hematoma segmentation module includes a shared encoder, two private segmentation decoders, a hematoma growth prediction head, a hematoma segmentation head and an initial hematoma expansion risk prediction head. Among them, the shared encoder, the first private segmentation encoder and the hematoma growth prediction head are connected in sequence to form a hematoma growth prediction subnetwork; the shared encoder, the second private segmentation encoder and the hematoma segmentation head are connected in sequence to form a baseline hematoma segmentation subnetwork; the initial hematoma expansion risk prediction head is connected to the private segmentation decoder of the hematoma growth prediction subnetwork and the private segmentation decoder of the baseline hematoma segmentation subnetwork respectively to form an initial hematoma expansion risk prediction subnetwork. The hematoma growth prediction subnetwork is used to perform hematoma growth prediction according to the input baseline CT image, obtain hematoma growth prediction results, and extract multi-scale hematoma growth depth features; the baseline hematoma segmentation subnetwork is used to perform baseline hematoma segmentation according to the input baseline CT image, obtain baseline hematoma segmentation results, and extract multi-scale hematoma region depth features; the initial hematoma expansion risk prediction subnetwork is used to perform initial hematoma expansion risk prediction according to the input multi-scale hematoma growth depth features and multi-scale hematoma region depth features, and obtain initial hematoma expansion risk prediction results.

[0073] Specifically, the shared encoder includes five sequentially connected downsampling modules, each downsampling module is composed of two repeated convolution blocks, each convolution block includes a 3D convolutional neural network with a convolution kernel size of 3, an instance normalization layer and a leakage rectified linear unit activation layer; the private segmentation decoder includes five sequentially connected upsampling modules, each upsampling module is composed of two repeated transposed convolution blocks, each transposed convolution block includes a 3D transposed convolutional neural network with a convolution kernel size of 3, an instance normalization layer and a leakage rectified linear unit activation layer; the hematoma growth prediction head and the hematoma segmentation head are both a convolutional neural network with a convolution kernel size of 1; the initial hematoma expansion risk prediction head is a fully connected layer.

[0074] Second, the hematoma feature hybrid module.

[0075] The method for determining the hematoma feature hybrid module includes: constructing a hematoma feature hybrid network; determining the baseline hematoma segmentation results, hematoma area segmentation features and hematoma expansion prediction labels of the sample case based on the hematoma segmentation module training set and the hematoma segmentation module; calculating the baseline hematoma imaging features of the sample case according to the baseline hematoma segmentation results of the sample case; calculating the baseline hematoma volume of the sample case according to the baseline hematoma segmentation results of the sample case; constructing the hematoma feature hybrid module training set according to the hematoma area segmentation features, baseline hematoma imaging features, baseline hematoma volume and hematoma expansion prediction labels of the sample case; using a five-fold cross-validation method to divide the hematoma feature hybrid module training set into five pairs of non-overlapping training sets and validation sets, and using a stochastic gradient descent algorithm to train the hematoma feature hybrid network, and obtaining the trained hematoma feature hybrid network as the hematoma feature hybrid module.

[0076] The hematoma feature hybrid module integrates multi-dimensional hematoma region features to build an automated and accurate hematoma expansion risk prediction tool. The input of this module consists of five parts: 1) baseline hematoma imaging features; 2) baseline hematoma volume; 3) multi-scale hematoma region depth features; 4) multi-scale hematoma growth depth features; 5) initial hematoma expansion risk prediction results of the hematoma segmentation module. The prediction process is as follows.

[0077] 1) Calculate the hematoma region segmentation features: Based on the constructed hematoma segmentation module and hematoma segmentation module training set, calculate the baseline hematoma segmentation result and hematoma region segmentation features of the input baseline CT image, as follows.

[0078] a) Preserve the multi-scale hematoma growth depth features produced by the hematoma growth private decoder in the forward process.

[0079] b) Preserve the multi-scale deep features of the hematoma region generated by the private decoder for hematoma segmentation in the forward process.

[0080] c) Calculate the initial hematoma expansion risk using multi-scale hematoma growth depth features and multi-scale hematoma area depth features.

[0081] 2) Calculation of baseline hematoma radiomics features: Based on the pyradiomics toolkit, multiple radiomics features were extracted according to Laplace transform, wavelet transform, gradient image and original image, including first-order grayscale features, shape features and gray-level co-occurrence matrix texture features, and finally 1379-dimensional features were extracted.

[0082] 3) Calculate the baseline hematoma volume feature: Based on the baseline hematoma segmentation result, according to the voxel size of 0.45×0.45×5mm 3 Calculate baseline hematoma volume.

[0083] 4) Constructing a hematoma feature mixed module training set: constructing a hematoma feature mixed module training set based on the five features extracted above and the hematoma expansion prediction label obtained by the hematoma segmentation module.

[0084] 5) Model construction: The constructed training set was divided into five pairs of non-overlapping training sets and validation sets using five-fold cross validation, and the hematoma feature hybrid network was trained using the stochastic gradient descent algorithm. The network was evaluated on the validation set to adjust the hyperparameters, and finally the hyperparameters with the best effect on the five validation sets were selected. The training was completed on the complete training set to obtain the hematoma feature hybrid module.

[0085] 6) Model generalization evaluation: The hematoma feature hybrid module trained with the complete training set is tested and evaluated on another test set to simulate the risk error of hematoma expansion prediction in real scenarios.

[0086] The overall architecture of the hematoma feature hybrid module is described in detail below. Figure 3 As shown in the figure, the input features of this module include the above five parts and a hematoma feature hybrid network, that is, a three-layer fully connected network. This module accepts the above five parts of feature input, performs high-dimensional hematoma expansion feature modeling in the feature hybrid network, and finally outputs accurate hematoma expansion risk prediction results.

[0087] In this module, the architecture of the three-layer fully connected network is shown in Table 1.

[0088] Table 1. Hybrid network architecture of hematoma features

[0089] Network components Number of input channels Number of output channels Remark Fully connected layer 3621 4096 Exponential Linear Unit / / The exponent parameter is 1 Fully connected layer 4096 2048 Exponential Linear Unit / / The exponent parameter is 1 Dropout layer / / Random dropout 50% Fully connected layer 2048 1 Hematoma expansion risk prediction strata

[0090] Specifically, the hematoma feature mixing module is a three-layer fully connected network; an exponential linear unit with an exponential parameter of 1 is set between the first fully connected layer and the second fully connected layer of the three-layer fully connected network, an exponential linear unit with an exponential parameter of 1 and a random inactivation layer with a random inactivation of 50% are set between the second fully connected layer and the third fully connected layer, and the third fully connected layer is the hematoma expansion risk prediction layer.

[0091] like Figure 4 As shown, the present application proposes a method for predicting whether the hematoma will expand in the future based on the baseline CT cerebral hemorrhage image, and the steps are as follows.

[0092] 1) Construct baseline CT images and follow-up CT images.

[0093] 2) Constructing hematoma region segmentation labels and hematoma expansion prediction labels: The hematoma regions of the baseline CT images and follow-up CT images were manually outlined and reviewed by experienced physicians to obtain binary mask hematoma segmentation labels. The baseline hematoma volume and the follow-up hematoma volume were then calculated using the voxel resolution of the original image and the hematoma segmentation label, and the two were subtracted. When the baseline hematoma volume increased by more than 6 ml on the follow-up CT, it was determined to be hematoma expansion, and the hematoma expansion prediction label was obtained.

[0094] 3) Construct the prediction label of hematoma area growth: including skull stripping, registration and label calculation. Skull stripping: Use the existing CT-BET tool to strip the skull of the baseline and follow-up CT brain images; CT image registration: Use the ANTs tool to align the follow-up CT image with the skull stripped to the baseline CT image through two rigid registrations and two affine registrations, and obtain the registered image deformation field obtained by registering the follow-up CT image to the baseline CT, removing the image that failed the registration; Hematoma segmentation label registration: Using the registered image deformation field, use the ANTs tool to further complete the registration of the hematoma segmentation label of the follow-up CT image to the baseline CT image; Calculate the hematoma growth segmentation label: Finally, use the follow-up hematoma segmentation label to subtract the baseline hematoma segmentation label to obtain the hematoma growth segmentation label.

[0095] 4) Image resampling: The baseline CT image after skull removal, the baseline hematoma segmentation label, the follow-up CT image after skull removal and registration, and the follow-up hematoma segmentation label after registration were resampled to 0.45×0.45×5mm. 3 , and crop the image matrix size to 320×320×24.

[0096] 5) Constructing a training set for the hematoma segmentation module: The preprocessed CT images, hematoma segmentation labels, hematoma expansion prediction labels, and hematoma growth prediction labels are organized to construct a data set, and the data set is divided into five pairs of training sets and test sets using a five-fold cross-validation method.

[0097] 6) Training and evaluation of hematoma segmentation module: The hematoma segmentation module is trained using the stochastic gradient descent algorithm and evaluated on the validation set to adjust the hyperparameters. Finally, the hyperparameters with the best effect on the five validation sets are selected and trained on the complete training set. Finally, the model is evaluated on another test set and the model error is measured.

[0098] 7) Constructing a training set for the mixed module of hematoma features: Based on the constructed hematoma segmentation module and data set, extract multidimensional hematoma features, and construct a training set for the mixed module of hematoma features in combination with the hematoma expansion prediction label. The multidimensional hematoma features include 1) baseline hematoma imaging features; 2) baseline hematoma volume; 3) multi-scale hematoma area depth features; 4) multi-scale hematoma growth depth features; 5) initial hematoma expansion risk prediction results.

[0099] 8) Training and evaluation of the hematoma feature hybrid module: The five-fold cross-validation method mentioned above was used to divide the data of the hematoma feature hybrid module training set, and the stochastic gradient descent algorithm was used to train the model and adjust the hyperparameters. Finally, the optimal hyperparameters were selected to build a hematoma expansion prediction model on the complete training set, and the model error was evaluated on another test set.

[0100] In summary, the present application uses a deep learning method to automatically predict the future risk of hematoma expansion in patients with cerebral hemorrhage using only baseline CT images. The present application simultaneously predicts the baseline hematoma segmentation result and the future expansion area of ​​the baseline hematoma to extract multi-scale hematoma expansion depth features. The present application adopts a two-stage risk prediction, combining multi-dimensional hematoma features to comprehensively predict the risk of hematoma expansion, wherein the multi-dimensional hematoma features include multi-scale hematoma growth depth features, multi-scale hematoma area depth features, initial hematoma expansion risk, hematoma area imaging omics features and baseline hematoma volume. Compared with the prior art, the present application has the following advantages:

[0101] 1. This application uses a deep learning method, without the help of other clinical information of patients with cerebral hemorrhage, and can quickly and accurately predict the patient's future risk of hematoma expansion only through baseline CT imaging data, providing strong support for promoting the verification of existing clinical anti-hematoma expansion treatment methods.

[0102] 2. This application combines CT registration imaging technology, which can be combined with existing follow-up CT images to assist in predicting the future growth area of ​​the hematoma, providing good interpretability for clinical decision-making.

[0103] 3. The two-stage modules of this application do not interfere with each other. The first stage is not interfered by auxiliary information such as imaging genomics, and effectively extracts multi-scale hematoma area depth features and multi-scale hematoma growth depth features.

[0104] 4. This application focuses on the hematoma area to extract multi-dimensional hematoma features to comprehensively predict the final risk of hematoma expansion, effectively avoiding the network's attention to other irrelevant areas when judging hematoma expansion, and has good robustness.

[0105] In an exemplary embodiment, the present application further provides a computer device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0106] In an exemplary embodiment, the present application further provides a computer-readable storage medium storing a computer program, which implements the steps in the above-mentioned method embodiments when executed by a processor.

[0107] In an exemplary embodiment, the present application further provides a computer program product, including a computer program, which implements the steps in the above-mentioned method embodiments when executed by a processor.

[0108] In this application, all actions to obtain signals, information or data are carried out in compliance with the relevant data protection laws and policies of the country where the application is located, and with the authorization of the corresponding device owner. It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant legal provisions.

[0109] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0110] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0111] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0112] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for predicting the risk of hematoma expansion, characterized in that: include: Inputting the baseline CT image of the target patient into the hematoma segmentation module to obtain the baseline hematoma segmentation result and hematoma region segmentation features of the target patient; the hematoma region segmentation features include multi-scale hematoma growth depth features, multi-scale hematoma region depth features and initial hematoma expansion risk prediction results; Calculate the target patient's baseline hematoma radiomics features based on the target patient's baseline hematoma segmentation results; The baseline hematoma imaging features include: first-order grayscale features, shape features, and gray-level co-occurrence matrix texture features; Calculate the target patient's baseline hematoma volume according to the target patient's baseline hematoma segmentation result; The target patient's baseline hematoma radiomics features, baseline hematoma volume, and hematoma area segmentation features are input into the hematoma feature mixing module to obtain the final hematoma expansion risk prediction results for the target patient.

2. The method for predicting the risk of hematoma expansion according to claim 1, characterized in that: The baseline CT image of the target patient is input into the hematoma segmentation module to obtain the baseline hematoma segmentation result and hematoma area segmentation features of the target patient, which also includes: Construct a hematoma segmentation network; A training set for the hematoma segmentation module was constructed based on the baseline CT images and follow-up CT images of the sample cases; The five-fold cross-validation method was used to divide the hematoma segmentation module training set into five pairs of non-overlapping training sets and validation sets, and the stochastic gradient descent algorithm was used to train the hematoma segmentation network to obtain the trained hematoma segmentation network as the hematoma segmentation module.

3. The method for predicting the risk of hematoma expansion according to claim 2, characterized in that: The hematoma segmentation module training set was constructed based on the baseline CT images and follow-up CT images of the sample cases, including: Obtain baseline CT images and follow-up CT images of sample cases, and determine hematoma segmentation labels; Calculate the baseline hematoma volume and follow-up hematoma volume of the sample case according to the hematoma segmentation label of the sample case, and determine the hematoma expansion prediction label; Perform skull stripping on the baseline CT images and follow-up CT images of the sample cases to obtain the skull stripping baseline CT images and skull stripping follow-up CT images of the sample cases; The skull stripping follow-up CT image of the sample case is registered with the skull stripping baseline CT image to obtain the registered follow-up CT image of the sample case, and the registered image deformation field is determined; The registered image deformation field is used to register the hematoma segmentation label of the follow-up CT image of the sample case with the hematoma segmentation label of the baseline CT image to obtain the registered hematoma segmentation label of the follow-up CT image of the sample case; The skull stripping baseline CT image of the sample case, the hematoma segmentation label of the baseline CT image, the registered follow-up CT image, and the registered hematoma segmentation label of the follow-up CT image are resampled and cropped to obtain the processed baseline CT image, the processed follow-up CT image, and the corresponding hematoma segmentation label of the sample case; A hematoma segmentation module training set is constructed based on the processed baseline CT images of the sample cases, the processed baseline CT images and the corresponding hematoma segmentation labels, and the hematoma expansion prediction labels.

4. The method for predicting the risk of hematoma expansion according to claim 2, characterized in that: The hematoma segmentation module includes a shared encoder, two private segmentation decoders, a hematoma growth prediction head, a hematoma segmentation head and an initial hematoma expansion risk prediction head; the shared encoder, the first private segmentation encoder and the hematoma growth prediction head are connected in sequence to form a hematoma growth prediction subnetwork; the shared encoder, the second private segmentation encoder and the hematoma segmentation head are connected in sequence to form a baseline hematoma segmentation subnetwork; the initial hematoma expansion risk prediction head is respectively connected to the private segmentation decoder of the hematoma growth prediction subnetwork and the private segmentation decoder of the baseline hematoma segmentation subnetwork to form an initial hematoma expansion risk prediction subnetwork; The hematoma growth prediction subnetwork is used to predict hematoma growth according to the input baseline CT image, obtain hematoma growth prediction results, and extract multi-scale hematoma growth depth features; The baseline hematoma segmentation subnetwork is used to perform baseline hematoma segmentation according to the input baseline CT image, obtain the baseline hematoma segmentation result, and extract the multi-scale hematoma region depth features; The initial hematoma expansion risk prediction subnetwork is used to perform initial hematoma expansion risk prediction based on the input multi-scale hematoma growth depth features and multi-scale hematoma area depth features to obtain an initial hematoma expansion risk prediction result.

5. The method for predicting the risk of hematoma expansion according to claim 4, characterized in that: The shared encoder includes five downsampling modules connected in sequence, each of which is composed of two repeated convolution blocks, each of which includes a 3D convolutional neural network with a convolution kernel size of 3, an instance normalization layer and a leakage corrected linear unit activation layer; the private segmentation decoder includes five upsampling modules connected in sequence, each of which is composed of two repeated transposed convolution blocks, each of which includes a 3D transposed convolutional neural network with a convolution kernel size of 3, an instance normalization layer and a leakage corrected linear unit activation layer; the hematoma growth prediction head and the hematoma segmentation head are both convolutional neural networks with a convolution kernel size of 1; the initial hematoma expansion risk prediction head is a fully connected layer.

6. The method for predicting the risk of hematoma expansion according to claim 2, characterized in that: The target patient’s baseline hematoma radiomics features, baseline hematoma volume, and hematoma area segmentation features are input into the hematoma feature mixing module to obtain the final hematoma expansion risk prediction result for the target patient, which also includes: Construct a hybrid network of hematoma features; Based on the hematoma segmentation module training set and the hematoma segmentation module, the baseline hematoma segmentation results, hematoma area segmentation features and hematoma expansion prediction labels of the sample cases are determined; The baseline hematoma radiomics features of the sample cases were calculated according to the baseline hematoma segmentation results of the sample cases; The baseline hematoma volume of the sample case is calculated according to the baseline hematoma segmentation results of the sample case; A hematoma feature hybrid module training set was constructed based on the hematoma area segmentation features, baseline hematoma radiomics features, baseline hematoma volume, and hematoma expansion prediction labels of the sample cases. The five-fold cross-validation method was used to divide the hematoma feature hybrid module training set into five pairs of non-overlapping training sets and validation sets, and the stochastic gradient descent algorithm was used to train the hematoma feature hybrid network to obtain the trained hematoma feature hybrid network as the hematoma feature hybrid module.

7. The method for predicting the risk of hematoma expansion according to claim 6, characterized in that: The hematoma feature mixing module is a three-layer fully connected network; an exponential linear unit is provided between the first fully connected layer and the second fully connected layer of the three-layer fully connected network, and an exponential linear unit and a random inactivation layer are provided between the second fully connected layer and the third fully connected layer.

8. The method for predicting the risk of hematoma expansion according to claim 1, characterized in that: The baseline hematoma radiomics features of the target patient are calculated according to the baseline hematoma segmentation results of the target patient, including: Based on the pyradiomics toolkit, the baseline hematoma radiomics features of the target patient were calculated according to the baseline hematoma segmentation results of the target patient.

9. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method for predicting the risk of hematoma expansion as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the method for predicting the risk of hematoma expansion as described in any one of claims 1-8.

Images