Cerebral stroke prognosis prediction method and system

Abstract

The invention discloses a stroke prognosis prediction method and system, and the method comprises the steps: S1, obtaining a CTP image of a whole cerebral vessel; and S2, the CTP image is segmented, and the segmented CTP image is obtained. And S3, carrying out feature extraction on the segmented CTP image to respectively obtain morphological features of whole cerebral vessels, perfusion features of an infarction core region and blood flow difference features of key blood vessel points. And S4, constructing a stroke prognosis prediction model based on machine learning according to the morphological characteristics of the whole brain blood vessels, the infarct core region perfusion characteristics and the blood flow difference characteristics of the key blood vessel points. According to the method, the blood flow and perfusion characteristics of the blood vessel are extracted more comprehensively, finely and quantitatively, and the whole method and system are full-automatic and do not need manual participation.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a method and system for predicting the prognosis of stroke. Background Art

[0002] By 2019, the absolute number of stroke events worldwide increased by 70.0%, and the stroke prevalence increased by 85.0%. In addition to its high mortality rate, stroke is also characterized by high disability and recurrence rates. A large number of stroke patients have brought huge economic burdens to their families and society.

[0003] Generally speaking, reperfusion therapy is the most effective treatment for acute ischemic stroke and has become the focus of current cerebral infarction research. Although the progress of these studies has unquestionable value, the narrow treatment window means that few patients have the opportunity to receive reperfusion therapy. If the factors that affect the prognosis of these patients who cannot be reperfused can be identified, it may become the basis for late-stage treatment of these patients.

[0004] Collateral circulation affects whether a patient will develop into a cerebral infarction or get good enough perfusion for brain cells to survive. Differences in collateral circulation in patients with cerebral infarction largely determine the final infarct size and the size of the ischemic penumbra. However, the existing collateral circulation classifications (such as the Miteff scoring system, the Maas scoring system, the modified Tan scoring system, the ASPECTS scoring system, and the human eye recognition system) are not ideal for predicting clinical outcomes and cannot be used as independent predictors of clinical outcomes. Therefore, manually identifying the number of collateral vessels is not enough for the stroke prognosis prediction model. To establish a reliable and effective prognosis prediction model, a more accurate, quantitative, and comprehensive description of the entire cerebral vascular and blood flow information is required.

[0005] Therefore, based on a comprehensive analysis of the relationship between vascular structure, blood flow changes and brain parenchymal perfusion during the occurrence, development and prognosis of stroke, it is urgent to propose a stroke prognosis prediction method and system to reduce the mortality, disability and recurrence rates of stroke patients. Summary of the invention

[0006] The purpose of the present invention is to provide a method and system for predicting the prognosis of stroke, which are used to predict the development of the disease in the prognosis stage of stroke patients and reduce the mortality, disability and recurrence rates of stroke patients.

[0007] In order to achieve the above objectives, the present invention is implemented by the following technical solutions:

[0008] A method for predicting the prognosis of stroke comprises: step S1, obtaining a CTP image of the whole brain blood vessels. Step S2, segmenting the CTP image to obtain the segmented CTP image. Step S3, extracting features from the segmented CTP image to obtain the morphological features of the whole brain blood vessels, the perfusion features of the infarct core area and the blood flow difference features of key blood vessel points. Step S4, constructing a stroke prognosis prediction model based on machine learning according to the morphological features of the whole brain blood vessels, the perfusion features of the infarct core area and the blood flow difference features of the key blood vessel points.

[0009] Optionally, the step S2 includes: step S21, normalizing the grayscale of the CTP image, and dividing the normalized CTP image into a training image and a test image. Step S22, constructing an initial 3D-Unet segmentation model under the Tensorflow architecture. Step S23, inputting the training image into the initial 3D-Unet segmentation model for training and verification, and adding the centerline loss to the model training loss part to obtain a segmentation model. Step S24, inputting the test image into the segmentation model, and automatically segmenting the test image to obtain the segmented CTP image; the segmented CTP image includes the whole brain blood vessel segmentation result.

[0010] Optionally, the step S3 includes: step S31, using the centerline extraction algorithm to extract the 3D blood vessel centerline corresponding to the whole brain blood vessels in the whole brain blood vessel segmentation result. Step S32, calculating the morphological features of the 3D blood vessel centerline to obtain a plurality of 3D morphological features. Step S33, projecting the 3D blood vessels in the whole brain blood vessel segmentation result to the 2D axial plane, and performing blood vessel enhancement and 2D threshold segmentation to obtain a 2D blood vessel segmentation result. Step S34, using the centerline extraction algorithm to extract the 2D blood vessel centerline corresponding to the whole brain blood vessels in the 2D blood vessel segmentation result. Step S35, calculating the morphological features of the 2D blood vessel centerline to obtain a plurality of 2D morphological features. Step S36, the plurality of 3D morphological features include: a plurality of healthy side 3D morphological features and a plurality of diseased side 3D morphological features; the plurality of 2D morphological features include: a plurality of healthy side 2D morphological features and a plurality of diseased side 2D morphological features. A number of the healthy side 3D morphological features are compared with a number of the diseased side 3D morphological features, and a number of the healthy side 2D morphological features are compared with a number of the diseased side 2D morphological features, to obtain a number of 3D healthy and diseased side difference features and a number of 2D healthy and diseased side difference features. The morphological features of the whole brain blood vessels include: a number of 3D morphological features, a number of 2D morphological features, a number of 3D healthy and diseased side difference features, and a number of 2D healthy and diseased side difference features.

[0011] Optionally, step S3 also includes: step S37, drawing a time-density curve according to the segmented CTP image. Step S38, performing Gaussian fitting on the time-density curve to obtain a first Gaussian time-density curve. Step S39, constructing a pseudo-color map based on the first Gaussian time-density curve. Step S310, acquiring a DWI image and marking the stroke area. Step S311, aligning the stroke area marked on the DWI image to the pseudo-color map to obtain the infarct core area on the pseudo-color map. Step S312, extracting the substantial high-throughput features of the infarct core area on the pseudo-color map to obtain the high-throughput features of the perfusion area; the perfusion features of the infarct core area include the high-throughput features of the perfusion area.

[0012] Optionally, the step S3 also includes: step S313, marking a number of key vascular points in the segmented CTP image. Step S314, drawing the time-density curve of the key vascular points according to the pixel values ​​of the marked key vascular points on the segmented CTP image. Step S315, performing Gaussian fitting on the time-density curve of the key vascular points to obtain a second Gaussian time-density curve. Step S316, calculating the TTP, CBF and CBV of each key vascular point according to the second Gaussian time-density curve, and using the TTP, the CBF and the CBV as blood flow features. Step S317, calculating the difference between the key blood flow point features of different key vascular points to obtain key blood flow point difference features.

[0013] Optionally, step S4 includes: step S41, dividing the data set into a cross-validation set and an independent test set; and using the model established by the cross-validation set to test the independent test set to obtain test results. Step S42, according to the test results and based on the cross-validation set, a sparse representation feature selection model is established using a minimum shrinkage operator and a selection operator, and the sparse representation feature selection model is used to perform feature selection on the morphological characteristics of the entire cerebral blood vessels, the perfusion characteristics of the infarct core area, and the blood flow difference characteristics of the key points to obtain high-resolution features. Step S43, step S43, obtain a 3-month modified Rankin scale, and use the 3-month modified Rankin scale as a label to establish a 2-classification neural network model based on a support vector machine; divide the high-resolution features into a training set, train the 2-classification neural network model to establish the stroke prognosis prediction model.

[0014] Optionally, it also includes step S5: using the stroke prognosis prediction model to predict cases with unknown prognosis results. The step S5 includes: step S51, after calculating the morphological characteristics of the whole cerebral blood vessels of the case with unknown prognosis results, the perfusion characteristics of the infarct core area and the key blood flow point difference characteristics, feature screening is performed to obtain the high-resolution features of the unknown case. Step S52, inputting the high-resolution features of the unknown case into the stroke prognosis prediction model, and a result of 1 represents that the prediction result of the case is good; a result of 0 represents that the prediction result of the case is bad.

[0015] Optionally, the stroke prognosis prediction method further includes step S5, evaluating the stroke prognosis prediction model using accuracy, sensitivity, specificity, positive predictive value and negative predictive value.

[0016] On the other hand, the present invention also provides a stroke prognosis prediction system, including: a whole brain blood vessel segmentation module, used to segment the CTP image of the whole brain blood vessels to obtain the segmented CTP image. A feature extraction module, used to extract features from the segmented CTP image to obtain the morphological characteristics of the whole brain blood vessels, the perfusion characteristics of the infarct core area, and the blood flow difference characteristics of key blood vessel points. A prognosis prediction model establishment module based on machine learning, used to construct a stroke prognosis prediction model based on the morphological characteristics of the whole brain blood vessels, the perfusion characteristics of the infarct core area, and the blood flow difference characteristics of key blood vessel points.

[0017] In other aspects, the present invention further provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the method described above is implemented.

[0018] The present invention has at least one of the following technical effects:

[0019] The present invention utilizes machine learning to construct a stroke prognosis prediction model, thereby reducing the subjective differences in manual identification; based on the pathophysiology of the occurrence and development of stroke, it comprehensively incorporates vascular morphology, infarct core area perfusion and key vascular point characteristics to predict the prognosis of stroke; more comprehensive and refined quantitative extraction of vascular, blood flow and perfusion characteristics; the entire method is fully automatic and does not require manual participation. The stroke prognosis prediction model has high robustness and repeatability, high prediction accuracy and strong practicality.

[0020] The overall system of the present invention can realize automatic quantification of the comprehensive characteristics of the whole brain blood vessels, especially quantification of the difference characteristics of the healthy and affected side blood vessels, which is different from the traditional manual method of only evaluating the blood vessel density of a specific area. It also reduces labor costs while utilizing more comprehensive blood vessel characteristics.

[0021] The present invention uses high-throughput imaging genomics features including grayscale features and texture features to describe the parenchymal perfusion information of the infarct core area, and completely describes the changes in the perfusion information of the infarct core area.

[0022] The present invention uses the perfusion characteristics of key venous vascular points to describe the blood flow information of the downstream area of ​​the infarction, and uses the perfusion characteristics of a few key vascular points to describe the blood flow information of the entire downstream area of ​​the infarction, which effectively reflects the blood flow situation of the entire brain. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A flowchart of a method for predicting stroke prognosis provided by one embodiment of the present invention;

[0024] Figure 2 A detailed flow chart of a method for predicting stroke prognosis provided by an embodiment of the present invention;

[0025] Figure 3 This is a ROC curve (receiver operating characteristic curve) diagram of the mRS (modified Rankin scale) prediction results provided by one embodiment of the present invention.

[0026] Figure 4 A schematic diagram of a time-density curve for calculating peak time, cerebral blood flow and cerebral blood volume provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0027] The following is a further detailed description of a stroke prognosis prediction method and system proposed by the present invention in conjunction with the accompanying drawings and specific embodiments. According to the following description, the advantages and features of the present invention will be clearer. It should be noted that the accompanying drawings are in a very simplified form and use non-precise proportions, which are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention. In order to make the purposes, features and advantages of the present invention more obvious and easy to understand, please refer to the accompanying drawings. It should be noted that the structures, proportions, sizes, etc. illustrated in the accompanying drawings of this specification are only used to match the contents disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the limiting conditions for the implementation of the present invention, so they have no technical substantive significance. Any structural modification, change in proportional relationship or adjustment of size, without affecting the effects that the present invention can produce and the purposes that can be achieved, should still fall within the scope of the technical content disclosed by the present invention.

[0028] As described in the background art, previous studies have shown that there is a correlation between changes in cerebrovascular morphology and cerebral infarction volume. Changes in cerebrovascular morphology precede changes in cerebral infarction volume. Fully automated image analysis to diagnose vascular abnormalities and quantify cerebrovascular status (arterial vascular density or arterial inner diameter, etc.) can facilitate the discovery of potentially effective diagnostic and predictive biomarkers.

[0029] Perfusion images contain spatial information, showing the close interdependence between the ischemic penumbra and the infarct area, and indirectly reflecting the physiological characteristics of vascular supply and collateral blood flow. Its parameters include time to peak (TTP), cerebral blood flow (CBF), and cerebral blood volume (CBV), which are used as blood flow characteristics; they can comprehensively characterize the state of collateral vessels, predict early changes in infarct volume, and predict the outcome of lesion tissue through the dynamic characteristics of infarct growth. Therefore, the perfusion model can better integrate all available image feature statistical models and accurately predict the final changes in lesion volume.

[0030] The ischemic penumbra is the result of blood perfusion, which comes from the pressure difference generated by the collateral circulation under the interaction of different tissues such as arteries, brain tissue, capillaries, and collateral components. The dynamic balance of this pressure difference determines the instability of the ischemic penumbra, that is, the risk of the ischemic penumbra transforming into infarction, or the possibility of recovery to asymptomatic ischemic areas or even normal blood supply. Dynamic blood flow monitoring plays a key role in predicting prognosis.

[0031] This embodiment uses the 3D-Unet model with the advantage of prior vascular field segmentation to perform whole-brain vascular segmentation on CTP images. After obtaining the whole-brain vascular segmentation results, the 3D and 2D whole-brain vascular features and the difference features between the healthy and affected sides are extracted from the results; then, the perfusion index pseudo-color map is constructed based on the time-density curve of each pixel in the CTP image, and the grayscale change characteristics and texture characteristics of the infarct core area in the pseudo-color map are calculated; in order to obtain the blood flow change characteristics of the key venous points of the whole brain, 16 key vascular points are determined in the CTP image (or CTP image), and their TTP, CBV and CBF characteristics are calculated according to the corresponding time-density curves, and finally the blood flow difference characteristics of the key vascular points are obtained according to the corresponding definitions; finally, a feature screening method based on minimum absolute shrinkage and selection operator is used to extract a few high-resolution features and establish a prognosis prediction model based on machine learning. The specific process is as follows:

[0032] Figure 1 A flow chart of a method for predicting stroke prognosis provided by one embodiment of the present invention. Figure 1As shown, this embodiment provides a method for predicting the prognosis of stroke, including: step S1, obtaining a CTP (computed tomography perfusion) image of the whole brain blood vessels. Step S2, segmenting the CTP image to obtain the segmented CTP image. Step S3, extracting features from the segmented CTP image to obtain the morphological characteristics of the whole brain blood vessels, the perfusion characteristics of the infarct core area, and the blood flow difference characteristics of the key blood vessel points. Step S4, based on the morphological characteristics of the whole brain blood vessels, the perfusion characteristics of the infarct core area, and the blood flow difference characteristics of the key blood vessel points, and based on machine learning, a stroke prognosis prediction model is constructed.

[0033] This embodiment uses machine learning to construct a stroke prognosis prediction model, which reduces the subjective differences in manual identification; it comprehensively incorporates vascular morphology, infarct core area perfusion and key vascular point characteristics to predict the prognosis of stroke based on the pathophysiology of the occurrence and development of stroke; it extracts vascular, blood flow and perfusion characteristics in a more comprehensive and refined quantitative manner; the entire method is fully automatic and does not require manual participation, and the stroke prognosis prediction model has high robustness and repeatability.

[0034] In this embodiment, the step S2 includes: step S21, normalizing the grayscale of the CTP image, and dividing the normalized CTP image into a training image and a test image.

[0035] Step S22: construct an initial 3D-Unet segmentation model under the Tensorflow (Tensorflow is a deep learning architecture) architecture.

[0036] Step S23, inputting the training image into the initial 3D-Unet (3D-Unet is a 3D-Unet convolutional neural network) segmentation model for training and verification, and adding the centerline loss to the model training loss part to obtain the segmentation model.

[0037] Step S24: input the test image into the segmentation model, and automatically segment the test image to obtain the whole brain blood vessel segmentation result.

[0038] Specifically, Figure 2 As shown, first, the CTP images of the original data set were screened and preprocessed, the CTP images of the original data were normalized to the range of 0-600, and randomly divided into a training set (n=74) and a test set (n=37).

[0039] Secondly, the 3D-Unet model was built using the tensorflow architecture, and the 3D image block sizes for input and output during the training phase were both 128×128×128. Both the encoding and decoding phases contain 4 layers of convolution. In order to better train the network, this embodiment adds batch normalization (BatchNormalization, BN) before each ReLU (Rectified Linear Unit). During model training, the weights of the cross entropy loss and the centerline loss are both 0.5. The segmentation results are evaluated using DSC (DiceSimilarity Coefficient, Dice / DSC is a commonly used segmentation evaluation indicator) and TPR (True Positive Rate) indicators.

[0040] Finally, image patch pairs of size 128×128×128 were extracted from the blood vessels and background areas of the training data. 100 image patches were randomly extracted for each case. All extracted image patches were randomly shuffled 8 times and input into the network training. During the training process, the batch size was set to 2, the convolution kernel L2 norm constraint weight was 1e-5, and the learning rate of the adaptive moment estimation (Adam) optimizer was 1e-4. The model was trained 42,000 times. In the test phase, the entire image was input into the training model to obtain the whole image segmentation result. The model was implemented using the TensorFlow (TF) library written in Python, and the computing platform was a 3.00GHz Intel Xeon E5-2587W (v4) CPU (a high-performance processor) and an Nvidia TitanXp GPU (a high-end graphics processing unit).

[0041] In this embodiment, the step S3 includes: step S31, extracting the 3D blood vessel centerline corresponding to the whole brain blood vessel in the whole brain blood vessel segmentation result by using a centerline extraction algorithm.

[0042] Step S32: Calculate morphological features of the 3D blood vessel centerline to obtain a number of 3D morphological features.

[0043] Step S33: Project the 3D blood vessels in the whole brain blood vessel segmentation result onto a 2D axial plane, and perform blood vessel enhancement and 2D threshold segmentation to obtain a 2D blood vessel segmentation result.

[0044] Step S34: extracting the 2D blood vessel centerline corresponding to the whole brain blood vessels in the 2D blood vessel segmentation result by using the centerline extraction algorithm.

[0045] Step S35: Calculate morphological features of the 2D blood vessel centerline to obtain a plurality of 2D morphological features.

[0046] Step S36, the several 3D morphological features include: several healthy side 3D morphological features and several diseased side 3D morphological features; the several 2D morphological features include: several healthy side 2D morphological features and several diseased side 2D morphological features. Compare and subtract the several healthy side 3D morphological features with the several diseased side 3D morphological features, and the several healthy side 2D morphological features with the several diseased side 2D morphological features, to obtain several 3D healthy and diseased side difference features and several 2D healthy and diseased side difference features. The morphological features of the whole brain blood vessels include: several 3D morphological features, several 2D morphological features, several 3D healthy and diseased side difference features and several 2D healthy and diseased side difference features.

[0047] Specifically, first, the centerline extraction algorithm (combined algorithms such as image expansion, corrosion, and opening and closing operations) is used to extract the 3D vascular centerline corresponding to the entire brain blood vessels. Vascular morphological feature algorithm design: Based on the 3D centerline of the blood vessel, 11 morphological features describing the vascular morphological structure are extracted, including vascular volume, number of vascular branches, total vascular length, average vascular length, average distance factor (Distance Factor, DF), average sum of angles (Sum of Angles Metric, SOAM), number of edge nodes, number of network nodes, clustering coefficient, structural entropy and average path length. The vascular morphological feature algorithm is used to calculate the 11 3D morphological features of the 3D centerline of the entire brain blood vessels.

[0048] Among them, DF is the ratio of the total length of the blood vessel to the straight-line distance between the endpoints, which is used to measure the curvature of the blood vessel. A DF value can be calculated for each blood vessel branch, and the average DF refers to the average DF of all blood vessel branches. The average SOAM is similar to DF. It can measure the tortuosity of the blood vessel. It measures the tortuosity of the blood vessel in 3D space through three vectors between four nodes. After extracting the 3D whole brain vascular features, the 3D blood vessels are projected onto the 2D axial plane and the 2D blood vessel centerline is extracted. The same method is used to extract 11 2D whole brain vascular features.

[0049] Secondly, the 3D blood vessels are projected onto the 2D axial plane, and the Hessian matrix is ​​used to enhance the blood vessels. The enhanced image is subjected to 2D threshold segmentation and the blood vessel centerline extraction operation to extract 11 morphological features of the 2D blood vessels. The steps for obtaining 2D blood vessel morphological features are the same as those for obtaining 3D blood vessel morphological features. After extracting the centerline of the 2D blood vessel segmentation result, its 11 2D morphological features are calculated.

[0050] Finally, the characteristic values ​​of the blood vessels on the healthy side were subtracted from the characteristic values ​​of the diseased side in 3D and 2D forms, respectively, to obtain 11 morphological feature differences of the blood vessels in 3D and 2D forms respectively; ultimately, 44 vascular morphological features were extracted for each case, including 22 2D and 3D features each (11 whole brain vascular features and 11 healthy and affected side morphological difference features).

[0051] In this embodiment, the step S3 further includes: step S37, drawing a time-density curve according to the segmented CTP image.

[0052] Step S38, performing Gaussian fitting on the time-density curve to obtain a first Gaussian time-density curve.

[0053] Step S39: construct a pseudo-color image according to the first Gaussian time-density curve.

[0054] Step S310: Obtain a DWI (diffusion weighted imaging) image and mark the stroke area.

[0055] Step S311 , registering the stroke area marked on the DWI image to the pseudo-color image to obtain the infarct core area on the pseudo-color image.

[0056] Step S312, extracting substantial high-throughput features of the infarct core region on the pseudo-color image to obtain the high-throughput features of the perfusion region; the perfusion features of the infarct core region include the high-throughput features of the perfusion region.

[0057] It should be noted that the perfusion index of the infarct core area directly determines the changes in stroke prognosis. However, the existing evaluation method only extracts the grayscale mean of the core area as a feature to evaluate the stroke prognosis, ignoring the structure and spatial distribution inside the core area, that is, the changes in the perfusion index inside the core area. Therefore, it is difficult to effectively obtain a higher prediction performance.

[0058] Specifically, the present embodiment first calculates the time-density curve of each pixel of the CTP image, calculates the time to peak (TTP) according to the time when the maximum contrast concentration appears on the time-density curve, calculates the cerebral blood flow (CBF) according to the maximum rising slope, and calculates the cerebral blood volume (CBV) according to the area under the curve; then constructs TTP, CBF and CBV pseudo-color maps according to the TTP, CBF and CBV values ​​of each pixel; then aligns the stroke area marked on the DWI to the perfusion pseudo-color map to obtain the infarct core area on the pseudo-color map; finally, extracts the high-throughput omics features of the stroke infarct core area on the TTP, CBF and CBV perfusion pseudo-color maps, including: 18 grayscale change features (grayscale features describe the statistical distribution of voxel intensities in the infarct and surrounding areas) and 39 texture features (texture features quantify the spatial arrangement of the infarct and surrounding areas), so as to more comprehensively describe the overall perfusion change information in the infarct core area, and the feature list is shown in Table 1.

[0059] More specifically, first, Matlab software was used to plot the time-density scatter plot of each pixel point on a single scanning layer on the CTP image over time, and the scatter plot was fitted into a Gaussian curve using Matlab's fit function. On the time-density curve, the TTP, CBF, and CBV of the pixel point could be calculated based on the time of maximum contrast concentration, the maximum rising slope, and the area under the curve.

[0060] It should be noted that the pseudo-color images of the peak time TTP, relative cerebral blood volume CBF and relative cerebral blood flow CBV indicators of the Gaussian time-density curve fitted in step S38 are specifically defined as follows: Peak time (TTP): the time (s) from the time when the contrast agent reaches the main artery of the imaging brain area to the time when the contrast agent reaches the maximum amount. Cerebral blood flow (CBF): the number of milliliters of blood flow per minute per 100g of brain tissue [ml / (100g.min)]. Cerebral blood volume (CBV): the amount of blood volume per 100g of brain tissue [ml / 100g].

[0061] like Figure 4 As shown, on the fitted time-density curve, TTP corresponds to the time from the starting point to the highest density point, CBF corresponds to the maximum slope value in the density rising stage, and CBV corresponds to the area value under the time-density curve.

[0062] Secondly, after calculating the TTP, CBF, and CBV values ​​of all pixels, the whole numerical image of the scan layer was visualized using Matlab to construct the TTP, CBF, and CBV pseudo-color map of the scan layer. The patient's DWI image was registered with the obtained TTP, CBF, and CBV pseudo-color map using the spm12 tool (Statistical Parametric Mapping, a tool for brain imaging data analysis), and the infarct core area corresponding to the pseudo-color map was found from the registration result. The lesion mask in the DWI image was jointly determined by two senior radiologists on the ITK-SNAP software (version 3.6.0; www.itk-snap.org).

[0063] Finally, the calculate_glcm function of Python was used to calculate the grayscale co-occurrence matrix of the image, from which high-throughput features of the stroke area were extracted: 18 grayscale features were calculated using functions such as mean variance, and 39 texture features were extracted using the extract_texture_features function. Finally, 171 CTP imaging features were extracted for each case, that is, 171 high-throughput features of the perfusion area were obtained for each case, but the present invention is not limited to this.

[0064] Table 1 List of image features for high-throughput feature extraction

[0065]

[0066] This embodiment uses high-throughput imaging omics features including grayscale features and texture features to describe the parenchymal perfusion information of the infarct core area, and fully describes the changes in the perfusion information of the infarct core area.

[0067] In this embodiment, the step S3 further includes: step S313, marking a number of key blood vessel points in the segmented CTP image.

[0068] Step S314: Draw a time-density curve of the key blood vessel points according to the pixel values ​​of the marked key blood vessel points on the segmented CTP image.

[0069] Step S315 , performing Gaussian fitting on the time-density curve of the key blood vessel point to obtain a second Gaussian time-density curve.

[0070] Step S316: Calculate the TTP, CBF and CBV of each of the key blood vessel points according to the second Gaussian time-density curve, and use the TTP, CBF and CBV as key blood flow point features.

[0071] Step S317: Calculate the difference between the key blood flow point features of different key blood vessel points.

[0072] Specifically, first, two senior radiologists annotated 16 key vascular points in the CTP images, including: 1. The end of the internal carotid artery - healthy side (P1); 2. The end of the internal carotid artery - affected side (P2); 3. The middle cerebral artery - healthy side (P3); 4. The middle cerebral artery - affected side (P4); 5. The origin of the basilar artery (P5); 6. The end of the basilar artery (P6); 7. The apex of the superior sagittal sinus (P7); 8. The lower The apex of the sagittal sinus (P8); 9. The junction of the inferior sagittal sinus and the straight sinus (P9); 10. The confluence of the sinuses (P10); 11. The junction of the transverse sinus and the sigmoid sinus - the healthy side (P11); 12. The junction of the transverse sinus and the sigmoid sinus - the affected side (P12); 13. The thalamostriate vein - the healthy side (P13); 14. The thalamostriate vein - the affected side (P14); 15. The internal cerebral vein - the healthy side (P15); 16. The internal cerebral vein - the affected side (P16). All the data were determined on the ITK-SNAP software.

[0073] Secondly, according to the changes in the blood vessel pixels detected when the contrast agent passes through, the time-density curves of the blood flow at 16 blood vessel points are fitted, and the TTP, CBV and CBF indicators of each blood vessel point are calculated. For each indicator, the difference between different blood vessel points is calculated according to Table 2. Finally, 249 (3*83) key blood flow difference features of blood vessel points are extracted.

[0074] Specifically, (1) drawing the key point blood vessel time-density curve: drawing the key point blood vessel time-density curve according to the pixel value of the marked key point blood vessel on the CTP image.

[0075] (2) Fitting Gaussian curve: Perform Gaussian fitting on the time-density curve obtained in (1).

[0076] (3) Obtaining blood flow characteristics at key points: Based on the curve fitted in (2) (i.e., the second Gaussian time-density curve), calculate TTP, CBF, and CBV as blood flow characteristics. Figure 4 As shown, on the fitted time-density curve, TTP corresponds to the time from the starting point to the highest density point, CBF corresponds to the maximum slope value in the density rising stage, and CBV corresponds to the area value under the time-density curve;

[0077] (4) Obtaining key point difference features: For each indicator, the difference between different vascular points is calculated according to Table 2. The difference of the key blood flow point feature is the difference between the TTP, CBF and CBV of different key vascular points. It should be noted that according to the definition of key blood flow point difference features, the difference between the TTP, CBF and CBV corresponding to the key blood flow point is the key blood flow point difference feature.

[0078] Finally, 249 (3*83) key vascular point blood flow difference features were extracted, and the definitions of 83 features are shown in Table 2.

[0079] Table 2 Definition of key vascular point difference features

[0080]

[0081] This embodiment uses the perfusion characteristics of key venous vascular points to describe the blood flow information of the downstream area of ​​the infarction, and uses the perfusion characteristics of a few key vascular points to describe the blood flow information of the entire downstream area of ​​the infarction, which can effectively reflect the blood flow situation of the entire brain.

[0082] In this embodiment, the step S4 includes: step S41, dividing the data set into a cross-validation set and an independent test set; and using the model established by the cross-validation set to test the independent test set to obtain a test result.

[0083] Step S42: According to the test results and based on the cross-validation set, a sparse representation feature selection model is established using a minimum shrinkage operator and a selection operator, and the sparse representation feature selection model is used to perform feature selection on the morphological characteristics of the entire cerebral blood vessels, the perfusion characteristics of the infarct core area, and the blood flow difference characteristics of the key points to obtain high-resolution features.

[0084] Step S43, obtain the 3-month modified Rankin scale, use the 3-month modified Rankin scale as a label to establish a 2-classification neural network model based on a support vector machine; divide the high-resolution features into a training set, train the 2-classification neural network model to establish the stroke prognosis prediction model. It should be noted that, based on the patient's high-resolution features as features, the 3-month modified Rankin scale (the modified Rankin scale is a simplified overall assessment of the patient's neurological function scale, and the patient can determine the stroke recovery status based on this scale) is used as a label to establish a 2-classification model (Rankin scale value less than 3 means good prognosis, and the classification result is counted as 1; Rankin scale value greater than 3 means poor prognosis, and the classification result is counted as 0). Therefore, the model has the ability to predict the patient's prognosis based on the input high-resolution features (prediction of 1 represents a good prognosis, and prediction of 0 represents a bad prognosis).

[0085] In this embodiment or some other embodiments, step S5 is further included: using the stroke prognosis prediction model to predict cases with unknown prognosis results.

[0086] The step S5 comprises: step S51, after calculating the morphological characteristics of the whole cerebral blood vessels, the perfusion characteristics of the infarct core area and the blood flow difference characteristics of the key points of the case with unknown prognosis, feature screening is performed to obtain the high-resolution characteristics of the unknown case.

[0087] Step S52, inputting the high-resolution features of the unknown case into the stroke prognosis prediction model, and obtaining a result of 1 represents that the case prediction result is good; and obtaining a result of 0 represents that the case prediction result is bad.

[0088] The method further comprises step S6: evaluating the stroke prognosis prediction model by using accuracy, sensitivity, specificity, positive predictive value and negative predictive value.

[0089] Specifically, Figure 3 As shown in the figure, the extracted high-dimensional features have certain redundancy, so a feature screening method based on minimum absolute shrinkage and selection operator is established to select a few high-resolution features; the modified Rankin scale prediction model for stroke prognosis 3 months is modeled as a 2-classification problem, that is, the scale value less than 3 indicates a good prognosis, and greater than 3 indicates a bad prognosis. A classification model based on support vector machine is established to classify and predict it. The receiver operating characteristic (ROC) curve is used to evaluate the performance of the four prediction models to compare the effectiveness of the models.

[0090] It should be noted that before establishing the machine learning prediction model, an independent sample T test was performed using SPSS (Statistical Product and Service Solutions) software to evaluate the differences between the good and poor prognosis groups for each feature. The false discovery rate (FDR) method was used to correct the test results to improve the reliability of the test results. The FDR correction was performed by the p.adjust function in the R language package. After the correction, the 3-month modified Rankin scale value was less than 3 for a good prognosis and greater than 3 for a bad prognosis. The stroke prognosis prediction model was modeled as a 2-classification prognosis prediction problem, and a classifier based on a support vector machine was established.

[0091] More specifically, 1) Clinical task definition: The 3-month modified Rankin Scale for stroke prognosis, i.e., the 90-day mRS score, was predicted and modeled as a 2-classification problem, i.e., a scale value less than 3 indicated a good prognosis, and a scale value greater than 3 indicated a poor prognosis.

[0092] 2) Dataset division: The data set is divided into a cross-validation set and an independent test set in a ratio of 2:1. The cross-validation set is used for ten-fold cross-validation. Finally, the model built using all the data in the validation set is used to test the test set.

[0093] 3) Feature screening: Based on the cross-validation set data, a sparse representation feature selection model is established using the minimum shrinkage operator and selection operator to perform feature selection and select a small number of high-resolution features.

[0094] 4) Classification model construction: Establish a 2-classification model based on support vector machine.

[0095] 5) Model performance evaluation indicators: Accuracy (ACC), sensitivity (SEN), specificity (SPE), positive predictive value (PPV), and negative predictive value (NPV) are used to evaluate the performance of the prediction model. The calculation formulas for ACC, SEN, SPE, PPV, and NPV are as follows:

[0096]

[0097] In the formula, "True Positive" (TP) indicates the case of correct identification of infarction enlargement, "True Negative" (TN) indicates the case of correct identification of infarction reversal, "False Positive" (FP) indicates the case of actual infarction reversal being identified as infarction enlargement, and "False Negative" (FN) indicates the case of actual infarction enlargement being identified as infarction reversal.

[0098] The results of this embodiment are analyzed as follows:

[0099] Finally, 111 patients were enrolled in the clinical CTP images. For the vascular segmentation, the 3D-Unet model achieved a DSC (Dice Similarity Coefficient) of 0.83 and a TPR (True Positive Rate) of 78.96%. For the feature part, a total of 464 features describing blood vessels and blood flow were extracted from each patient, including 44 vascular morphology features, 249 vascular perfusion features, and 171 parenchymal perfusion features.

[0100] Table 3 shows the classification prediction results of 90-day mRS by the method of this embodiment and the comparative method. The method of this embodiment achieved an AUC of 0.889, which is about 8% higher than the comparative method. In addition, the sensitivity and specificity of the prediction results of the method of this embodiment are within 5%, indicating that the method of this embodiment has good prediction performance for each type of sample. Figure 3 As shown, the ROC curve of the prognosis prediction result of the method of this embodiment is displayed.

[0101] This embodiment uses the combination of global cerebral vascular blood flow and parenchymal perfusion characteristics to predict the prognosis of stroke and the modified Rankin scale value (90-day mRS score) of stroke prognosis 3 months later, and the model has achieved satisfactory results, reflecting the practicality and universal applicability of the method of this embodiment. In addition, since the entire prediction process of the method of this embodiment is automatically completed by the algorithm, there is no need for manual intervention to extract prediction parameters, so it also has strong robustness in practical applications.

[0102] Table 3 90-day mRS prediction results of different methods. []: 95% confidence interval

[0103]

[0104] On the other hand, the present embodiment also provides a stroke prognosis prediction system, including: a whole brain blood vessel segmentation module, used to segment the CTP image of the whole brain blood vessels to obtain the segmented CTP image. A feature extraction module, used to extract features from the whole brain blood vessel segmentation result, and obtain the morphological characteristics of the whole brain blood vessels, the perfusion characteristics of the infarct core area, and the blood flow difference characteristics of key blood vessel points. A prognosis prediction model establishment module based on machine learning, used to construct a stroke prognosis prediction model based on the morphological characteristics of the whole brain blood vessels, the perfusion characteristics of the infarct core area, and the blood flow difference characteristics of key blood vessel points.

[0105] In other aspects, this embodiment further provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the method described above is implemented.

[0106] This embodiment first uses the 3D-Unet network model to automatically segment the blood vessels in the computed tomography perfusion (CTP) image; then extracts the morphological characteristics of the whole brain blood vessels, the parenchymal perfusion characteristics of the infarct core area, and the blood flow difference characteristics of key vascular points. In order to reduce feature redundancy and improve the accuracy of prognosis prediction, the minimum absolute shrinkage and selection operator algorithm are used to select a small number of high-resolution features; finally, a stroke prognosis prediction model based on a support vector machine is established based on the selected features. The method of this embodiment not only has a high prediction accuracy, but the entire process is automatically completed by a computer without manual intervention, and can be used for clinical stroke 3-month modified Rankin scale prediction.

[0107] In summary, this embodiment studies the prediction of stroke prognosis based on the characteristics of whole brain vascular blood flow and parenchymal perfusion. The deep learning network is used to automatically segment the whole brain blood vessels to reduce the subjective differences in manual identification. From the pathophysiology of the occurrence and development of stroke, the vascular morphology, infarct area perfusion and key venous vascular point characteristics are comprehensively incorporated to predict the prognosis of stroke. The vascular, blood flow and perfusion characteristics are extracted in a more comprehensive and refined quantitative manner. The entire method and system are fully automatic and do not require manual participation. The model has high robustness and repeatability, high prediction accuracy, and strong practicality and universal applicability.

[0108] The overall system of this embodiment can realize automatic quantification of the comprehensive characteristics of the whole brain blood vessels, especially quantification of the difference characteristics of the blood vessels on the healthy and affected sides. Different from the traditional manual method of only evaluating the blood vessel density of a specific area, the method of the present invention reduces the labor cost while utilizing more comprehensive blood vessel characteristics.

[0109] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0110] It should be noted that the devices and methods disclosed in the embodiments of this article can also be implemented in other ways. The device implementation described above is only schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the devices, methods and computer program products according to multiple embodiments of this article. In this regard, each box in the flowchart or block diagram can represent a part of a module, program or code, and the module, program segment or part of the code contains one or more executable instructions for implementing the specified logical function, and the module, program segment or part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented by a dedicated hardware-based system for performing a specified function or action, or can be implemented by a combination of dedicated hardware and computer instructions.

[0111] In addition, the functional modules in the various embodiments of this document may be integrated together to form an independent part, or each module may exist independently, or two or more modules may be integrated to form an independent part.

[0112] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be appreciated that the above description should not be considered as a limitation of the present invention. After reading the above content, it will be apparent to those skilled in the art that various modifications and substitutions of the present invention will occur. Therefore, the protection scope of the present invention should be limited by the appended claims.

Claims

1. A method for predicting the prognosis of stroke, characterized in that: include: Step S1, obtaining a CTP image of the entire cerebral blood vessels; Step S2, segmenting the CTP image to obtain the segmented CTP image; Step S3, extracting features from the segmented CTP image to obtain morphological features of the entire cerebral blood vessels, perfusion features of the infarct core area, and blood flow difference features of key vascular points; Step S4: constructing a stroke prognosis prediction model based on machine learning according to the morphological characteristics of the entire brain blood vessels, the perfusion characteristics of the infarct core area and the blood flow difference characteristics of the key vascular points.

2. The method for predicting stroke prognosis according to claim 1, characterized in that: The step S2 comprises: Step S21, normalizing the grayscale of the CTP image, and dividing the normalized CTP image into a training image and a test image; Step S22: construct an initial 3D-Unet segmentation model under the Tensorflow architecture; Step S23, inputting the training image into the initial 3D-Unet segmentation model for training and verification, and adding the centerline loss to the model training loss part to obtain a segmentation model; Step S24, inputting the test image into the segmentation model, and automatically segmenting the test image to obtain the segmented CTP image; the segmented CTP image includes the whole brain blood vessel segmentation result.

3. The method for predicting stroke prognosis according to claim 2, characterized in that: The step S3 comprises: Step S31, extracting the 3D blood vessel centerline corresponding to the whole brain blood vessel in the whole brain blood vessel segmentation result by using a centerline extraction algorithm; Step S32, calculating morphological features of the 3D blood vessel centerline to obtain a plurality of 3D morphological features; Step S33, projecting the 3D blood vessels in the whole brain blood vessel segmentation result to a 2D axial plane, and performing blood vessel enhancement and 2D threshold segmentation to obtain a 2D blood vessel segmentation result; Step S34, extracting the 2D blood vessel centerline corresponding to the whole brain blood vessels in the 2D blood vessel segmentation result by using the centerline extraction algorithm; Step S35, calculating morphological features of the 2D blood vessel centerline to obtain a plurality of 2D morphological features; Step S36, the plurality of 3D morphological features include: a plurality of healthy-side 3D morphological features and a plurality of diseased-side 3D morphological features; the plurality of 2D morphological features include: a plurality of healthy-side 2D morphological features and a plurality of diseased-side 2D morphological features; Compare and subtract the healthy side 3D morphological features with the diseased side 3D morphological features, and the healthy side 2D morphological features with the diseased side 2D morphological features, to obtain a number of 3D healthy-and-affected side difference features and a number of 2D healthy-and-affected side difference features; The morphological features of the whole brain blood vessels include: a number of 3D morphological features, a number of 2D morphological features, a number of 3D healthy-side and affected-side difference features, and a number of 2D healthy-side and affected-side difference features.

4. The method for predicting stroke prognosis according to claim 1, characterized in that: Step S3 also includes: Step S37, drawing a time-density curve according to the segmented CTP image; Step S38, performing Gaussian fitting on the time-density curve to obtain a first Gaussian time-density curve; Step S39, constructing a pseudo-color image according to the first Gaussian time-density curve; Step S310, obtaining a DWI image and marking the stroke area; Step S311, registering the stroke area marked on the DWI image to the pseudo-color image to obtain the infarct core area on the pseudo-color image; Step S312, extracting substantial high-throughput features of the infarct core region on the pseudo-color image to obtain the high-throughput features of the perfusion region; the perfusion features of the infarct core region include the high-throughput features of the perfusion region.

5. The method for predicting stroke prognosis according to claim 1, characterized in that: The step S3 further comprises: Step S313, marking a number of key blood vessel points in the segmented CTP image; Step S314, drawing a time-density curve of the key blood vessel points according to the pixel values ​​of the marked key blood vessel points on the segmented CTP image; Step S315, performing Gaussian fitting on the time-density curve of the key blood vessel point to obtain a second Gaussian time-density curve; Step S316, calculating the TTP, CBF and CBV of each of the key blood vessel points according to the second Gaussian time-density curve, and using the TTP, CBF and CBV as key blood flow point features; Step S317: Calculate the difference between the key blood flow point features of different key blood vessel points to obtain key blood flow point difference features.

6. The method for predicting stroke prognosis according to claim 1, characterized in that: The step S4 comprises: Step S41, dividing the data set into a cross-validation set and an independent test set; and using the model established by the cross-validation set to test the independent test set to obtain a test result; Step S42: according to the test results and based on the cross-validation set, a sparse representation feature selection model is established using a minimum shrinkage operator and a selection operator, and the sparse representation feature selection model is used to perform feature selection on the morphological features of the whole cerebral blood vessels, the perfusion features of the infarct core area, and the blood flow difference features of the key points to obtain high-resolution features; Step S43, obtaining a 3-month modified Rankin scale, and using the 3-month modified Rankin scale as a label to establish a 2-classification neural network model based on a support vector machine; dividing the high-resolution features into a training set, and training the 2-classification neural network model to establish the stroke prognosis prediction model.

7. The method for predicting stroke prognosis according to claim 1, characterized in that: The method further comprises step S5: using the stroke prognosis prediction model to predict cases with unknown prognosis results; The step S5 comprises: Step S51, after calculating the morphological features of the whole cerebral blood vessels, the perfusion features of the infarct core area and the key blood flow point difference features of the case with unknown prognosis, feature screening is performed to obtain the high-resolution features of the unknown case; Step S52, inputting the high-resolution features of the unknown case into the stroke prognosis prediction model, and obtaining a result of 1 represents that the prediction result of the case is good; and obtaining a result of 0 represents that the prediction result of the case is bad.

8. The method for predicting stroke prognosis according to claim 1, characterized in that: The method further comprises step S6, evaluating the stroke prognosis prediction model by using accuracy, sensitivity, specificity, positive predictive value and negative predictive value.

9. A stroke prognosis prediction system, characterized in that: include: A whole brain blood vessel segmentation module is used to segment the CTP image of the whole brain blood vessels to obtain the segmented CTP image; A feature extraction module is used to extract features from the segmented CTP image to obtain the morphological features of the whole brain blood vessels, the perfusion features of the infarct core area and the blood flow difference features of key vascular points; The prognosis prediction model establishment module based on machine learning is used to construct a stroke prognosis prediction model based on the morphological characteristics of the entire brain blood vessels, the perfusion characteristics of the infarct core area and the blood flow difference characteristics of key vascular points.

10. An electronic device, characterized in that: The method comprises a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the method according to any one of claims 1 to 8 is implemented.

Images