Method and device for grading arteriovenous malformations based on time-enhanced radiomics

Abstract

The present invention provides a method and device for grading arteriovenous malformations based on time-enhanced radiomics. First, a neural network is used to detect blood vessels in the DSA images of patients, and the frame numbers at which the blood vessels start and end to appear are recorded; the frame numbers at which the blood vessels start and end to appear are divided by the DSA sampling frame rate to obtain the time points at which the detected blood vessels start and end to appear, and time features are obtained through calculation; since multiple frames of DSA images can be synthesized into a minimum intensity map by taking the minimum pixel value of each pixel point on the time axis, a corresponding minimum intensity map is synthesized for the DSA images corresponding to the time features, and the time features are used for weighting to obtain a time-enhanced minimum intensity map; time-enhanced radiomics features are obtained by calculating the time-enhanced minimum intensity map, and a prediction model for grading arteriovenous malformations is constructed by means of machine learning. This method and device can accurately and simply evaluate the high and low grades of arteriovenous malformations.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer-aided diagnosis, and particularly relates to a method and device for grading arteriovenous malformations based on time-enhanced radiomics. Background Art

[0002] Due to the lack of capillaries in the lesion area of arteriovenous malformations, the direct connection between arteries and veins in the lesion area will cause arteriovenous shunts, increasing the risk of bleeding and steal syndrome. The clinical manifestations of arteriovenous malformations are diverse, such as neurological dysfunction, increased intracranial pressure, epilepsy, headache, etc. Intracerebral hemorrhage is the most common symptom in young and middle-aged patients with arteriovenous malformations, accounting for 52-77%, and the mortality rate is 10-15%.

[0003] DSA images are the gold standard for the diagnosis of arteriovenous malformations. DSA has high temporal and spatial resolution and can accurately describe the location, size, number of feeding vessels, drainage pattern of the lesion, and determine whether there are hemodynamic-related aneurysms, etc. The Spetzler-Martin (SM) grading system is a commonly used grading system based on DSA images, which is used to evaluate the size of arteriovenous malformation lesions, the brain functional areas adjacent to arteriovenous malformations, and the venous drainage pattern. Clinical decisions regarding the treatment of arteriovenous malformations are usually made based on the SM score. Specifically, low-grade arteriovenous malformations can be resected by single-stage surgery with low risk, while high-grade arteriovenous malformations require multi-modal and multi-stage surgical resection.

[0004] However, the clinical discrimination of DSA images is difficult, and it is difficult for primary and inexperienced neuroradiologists and neurosurgeons to accurately interpret them alone. Therefore, a method for grading low-grade (SM I, II, III) and high-grade (SM IV, V) arteriovenous malformations based on DSA images is needed to provide auxiliary treatment suggestions for doctors. Summary of the Invention

[0005] The present invention is made to solve the above problems, and aims to provide a method and device that can classify low-grade and high-grade arteriovenous malformations, thereby providing auxiliary treatment suggestions. The aim is to construct a grading prediction model for arteriovenous malformations, utilize the time information obtained by DSA, and weight the radiomics so that the extracted features integrate time information and spatial features to more accurately and simply evaluate the high and low grades of arteriovenous malformations.

[0006] The present invention adopts the following technical solutions:

[0007] The present invention provides a method for grading arteriovenous malformations based on time-enhanced radiomics, which is characterized by including the following steps:

[0008] Step S1: Use a convolutional neural network to detect blood vessels in the patient's DSA images, and record the frames at which the blood vessels start and end developing.

[0009] Step S2: Based on the frames at which the blood vessels start and end developing and the DSA sampling frame rate, calculate the time points at which the blood vessels start and end developing.

[0010] Step S3: Calculate the time characteristics of blood vessel development based on the time points at which the blood vessels start and end developing.

[0011] Step S4: Synthesize multiple frames of the DSA images corresponding to the time characteristics into a minimum intensity map, and use the time characteristics to weight the minimum intensity map to obtain a time-enhanced minimum intensity map.

[0012] Step S5: Extract radiomics features from the time-enhanced minimum intensity map.

[0013] Step S6: Based on the radiomics features and the time characteristics, construct an arteriovenous malformation grading prediction model through a machine learning method for high and low level evaluation of the arteriovenous malformation grade.

[0014] The method for grading arteriovenous malformations based on time-enhanced radiomics provided by the present invention may also have the following technical characteristics. Among them, step S1 includes the following sub-steps:

[0015] Step S1-1: Annotate the blood vessel structure information of the DSA images of multiple arteriovenous malformation cases.

[0016] Step S1-2: Use the DSA images with annotations to train the YOLOv5 convolutional neural network to obtain a target detection network for detecting blood vessel structures.

[0017] Step S1-3: Use the target detection network to detect blood vessels in the DSA images, detect the blood vessels, and record the frames at which the blood vessels start developing and the frames at which the blood vessels end developing.

[0018] The method for grading arteriovenous malformations based on time-enhanced radiomics provided by the present invention may also have the following technical characteristics. Among them, in step S1-1, each case uses the DSA images of two sections, the anteroposterior view of the internal carotid artery and the lateral view of the internal carotid artery. The DSA images of each section are respectively annotated with four blood vessel structures: the internal carotid artery, artery, vein, and venous sinus. In step S1-2, the YOLOv5 convolutional neural network is YOLOv5s. During the training process, a pre-trained model is used to fine-tune the blood vessel detection task to obtain a convolutional neural network for detecting blood vessel structures, that is, the target detection network.

[0019] The arteriovenous malformation grading method based on time-enhanced radiomics provided by the present invention may further have the following technical feature: in step S1, the frame numbers at which the blood vessels start and end developing are marked as a vector:

[0020] α = [ICA sf , ICA ef , A sf , A ef , V sf , V ef , VS sf , VS ef

[0021] In step S2, the time points at which the blood vessels start and end developing are calculated as:

[0022]

[0023] β = [ICA s , ICA e , A s , A e , V s , V e , VS s , VS e

[0024] In the formula, ICA is the internal carotid artery, A is the artery, V is the vein, VS is the venous sinus, and the subscript sf or ef indicates the frame number at which the blood vessel starts or ends developing, is the DSA sampling frame rate, and the subscript s or e indicates the time at which the blood vessel starts or ends developing.

[0025] The arteriovenous malformation grading method based on time-enhanced radiomics provided by the present invention may further have the following technical feature: in step S3, the time feature is calibrated as t = [t0, t1, L, t 21 , and the time feature is calculated as:

[0026] t0 = ICA s

[0027] t1 = ICA e -t0

[0028] t2 = A s -t0

[0029] t3 = A e -t0

[0030] t4 = V s -t0

[0031] t5 = V​​e -t0

[0032] t6 = VS s -t0

[0033] t7 = VS e -t0

[0034]

[0035]

[0036]

[0037]

[0038] t 12 = t4 - t2

[0039] t 13 = t6 - t2

[0040] t 14 = L(t4 - t3)

[0041] t 15 = L(t3 - t4)

[0042] t 16 = L(t6 - t3)

[0043] t 17 = L(t3 - t6)

[0044] t 18 = L(t6 - t4)

[0045] t 19 = L(t4 - t6)

[0046] t 20 = L(t7 - t5)

[0047] t 21 = L(t5 - t7)

[0048] wherein, L(x) = x · I(x), and I(·) is an indicator function;

[0049] The method for grading arteriovenous malformations based on time-enhanced radiomics provided by the present invention may further have the following technical feature: in step S4, for the time feature τ i , multiple frames of the DSA images related to it are synthesized into a minimum intensity map For each frame of the DSA image with a size of M × N, the minimum intensity map is expressed as:

[0050]

[0051] Wherein, m ∈ (1, M), n ∈ (1, N), and p(m, n, f) is the pixel value of the f-th frame image at the pixel point (m, n), F si and F ei are respectively the start and end frame numbers of the images related to the time feature τ i , τ = [t8, t9, t 14 , t 15 , t 16 , t 17 , t 18 , t 19 , and then perform weighted calculation on the time feature τ i and the corresponding minimum intensity map to obtain the time-enhanced minimum intensity map:

[0052]

[0053] Wherein, τ i ∈ τ, 0 ≤ i < 7.

[0054] The arteriovenous malformation grading method based on time-enhanced radiomics provided by the present invention may also have the following technical feature: in step S5, for each of the time-enhanced minimum intensity maps, high-throughput features in three aspects of the first-order histogram, texture, and wavelet are extracted.

[0055] The arteriovenous malformation grading method based on time-enhanced radiomics provided by the present invention may also have the following technical feature: step S6 includes the following sub-steps: step S6-1, splice the extracted radiomics features with the corresponding time features to obtain a plurality of spliced features; step S6-2, perform feature screening on the spliced features by the sparse representation method to obtain different classification categories; step S6-3, perform classification discrimination based on the support vector machine so that the classification categories correspond to the arteriovenous malformation levels, and learn the classification method by machine learning to construct the arteriovenous malformation grading prediction model.

[0056] The arteriovenous malformation grading method based on time-enhanced radiomics provided by the present invention may also have the following technical feature: in step S6-2, the following expression is used to perform feature screening on the spliced features:

[0057]

[0058] Wherein, y is the classification category, D = [D1, D2,..., D ILet \(\Omega\) be the set of all dictionaries, \(a\) be the sparse coefficient, \(\mu\) be the regularization parameter greater than 0, and \(\|\cdot\|\) p denote \(l\) p regularization, \(\hat{a}\) denote the estimated value of the sparse coefficient \(a\),

[0059] In step S6 - 3, the computational model of the support vector machine is:

[0060]

[0061] y i (w T x i +b)\(\geq1 - \xi\) i , \(i = 1,\cdots,n\)

[0062] \(\xi\) i \(\geq0\), \(i = 1,\cdots,n\)

[0063] where \((x\) i ,y i ) is a sample in the given training set \(R=[(x_1,y_1),(x_2,y_2),\cdots,(x\) n ,y n )], \(y\) i \(\in(-1, + 1)\), \(w=(w_1;w_2;\cdots;w\) d ) is the normal vector that determines the direction of the hyperplane, \(b\) is the displacement term that determines the distance between the hyperplane and the origin, \(\xi\) i is the slack variable, and \(C\) is the penalty factor.

[0064] The radial basis function is used as the kernel function of the support vector machine, and the radial basis function is expressed as:

[0065]

[0066] where \(\|x\) i -y i \| represents the squared Euclidean distance between two feature vectors, and \(\sigma\) is a free parameter.

[0067] The present invention provides a device for grading arteriovenous malformations based on time-enhanced radiomics, which is characterized by including: a DSA image acquisition module for acquiring DSA images of a patient; a vascular imaging frame number acquisition module that uses a convolutional neural network to detect blood vessels in the DSA images and records the frame numbers at which the blood vessels start and end imaging; a vascular imaging time point acquisition module that calculates the start and end imaging time points of the blood vessels based on the frame numbers at which the blood vessels start and end imaging and the DSA sampling frame rate; a time feature acquisition module that calculates the time features of the blood vessel imaging based on the start and end imaging time points of the blood vessels; a minimum intensity map acquisition module that synthesizes multiple frames of the DSA images corresponding to the time features into a minimum intensity map and weights the minimum intensity map using the time features to obtain a time-enhanced minimum intensity map; a radiomics feature acquisition module that calculates time-enhanced radiomics features based on the time-enhanced minimum intensity map; and a feature screening and classification discrimination module that constructs a prediction model for grading arteriovenous malformations based on the time-enhanced radiomics features and the time features using a machine learning method for high and low level evaluation of the arteriovenous malformation grade.

[0068] Functions and effects of the invention

[0069] According to the method and device for grading arteriovenous malformations based on time-enhanced radiomics of the present invention, first, a neural network is used to detect blood vessels in the DSA images of a patient, and the frame numbers at which the blood vessels start and end imaging are recorded; the frame numbers at which the blood vessels start and end imaging are divided by the DSA sampling frame rate to obtain the start and end imaging time points of the detected blood vessels, and time features are calculated through calculation; since multiple frames of DSA images can be synthesized into a minimum intensity map by taking the minimum pixel value of each pixel point on the time axis, the DSA images corresponding to the time features are synthesized into corresponding minimum intensity maps, and the time features are used for weighting to obtain a time-enhanced minimum intensity map; time-enhanced radiomics features are calculated from the time-enhanced minimum intensity map, and a prediction model for grading arteriovenous malformations is constructed by means of machine learning. This method and device can accurately and simply evaluate the high and low grades of arteriovenous malformations. Description of the drawings

[0070] Figure 1 is the schematic diagram of the method for grading arteriovenous malformations based on time-enhanced radiomics in an embodiment of the present invention;

[0071] Figure 2 is the flowchart of the method for grading arteriovenous malformations based on time-enhanced radiomics in an embodiment of the present invention;

[0072] Figure 3 is the structural block diagram of the device for grading arteriovenous malformations based on time-enhanced radiomics in an embodiment of the present invention. Detailed implementation manners

[0073] In order to make the technical means, creative features, achieved objectives and functions realized by the present invention easy to understand, the following specifically describes the arteriovenous malformation grading method and device based on time-enhanced radiomics of the present invention in conjunction with embodiments and drawings.

[0074] For ease of understanding, first, a brief description of the technical terms related to the present invention is given.

[0075] DSA: Digital Subtraction Angiography (DSA) images have high temporal and spatial resolutions and can accurately describe the location, size, number of feeding vessels, drainage patterns of lesions, and determine whether there are hemodynamic-related aneurysms, etc. Convolutional Neural Network: It is a class of feedforward neural networks (Feedforward Neural Networks) that contain convolutional calculations and have a deep structure, and is one of the representative algorithms of deep learning; it can perform supervised learning and unsupervised learning and is applied in fields such as computer vision and natural language processing. YOLOv5: YOLO is a convolutional neural network applied to object detection tasks in computer vision. It solves object detection as a regression problem and realizes end-to-end detection. YOLOv5 is the fifth version of YOLO released in 2020. Minimum intensity map: Multiple consecutive DSA images can be synthesized into a minimum intensity map by taking the minimum pixel value of each pixel point on the time axis. Since DSA images are grayscale images, obtaining the minimum pixel value on the time axis can make all the blood vessels appearing in the consecutive multiple DSA images appear in the minimum intensity map. Radiomics: Radiomics is an interdisciplinary research direction between computer science and medicine, which refers to the high-throughput extraction of a large amount of image information from images (CT, MRI, PET, etc.) to achieve tumor segmentation, feature extraction, and model establishment, and assist physicians in making the most accurate diagnosis by deeper mining, prediction, and analysis of massive image data information.

[0076] Sparse representation: The purpose of signal sparse representation is to represent a signal with as few atoms as possible in a given overcomplete dictionary, which can obtain a more concise representation of the signal, so that it is easier for us to obtain the information contained in the signal and is more convenient for further processing of the signal, such as compression, encoding, etc.

[0077] Support Vector Machine: A support vector machine (SVM) is a type of generalized linear classifier that performs binary classification on data in a supervised learning manner. Its decision boundary is the maximum margin hyperplane solved for the learning samples.

[0078] Radial Basis Function: A radial basis function is a real-valued function whose value depends only on the distance from the origin, that is, Φ(x) = Φ(‖x‖), or it can also be the distance to an arbitrary point c, where point c is called the center point, that is, Φ(x, c) = Φ(‖x - c‖); any function Φ that satisfies the property Φ(x) = Φ(‖x‖) is called a radial basis function. The standard one generally uses the Euclidean distance (also called the Euclidean radial basis function), although other distance functions are also possible; in a neural network structure, it can be the main function of the fully connected layer and the ReLU layer.

[0079] <Example>

[0080] Figure 1 and Figure 2 are respectively the schematic diagram and the flowchart of the arteriovenous malformation grading method based on time-enhanced radiomics in this example.

[0081] As Figure 1-2 shown, the arteriovenous malformation grading method based on time-enhanced radiomics pre-determines samples of arteriovenous malformation grades evaluated under the known Spetzler-Martin scoring standard, and constructs an arteriovenous malformation grading prediction model through machine learning, and then performs high and low grade grading of arteriovenous malformations through the constructed venous malformation grading prediction model.

[0082] The method specifically includes the following steps:

[0083] Step S1, use a convolutional neural network to detect blood vessels in the patient's DSA image, and record the frames where the blood vessels start to appear and end to appear.

[0084] In this example, the YOLOv5s convolutional neural network is used for blood vessel detection. Taking the DSA images of 232 arteriovenous malformations as data for experiments, according to the Spetzler-Martin scoring standard, the liver fibrosis grades are divided into a total of 5 different grades: I, II, III, IV, and V. Among them, grades I, II, and III are low grades, and grades IV and V are high grades. The data distribution is: 130 cases of low grades and 102 cases of high grades.

[0085] Step S1 includes the following sub-steps:

[0086] Step S1-1: Annotate the vascular structure information in the DSA images of multiple arteriovenous malformation cases.

[0087] In this embodiment, 100 DSA images of arteriovenous malformations in the above data are annotated with vascular structure information, and the annotation result is the target box of a specific blood vessel; that is, each case uses DSA images of two sections, the anteroposterior view of the internal carotid artery and the lateral view of the internal carotid artery. Among them, the DSA images of each section are respectively annotated with the target boxes of four vascular structures: the internal carotid artery, artery, vein, and venous sinus.

[0088] Step S1-2: Use the DSA images with annotations to train the YOLOv5 convolutional neural network to obtain a target detection network that can detect vascular structures.

[0089] In this embodiment, the YOLOv5 convolutional neural network is YOLOv5s. During the training process, the pre-trained model provided by the official on the COCO128 dataset is used to fine-tune the vascular detection task, and finally a convolutional neural network that can detect vascular structures is obtained, that is, the above target detection network.

[0090] Step S1-3: Use the above target detection network to detect blood vessels in the patient's DSA images, detect the blood vessels, and record the frames at which the blood vessels start and end to appear.

[0091] In this embodiment, recording the frames at which the blood vessels start and end to appear specifically includes:

[0092] Mark the frames at which the internal carotid artery, artery, vein, and venous sinus start and end to appear as a vector:

[0093] α = [ICA sf ,ICA ef ,A sf ,A ef ,V sf ,V ef ,VS sf ,VS ef

[0094] In the formula, ICA is the internal carotid artery, A is the artery, V is the vein, VS is the venous sinus, and the subscript sf or ef indicates the frame number at which the blood vessel starts or ends to appear.

[0095] Step S2: Based on the above frames at which the blood vessels start and end to appear and the DSA sampling frame rate, calculate the start and end appearance time points of the detected blood vessels.

[0096] In this embodiment, calculating the start and end appearance time points of the detected blood vessels specifically includes:

[0097] ​

[0098] β = [ICA s , ICA e , A s , A e , V s , V e , VS s , VS e

[0099] Wherein, is the DSA sampling frame rate, ICA is the internal carotid artery, A is the artery, V is the vein, VS is the venous sinus, and the subscript s or e indicates the start or end imaging time of the blood vessel.

[0100] Step S3, calculate the time characteristics of the blood vessel imaging based on the start and end imaging time points of the blood vessel.

[0101] In this embodiment, calculate the time characteristics, and calibrate the time characteristics as t = [t0, t1, L, t 21 , specifically including:

[0102] Table 1 Time Characteristic Calculation Table

[0103]

[0104] where L(x) = x · I(x), and I(·) is the indicator function;

[0105] Step S4, since multiple frames of DSA images can be synthesized into a minimum intensity map by taking the minimum pixel value of each pixel point on the time axis, therefore, synthesize the multiple frames of DSA image frames corresponding to the time characteristics into the corresponding minimum intensity map, and weight the minimum intensity map with the time characteristics to obtain a time-enhanced minimum intensity map.

[0106] In this embodiment, obtain the time-enhanced minimum intensity map for the DSA images of two sections respectively. Synthesize the multiple frames of DSA images into a minimum intensity map by taking the minimum pixel value of each pixel point on the time axis, specifically including:

[0107] Define τ = [t8, t9, t 14 , t 15 , t 16 , t 17 , t 18 , t 19 , for the feature τ i , the multiple frames of images related to it can be synthesized into a minimum intensity map For each frame of DSA image with a size of M × N, can be expressed as: ​

[0108]

[0109] Wherein, m ∈ (1, M), n ∈ (1, N), and p(m, n, f) is the pixel value of the f-th frame image at the pixel point (m, n), F si and F ei are respectively the start and end frame numbers of the images related to the feature τ i

[0110] In this embodiment, after obtaining the minimum intensity map, it is also necessary to perform a weighted calculation on the temporal feature τ i and its corresponding minimum intensity map to obtain a temporally enhanced minimum intensity map:

[0111]

[0112] Wherein, τ i ∈ τ, 0 ≤ i < 7.

[0113] Step S5: Extract imaging features from the temporally enhanced minimum intensity map.

[0114] In this embodiment, the imaging features at least include three aspects: first-order histogram, texture, and wavelet. Extracting the imaging features specifically includes:

[0115] For each temporally enhanced minimum intensity map, extract high-throughput features in three aspects: first-order histogram, texture, and wavelet, a total of 350. The specific features and corresponding numbers are shown in Table 2 below:

[0116] Table 2 Imaging Feature Table

[0117]

[0118] The method for extracting imaging features is specifically the prior art and will not be elaborated here.

[0119] Step S6: Based on the above imaging features and temporal features, construct a prediction model for arteriovenous malformation grading through machine learning. This model is used to evaluate the high and low levels of arteriovenous malformation grades.

[0120] The construction of the prediction model for arteriovenous malformation grading includes:

[0121] Step S6-1: Concatenate the extracted imaging features with their corresponding temporal features to obtain multiple concatenated features, which are also temporally enhanced imaging features.

[0122] In this embodiment, the respective imaging features extracted from the two temporally enhanced minimum intensity maps are concatenated with their respective temporal features together, and a total of 744 concatenated features are obtained.​

[0123] Step S6-2: Screen the splicing features by sparse representation to obtain different classification categories.

[0124] In this embodiment, the screening process is expressed as:

[0125]

[0126] In the formula, represents the strain information, y is the classification category, D = [D1, D2,..., D I is the set of all dictionaries, a is the sparse coefficient, μ is a regularization parameter greater than 0, and ||·|| p represents l p regularization.

[0127] Step S6-3: Perform classification discrimination based on the support vector machine so that the classification category corresponds to the arteriovenous malformation level, and learn the classification method through machine learning to construct an arteriovenous malformation grading prediction model.

[0128] In this embodiment, the calculation model of the support vector machine is:

[0129]

[0130] y i (w T x i +b)≥1 - ξ i , i = 1,..., n

[0131] ξ i ≥0, i = 1,..., n

[0132] In the formula, (x i , y i ) is the given training set R = [(x1, y1), (x2, y2),...,(x n , y n )], y i ∈(-1, +1) is the sample in it; w = (w1; w2;...; w d ) is the normal vector, which determines the direction of the hyperplane; b is the displacement term, which determines the distance between the hyperplane and the origin; ξ i is the slack variable, and C is the penalty factor.

[0133] Moreover, the radial basis function is used as the kernel function of the support vector machine, and the radial basis function is expressed as:

[0134]

[0135] In the formula, ||xi -y i || represents the squared Euclidean distance between two feature vectors, and σ is a free parameter.

[0136] In this embodiment, the above 100 DSA images of arteriovenous malformations with annotation information are used to train the arteriovenous malformation grading prediction model. After the training is completed, the 232 above-mentioned samples are input into the trained arteriovenous malformation grading prediction model for grading prediction, and finally the classification performance results of the 232 samples are obtained. Table 3 below shows the classification performance results.

[0137] Table 3 Classification Performance Table of Arteriovenous Malformation Grading Prediction Model

[0138]

[0139] Among them, the calculation methods of accuracy (ACC), sensitivity (SENS), and specificity (SPEC) are as follows:

[0140]

[0141]

[0142]

[0143] In the formula, TP, TN, FP, and FN are the numbers of true positives, true negatives, false positives, and false negatives, respectively.

[0144] Figure 3 is the structural block diagram of the arteriovenous malformation grading device based on time-enhanced radiomics in this embodiment.

[0145] As Figure 3 shown, this embodiment also provides an arteriovenous malformation grading device 10 based on time-enhanced radiomics corresponding to the above method. The device 10 includes a DSA image acquisition module 11, a vascular imaging frame number acquisition module 12, a vascular imaging time point acquisition module 13, a time feature acquisition module 14, a minimum intensity map acquisition module 15, a radiomics feature acquisition module 16, a feature screening and classification discrimination module 17, and a control module 18.

[0146] Among them, the DSA image acquisition module 11 is used to acquire the DSA image of the patient.

[0147] The vascular imaging frame number acquisition module 12 acquires the frame numbers where the blood vessels start and end imaging in the DSA image according to the method of step S1 above.

[0148] The vascular imaging time point acquisition module 13 calculates the time points when the blood vessels start and end imaging according to the method of step S2 above.

[0149] The time feature acquisition module 14 calculates time features according to the method of step S3 above.

[0150] The minimum intensity map acquisition module 15 synthesizes a time-enhanced minimum intensity map according to the method of step S4 above.

[0151] The radiomics feature acquisition module 16 obtains time-enhanced radiomics features according to the method of step S5 above.

[0152] The feature screening and classification discrimination module 17 performs feature screening and classification discrimination according to the method of step S6 above, so as to obtain the arteriovenous malformation grading prediction result.

[0153] The control module 18 controls the work of the above-mentioned modules.

[0154] In this embodiment, the parts not described in detail are well-known technologies in the art.

[0155] Functions and effects of the embodiment

[0156] According to the arteriovenous malformation grading method and device based on time-enhanced radiomics provided in this embodiment, first use a neural network to detect blood vessels in the patient's DSA images, and record the frame numbers where the blood vessels start and end to appear; divide the frame numbers where the blood vessels start and end to appear by the DSA sampling frame rate to obtain the time points when the detected blood vessels start and end to appear, and calculate time features; since multiple frames of DSA images can synthesize a minimum intensity map by taking the minimum pixel value of each pixel point on the time axis, synthesize the corresponding minimum intensity map for the DSA images corresponding to the time features, and use the time features for weighting to obtain a time-enhanced minimum intensity map; calculate time-enhanced radiomics features from the time-enhanced minimum intensity map, and construct an arteriovenous malformation grading prediction model by machine learning methods. This method and device can accurately and simply evaluate the high and low grades of arteriovenous malformations.

[0157] The above embodiments are only used to illustrate the specific implementation manners of the present invention, and the present invention is not limited to the description scope of the above embodiments.

Claims

1. A method for grading arteriovenous malformations based on time-enhanced radiomics, characterized in that It includes the following steps: Step S1: Use a convolutional neural network to detect blood vessels in the patient's DSA images, and record the frames at which the blood vessels start to appear and end to appear; Step S2: Based on the frames at which the blood vessels start to appear and end to appear and the DSA sampling frame rate, calculate the time points at which the blood vessels start to appear and end to appear; Step S3: Calculate the time characteristics of blood vessel visualization based on the time points at which the blood vessels start to appear and end to appear; Step S4: Synthesize multiple frames of the DSA images corresponding to the time characteristics into a minimum intensity map, and use the time characteristics to weight the minimum intensity map to obtain a time-enhanced minimum intensity map; Step S5: Extract radiomics features from the time-enhanced minimum intensity map; Step S6: Based on the radiomics features and the time characteristics, construct a arteriovenous malformation grading prediction model through a machine learning method for high and low level evaluation of the arteriovenous malformation grade.

2. The arteriovenous malformation grading method based on time-enhanced radiomics according to claim 1, wherein: Among them, Step S1 includes the following sub-steps: Step S1-1: Annotate the blood vessel structure information of the DSA images of multiple arteriovenous malformation cases; Step S1-2: Use the DSA images with annotations to train the YOLOv5 convolutional neural network to obtain an object detection network for detecting blood vessel structures; Step S1-3: Use the object detection network to detect blood vessels in the DSA images, detect the blood vessels, and record the frames at which the blood vessels start to appear and the frames at which the blood vessels end to appear.

3. The arteriovenous malformation grading method based on time-enhanced radiomics according to claim 2, wherein: Among them, In step S1-1, each case uses the DSA images of the anteroposterior view of the internal carotid artery and the lateral view of the internal carotid artery. The DSA images of each section are respectively annotated with four blood vessel structures: the internal carotid artery, artery, vein, and venous sinus, In step S1-2, the YOLOv5 convolutional neural network is YOLOv5s. During the training process, a pre-trained model is used to fine-tune the blood vessel detection task to obtain a convolutional neural network for detecting blood vessel structures, that is, the object detection network.

4. The arteriovenous malformation grading method based on time-enhanced radiomics according to claim 1, wherein: Among them, In step S1, mark the frames at which the blood vessels start to appear and end to appear as vectors: α = [ICA sf , ICA ef , A sf , A ef , V sf , V ef , VS sf , VS ef ​ In step S2, calculate the time points at which the blood vessels start to appear and end to appear as: β = [ICA s , ICA e , A s , A e , V s , V e , VS s , VS e ​ Wherein, ICA is the internal carotid artery, A is the artery, V is the vein, VS is the venous sinus, and the subscript sf or ef indicates the frame number at which the vessel starts or ends to be visualized. is the DSA sampling frame rate, and the subscript s or e indicates the time at which the vessel starts or ends to be visualized.

5. The arteriovenous malformation grading method based on time-enhanced radiomics according to claim 1, wherein: Among them, In step S3, calibrate the time feature as t = [t0, t1, L, t 21 , and calculate the time feature as: t0 = ICA s t1 = ICA e - t0 t2 = A s -t0 t3 = A e -t0 t4 = V s - t0 t5 = V e - t0 t6 = VS s -t0 t7 = VS e -t0 t 12 = t4 - t2 t 13 = t6 - t2 t 14 = L(t4 - t3) t 15 = L(t3 - t4) t 16 = L(t6 - t3) t 17 = L(t3 - t6) t 18 = L(t6 - t4) t 19 = L(t4 - t6) t 20 = L(t7 - t5) t 21 = L(t5 - t7) where \(L(x) = x\cdot I(x)\), and \(I(\cdot)\) is the indicator function; 6. The arteriovenous malformation grading method based on time-enhanced radiomics according to claim 5, wherein: Among them, In step S4, for the time feature τ i , multiple frames of the DSA images related to it are synthesized into a minimum intensity map For each frame of the DSA image with a size of M×N, the minimum intensity map is expressed as: where m ∈ (1, M), n ∈ (1, N), p(m, n, f) is the pixel value of the f-th frame image at the pixel point (m, n), F si and F ei are the start and end frame numbers of the images related to the time feature τ i respectively, τ = [t8, t9, t 14 , t 15 , t 16 , t 17 , t 18 , t 19 , Then, perform weighted calculation on the time feature τ i and the corresponding minimum intensity map to obtain the time-enhanced minimum intensity map: where τ i ∈ τ, 0 ≤ i < 7.

7. The arteriovenous malformation grading method based on time-enhanced radiomics according to claim 1, wherein: Among them, In step S5, for each time-enhanced minimum intensity map, extract high-throughput features in three aspects: first-order histogram, texture, and wavelet.

8. The arteriovenous malformation grading method based on time-enhanced radiomics according to claim 1, wherein: wherein, Step S6 includes the following sub-steps: Step S6-1: Concatenate the extracted radiomics features with the corresponding time features to obtain a plurality of concatenated features; Step S6-2: Screen the concatenated features through a sparse representation method to obtain different classification categories; Step S6-3: Perform classification discrimination based on a support vector machine so that the classification categories correspond to the arteriovenous malformation levels, and learn the classification method through machine learning to construct the arteriovenous malformation grading prediction model.

9. The arteriovenous malformation grading method based on time-enhanced radiomics according to claim 8, wherein: Among them, In step S6-2, the following expression is used to screen the concatenated features: where y is the classification category, D = [D1, D2,..., D I is the set of all dictionaries, a is the sparse coefficient, μ is a regularization parameter greater than 0, ||·|| p denotes l p regularization, denotes the evaluation value of the sparse coefficient a, In step S6-3, the calculation model of the support vector machine is: y i (w T x i +b)≥1 - ξ i , i = 1, ..., n ξ i ≥0, i = 1, ..., n where (x i , y i ) is a sample in the given training set R = [(x1, y1), (x2, y2),..., (x n , y n ), y i ∈ (-1, +1), w = (w1; w2;...; w d ) is the normal vector that determines the direction of the hyperplane, b is the displacement term that determines the distance between the hyperplane and the origin, ξ i is the slack variable, and C is the penalty factor. Use the radial basis function as the kernel function of the support vector machine, and the radial basis function is expressed as: where ||x i - y i || represents the squared Euclidean distance between two feature vectors, and σ is a free parameter.

10. An arteriovenous malformation grading device based on time-enhanced radiomics, characterized in that, Including: DSA image acquisition module, used to acquire the patient's DSA image; Vascular imaging frame number acquisition module, which uses a convolutional neural network to detect blood vessels in the DSA image and records the frames where blood vessels start and end imaging; Vascular imaging time point acquisition module, based on the frames where blood vessels start and end imaging and the DSA sampling frame rate, calculates the time points when blood vessels start and end imaging; Time feature acquisition module, calculates the time features of blood vessel imaging based on the time points when blood vessels start and end imaging; Minimum intensity map acquisition module, synthesizes multiple frames of the DSA image corresponding to the time features into a minimum intensity map, and weights the minimum intensity map with the time features to obtain a time-enhanced minimum intensity map; Radiomics feature acquisition module, calculates time-enhanced radiomics features based on the time-enhanced minimum intensity map; and Feature screening and classification discrimination module, constructs an arteriovenous malformation grading prediction model based on the time-enhanced radiomics features and the time features through a machine learning method, and is used to evaluate the high and low levels of arteriovenous malformation grades.

Images