Carotid artery ultrasound automatic Doppler method, ultrasound device and storage medium

Abstract

The present invention provides a carotid ultrasound automatic Doppler method, the method comprising: acquiring a carotid ultrasound image; randomly extracting a preset number of image blocks on the carotid ultrasound image; inputting the preset number of image blocks into a convolutional neural network classification model to obtain the blood vessel angle in the carotid ultrasound image; inputting the preset number of image blocks into a convolutional neural network regression model to obtain the blood vessel position and blood vessel width in the carotid ultrasound image; based on the obtained blood vessel angle, blood vessel position and blood vessel width, automatically generating a sampling gate on the carotid ultrasound image. The present invention can simplify the doctor's operation during carotid ultrasound Doppler examination and improve the doctor's work efficiency.

Description

Technical Field

[0001] The invention belongs to the technical field of ultrasonic equipment, in particular to an automatic carotid artery ultrasonic Doppler method. Background Art

[0002] The incidence of cerebrovascular disease is increasing year by year. The use of non-invasive examination methods to detect the presence of extracranial arterial vascular lesions early and treat them in time is an effective means to prevent and reduce the incidence of cerebrovascular disease, which has important clinical significance. The above extracranial arteries mainly refer to the carotid arteries. The carotid arteries detected by carotid ultrasound include the common carotid artery, internal carotid artery, and external carotid artery. It is one of the effective means to diagnose and evaluate carotid artery lesions. Carotid ultrasound examination has the characteristics of simple operation, strong repeatability, economy and practicality, and easy acceptance by patients; therefore, it is often used for screening of normal people and examination of the condition of people at high risk of cerebrovascular disease, including preoperative, intraoperative, and postoperative evaluation and follow-up.

[0003] Carotid artery ultrasound examination includes comprehensive analysis such as two-dimensional structure and Doppler spectrum. Two-dimensional structural analysis can observe the inner diameter of the blood vessel, which is used to determine whether the lumen is dilated, narrowed, twisted, whether the course is normal, the thickness of the intima, and whether there are plaques. Doppler spectrum analysis obtains the blood flow spectrum to observe the blood flow velocity and determine whether the blood flow velocity increases or decreases. The blood flow velocity often corresponds to the two-dimensional structure. For example, an increase in blood flow velocity is often due to the narrowing of the carotid artery lumen. Doppler spectrum analysis is also used to obtain the peak systolic blood flow velocity, the end-diastolic blood flow velocity, and the ratio of the flow velocity between the internal carotid artery and the common carotid artery.

[0004] At present, when doctors use ultrasound equipment to perform carotid Doppler examination, the first step is to adjust the probe position to make the carotid ultrasound long-axis image clearest; then adjust the Doppler angle (that is, the angle of the sampling gate), which is generally between plus and minus 30°; then adjust the width of the parallelogram sampling gate, which is generally 1 / 3 of the blood vessel width; finally, record a frame of carotid image, a frame of representative Doppler spectrum, and measure the systolic peak velocity at the same time. Doctors must perform the above steps to ensure accurate measurement of the systolic peak velocity. If doctors do not perform the above steps, blood flow sampling will be incorrect, which will artificially increase or decrease the peak velocity, and errors will occur when calculating the systolic peak velocity ratio to evaluate lumen stenosis. At this time, if there is clinically significant carotid stenosis, it may lead to misjudgment.

[0005] However, doctors often have a large number of patients to be examined and a heavy workload. If the process of carotid ultrasound examination can be simplified using automated algorithms, it can directly improve the doctors' work efficiency. Summary of the invention

[0006] The purpose of this application is to overcome the deficiencies in the prior art and provide a method for automatic carotid artery ultrasound Doppler examination to simplify the doctor's operation and improve the doctor's work efficiency during carotid artery ultrasound Doppler examination; this application also provides an ultrasound device and storage medium corresponding to the above method. To achieve the above technical objectives, the technical solution adopted in the embodiment of the present invention is:

[0007] In a first aspect, an embodiment of the present invention provides a method for automatic carotid artery ultrasound Doppler, the method comprising:

[0008] Obtain carotid artery ultrasound images;

[0009] Randomly extracting a preset number of image blocks on the carotid artery ultrasound image, so that at least one image block includes a section of the carotid artery;

[0010] Inputting the preset number of image blocks into a convolutional neural network classification model to obtain the blood vessel angle in the carotid artery ultrasound image;

[0011] Inputting the preset number of image blocks into a convolutional neural network regression model to obtain the blood vessel position and blood vessel width in the carotid artery ultrasound image;

[0012] Based on the obtained blood vessel angle, blood vessel position and blood vessel width, a sampling gate is automatically generated on the carotid artery ultrasound image.

[0013] Furthermore, randomly extracting a preset number of image blocks from the carotid artery ultrasound image includes:

[0014] A preset number of image blocks are extracted from the carotid artery ultrasound image according to a preset extraction standard, and each of the image blocks records position information on the carotid artery ultrasound image; the preset extraction standard includes: the aspect ratio of the image block is one or more preset aspect ratios, and / or the extraction length is a preset length and the width is a preset width.

[0015] Specifically, inputting the preset number of image blocks into a convolutional neural network classification model to obtain the blood vessel angle in the carotid artery ultrasound image specifically includes:

[0016] Outputting a first prediction output vector including a plurality of blood vessel angle prediction probabilities for each of the image blocks through the convolutional neural network classification model;

[0017] Determine the image block with the highest blood vessel angle prediction probability among all first prediction output vectors;

[0018] The blood vessel angle corresponding to the category with the highest blood vessel angle prediction probability in the first prediction output vector of the image block with the highest blood vessel angle prediction probability is used as the blood vessel angle in the carotid artery ultrasound image.

[0019] Specifically, inputting the preset number of image blocks into a convolutional neural network regression model to obtain the blood vessel position and blood vessel width in the carotid artery ultrasound image includes:

[0020] Outputting a second prediction output vector including a predicted blood vessel position and a blood vessel width for each of the image blocks through the convolutional neural network regression model;

[0021] The image block with the highest prediction probability of the blood vessel angle is restored to the carotid artery ultrasound image, and the blood vessel position and blood vessel width predicted in the image block are combined to obtain the blood vessel position and blood vessel width in the carotid artery ultrasound image.

[0022] Specifically, the automatically generating a sampling gate on the carotid artery ultrasound image based on the obtained blood vessel angle, blood vessel position and blood vessel width includes:

[0023] The center of the sampling gate is set at the center point of the blood vessel position in the carotid artery ultrasound image; the angle of the sampling gate is the obtained blood vessel angle; and the width of the sampling gate is less than or equal to the obtained blood vessel width.

[0024] Specifically, the convolutional neural network classification model and the convolutional neural network regression model are trained in the following manner:

[0025] Acquire a sample carotid artery ultrasound image, wherein the sample carotid artery ultrasound image is marked with blood vessel angle information, blood vessel position mark, and blood vessel width information;

[0026] randomly extracting a preset number of sample image blocks on the sample carotid artery ultrasound image, and recording position information of each sample image block on the carotid artery ultrasound image;

[0027] The sample image block is divided into a positive sample image block and a negative sample image block according to a predetermined rule; the blood vessel angle of the positive sample image block is the blood vessel angle marked by the sample carotid artery ultrasound image corresponding to the positive sample image block;

[0028] Training the convolutional neural network classification model according to the positive sample image blocks and the negative sample image blocks;

[0029] The convolutional neural network regression model is trained according to the positive sample image blocks and the negative sample image blocks.

[0030] Furthermore, dividing the sample image blocks into positive sample image blocks and negative sample image blocks according to a predetermined rule comprises:

[0031] Calculating the intersection-and-union ratio of each sample image block and the blood vessel position marker in the corresponding sample carotid artery ultrasound image, where the intersection-and-union ratio is the ratio of the intersection to the union;

[0032] Take the sample image patches with the intersection over union greater than the set threshold as positive sample image patches, and otherwise as negative sample image patches.

[0033] Specifically, training the convolutional neural network classification model based on the positive sample image patches and negative sample image patches includes:

[0034] Process the positive sample image patches and negative sample image patches through the first layer group composed of several convolutional layers, batch normalization layers and activation function layers, then connect the global pooling layer, and finally connect a fully connected layer. The layers in the convolutional neural network classification model are connected by weight parameters to output the first predicted output vector. The value of each dimension in the first predicted output vector represents the prediction probability of the input sample image patch corresponding to that dimension. The prediction probability includes the predicted probability of the blood vessel angle of the sample image patch.

[0035] Calculate the first loss function in the convolutional neural network classification model;

[0036] Calculate the partial derivative values of the first loss function with respect to each weight parameter, and backpropagate the partial derivative values back to the convolutional neural network classification model for update iteration to obtain the trained convolutional neural network classification model.

[0037] Preferably, the formula of the first loss function in the convolutional neural network classification model is:

[0038] Loss=α(1-y’) r (-logy’) (1)

[0039] Where α and r are hyperparameters, 0 < α < 0.8, 1 < r < 3; y’ represents the prediction probability with the largest value in the output first predicted output vector; α(1-y’) r is the proportion of the entire first loss function, and (-logy’) is the error loss generated by a sample image patch.

[0040] Specifically, training the convolutional neural network regression model based on the positive sample image patches and negative sample image patches includes:

[0041] Process the positive sample image patches and negative sample image patches through the second layer group composed of several convolutional layers, batch normalization layers and activation function layers, then connect the global pooling layer, and finally connect a fully connected layer to output the second predicted output vector. The second predicted output vector contains the predicted blood vessel position information and blood vessel width information of the sample image patch;

[0042] Calculate the second loss function in the convolutional neural network regression model;

[0043] The function value of the second loss function is back-propagated to the convolutional neural network regression model for updating and iteration to obtain a trained convolutional neural network regression model.

[0044] Preferably, the formula of the second loss function in the convolutional neural network regression model is:

[0045]

[0046] Among them, out represents the second predicted output vector of the convolutional neural network regression model, and truth represents the actual blood vessel position and blood vessel width.

[0047] In a second aspect, an embodiment of the present invention provides an ultrasonic device, comprising:

[0048] a memory storing a computer program;

[0049] The processor is used to run the computer program, and when the computer program is run, the steps of the automatic carotid ultrasound Doppler method as described above are performed.

[0050] In a third aspect, an embodiment of the present invention provides a storage medium, wherein a computer program is stored in the storage medium, and the computer program is configured to execute the steps of the automatic carotid ultrasound Doppler method as described above when running.

[0051] The technical solution provided by the embodiment of the present invention has the following beneficial effects:

[0052] 1) Automated processing can simplify doctors' operations during carotid artery Doppler ultrasound examination and significantly improve doctors' work efficiency.

[0053] 2) Through artificial intelligence methods, doctors can greatly save time in performing carotid artery ultrasound Doppler examinations. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 Schematic diagram of a carotid artery ultrasound long-axis image in an embodiment of the present invention.

[0055] Figure 2 Schematic diagram of extracted image blocks in an embodiment of the present invention.

[0056] Figure 3 Schematic diagram of a sampling gate generated in an embodiment of the present invention.

[0057] Figure 4 4 is a flow chart of a method in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0059] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, it can also be the internal connection of two components, it can be a wireless connection, or it can be a wired connection. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0060] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0061] In a first aspect, an embodiment of the present invention provides a method for automatic carotid artery ultrasound Doppler, the method comprising:

[0062] Acquire a carotid artery ultrasound image; here, the carotid artery ultrasound image is preferably a carotid artery ultrasound long-axis image, in which the carotid artery blood vessels are clearer;

[0063] Randomly extracting a preset number of image blocks on the carotid artery ultrasound image; so that at least one image block contains a section of the carotid artery; the preset number is usually tens to thousands;

[0064] Inputting the preset number of image blocks into a convolutional neural network classification model to obtain the blood vessel angle in the carotid artery ultrasound image;

[0065] Inputting the preset number of image blocks into a convolutional neural network regression model to obtain the blood vessel position and blood vessel width in the carotid artery ultrasound image;

[0066] Based on the obtained blood vessel angle, blood vessel position and blood vessel width, a sampling gate is automatically generated on the carotid artery ultrasound image; the sampling gate position is set within the obtained blood vessel position;

[0067] The embodiment of the present invention obtains the blood vessel angle in the carotid artery ultrasound image through a convolutional neural network classification model, and obtains the blood vessel position and blood vessel width in the carotid artery ultrasound image through a convolutional neural network regression model. Separate processing can improve the accuracy of prediction.

[0068] The embodiment of the present invention can simplify the doctor's operation during the carotid artery ultrasound Doppler examination through the above automated process of automatically acquiring vascular information and automatically generating a sampling gate, thereby significantly improving the doctor's work efficiency; and greatly save the doctor's time for the carotid artery ultrasound Doppler examination through the artificial intelligence method;

[0069] Preferably, the randomly extracting a preset number of image blocks on the carotid artery ultrasound image comprises: extracting a preset number of image blocks on the carotid artery ultrasound image according to a preset extraction standard, each of the image blocks recording position information on the carotid artery ultrasound image; the preset extraction standard comprises: the aspect ratio of the image block is one or more preset aspect ratios, and / or the extraction length is a preset length and the width is a preset width; extracting the image blocks according to the preset standard can obtain more accurate results when obtaining the blood vessel angle, blood vessel position and blood vessel width in the carotid artery ultrasound image based on each image block;

[0070] Specifically, the inputting the preset number of image blocks into the convolutional neural network classification model to obtain the vascular angle in the carotid artery ultrasound image specifically includes: outputting a first prediction output vector containing multiple vascular angle prediction probabilities for each of the image blocks through the convolutional neural network classification model; determining the image block with the highest vascular angle prediction probability among all the first prediction output vectors; and taking the vascular angle corresponding to the category with the highest vascular angle prediction probability in the first prediction output vector of the image block with the highest vascular angle prediction probability as the vascular angle in the carotid artery ultrasound image;

[0071] Specifically, the inputting the preset number of image blocks into the convolutional neural network regression model to obtain the blood vessel position and blood vessel width in the carotid artery ultrasound image includes: outputting a second prediction output vector containing the predicted blood vessel position and blood vessel width for each of the image blocks through the convolutional neural network regression model; restoring the image block with the highest probability of blood vessel angle prediction to the carotid artery ultrasound image, and combining the predicted blood vessel position and blood vessel width in the image block to obtain the blood vessel position and blood vessel width in the carotid artery ultrasound image;

[0072] The processing processes of the above convolutional neural network classification model and convolutional neural network regression model have both independent processing processes and mutually coordinated processing processes, which can obtain more accurate processing results;

[0073] Specifically, the blood vessel position in the obtained carotid artery ultrasound image can be represented by a blood vessel prediction rectangular frame;

[0074] Specifically, the automatically generating a sampling gate on the carotid artery ultrasound image based on the obtained blood vessel angle, blood vessel position and blood vessel width includes: setting the center of the sampling gate at the center point of the blood vessel position in the carotid artery ultrasound image, so that the sampling gate is located in the middle area of ​​the blood vessel prediction rectangular frame; the angle of the sampling gate is the obtained blood vessel angle; the width of the sampling gate is less than or equal to the obtained blood vessel width, for example, the width of the sampling gate is 1 / 3 to 1 / 2 of the blood vessel width in the obtained carotid artery ultrasound image, so that blood flow sampling can be in an optimal area;

[0075] The following describes the automation process of the carotid ultrasound automatic Doppler method with examples;

[0076] First, obtain an ultrasound image of the carotid artery. Figure 1 An acquired carotid ultrasound long-axis image is displayed, in which a dark black carotid blood vessel (roughly in a transverse direction) can be seen in the middle and upper region of the image, and then a preset number of image blocks are randomly extracted from the carotid ultrasound image according to a preset extraction standard and the position information of each image block on the carotid ultrasound image is recorded; the preset extraction standard includes: the aspect ratio of the image block is one or more preset aspect ratios, such as 2:1, 3:1, 4:1, 5:1, etc., and / or the extraction length is a preset length and the width is a preset width; Figure 2Three extracted image blocks are shown exemplarily; since they are randomly extracted and the number is large, it is guaranteed that at least one image block can contain a complete section of the carotid artery; then a preset number of image blocks are input into the convolutional neural network classification model and the convolutional neural network regression model respectively; the convolutional neural network classification model outputs a first prediction output vector for each image block, which includes a negative sample image block prediction probability and a plurality of vascular angle prediction probabilities of the image block, for example, [0.02, 0.8, 0.1, ..., 0, 0.01], and the first prediction The prediction output vector includes 14 categories, the negative sample image block is taken as the 0th category, and the 13 vascular angles of -30°, -25°, -20°, -15°, -10°, -5°, 0°, 5°, 10°, 15°, 20°, 25°, and 30° are taken as the remaining 13 categories; in the first prediction output vector, the prediction probability of the negative sample image block is 0.02, the prediction probability of the -30° vascular angle is 0.8, the prediction probability of the -25° vascular angle is 0.1... The prediction probability of the 25° vascular angle is 0, and the prediction probability of the 30° vascular angle is 0. The probability is 0.01, and the sum of all prediction probabilities is 1; assuming that the maximum prediction probabilities in some other first prediction output vectors are all less than 0.8, then the image block with the first prediction output vector of [0.02, 0.8, 0.1, ..., 0, 0.01] can be determined; then the -30° vascular angle corresponding to the category (the second category) with the vascular angle prediction probability of 0.8 in the first prediction output vector of the image block is used as the vascular angle in the obtained carotid ultrasound image; the convolutional neural network regression model is used to output a second prediction output vector containing the predicted vascular position and vascular width for each of the image blocks; after the above image blocks are determined, the image block with the highest vascular angle prediction probability is restored to the carotid ultrasound image, and the vascular position in the carotid ultrasound image is obtained in combination with the predicted vascular position in the image block, and the predicted vascular width in the image block is used as the vascular width in the obtained carotid ultrasound image; in this embodiment, the predicted vascular position in the image block is represented by a vascular prediction rectangular box, so the vascular position in the obtained carotid ultrasound image is also represented by a vascular prediction rectangular box; Figure 3Two rectangular boxes are shown in the figure. The predicted blood vessel position in the image block has been transformed into the carotid artery ultrasound image, where the larger rectangular box represents the blood vessel prediction rectangular box and the smaller rectangular box represents the sampling gate. The second prediction output vector output by the convolutional neural network regression model uses a 5-dimensional vector to represent the predicted blood vessel position and blood vessel width in the image block. The first 4 dimensions in the second prediction output vector represent the position information of the blood vessel prediction rectangular box, and the 5th dimension represents the predicted blood vessel width information. There are two ways to represent the second prediction output vector. For example, a second prediction output vector [100,10 For example, in the second prediction output vector [0,200,200,200,20], the first and second values ​​represent the coordinates of the upper left corner of the blood vessel prediction rectangular box, the third and fourth values ​​represent the coordinates of the lower right corner of the blood vessel prediction rectangular box, and the fifth value represents the predicted blood vessel width is 20 pixels; for another example, in the second prediction output vector [150,150,50,50,20], the first and second values ​​represent the coordinates of the center of the blood vessel prediction rectangular box, the third and fourth values ​​represent the length and width of the blood vessel prediction rectangular box, and the fifth value represents the predicted blood vessel width is 20 pixels;

[0077] Finally, based on the obtained blood vessel angle, blood vessel position and blood vessel width, a sampling gate is automatically generated on the carotid artery ultrasound image; the sampling gate position is set within the obtained blood vessel position (blood vessel prediction rectangular frame), such as Figure 3 The smaller rectangular frame (sampling gate) is located in the larger rectangular frame (blood vessel prediction rectangular frame) as shown; of course, the blood vessel prediction rectangular frame does not necessarily need to be displayed, and can also be hidden, and only four position values ​​are needed to represent it; the angle of the sampling gate is the obtained blood vessel angle, and the width of the sampling gate is less than or equal to the predicted blood vessel width; preferably, the center point of the sampling gate is set to the center point of the blood vessel position in the obtained carotid artery ultrasound image, so that the sampling gate is located in the middle area of ​​the blood vessel prediction rectangular frame, thereby making blood flow sampling more accurate; the width of the sampling gate is preferably 1 / 3 of the blood vessel width in the obtained carotid artery ultrasound long axis image, so that blood flow sampling can be in the best area;

[0078] The following will introduce how to train a convolutional neural network classification model and a convolutional neural network regression model; the convolutional neural network classification model and the convolutional neural network regression model are trained in the following way:

[0079] To obtain a sample carotid ultrasound image, as comprehensive a sample carotid ultrasound image as possible should be collected, such as sample carotid ultrasound images of different image depths and different collection objects, etc.; the sample carotid ultrasound image is marked with vascular angle information, vascular position mark, and vascular width information; specifically, the collected sample carotid ultrasound images can be marked by image annotation personnel, firstly, the sample carotid ultrasound image is marked with vascular angle information, for example, a total of 13 categories, including -30°, -25°, -20°, -15°, -10°, -5°, 0°, 5°, 10°, 15°, 20°, 25°, and 30°; then the sample carotid ultrasound image is marked with a vascular position mark, that is, a complete section of the carotid blood vessel in the sample carotid ultrasound image is marked with a vascular marking rectangular frame; then the distance between two corresponding points on the two long sides of the vascular marking rectangular frame is calculated and marked as the vascular width;

[0080] Then, a preset number of sample image blocks are randomly extracted from the sample carotid artery ultrasound image, and the position information of each sample image block on the carotid artery ultrasound image is recorded to obtain a sample image block corresponding to each sample carotid artery ultrasound image; preferably, a preset number of image blocks can be extracted from the sample carotid artery ultrasound image according to a preset extraction standard, and the preset extraction standard includes: the aspect ratio of the sample image block is one or more preset aspect ratios, such as 2:1, 3:1, 4:1, 5:1, etc., and / or the extraction length is a preset length and the width is a preset width; the length and width of the extracted sample image block need to be respectively smaller than the length and width of the original sample carotid artery ultrasound image;

[0081] The sample image blocks are divided into positive sample image blocks and negative sample image blocks according to a predetermined rule; the vascular angle of the positive sample image block is the vascular angle of the sample carotid artery ultrasound image corresponding to the positive sample image block; the sample image blocks are divided into positive sample image blocks and negative sample image blocks according to the predetermined rule, specifically including: calculating the vascular intersection and union ratio of each sample image block and the vascular position mark in the corresponding sample carotid artery ultrasound image, and taking the sample image blocks with the intersection and union ratio greater than the set threshold value of 0.8 as positive sample image blocks, otherwise as negative sample image blocks; the intersection and union ratio is the ratio of the intersection and union of blood vessels; the above operation will cause the number of positive sample image blocks to be much less than the number of negative sample image blocks, which is not conducive to the training of the convolutional neural network classification model and / or the convolutional neural network regression model. Therefore, before training the convolutional neural network classification model and / or the convolutional neural network regression model, a set number of negative sample image blocks are randomly retained, and the rest are discarded; for example, assuming that the number of positive sample image blocks is N1, 2N1 to 3N1 negative sample image blocks are randomly retained;

[0082] Then, the convolutional neural network classification model is trained according to the positive sample image blocks and the negative sample image blocks; the convolutional neural network regression model is trained according to the positive sample image blocks and the negative sample image blocks;

[0083] (1) The specific process of training the convolutional neural network classification model according to the positive sample image blocks and the negative sample image blocks is as follows:

[0084] The positive sample image block and the negative sample image block are input into the convolutional neural network classification model; in the convolutional neural network classification model, a corresponding sample category vector is set for each sample image block, and the sample category vector includes a negative sample category and a plurality of vascular angle categories of the predefined sample image block; for example, the 0th category in the sample category vector is a negative sample category, and the remaining 13 categories are -30°, -25°, -20°, -15°, -10°, -5°, 0°, 5°, 10°, 15°, 20°, 25°, and 30 vascular angle categories; the sample category vector [0,0,0,0,0,0,1,0,0,0,0,0,0,0,0] indicates that this is a positive sample image block with a vascular angle of -5°, and the sample category vector [1,0,0,0,0,0,0,0,0,0,0,0,0,0,0] indicates that this is a negative sample image block;

[0085] The positive sample image block and the negative sample image block are fixed to a uniform pixel size, processed by a first layer group consisting of several convolutional layers, batch normalization layers and activation function layers, then connected to a global pooling layer, and finally connected to a fully connected layer. The layers in the convolutional neural network classification model are connected by weight parameters, and a first prediction output vector is output. The value of each dimension in the first prediction output vector represents the prediction probability of the input sample image block corresponding to the dimension, and the prediction probability includes the prediction probability of the blood vessel angle of the sample image block; for example, the value of the first dimension in the first prediction output vector represents the prediction probability of the sample image block corresponding to the negative sample category, and the value of the second dimension in the first prediction output vector represents the prediction probability of the sample image block corresponding to the -30° blood vessel angle; when the step size of the convolution layer in the first layer group is selected as 1, it needs to be used together with the maximum pooling layer, that is, the structure of the first layer group is a convolutional layer, a maximum pooling layer, a batch normalization layer and an activation function layer; during the training process, a learning rate of 0.001 is used, and the saturation, hue and brightness of the input image block samples are randomly changed;

[0086] After obtaining the predicted probability, the first loss function in the convolutional neural network classification model is calculated, and the formula is:

[0087] Loss = α(1-y') r (-logy') (1)

[0088] Among them, α and r are hyperparameters with values ​​of 0.25 and 2; y' represents the predicted probability with the largest value in the first predicted output vector; α(1-y') r is the weight of the entire first loss function. If y' is larger, its weight in the first loss function is smaller. On the contrary, if the predicted probability is smaller, the weight of the sample image block in the first loss function is larger. This can effectively suppress the influence of a large number of simple samples on the neural network, and make the convolutional neural network classification model focus on samples with greater difficulty in distinguishing. (-logy') is the error loss generated by a sample image block. The value range of y' is 0 to 1. When the predicted probability is smaller, the error loss is larger, that is, the convolutional neural network classification model can finally judge the blood vessel angle of a sample image block with a higher probability.

[0089] Finally, the partial derivative value of the first loss function with respect to each weight parameter is calculated, and the partial derivative value is back-propagated back to the convolutional neural network classification model for updating and iteration to obtain a trained convolutional neural network classification model;

[0090] (2) The specific process of training the convolutional neural network regression model according to the positive sample image blocks and the negative sample image blocks is as follows:

[0091] Input the positive sample image patches and the negative sample image patches into the convolutional neural network regression model;

[0092] The positive sample image block and the negative sample image block are fixed to a uniform pixel size, processed by a second layer group consisting of a plurality of convolutional layers, a batch normalization layer, and an activation function layer, and then connected to a global pooling layer, and finally connected to a fully connected layer, and a second prediction output vector is output, wherein the second prediction output vector contains the blood vessel position information and blood vessel width information predicted for the sample image block, for example, a 5-dimensional second prediction output vector, wherein the first 4 dimensions represent the position information of the blood vessel prediction rectangular box, and the 5th dimension represents the blood vessel width information; wherein the bottom convolutional layer can extract the edge features of the blood vessel, the middle convolutional layer can extract the texture features of the blood vessel, and the high-level convolutional layer can extract the global shape features of the blood vessel;

[0093] Calculate the second loss function in the convolutional neural network regression model, the formula is:

[0094]

[0095] Wherein, out represents the second predicted output vector of the convolutional neural network regression model, truth represents the true blood vessel position and blood vessel width (that is, the position information and blood vessel width of the blood vessel marking rectangular box marked in the carotid artery ultrasound long axis image sample), and the second loss function calculates the square error of each dimension in the second predicted output vector to guide the neural network training;

[0096] The function value of the second loss function is back-propagated to the convolutional neural network regression model for updating and iteration to obtain a trained convolutional neural network regression model.

[0097] The reason why the embodiment of the present invention adopts the second neural network model, namely the convolutional neural network regression model, and adopts another loss function is that the numerical range of blood vessel position and width is much larger than the predicted probability values ​​output by the convolutional neural network classification model. If they are mixed into one neural network model, the loss function will be dominated by the errors in blood vessel position and width, making it difficult to accurately predict the blood vessel angle.

[0098] In a second aspect, an embodiment of the present invention further proposes an ultrasound device, comprising: a processor and a memory; the processor and the memory communicate with each other, and a computer program is stored in the memory; the processor is used to run the computer program, and the computer program executes the steps of the method described above when it runs; the processor can be a CPU, or other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), field programmable gate arrays (Field-Programmable Gate Array, FPGA) or other programmable logic devices, etc.; the memory can include volatile memory, such as random access memory; the memory can also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory can also include a combination of the above types of memory.

[0099] In a third aspect, an embodiment of the present invention further provides a storage medium, wherein a computer program is stored in the storage medium, and the computer program is configured to execute the steps of the method described above when running. The storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk drive (HDD) or a solid-state drive (SSD), etc.; the storage medium may also include a combination of the above-mentioned types of memory.

[0100] Finally, it should be noted that the above specific implementation methods are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. An automatic Doppler method for carotid ultrasound, It is characterized in that The method comprises: Obtain carotid artery ultrasound images; randomly extracting a preset number of image blocks from the carotid artery ultrasound image; Inputting the preset number of image blocks into a convolutional neural network classification model to obtain the blood vessel angle in the carotid artery ultrasound image; Inputting the preset number of image blocks into a convolutional neural network regression model to obtain the blood vessel position and blood vessel width in the carotid artery ultrasound image; Automatically generate a sampling gate on the carotid artery ultrasound image based on the obtained blood vessel angle, blood vessel position and blood vessel width; Inputting the preset number of image blocks into a convolutional neural network classification model to obtain the blood vessel angle in the carotid artery ultrasound image specifically includes: Outputting a first prediction output vector including a plurality of blood vessel angle prediction probabilities for each of the image blocks through the convolutional neural network classification model; Determine the image block with the highest blood vessel angle prediction probability among all first prediction output vectors; The vascular angle corresponding to the category with the highest vascular angle prediction probability in the first prediction output vector of the image block with the highest vascular angle prediction probability is used as the vascular angle in the carotid artery ultrasound image; The step of inputting the preset number of image blocks into a convolutional neural network regression model to obtain the blood vessel position and blood vessel width in the carotid artery ultrasound image comprises: Outputting a second prediction output vector including a predicted blood vessel position and a blood vessel width for each of the image blocks through the convolutional neural network regression model; The image block with the highest prediction probability of the blood vessel angle is restored to the carotid artery ultrasound image, and the blood vessel position and blood vessel width predicted in the image block are combined to obtain the blood vessel position and blood vessel width in the carotid artery ultrasound image.

2. The method for automatic carotid ultrasound Doppler according to claim 1, It is characterized in that The randomly extracting a preset number of image blocks from the carotid artery ultrasound image comprises: A preset number of image blocks are extracted from the carotid artery ultrasound image according to a preset extraction standard, and each of the image blocks records position information on the carotid artery ultrasound image; the preset extraction standard includes: the aspect ratio of the image block is one or more preset aspect ratios, and / or the extraction length is a preset length and the width is a preset width.

3. The automatic carotid artery ultrasound Doppler method as claimed in claim 1, It is characterized in that The method of automatically generating a sampling gate on the carotid artery ultrasound image based on the obtained blood vessel angle, blood vessel position and blood vessel width comprises: The center of the sampling gate is set at the center point of the blood vessel position in the carotid artery ultrasound image; the angle of the sampling gate is the obtained blood vessel angle; and the width of the sampling gate is less than or equal to the obtained blood vessel width.

4. The automatic carotid ultrasound Doppler method according to claim 1 or 2, It is characterized in that The convolutional neural network classification model and the convolutional neural network regression model are trained in the following manner: Acquire a sample carotid artery ultrasound image, wherein the sample carotid artery ultrasound image is marked with blood vessel angle information, blood vessel position mark, and blood vessel width information; randomly extracting a preset number of sample image blocks on the sample carotid artery ultrasound image, and recording position information of each sample image block on the carotid artery ultrasound image; The sample image block is divided into a positive sample image block and a negative sample image block according to a predetermined rule; the blood vessel angle of the positive sample image block is the blood vessel angle marked by the sample carotid artery ultrasound image corresponding to the positive sample image block; Training the convolutional neural network classification model according to the positive sample image blocks and the negative sample image blocks; The convolutional neural network regression model is trained according to the positive sample image blocks and the negative sample image blocks.

5. The method for automatic carotid ultrasound Doppler according to claim 4, It is characterized in that The step of dividing the sample image blocks into positive sample image blocks and negative sample image blocks according to a predetermined rule comprises: Calculate the intersection-and-union ratio of each sample image block and the blood vessel position marker in the corresponding sample carotid artery ultrasound image, where the intersection-and-union ratio is the ratio of the intersection to the union; The sample image blocks whose intersection-over-union ratio is greater than a set threshold are taken as positive sample image blocks, and otherwise are taken as negative sample image blocks.

6. The method for automatic carotid artery ultrasound Doppler according to claim 4, It is characterized in that The step of training the convolutional neural network classification model according to the positive sample image blocks and the negative sample image blocks includes: The positive sample image block and the negative sample image block are processed by a first layer group consisting of several convolution layers, batch normalization layers and activation function layers, and then connected to a global pooling layer, and finally connected to a fully connected layer. The layers in the convolutional neural network classification model are connected through weight parameters, and a first prediction output vector is output. The value of each dimension in the first prediction output vector represents the prediction probability of the input sample image block corresponding to the dimension, and the prediction probability includes the prediction probability of the blood vessel angle of the sample image block; Calculate the first loss function in the convolutional neural network classification model; Calculate the partial derivative value of the first loss function with respect to each weight parameter, and back-propagate the partial derivative value back to the convolutional neural network classification model for updating and iteration to obtain a trained convolutional neural network classification model.

7. The method for automatic carotid ultrasound Doppler according to claim 4, It is characterized in that The step of training the convolutional neural network regression model according to the positive sample image blocks and the negative sample image blocks includes: Processing the positive sample image block and the negative sample image block through a second layer group consisting of a plurality of convolutional layers, a batch normalization layer, and an activation function layer, then connecting the global pooling layer, and finally connecting the fully connected layer, outputting a second prediction output vector, wherein the second prediction output vector includes a blood vessel position and a blood vessel width predicted for the sample image block; Calculate the second loss function in the convolutional neural network regression model; The function value of the second loss function is back-propagated to the convolutional neural network regression model for updating and iteration to obtain a trained convolutional neural network regression model.

8. An ultrasonic device, It is characterized in that include: a memory storing a computer program; A processor for running the computer program, and when the computer program runs, it executes the steps of the carotid artery ultrasound automatic Doppler method according to any one of claims 1 to 7.

9. A storage medium, characterized in that, the storage medium stores a computer program, and the computer program is configured to execute the steps of the carotid artery ultrasound automatic Doppler method according to any one of claims 1 to 7 when running.

Images