A method and system for early warning of acute kidney injury

Abstract

The present invention provides an acute kidney injury early warning method and system, which divides a renal color ultrasound image into multiple image blocks and encodes the image blocks, and inputs the image block results, the element encoding results corresponding to the first set containing renal function index data and user data, and the element encoding results corresponding to the second set where PSV and EDV are located into three encoders of the same structure respectively, and adds three classification identification vectors to the output results of each encoder and inputs them into the encoder of the next layer respectively; the second and third classification identification vectors in the output of the encoder of the next layer are replaced with the first classification identification vectors in the output of the other two encoders of the next layer until the last layer of encoders; the acute kidney injury level is obtained according to the classification identification vectors output by the three encoders of the last layer, and early warning information is issued to the user. The present invention adopts a cross-modal approach to obtain information on acute kidney injury from multiple angles, which effectively improves the judgment of acute kidney injury.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to an acute kidney injury early warning method and system. Background Art

[0002] Acute Kidney Injury (AKI) refers to a disease in which kidney function declines sharply, leading to the retention of urea and other nitrogenous wastes, as well as the imbalance of extracellular fluid volume and electrolytes, resulting in the inability of the kidneys to normally eliminate waste from the body and maintain electrolyte balance. Early warning of acute kidney injury is of great significance, because early diagnosis and intervention can help reduce the risk of worsening of the patient's condition and complications. AKI may be associated with a variety of causes, including infection, drug poisoning, hypovolemia, surgical complications, etc., so early detection and treatment may help prevent severe damage to kidney function. For medical professionals and patients, it is very important to warn of the signs and symptoms of AKI so that necessary treatment measures can be taken early to reduce the mortality and complications of AKI patients. The judgment of acute kidney injury is usually based on some indicator data, such as urine volume, blood creatinine level, etc., but these indicators have a lag. If these indicators are found to be abnormal and then intervened, the effect is limited, and these indicators are too single. Bedside renal ultrasound indicators have high sensitivity and specificity for determining whether AKI occurs, and their changes are earlier than conventional renal function indicators. However, there is no method in the prior art that can effectively combine bedside renal ultrasound images and renal function indicators to provide early warning for acute kidney injury, and the effect of prediction using existing models is not good. Summary of the invention

[0003] In view of the fact that there is no method in the prior art that can combine renal ultrasound images, especially bedside renal ultrasound images, and commonly used renal injury indicators to provide early warning for acute renal injury, the present invention provides an acute renal injury early warning method, the method comprising the following steps:

[0004] Obtain renal function index data and user data and put them into the first set; obtain color Doppler ultrasound images of the left and right kidneys taken by a bedside instant ultrasound device, and put the peak systolic velocity PSV and end diastolic velocity EDV corresponding to the left and right kidneys into the second set;

[0005] The color ultrasound image is divided into a plurality of image blocks and the image blocks are encoded, and the image block encoding results, the element encoding results corresponding to the first set, and the element encoding results corresponding to the second set are respectively input into three encoders of the same structure, and the output results of each encoder are added with three classification identification vectors and then respectively input into the encoder of the next layer; the second and third classification identification vectors in the output of the encoder of the next layer are replaced with the first classification identification vectors in the outputs of the other two encoders of the next layer, until the last encoder;

[0006] The acute kidney injury level is obtained according to the classification identification vectors output by the three encoders of the last layer, and warning information is issued to the user according to the acute kidney injury level.

[0007] Preferably, the acute kidney injury grade is obtained according to the classification identification vector output by the three encoders of the last layer, specifically:

[0008] The three classification identification vectors output by the last layer encoder corresponding to the image are recorded as the first vector set, the three classification identification vectors output by the last layer encoder corresponding to the first set are recorded as the second vector set, and the three classification identification vectors output by the last layer encoder corresponding to the second set are recorded as the third vector set;

[0009] The first classification identification vector in the first vector set is concatenated with the second classification identification vector and the third classification identification vector respectively, and then input into the respective MLPs to obtain two outputs S1 and S2; the first classification identification vector in the second vector set is concatenated with the second classification identification vector and the third classification identification vector respectively, and then input into the respective MLPs to obtain two outputs S3 and S4; the first classification identification vector in the third vector set is concatenated with the second classification identification vector and the third classification identification vector respectively, and then input into the respective MLPs to obtain two outputs S5 and S6;

[0010] S1, S2, S3, S4, S5, and S6 were input into the MLP to obtain the grade of acute kidney injury.

[0011] Preferably, three classification identification vectors are added to the output result of each encoder, specifically:

[0012] Obtain a matrix corresponding to the output result of the encoder, wherein the matrix size is M×N; add three row vectors before the first row of the matrix so that the size of the matrix is ​​(M+3)×N, and the added three row vectors are divided into the first classification identification vector, the second classification identification vector, and the third classification identification vector; wherein M and N are positive integers.

[0013] Preferably, the step of dividing the color ultrasound image into a plurality of image blocks and encoding the image blocks is specifically as follows:

[0014] Convert the left kidney color ultrasound image and the right kidney color ultrasound image to the same size, and divide the left kidney color ultrasound image and the right kidney color ultrasound image into multiple image blocks of the same size to obtain an image block sequence H1 of the left kidney color ultrasound image and an image block sequence H2 of the right kidney color ultrasound image, and if the sequence numbers of the image blocks in H1 and the sequence numbers of the image blocks in H2 satisfy a preset relationship and the similarity of the two image blocks is greater than a threshold, delete the image blocks in H1 and / or the image blocks in H2;

[0015] The sequence number is encoded to obtain a position code, and the remaining image blocks in H1 and the remaining image blocks in H2 are encoded according to the position code and the image block code to obtain an image block coding result.

[0016] Preferably, the encoder includes a multi-head attention layer, an LN layer, a Feed Forward layer and a residual connection layer, and the output of the encoder is: O=LayerNorm(A+FeedForward(A)), where A=LayerNorm(X+MultiHeadAttention(X)), X represents the input of the encoder.

[0017] In addition, the present invention provides an acute kidney injury early warning system, the system comprising the following units:

[0018] An information acquisition unit is used to acquire renal function index data and user data and put them into a first set; acquire color Doppler ultrasound images of the left and right kidneys taken by a bedside instant ultrasound device, and put the systolic peak velocity PSV and the diastolic end velocity EDV corresponding to the left kidney and the right kidney into a second set;

[0019] An encoding unit, used for dividing the color ultrasound image into a plurality of image blocks and encoding the image blocks, inputting the image block encoding results, the element encoding results corresponding to the first set, and the element encoding results corresponding to the second set into three encoders with the same structure respectively, adding three classification identification vectors to the output results of each encoder and inputting them into the encoder of the next layer respectively; replacing the second and third classification identification vectors in the output of the encoder of the next layer with the first classification identification vectors in the outputs of the other two encoders of the next layer, until the last encoder;

[0020] The decoding unit is used to obtain the acute kidney injury level according to the classification identification vectors output by the three encoders of the last layer, and to send warning information to the user according to the acute kidney injury level.

[0021] Preferably, the acute kidney injury grade is obtained according to the classification identification vector output by the three encoders of the last layer, specifically:

[0022] The three classification identification vectors output by the last layer encoder corresponding to the image are recorded as the first vector set, the three classification identification vectors output by the last layer encoder corresponding to the first set are recorded as the second vector set, and the three classification identification vectors output by the last layer encoder corresponding to the second set are recorded as the third vector set;

[0023] The first classification identification vector in the first vector set is concatenated with the second classification identification vector and the third classification identification vector respectively, and then input into the respective MLPs to obtain two outputs S1 and S2; the first classification identification vector in the second vector set is concatenated with the second classification identification vector and the third classification identification vector respectively, and then input into the respective MLPs to obtain two outputs S3 and S4; the first classification identification vector in the third vector set is concatenated with the second classification identification vector and the third classification identification vector respectively, and then input into the respective MLPs to obtain two outputs S5 and S6;

[0024] S1, S2, S3, S4, S5, and S6 were input into the MLP to obtain the grade of acute kidney injury.

[0025] Preferably, three classification identification vectors are added to the output result of each encoder, specifically:

[0026] Obtain a matrix corresponding to the output result of the encoder, wherein the matrix size is M×N; add three row vectors before the first row of the matrix so that the size of the matrix is ​​(M+3)×N, and the added three row vectors are divided into the first classification identification vector, the second classification identification vector, and the third classification identification vector; wherein M and N are positive integers.

[0027] Preferably, the step of dividing the color ultrasound image into a plurality of image blocks and encoding the image blocks is specifically as follows:

[0028] Convert the left kidney color ultrasound image and the right kidney color ultrasound image to the same size, and divide the left kidney color ultrasound image and the right kidney color ultrasound image into multiple image blocks of the same size to obtain an image block sequence H1 of the left kidney color ultrasound image and an image block sequence H2 of the right kidney color ultrasound image, and if the sequence numbers of the image blocks in H1 and the sequence numbers of the image blocks in H2 satisfy a preset relationship and the similarity of the two image blocks is greater than a threshold, delete the image blocks in H1 and / or the image blocks in H2;

[0029] The sequence number is encoded to obtain a position code, and the remaining image blocks in H1 and the remaining image blocks in H2 are encoded according to the position code and the image block code to obtain an image block coding result.

[0030] The encoder includes a multi-head attention layer, an LN layer, a Feed Forward layer and a residual connection layer. The output of the encoder is: O=LayerNorm(A+FeedForward(A)), where A=LayerNorm(X+MultiHeadAttention(X)), X represents the input of the encoder.

[0031] Finally, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the method described above is implemented.

[0032] In the prior art, the judgment of acute kidney injury is mostly based on blood creatinine level and urine volume. These indicators have lags and are limited in their effectiveness in early judgment of acute kidney injury. Color Doppler ultrasound, especially instant bedside color Doppler ultrasound, is increasingly becoming an important basis for doctors to judge acute kidney injury. However, the existing methods mostly rely on doctors to make manual judgments, which are highly subjective. Based on this, the present invention proposes an acute kidney injury early warning method and system, which uses a multimodal approach in artificial intelligence to warn of acute kidney injury, and combined with other patient data, the judgment result is more accurate. The main contributions of the present invention relative to the prior art are: 1) combining kidney color Doppler ultrasound images with other indicator data to warn of acute kidney injury; 2) proposing a multimodal mid-term feature fusion method; 3) processing kidney color Doppler ultrasound images on the left and right sides to reduce the amount of image data processing and improve training and prediction speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0034] Figure 1 This is a flow chart of Embodiment 1;

[0035] Figure 2 is a schematic diagram of the coding unit structure;

[0036] Figure 3 Schematic diagram of the decoding unit structure;

[0037] Figure 4 It is a structural diagram of the second embodiment. DETAILED DESCRIPTION

[0038] In this article, relational terms such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] Embodiment 1, the present invention provides an early warning method for acute kidney injury, such as Figure 1 As shown, the method comprises the following steps:

[0041] Step 1: Obtain renal function index data and user data and put them into the first set; obtain color Doppler ultrasound images of the left and right kidneys taken by a bedside instant ultrasound device, and put the peak systolic flow velocity PSV and end diastolic flow velocity EDV corresponding to the left kidney and the right kidney into the second set;

[0042] The earlier acute kidney injury is discovered, the earlier the intervention is, and the better the final effect is. However, judging only by blood creatinine level or urine volume has lag, and some kidney injuries cannot be determined by blood creatinine level alone, such as tubular epithelial cell damage. In the present invention, judgment is made based on three indicators. The first aspect is renal function index data and user data, the second aspect is the left and right kidney color Doppler ultrasound images, and the third aspect is the kidney's peak systolic flow velocity (PeakSystolic Flow Velocity, PSV) and end diastolic flow velocity (End Diastolic Flow Velocity, EDV). In a more specific embodiment, the renal function index includes blood creatinine and urine volume, and the user data includes information such as user age and weight, wherein the renal function index data and user data can be obtained through electronic medical records, and PSV and EDV can be obtained through color Doppler ultrasound.

[0043] Renal color ultrasound images include two types of images, namely, the renal ultrasound image taken on the left and the renal ultrasound image taken on the right. There are also two corresponding PSVs and two EDVs, corresponding to the ultrasound examination on the left and the ultrasound examination on the right.

[0044] Step 2: Divide the color ultrasound image into multiple image blocks and encode the image blocks, input the image block encoding results, the element encoding results corresponding to the first set, and the element encoding results corresponding to the second set into three encoders with the same structure respectively, and add three classification identification vectors to the output results of each encoder and input them into the encoder of the next layer respectively; replace the second and third classification identification vectors in the output of the encoder of the next layer with the first classification identification vectors in the outputs of the other two encoders of the next layer, until the last encoder;

[0045] Color ultrasound images and text information belong to two modal information forms, which requires the use of a multimodal method to identify the color ultrasound images and text information, and then fuse the information. There are many methods for information fusion, including early fusion, mid-term fusion, and late fusion. The present invention adopts a combination of mid-term fusion and late fusion. Specifically, the color ultrasound image is first segmented into multiple image blocks, the patches are encoded in combination with the positions of the patches, and then the encoding results are input into the encoder. In a more specific embodiment, the encoder uses a Transformer Encoder layer. In a specific embodiment, the encoding process of the image is the same as the encoding process in the ViT model. The encoder includes a multi-head attention layer, an LN layer, a Feed Forward layer, and a residual connection layer. The output of the encoder is: O = LayerNorm (A + FeedForward (A)), where A = LayerNorm (X + MultiHeadAttention (X)), and X represents the input of the encoder. In the following, unless otherwise specified, the encoder adopts the above structure.

[0046] Then, the elements in the first set are encoded and the elements in the second set are encoded. After, for example, gender is digitally represented, the elements in the first set and the second set are both numerical values. The numerical values ​​are embedded to obtain the encoding results of the elements in the first set and the encoding results of the elements in the second set. Then, the image encoding result, the element encoding result corresponding to the first set, and the element encoding result corresponding to the second set are input into three independent encoders respectively. From the perspective of the entire model, the first layer includes three independent encoders with the same structure, such as Figure 2As shown (the feature fusion part is not shown in the figure), of course, due to the different numbers of image blocks and the number of elements in the first set and the second set, the specific parameters of the encoder are different. It should be noted that, in order to facilitate subsequent fusion, the vector dimension of the encoding result of the image block is the same as the vector dimension of the encoding result of the element.

[0047] It should be noted that in the present invention, layers and encoders are two concepts. From the initial input to the final output, the model of acute kidney injury early warning is divided into multiple layers, and the numbering starts from 1. Each layer includes at least one encoder. In a more specific embodiment, the first layer includes three encoders, the second layer includes three encoders, the third layer includes three encoders, and the fourth layer includes three encoders; in another embodiment, the first layer includes three encoders, the second layer includes three encoders, the fourth layer includes three encoders, and the fifth layer includes three encoders. The above number of layers is just an example, and more layers may also be included.

[0048] For the three encoders in the first layer, their inputs are the image block encoding results, the element encoding results corresponding to the first set, and the element encoding results corresponding to the second set, respectively. Then, the outputs of the three encoders are obtained. For the output of each encoder in the first layer, three vectors are added. These three vectors are respectively recorded as the first classification identification vector, the second classification identification vector, and the third classification identification vector. The classification identification vector is the same as the CLS in the ViT model, and is used to obtain the global information of the encoder input. In this way, three classification identification vectors are added to the output of each of the three encoders in the first layer.

[0049] In a more specific embodiment, three classification identification vectors are added to the output result of each encoder, specifically:

[0050] Obtain a matrix corresponding to the output result of the encoder, wherein the matrix size is M×N; add three row vectors before the first row of the matrix so that the size of the matrix is ​​(M+3)×N, and the added three row vectors are divided into the first classification identification vector, the second classification identification vector, and the third classification identification vector; wherein M and N are positive integers.

[0051] For example, the output of the first encoder is 50×256, that is, the length of each image block after encoding is 256, and there are 50 image blocks in total. After adding three classification identification vectors to the output of the first encoder, the resulting matrix is ​​53×256, where the first row of the matrix is ​​the first classification identification vector CLS11, the second row of the matrix is ​​the second classification identification vector CLS12, and the third row of the matrix is ​​the third classification identification vector CLS13. The same operation is performed on the output of the second encoder. After adding three classification identification vectors to the output of the second encoder, the first row of the output of the second encoder is the first classification identification vector CLS21, the second row is the second classification identification vector CLS22, and the third row is the third classification identification vector CLS23. The third encoder of the first layer also adds three classification identification vectors, which will not be repeated here.

[0052] When training the model, the classification identification vector is randomly initialized and continuously learned during the training process, so that when using the model, the trained model can be directly used.

[0053] In the first layer, not using classification labels in the input of the three encoders can use the attention mechanism to extract features, avoiding the interference of the classification label vector on the above process and reducing the amount of calculation; introducing the classification label vector from the input of the three encoders in the second layer to perform multimodal feature fusion, that is, adopting a mid-term feature fusion method and using the classification label vector method.

[0054] The corresponding positions in the outputs of the three encoders of the second layer are still classification identification vectors. The second and third classification identification vectors of the matrix output by the first encoder of the second layer are replaced by the first classification identification vector of the matrix output by the second encoder of the second layer; the second and third classification identification vectors of the matrix output by the second encoder of the second layer are replaced by the first classification identification vector of the matrix output by the first encoder of the second layer; the second and third classification identification vectors of the matrix output by the third encoder of the second layer are replaced by the first classification identification vector of the matrix output by the first encoder of the second layer. Perform the same operation for each subsequent layer until the last layer. The output of the encoder of the last layer does not need to perform the above operation, and then obtain the three classification identification vectors in the output of each of the three encoders of the last layer, that is, 9 classification identification vectors can be obtained. It should be noted that since the above-mentioned replacement process involves matrix operations, it is necessary to ensure that the number of matrix columns output by the encoder is the same. Since the encoder does not change the size of the input content, that is, it does not change the size of the input matrix, when performing image block encoding and numerical encoding, it is necessary to ensure that the encoding lengths of the two are the same. For example, an image block is encoded as a length of 256, and the numerical value is also encoded as a length of 256. This involves the specific operation of embedding. Those skilled in the art know how to operate and will not be described in detail.

[0055] Step three, obtaining the acute kidney injury level according to the classification identification vectors output by the three encoders of the last layer, and issuing warning information to the user according to the acute kidney injury level.

[0056] The output of each of the three encoders of the last layer includes three classification identification vectors. These nine classification identification vectors contain information on renal function index data, user data, renal color Doppler ultrasound images, and systolic peak velocity PSV, diastolic end velocity EDV, and the relationship between them. In a specific embodiment, the acute kidney injury grade is obtained according to the classification identification vectors output by the three encoders of the last layer, specifically:

[0057] The three classification identification vectors output by the last layer encoder corresponding to the image are recorded as the first vector set, the three classification identification vectors output by the last layer encoder corresponding to the first set are recorded as the second vector set, and the three classification identification vectors output by the last layer encoder corresponding to the second set are recorded as the third vector set;

[0058] The early warning model for acute kidney injury includes multiple layers, each layer includes three encoders, and each column of encoders corresponds to a content, which is an ultrasound image or a first set or a second set.

[0059] The first classification identification vector in the first vector set is concatenated with the second classification identification vector and the third classification identification vector respectively, and then input into the respective MLPs to obtain two outputs S1 and S2; the first classification identification vector in the second vector set is concatenated with the second classification identification vector and the third classification identification vector respectively, and then input into the respective MLPs to obtain two outputs S3 and S4; the first classification identification vector in the third vector set is concatenated with the second classification identification vector and the third classification identification vector respectively, and then input into the respective MLPs to obtain two outputs S5 and S6;

[0060] For the first classification identification vector in each vector set, it does not directly participate in the feature fusion, that is, it is not replaced, and the second and third classification identification vectors contain information of other contents. For each vector set, the first classification identification is concatenated with the second and third classification identification vectors, respectively, and then input into the corresponding MLPs to obtain two outputs. For example, for the first vector set, its first classification identification vector is CLS11, with a size of 1×256; the second classification identification vector is CLS12, with a size of 1×256; the third classification identification vector is CLS13, with a size of 1×256. After concatenation, two vectors of size 1×512 are obtained, and then the two 1×512 vectors are respectively input into two multilayer perceptron MLPs to obtain two outputs S1 and S2, where the lengths of S1 and S2 are less than 512, for example, the size of S1 is 1×128, and the size of S2 is 1×64.

[0061] S1, S2, S3, S4, S5, and S6 were input into the MLP to obtain the grade of acute kidney injury.

[0062] Regarding the grading of acute kidney injury, in a specific embodiment, a higher grade indicates a greater possibility of acute kidney injury. For example, it is divided into 5 grades, with grade 1 indicating a very low possibility of acute kidney injury and grade 5 indicating an extremely high possibility of acute kidney injury.

[0063] The image contains many pixels. The training and recognition speed of the early warning model of acute kidney injury in the present invention mainly depends on the processing of the image. However, the color ultrasound image contains a lot of useless information, such as the surrounding area, which is of little significance for judging acute kidney injury and greatly increases the amount of calculation. The color ultrasound image is divided into multiple image blocks and the image blocks are encoded, specifically:

[0064] Convert the left kidney color ultrasound image and the right kidney color ultrasound image to the same size, and divide the left kidney color ultrasound image and the right kidney color ultrasound image into multiple image blocks of the same size to obtain an image block sequence H1 of the left kidney color ultrasound image and an image block sequence H2 of the right kidney color ultrasound image, and if the sequence numbers of the image blocks in H1 and the sequence numbers of the image blocks in H2 satisfy a preset relationship and the similarity of the two image blocks is greater than a threshold, delete the image blocks in H1 and / or the image blocks in H2;

[0065] The color Doppler ultrasound images of the kidneys on the left and right sides are converted to the same size, and the conversion method includes but is not limited to image cropping, convolution, etc. Then the two images are divided into multiple patches (image blocks), so as to obtain a left kidney color Doppler ultrasound image block sequence and a right kidney color Doppler ultrasound image block sequence. For the sequence number that satisfies the preset relationship and the similarity of the two image blocks is greater than the threshold value, the image block in H1 and / or the image block in H2 are deleted. The preset relationship is that the sequence number of the image block in H1 is the same as the sequence number of the image block in H2, and the sequence number indicates that the image block is located in the surrounding area of ​​the kidney color Doppler ultrasound image. In another embodiment, the preset relationship is that the sequence number of the image block in H1 is the same as the sequence number of the image block in H2.

[0066] The sequence number is encoded to obtain a position code, and the remaining image blocks in H1 and the remaining image blocks in H2 are encoded according to the position code and the image block code to obtain an image block coding result.

[0067] For the remaining image blocks, the original serial numbers of the image blocks are obtained, the original serial numbers are encoded to obtain the position codes, and then the image blocks are encoded, and the position codes and the image block codes are bitwise added to obtain the image block encoding results; and then the encoded image block results are spliced ​​into the image encoding results.

[0068] Embodiment 2, the present invention provides an early warning system for acute kidney injury, such as Figure 4 As shown, the system comprises the following units:

[0069] The information acquisition unit 101 is used to acquire renal function index data and user data and put them into a first set; acquire color Doppler ultrasound images of the left and right kidneys taken by a bedside instant ultrasound device, and put the peak systolic velocity PSV and end diastolic velocity EDV corresponding to the left kidney and the right kidney into a second set;

[0070] The encoding unit 102 is used to divide the color ultrasound image into a plurality of image blocks and encode the image blocks, and input the image block encoding results, the element encoding results corresponding to the first set, and the element encoding results corresponding to the second set into three encoders with the same structure respectively, and add three classification identification vectors to the output results of each encoder and input them into the encoder of the next layer respectively; replace the second and third classification identification vectors in the output of the encoder of the next layer with the first classification identification vectors in the outputs of the other two encoders of the next layer, until the last encoder;

[0071] The decoding unit 103 is used to obtain the acute kidney injury level according to the classification identification vectors output by the three encoders in the last layer, and send warning information to the user according to the acute kidney injury level.

[0072] Preferably, the acute kidney injury grade is obtained according to the classification identification vector output by the three encoders of the last layer, specifically:

[0073] The three classification identification vectors output by the last layer encoder corresponding to the image are recorded as the first vector set, the three classification identification vectors output by the last layer encoder corresponding to the first set are recorded as the second vector set, and the three classification identification vectors output by the last layer encoder corresponding to the second set are recorded as the third vector set;

[0074] The first classification identification vector in the first vector set is concatenated with the second classification identification vector and the third classification identification vector respectively, and then input into their respective MLPs to obtain two outputs S1 and S2; the first classification identification vector in the second vector set is concatenated with the second classification identification vector and the third classification identification vector respectively, and then input into their respective MLPs to obtain two outputs S3 and S4; the first classification identification vector in the third vector set is concatenated with the second classification identification vector and the third classification identification vector respectively, and then input into their respective MLPs to obtain two outputs S5 and S6; Figure 3 shown.

[0075] S1, S2, S3, S4, S5, and S6 were input into the MLP to obtain the grade of acute kidney injury.

[0076] Preferably, three classification identification vectors are added to the output result of each encoder, specifically:

[0077] Obtain a matrix corresponding to the output result of the encoder, wherein the matrix size is M×N; add three row vectors before the first row of the matrix so that the size of the matrix is ​​(M+3)×N, and the added three row vectors are divided into the first classification identification vector, the second classification identification vector, and the third classification identification vector; wherein M and N are positive integers.

[0078] Preferably, the step of dividing the color ultrasound image into a plurality of image blocks and encoding the image blocks is specifically as follows:

[0079] Convert the left kidney color ultrasound image and the right kidney color ultrasound image to the same size, and divide the left kidney color ultrasound image and the right kidney color ultrasound image into multiple image blocks of the same size to obtain an image block sequence H1 of the left kidney color ultrasound image and an image block sequence H2 of the right kidney color ultrasound image, and if the sequence numbers of the image blocks in H1 and the sequence numbers of the image blocks in H2 satisfy a preset relationship and the similarity of the two image blocks is greater than a threshold, delete the image blocks in H1 and / or the image blocks in H2;

[0080] The sequence number is encoded to obtain a position code, and the remaining image blocks in H1 and the remaining image blocks in H2 are encoded according to the position code and the image block code to obtain an image block coding result.

[0081] The encoder includes a multi-head attention layer, an LN layer, a Feed Forward layer and a residual connection layer. The output of the encoder is: O=LayerNorm(A+FeedForward(A)), where A=LayerNorm(X+MultiHeadAttention(X)), X represents the input of the encoder.

[0082] Embodiment 3, the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a computer, the method described in Embodiment 1 is implemented.

[0083] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by adding a necessary general hardware platform, and of course can also be implemented by combining hardware and software. Based on such an understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a computer product, and the present invention can be in the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for early warning of acute kidney injury, It is characterized in that The method comprises the following steps: Obtain renal function index data and user data and put them into the first set; obtain color Doppler ultrasound images of the left and right kidneys taken by a bedside instant ultrasound device, and put the peak systolic velocity PSV and end diastolic velocity EDV corresponding to the left and right kidneys into the second set; The color ultrasound image is divided into a plurality of image blocks and the image blocks are encoded, and the image block encoding results, the element encoding results corresponding to the first set, and the element encoding results corresponding to the second set are respectively input into three encoders of the same structure, and the output results of each encoder are added with three classification identification vectors and then respectively input into the encoder of the next layer; the second and third classification identification vectors in the output of the encoder of the next layer are replaced with the first classification identification vectors in the outputs of the other two encoders of the next layer, until the last encoder; Obtaining an acute kidney injury grade according to the classification identification vectors output by the three encoders of the last layer, and issuing a warning message to the user according to the acute kidney injury grade; The acute kidney injury grade is obtained according to the classification identification vector output by the three encoders in the last layer, specifically: The three classification identification vectors output by the last layer encoder corresponding to the image are recorded as the first vector set, the three classification identification vectors output by the last layer encoder corresponding to the first set are recorded as the second vector set, and the three classification identification vectors output by the last layer encoder corresponding to the second set are recorded as the third vector set; the first classification identification vector in the first vector set is spliced ​​with the second classification identification vector and the third classification identification vector, respectively, and input into the respective MLPs to obtain two outputs S1 and S2; the first classification identification vector in the second vector set is spliced ​​with the second classification identification vector and the third classification identification vector, respectively, and input into the respective MLPs to obtain two outputs S3 and S4; the first classification identification vector in the third vector set is spliced ​​with the second classification identification vector and the third classification identification vector, respectively, and input into the respective MLPs to obtain two outputs S5 and S6; S1, S2, S3, S4, S5, and S6 are input into the MLP to obtain the grade of acute kidney injury.

2. The method according to claim 1, It is characterized in that The output result of each encoder is added with three classification identification vectors, specifically: Obtain a matrix corresponding to the output result of the encoder, wherein the matrix size is M×N; add three row vectors before the first row of the matrix so that the size of the matrix is ​​(M+3)×N, and the added three row vectors are divided into the first classification identification vector, the second classification identification vector, and the third classification identification vector; wherein M and N are positive integers.

3. The method according to claim 1, It is characterized in that The step of dividing the color ultrasound image into a plurality of image blocks and encoding the image blocks is specifically as follows: Convert the left kidney color ultrasound image and the right kidney color ultrasound image to the same size, and divide the left kidney color ultrasound image and the right kidney color ultrasound image into multiple image blocks of the same size to obtain an image block sequence H1 of the left kidney color ultrasound image and an image block sequence H2 of the right kidney color ultrasound image, and if the sequence numbers of the image blocks in H1 and the sequence numbers of the image blocks in H2 satisfy a preset relationship and the similarity of the two image blocks is greater than a threshold, delete the image blocks in H1 and / or the image blocks in H2; The sequence number is encoded to obtain a position code, and the remaining image blocks in H1 and the remaining image blocks in H2 are encoded according to the position code and the image block code to obtain an image block coding result.

4. The method according to claim 1, It is characterized in that The encoder includes a multi-head attention layer, an LN layer, a FeedForward layer and a residual connection layer. The output of the encoder is: O=LayerNorm(A+FeedForward(A)), where A=LayerNorm(X+MultiHeadAttention(X)), X represents the input of the encoder.

5. An early warning system for acute kidney injury, It is characterized in that The system comprises the following units: An information acquisition unit is used to acquire renal function index data and user data and put them into a first set; acquire color Doppler ultrasound images of the left and right kidneys taken by a bedside instant ultrasound device, and put the systolic peak velocity PSV and the diastolic end velocity EDV corresponding to the left kidney and the right kidney into a second set; An encoding unit, used for dividing the color ultrasound image into a plurality of image blocks and encoding the image blocks, inputting the image block encoding results, the element encoding results corresponding to the first set, and the element encoding results corresponding to the second set into three encoders with the same structure respectively, adding three classification identification vectors to the output results of each encoder and inputting them into the encoder of the next layer respectively; replacing the second and third classification identification vectors in the output of the encoder of the next layer with the first classification identification vectors in the outputs of the other two encoders of the next layer, until the last encoder; A decoding unit, used for obtaining the acute kidney injury grade according to the classification identification vectors output by the three encoders of the last layer, and issuing warning information to the user according to the acute kidney injury grade; The acute kidney injury grade is obtained according to the classification identification vector output by the three encoders in the last layer, specifically: The three classification identification vectors output by the last layer encoder corresponding to the image are recorded as the first vector set, the three classification identification vectors output by the last layer encoder corresponding to the first set are recorded as the second vector set, and the three classification identification vectors output by the last layer encoder corresponding to the second set are recorded as the third vector set; the first classification identification vector in the first vector set is spliced ​​with the second classification identification vector and the third classification identification vector, respectively, and input into the respective MLPs to obtain two outputs S1 and S2; the first classification identification vector in the second vector set is spliced ​​with the second classification identification vector and the third classification identification vector, respectively, and input into the respective MLPs to obtain two outputs S3 and S4; the first classification identification vector in the third vector set is spliced ​​with the second classification identification vector and the third classification identification vector, respectively, and input into the respective MLPs to obtain two outputs S5 and S6; S1, S2, S3, S4, S5, and S6 are input into the MLP to obtain the grade of acute kidney injury.

6. The system according to claim 5, It is characterized in that The output result of each encoder is added with three classification identification vectors, specifically: Obtain a matrix corresponding to the output result of the encoder, wherein the matrix size is M×N; add three row vectors before the first row of the matrix so that the size of the matrix is ​​(M+3)×N, and the added three row vectors are divided into the first classification identification vector, the second classification identification vector, and the third classification identification vector; wherein M and N are positive integers.

7. The system according to claim 5, It is characterized in that The step of dividing the color ultrasound image into a plurality of image blocks and encoding the image blocks is specifically as follows: Convert the left kidney color ultrasound image and the right kidney color ultrasound image to the same size, and divide the left kidney color ultrasound image and the right kidney color ultrasound image into multiple image blocks of the same size to obtain an image block sequence H1 of the left kidney color ultrasound image and an image block sequence H2 of the right kidney color ultrasound image, and if the sequence numbers of the image blocks in H1 and the sequence numbers of the image blocks in H2 satisfy a preset relationship and the similarity of the two image blocks is greater than a threshold, delete the image blocks in H1 and / or the image blocks in H2; The sequence number is encoded to obtain a position code, and the remaining image blocks in H1 and the remaining image blocks in H2 are encoded according to the position code and the image block code to obtain an image block coding result.

8. A computer-readable storage medium, wherein a computer program is stored on the storage medium, and when the computer program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.

Images