Method and system for automatic detection of kidney disease based on machine learning

Abstract

The method and system for automatic detection of kidney diseases based on machine learning include a kidney image collection module, a kidney image management module, a kidney image intelligent diagnosis module and a kidney disease classification module. The kidney images with kidney disease diagnosis labels collected by the kidney image collection module are stored and updated in the kidney image management module according to the kidney disease type. The kidney disease type is inferred by the kidney image intelligent diagnosis module using a machine learning algorithm. The kidney disease classification module classifies the diagnosed kidney diseases according to different types of kidney diseases, thereby achieving efficient and accurate diagnosis of various types of kidney diseases, which is of great significance for the diagnosis and early screening of kidney diseases.

Description

Technical Field

[0001] The present disclosure relates to systems and methods for image analysis and medical diagnostic testing using image analysis, and more particularly, the present disclosure relates to a method and system for automatic detection of kidney disease based on machine learning. Background Art

[0002] Medical imaging refers to the technology and process of obtaining internal images of the human body or a part of the human body in a non-invasive manner for medical treatment or medical research. With the rapid development of various scientific and technological fields, medical image processing technology has also made rapid progress. However, with the development of science and technology and the promotion of medical imaging applications, more and more medical images need to be interpreted by doctors, and medical image interpretation has gradually become a challenging task. The traditional interpretation and diagnosis of medical images mainly rely on experienced doctors. This process is subjective. In addition, in the process of manual medical image interpretation, there will be interpretation errors due to the cognitive limitations or fatigue of doctors, resulting in misdiagnosis. These defects all show the importance of using effective medical image analysis technology to improve the accuracy of disease diagnosis results.

[0003] Artificial intelligence technology consists of different fields such as machine learning and computer vision, and its purpose is to produce intelligent machines similar to humans. In response to the problems existing in the interpretation of medical images, researchers have applied image processing technology and artificial intelligence technology to medical image diagnosis, thereby achieving the role of automatic analysis of medical images and assisting doctors in making medical diagnoses, which can effectively reduce the workload of doctors. Summary of the invention

[0004] In response to the above problems, the present invention aims to provide a method and system for automatic detection of kidney disease based on machine learning.

[0005] The purpose of the present invention is achieved through the following technical solutions:

[0006] A method and system for automatic detection of kidney diseases based on machine learning, including a kidney image collection module, a kidney image management module, a kidney impact intelligent diagnosis module, and a kidney disease classification module;

[0007] Kidney image collection module: used to collect various types of kidney disease images of patients. The sources of kidney images include InterVar (site pathogenicity judgment), GeneReviews (disease database) and kidney images of in-hospital patients;

[0008] Renal image management module: including a renal image pre-storage unit and a renal image update unit. The renal image pre-storage unit is used to pre-store various types of renal images collected by the renal image collection module. The renal image update unit is used to screen renal images with renal disease diagnosis labels, clean images with excessive noise, default or irrelevant images, and then update the renal image training set.

[0009] The renal image intelligent diagnosis module includes a renal image processing unit and a renal image database. The renal image processing unit extracts the disease characteristics of the renal image by analyzing the renal image with the renal disease diagnosis label and the plain scan image. The renal image updated by the renal image updating unit is used as a training set to establish a machine learning model to realize automatic disease detection of the renal image that needs to be identified. The trained renal image with the renal disease diagnosis label is saved to the renal image database to generate a model for automatic disease detection of renal images that can be identified so that it can be used directly next time without retraining.

[0010] The kidney disease classification module further classifies the kidney diseases identified by the kidney imaging intelligent diagnosis module to ensure that patients using the system can obtain diagnosis and treatment plans through self-examination of kidney images.

[0011] Furthermore, the renal image collection module collects data from InterVar (site pathogenicity assessment), GeneReviews (disease database), and renal images of in-hospital patients, and aggregates them into a sql database for online management.

[0012] Furthermore, the renal imaging update unit screens renal images with diagnostic labels for renal diseases, including calcifications, calyx and pelvis density, renal medulla, cysts, and cancer thrombi.

[0013] Furthermore, the kidney image processing unit uses a maximum entropy multi-threshold segmentation method to perform image segmentation on the kidney image to be diagnosed, and uses a particle swarm algorithm to determine the optimal combination threshold of the maximum entropy multi-threshold segmentation method used in the kidney image processing unit.

[0014] Furthermore, in the particle swarm algorithm, the particle swarm iteratively updates the particle solution in the following manner:

[0015] V i (t+1)=ωV i (t)+c i1 (t)r1(Pbest i (t)-X i (t))+c i2 (t)r2(Gbest(t)-X i (t))

[0016] Xi (t+1)=X i (t)+V i (t+1)

[0017] In the formula, X i (t+1) and V i (t+1) respectively represent the solution and velocity of particle i at the (t+1)th iteration, X i (t) and V i (t) represent the solution and velocity of particle i at the tth iteration, ω represents the inertia weight factor, Pbest i (t) represents the historical optimal solution of particle i at the tth iteration, Gbest(t) represents the global optimal solution of the particle swarm at the tth iteration, r1 and r2 represent randomly generated random numbers between 0 and 1, and c i1 (t) represents the cognitive learning factor of particle i at the tth iteration, c i2 (t) represents the social learning factor of particle i at the tth iteration.

[0018] Furthermore, the particle swarm is set to adjust the cognitive learning factor c of particle i at the tth iteration in the following way: i1 (t) and social learning factor c i2 The value of (t):

[0019] Assume Ω i (t) represents the local neighborhood of particle i at the tth iteration, and Ω i (t) is the solution to X i (t) is the center and R(t) is the radius. R(t) is the local neighborhood radius of the particle swarm at the tth iteration. The value of R(t) is set as: Among them, R i (t) represents the local neighborhood radius of particle i at the tth iteration, and Indicates the distance solution X in the particle swarm at the tth iteration i (t) The solution of the lth nearest particle, c represents a given positive integer, and c<N, N represents the number of particles in the particle group; Ω i,best (t) represents the historical optimal neighborhood of particle i at the tth iteration, and Ω i,best (t) is the solution of Pbest i (t) is the circular area with r(t) as the radius, r(t) is the historical optimal neighborhood radius of the particle swarm at the tth iteration, and Among them, r i (t) represents the historical optimal neighborhood radius of particle i at the tth iteration, and In the formula, Indicates the distance solution Pbest in the particle swarm at the tth iteration i (t) The solution of the ath nearest particle;

[0020] Define the function q i (t) is used to detect the optimization uniformity of particle i at the tth iteration, and set q i The calculation formula of (t) is:

[0021]

[0022] In the formula, q i (t) represents the optimization uniformity of particle i at the tth iteration, J i (t) represents the local optimization uniformity of particle i at the tth iteration, S i (t) represents the historical optimization uniformity of particle i at the tth iteration, J i (t) and S i The values ​​of (t) are:

[0023]

[0024]

[0025] In the formula, X j (t) represents the solution of particle j at the tth iteration, X b (t) represents the solution of particle b at the tth iteration, and X j (t)≠X b (t), X k (t) represents the solution of particle k at the tth iteration, X o (t) represents the solution of particle o at the tth iteration, and X k (t)≠X o (t), λ i,j (t) is the optimal judgment function between particles i and j at the tth iteration, let f i (t) and f j (t) represent the fitness function values ​​of particle i and particle j at the tth iteration respectively, then λ i,b (t) is the optimal judgment function between particle i and particle b at the tth iteration, let f b (t) represents the fitness function value of particle b at the tth iteration, then m i,best (t) represents the historical optimal neighborhood Ω i,best The number of particles in (t);

[0026] When i (t)<H q1 (t) or qi (t)>H q2 (t), then c i1 (t) and c i2 The values ​​of (t) are adjusted to:

[0027]

[0028]

[0029] When H q1 (t)≤q i (t)≤H q2 (t), then c i1 (t) and c i2 The values ​​of (t) are adjusted to:

[0030]

[0031]

[0032] In the above formula, α and β are two constants used to control c i1 (t) and c i2 (t) range, H q1 (t) is the lower limit of the global optimization uniformity of the particle swarm at the tth iteration, and H q2 (t) is the upper limit of the global optimization uniformity of the particle swarm at the tth iteration, and in, represents the mean value of the optimization uniformity of the particles in the particle swarm at the tth iteration, and T max is the maximum number of iterations given.

[0033] Furthermore, the back propagation algorithm is used to further classify the kidney diseases identified by the kidney imaging intelligent diagnosis module to ensure that patients using the system can obtain diagnosis and treatment plans through self-examination of kidney images. The specific steps are as follows:

[0034] First, build an M-layer neural network. For the output of the m+1th layer network, we have:

[0035] a m+1 =f m+1 (W m+1 a m +b m+1 ),m=0,1,…,M-1

[0036] Among them, M is the number of network layers, a is the output of the neural network, that is, the category of kidney disease, and there is a 0 =p and a=a M , a m+1is the output of the m+1th layer of the neural network, a m is the output of the mth layer of the neural network, f m+1 is the activation function of the m+1th layer of the neural network, W m+1 is the weight of the m+1th layer of the neural network, b m+1 is the bias of the m+1th layer of the neural network. Assuming that for each type of kidney disease feature input p, there is a corresponding kidney disease type output t, the following corresponding relationship can be established:

[0037] {p1,t1},{p2,t2},…,{p q ,t q},…,{p Q ,t Q}

[0038] Among them, p1 is the input of the first kidney image feature, t1 is the expected output corresponding to the first classification, p2 is the input of the second kidney image feature, t2 is the expected output corresponding to the second classification, and p q is the input of the qth kidney image feature, t q is the expected output corresponding to the qth classification, p Q is the input of the Qth kidney image feature, t Q is the expected output corresponding to the Qth classification, and the algorithm will adjust the network parameters to minimize the mean square error: F(x) = E[(ta) 2 ]=E[(ta) T (ta)], we can let e = ta represent the error between the expected output and the predicted output, where F represents the mean square error, x represents the independent variable of F, in which case the independent variables include t and a, and E represents the mathematical expectation, (ta) T Denote the transpose of the error, and use Substitute F(x) to approximate the mean square error: Where k is the kth iteration, and the steepest descent method for approximate mean square error is: Where α is the learning rate, represents the weight of the i-th element to the j-th element in the k+1-th iteration of the m-th layer neural network, It represents the weight of the i-th element to the j-th element in the k-th iteration of the m-th layer neural network, and the chain rule is defined as:

[0039]

[0040] If f(n)=e^n and n=2w, then f(n(w))=e^{2w}, we have Then using the chain rule and It can be written as: For the net input n of the mth layer, we have: Among them, s represents sensitivity, s m-1 represents the sensitivity of the m-1th layer of the neural network, represents the net input of the i-th element in an m-layer neural network, represents the output of the jth element in the m-1 layer neural network, represents the bias of the i-th element in an m-layer neural network, so: If defined: but and It can be simplified to: Using the approximate steepest descent method, it can be expressed as:

[0041]

[0042]

[0043] in, represents the bias of the i-th element in the k+1th iteration of the m-layer neural network, then the matrix form can be expressed as:

[0044] W n (k+1)=w m (k)-αs m (a m-1 ) T

[0045] b m (k+1)=b m (k)-αs m

[0046] Among them, W m (k+1) represents w m The matrix form of (k+1) is the weight set after the k+1th iteration of the m-layer neural network. Here s m have:

[0047]

[0048] in, It represents the partial derivative of the approximate error with respect to the net input of the mth layer, It represents the partial derivative of the approximate error with respect to the first net input of the mth layer. It represents the partial derivative of the approximate error with respect to the second net input of the mth layer, represents the approximation error for the sth layer of the mth layer m Find the partial derivative of the net input and record the Jacobian matrix as:

[0049]

[0050] in, represents the partial derivative of the net input of the m+1th layer of the neural network to the net input of the mth layer of the neural network, represents the partial derivative of the first net input of the m+1th layer of the neural network to the first net input of the mth layer of the neural network, represents the partial derivative of the first net input of the m+1th layer of the neural network to the second net input of the mth layer of the neural network, It represents the first net input of the m+1th layer of neural network to the sth layer of the mth layer of neural network. m The partial derivative of the net input, represents the partial derivative of the second net input of the m+1th layer of the neural network to the first net input of the mth layer of the neural network, represents the partial derivative of the second net input of the m+1th layer of the neural network to the second net input of the mth layer of the neural network, It represents the second net input of the m+1th layer of neural network to the sth layer of the mth layer of neural network m The partial derivative of the net input, represents the sth layer of the m+1th neural network m+1 The partial derivative of the net input to the net input of the mth layer of the neural network, represents the sth layer of the m+1th neural network m+1 The partial derivative of the net input to the first net input of the m-th layer of the neural network, represents the sth layer of the m+1th layer of the neural network m+1 The net input to the sth layer of the mth neural network m The partial derivative of the net input, represents the sth layer of the m+1th layer of the neural network m+1 The net input to the sth layer of the mth neural network m The partial derivative of the net input, considering the i, j elements of the matrix: in: Then the Jacobian matrix can be written as: in:

[0051] F m (n m ) indicates that the net input of the m-th layer neural network is n m The objective function when Indicates that the net input of the m-th layer neural network is The objective function when Indicates that the net input of the m-th layer neural network is The objective function when Indicates that the net input of the m-th layer neural network is The objective function when , the recursive relation of sensitivity is written by the chain rule in matrix form:

[0052] Furthermore, the back-propagation algorithm and chain rule are used to obtain the optimal training weights, and the patient's disease information is input into the trained machine learning model, which can automatically detect whether the patient suffers from calcification, renal calyx and pelvis density, renal medulla, cysts and cancer thrombus diseases.

[0053] Furthermore, for the established machine learning model, automatic detection of diseases in kidney images that need to be identified is achieved, and the trained kidney images with kidney disease diagnosis labels are saved in the kidney image database, generating a model for automatic detection of diseases in kidney images that can be identified so that it can be used directly next time without re-training. The kidney disease classification module further classifies the kidney diseases identified by the kidney image intelligent diagnosis module to ensure that patients using the system can obtain diagnosis and treatment plans by self-examination of kidney images.

[0054] Beneficial effects of the present invention:

[0055] (1) Applying artificial intelligence technology to kidney image analysis, establishing tumor diagnosis models corresponding to various parts of the human body based on kidney images, which can effectively diagnose tumors in various parts of the human body, realize automatic analysis of kidney images and assist doctors in making medical tumor diagnosis;

[0056] (2) The maximum entropy multi-threshold segmentation method is used to segment the kidney image to be diagnosed, thereby obtaining the human body part image contained in the kidney image to be diagnosed, and the particle swarm algorithm is used to determine the optimal combination threshold of the maximum entropy multi-threshold segmentation method, thereby improving the accuracy of kidney image segmentation and laying the foundation for the subsequent tumor diagnosis;

[0057] (3) The particle swarm algorithm is set to use adaptively changing cognitive learning factors and social learning factors to adjust the iterative update method of the particle swarm, thereby balancing the global exploration and local mining of the particle swarm algorithm, improving the optimization accuracy and convergence speed of the particle swarm algorithm, and making the optimal combination threshold determined by the particle swarm algorithm have higher accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The invention is further illustrated by the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0059] Figure 1 It is a schematic diagram of the structure of the present invention. DETAILED DESCRIPTION

[0060] A method and system for automatic detection of kidney diseases based on machine learning, including a kidney image collection module, a kidney image management module, a kidney impact intelligent diagnosis module, and a kidney disease classification module;

[0061] Kidney image collection module: used to collect various types of kidney disease images of patients. The sources of kidney images include InterVar (site pathogenicity assessment), GeneReviews (disease database) and kidney images of in-hospital patients.

[0062] Renal image management module: including a renal image pre-storage unit and a renal image update unit. The renal image pre-storage unit is used to pre-store various types of renal images collected by the renal image collection module. The renal image update unit is used to screen renal images with renal disease diagnosis labels, clean images with excessive noise, default or irrelevant images, and then update the renal image training set.

[0063] The renal image intelligent diagnosis module includes a renal image processing unit and a renal image database. The renal image processing unit extracts the disease characteristics of the renal image by analyzing the renal image with the renal disease diagnosis label and the plain scan image. The renal image updated by the renal image updating unit is used as a training set to establish a machine learning model to realize automatic disease detection of the renal image that needs to be identified. The trained renal image with the renal disease diagnosis label is saved to the renal image database to generate a model for automatic disease detection of renal images that can be identified so that it can be used directly next time without retraining.

[0064] The kidney disease classification module further classifies the kidney diseases identified by the kidney imaging intelligent diagnosis module to ensure that patients using the system can obtain diagnosis and treatment plans through self-examination of kidney images.

[0065] Specifically, the renal image collection module collects data from InterVar (site pathogenicity assessment), GeneReviews (disease database), and renal images of in-hospital patients, and aggregates them into a SQL database for online management.

[0066] Specifically, the renal imaging update unit screens renal images with kidney disease diagnostic labels, including calcifications, calyx and pelvis density, renal medulla, cysts, and cancer thrombi.

[0067] Preferably, the medical image processing unit uses a maximum entropy multi-threshold segmentation method to perform image segmentation on the medical image to be diagnosed, and uses a particle swarm algorithm to determine the optimal combination threshold of the maximum entropy multi-threshold segmentation method used in the medical image processing unit.

[0068] Preferably, the fitness function of the particle swarm is set to maximum entropy. The larger the fitness function of the particles in the particle swarm is, the better the solution obtained by particle optimization is.

[0069] Specifically, this embodiment uses the maximum entropy multi-threshold segmentation method to perform image segmentation on the medical image to be diagnosed, thereby obtaining the image of the human body parts contained in the medical image to be diagnosed, and uses the particle swarm algorithm to determine the optimal combination threshold of the maximum entropy multi-threshold segmentation method, thereby improving the accuracy of medical image segmentation and laying the foundation for subsequent tumor diagnosis.

[0070] Preferably, in the particle swarm algorithm, the particle swarm iteratively updates the solution of the particle in the following manner:

[0071] V i (t+1)=ωV i (t)+c i1 (t)r1(Pbest i (t)-X i (t))+c i2 (t)r2(Gbest(t)-X i (t))

[0072] X i (t+1)=X i (t)+V i (t+1)

[0073] In the formula, X i (t+1) and V i (t+1) respectively represent the solution and velocity of particle i at the (t+1)th iteration, X i (t) and V i (t) represent the solution and velocity of particle i at the tth iteration, ω represents the inertia weight factor, Pbest i (t) represents the historical optimal solution of particle i at the tth iteration, Gbest(t) represents the global optimal solution of the particle swarm at the tth iteration, r1 and r2 represent randomly generated random numbers between 0 and 1, and c i1 (t) represents the cognitive learning factor of particle i at the tth iteration, c i2 (t) represents the social learning factor of particle i at the tth iteration.

[0074] Preferably, the particle swarm is set to adjust the cognitive learning factor c of particle i at the tth iteration in the following manner: i1 (t) and social learning factor c i2 The value of (t):

[0075] Assume Ω i (t) represents the local neighborhood of particle i at the tth iteration, and Ωi (t) is the solution to X i (t) is the center and r(t) is the radius. R(t) is the local neighborhood radius of the particle swarm at the tth iteration. The value of R(t) is set as: Among them, R i (t) represents the local neighborhood radius of particle i at the tth iteration, and Indicates the distance solution X in the particle swarm at the tth iteration i (t) The solution of the nearest particle, c represents a given positive integer, and c<N, the value of c can be 10, N represents the number of particles in the particle group; Ω i,best (t) represents the historical optimal neighborhood of particle i at the tth iteration, and Ω i,best (t) is the solution of Pbest i (t) is the circular area with r(t) as the radius, r(t) is the historical optimal neighborhood radius of the particle swarm at the tth iteration, and Among them, r i (t) represents the historical optimal neighborhood radius of particle i at the tth iteration, and In the formula, Indicates the distance solution Pbest in the particle swarm at the tth iteration i (t) The solution of the ath nearest particle;

[0076] Define the function q i (t) is used to detect the optimization uniformity of particle i at the tth iteration, and set q i The calculation formula of (t) is:

[0077]

[0078] In the formula, q i (t) represents the optimization uniformity of particle i at the tth iteration, J i (t) represents the local optimization uniformity of particle i at the tth iteration, S i (t) represents the historical optimization uniformity of particle i at the tth iteration, J i (t) and S i The values ​​of (t) are:

[0079]

[0080]

[0081] In the formula, X j (t) represents the solution of particle j at the tth iteration, X b (t) represents the solution of particle b at the tth iteration, and X j(t)≠X b (t), X k (t) represents the solution of particle k at the tth iteration, X o (t) represents the solution of particle o at the tth iteration, and X k (t)≠X o (t), λ i,j (t) is the optimal judgment function between particles i and j at the tth iteration, let f i (t) and f j (t) represent the fitness function values ​​of particle i and particle j at the tth iteration respectively, then λ i,b (t) is the optimal judgment function between particle i and particle b at the tth iteration, let f b (t) represents the fitness function value of particle b at the tth iteration, then m i,best (t) represents the historical optimal neighborhood Ω i,best The number of particles in (t);

[0082] When i (t)<H q1 (t) or q i (t)>H q2 (t), then c i1 (t) and c i2 The values ​​of (t) are adjusted to:

[0083]

[0084]

[0085] When H q1 (t)≤q i (t)≤H q2 (t), then c i1 (t) and c i2 The values ​​of (t) are adjusted to:

[0086]

[0087]

[0088] In the above formula, α and β are two constants used to control c i1 (t) and c i2 (t) range, where α can be 1, β can be 0.5, H q1 (t) is the lower limit of the global optimization uniformity of the particle swarm at the tth iteration, and H q2(t) is the upper limit of the global optimization uniformity of the particle swarm at the tth iteration, and in, represents the mean value of the optimization uniformity of the particles in the particle swarm at the tth iteration, and T max is the maximum number of iterations given.

[0089] Specifically, the back propagation algorithm is used to further classify the kidney diseases identified by the kidney imaging intelligent diagnosis module to ensure that patients using the system can obtain diagnosis and treatment plans through self-examination of kidney images. The specific steps are as follows:

[0090] First, build an M-layer neural network. For the output of the m+1th layer network, we have:

[0091] a m+1 =f m+1 (W m+1 a m +b m+1 ),m=0,1,…,M-1

[0092] Among them, M is the number of network layers, a is the output of the neural network, that is, the category of kidney disease, and there is a 0 =p and a=a M , a m+1 is the output of the m+1th layer of the neural network, a m is the output of the mth layer of the neural network, f m+1 is the activation function of the m+1th layer of the neural network, W m+1 is the weight of the m+1th layer of the neural network, b m+1 is the bias of the m+1th layer of the neural network. Assuming that for each type of kidney disease feature input p, there is a corresponding kidney disease type output t, the following corresponding relationship can be established:

[0093] {p1,t1},{p2,t2},…,{p q ,t q},…,{p Q ,t Q}

[0094] Among them, p1 is the input of the first kidney image feature, t1 is the expected output corresponding to the first classification, p2 is the input of the second kidney image feature, t2 is the expected output corresponding to the second classification, and p q is the input of the qth kidney image feature, t q is the expected output corresponding to the qth classification, p Q is the input of the Qth kidney image feature, t Qis the expected output corresponding to the Qth classification, and the algorithm will adjust the network parameters to minimize the mean square error: F(x) = E[(ta) 2 ]=E[(ta) T (ta)], we can let e = ta represent the error between the expected output and the predicted output, where F represents the mean square error, x represents the independent variable of F, in which case the independent variables include t and a, and E represents the mathematical expectation, (ta) T Denote the transpose of the error, and use Substitute F(x) to approximate the mean square error: Where k is the kth iteration, and the steepest descent method for approximate mean square error is: Where α is the learning rate, represents the weight of the i-th element to the j-th element in the k+1-th iteration of the m-th layer neural network, It represents the weight of the i-th element to the j-th element in the k-th iteration of the m-th layer neural network, and the chain rule is defined as:

[0095]

[0096] If f(n)=e^n and n=2w, then f(n(w))=e^{2w}, we have Then using the chain rule and It can be written as: For the net input n of the mth layer, we have: Among them, s represents sensitivity, s m-1 represents the sensitivity of the m-1th layer of the neural network, represents the net input of the i-th element in an m-layer neural network, represents the output of the jth element in the m-1 layer neural network, represents the bias of the i-th element in an m-layer neural network, so: If defined: but and It can be simplified to: Using the approximate steepest descent method, it can be expressed as:

[0097]

[0098]

[0099] in, represents the bias of the i-th element in the k+1th iteration of the m-layer neural network, then the matrix form can be expressed as:

[0100] W m(k+1)=w m (k)-αs m (a m-1 ) T

[0101] b m (k+1)=b m (k)-αs m

[0102] Among them, W m (k+1) represents w m The matrix form of (k+1) is the weight set after the k+1th iteration of the m-layer neural network. Here s m have:

[0103]

[0104] in, It represents the partial derivative of the approximate error with respect to the net input of the mth layer, It represents the partial derivative of the approximate error with respect to the first net input of the mth layer. It represents the partial derivative of the approximate error with respect to the second net input of the mth layer, represents the approximation error for the sth layer of the mth layer m Find the partial derivative of the net input and record the Jacobian matrix as:

[0105]

[0106] in, represents the partial derivative of the net input of the m+1th layer of the neural network to the net input of the mth layer of the neural network, represents the partial derivative of the first net input of the m+1th layer of the neural network to the first net input of the mth layer of the neural network, represents the partial derivative of the first net input of the m+1th layer of the neural network to the second net input of the mth layer of the neural network, It represents the first net input of the m+1th layer of neural network to the sth layer of the mth layer of neural network. m The partial derivative of the net input, represents the partial derivative of the second net input of the m+1th layer of the neural network to the first net input of the mth layer of the neural network, represents the partial derivative of the second net input of the m+1th layer of the neural network to the second net input of the mth layer of the neural network, It represents the second net input of the m+1th layer of neural network to the sth layer of the mth layer of neural network m The partial derivative of the net input, represents the sth layer of the m+1th layer of the neural network m+1 The partial derivative of the net input to the net input of the mth layer of the neural network, represents the sth layer of the m+1th layer of the neural network m+1 The partial derivative of the net input to the first net input of the m-th layer of the neural network, represents the sth layer of the m+1th layer of the neural network m+1 The net input to the sth layer of the mth neural network m The partial derivative of the net input, represents the sth layer of the m+1th layer of the neural network m+1 The net input to the sth layer of the mth neural network m The partial derivative of the net input, considering the i, j elements of the matrix: in: Then the Jacobian matrix can be written as: in:

[0107] F m (n m ) indicates that the net input of the m-th layer neural network is n m The objective function when Indicates that the net input of the m-th layer neural network is The objective function when Indicates that the net input of the m-th layer neural network is The objective function when Indicates that the net input of the m-th layer neural network is The objective function when , the recursive relation of sensitivity is written by the chain rule in matrix form:

[0108] Specifically, the back-propagation algorithm and chain rule are used to obtain the optimal training weights, and the patient's disease information is input into the trained machine learning model, which can automatically detect whether the patient suffers from calcifications, renal calyx and pelvis density, renal medulla, cysts and cancer thrombi.

[0109] Specifically, for the established machine learning model, automatic detection of diseases in kidney images that need to be identified is achieved, and the trained kidney images with kidney disease diagnosis labels are saved in the kidney image database, and a model for automatic detection of diseases in kidney images that can be identified is generated for the convenience of direct use next time without the need for re-training. The kidney disease classification module further classifies the kidney diseases identified by the kidney image intelligent diagnosis module to ensure that patients using the system can obtain diagnosis and treatment plans by self-examination of kidney images.

[0110] Specifically, when the particle swarm algorithm is applied to determine the optimal combination threshold of the maximum entropy multi-threshold segmentation method, the particle swarm algorithm is prone to poor diversity of the particle swarm, resulting in a "precocious" phenomenon, or while strengthening the diversity of the particle swarm, slowing down the convergence speed of the particle swarm, thereby wasting resources. That is, in the particle swarm algorithm, increasing the diversity of particles and improving the convergence speed of particles are contradictory, and it is difficult to balance the two. Considering that in the particle swarm algorithm, maintaining a large value of the cognitive learning factor helps particles to search over a large range, thereby improving the diversity of the particle swarm, but will lead to a slow convergence speed of the algorithm, while maintaining a large value of the social learning factor helps to learn social experience, so that the search results quickly converge to the vicinity of the optimal solution, thereby improving the convergence speed of the algorithm, but the search space ability is weak. Therefore, this embodiment balances the global exploration and local mining of the particle swarm algorithm by using adaptively changing cognitive learning factor and social learning factor values, thereby playing a role in improving the diversity of the particle swarm and improving the convergence speed. In the iterative update formula of particles, the cognitive part reflects the self-learning ability of particles in exploration, reflects the tendency of particles to follow their best characteristics, and the social part reflects the ability of particles to learn from the outside world, and reflects the group behavior of collaborative cooperation and knowledge sharing among particles. The cognitive learning factor and the social learning factor respectively determine the contribution of the cognitive part and the social part in the iterative update process of particles. Considering that when the particles in the particle swarm are more uniform in optimizing the current solution space, the diversity of the particle swarm is better at this time, and there is a greater probability of finding the global optimal solution, therefore, this embodiment defines the optimization uniformity to measure the uniformity of the particle's optimization of the solution space in which it is located. When measuring the uniformity of the particle's optimization of the solution space in which it is located, this article not only measures the uniformity of the optimization result of the particle's current solution space through the local optimization uniformity in the optimization uniformity, but also measures the uniformity of the optimization result of the particle's historical optimal solution in the optimization uniformity. Therefore, when the value of the particle's cognitive learning factor is increased according to the particle's optimization uniformity, the particle's cognitive part can better increase the diversity of the particle. In a particle swarm, when the optimization uniformity of an individual particle is smaller or larger than the global optimization uniformity, it indicates that the particle has not fully learned the solution space in which it is located in the current exploration compared with the global one. Therefore, the cognitive learning factor of the particle is increased, and the social learning factor of the particle is reduced, that is, the contribution of its cognitive part in the iterative update process is strengthened, thereby improving the diversity of the algorithm. When the optimization uniformity of an individual particle is more moderate than the global optimization uniformity, it indicates that the particle has fully learned the solution space in which it is located in the current exploration compared with the global one. Therefore, the cognitive learning factor of the particle is reduced, and the social learning factor of the particle is increased, that is, the contribution of its social part in the iterative update process is strengthened, thereby improving the convergence speed of the algorithm.That is, by adjusting the cognitive learning factor and social learning factor of the optimization uniformity, the convergence speed of the particle swarm algorithm can be improved while ensuring that the diversity of the particle swarm in the current solution space is good, thereby improving the optimization accuracy of the particle swarm.

[0111] The beneficial effects created by the present invention are as follows: applying artificial intelligence technology to kidney image analysis, establishing a tumor diagnosis model corresponding to each human body part based on kidney images, being able to effectively diagnose tumors in various parts of the human body, realizing automatic analysis of kidney images and assisting doctors in making medical tumor diagnosis; adopting the maximum entropy multi-threshold segmentation method to perform image segmentation on the kidney image to be diagnosed, thereby obtaining the human body part image contained in the kidney image to be diagnosed, and adopting the particle swarm algorithm to determine the optimal combination threshold of the maximum entropy multi-threshold segmentation method, thereby improving the accuracy of kidney image segmentation and laying a foundation for subsequent tumor diagnosis; setting the particle swarm algorithm to use adaptively changing cognitive learning factors and social learning factors to adjust the iterative update mode of the particle swarm, thereby balancing the global exploration and local mining of the particle swarm algorithm, improving the optimization accuracy and convergence speed of the particle swarm algorithm, and making the optimal combination threshold determined by the particle swarm algorithm have higher accuracy.

[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, rather than to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, 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 essence and scope of the technical solution of the present invention.

Claims

1. A system for automatic detection of kidney diseases based on machine learning, including a kidney image collection module, a kidney image management module, a kidney image intelligent diagnosis module, and a kidney disease classification module; Kidney image collection module: used to collect various types of kidney disease images of patients. The sources of kidney images include site pathogenicity evaluation, disease database and kidney images of in-hospital patients; Renal image management module: including a renal image pre-storage unit and a renal image update unit. The renal image pre-storage unit is used to pre-store various types of renal images collected by the renal image collection module. The renal image update unit is used to screen renal images with renal disease diagnosis labels, clean images with excessive noise, default or irrelevant images, and then update the renal image training set. The renal image intelligent diagnosis module includes a renal image processing unit and a renal image database. The renal image processing unit extracts the disease characteristics of the renal image by analyzing the renal image with the renal disease diagnosis label and the plain scan image. The renal image updated by the renal image updating unit is used as a training set to establish a machine learning model to realize automatic disease detection of the renal image that needs to be identified. The trained renal image with the renal disease diagnosis label is saved to the renal image database to generate a model for automatic disease detection of renal images that can be identified so that it can be used directly next time without retraining. The kidney disease classification module further classifies the kidney diseases identified by the kidney imaging intelligent diagnosis module to ensure that patients using the system can obtain diagnosis and treatment plans through self-examination of kidney images; The kidney image processing unit uses the maximum entropy multi-threshold segmentation method to perform image segmentation on the kidney image to be diagnosed, and uses the particle swarm algorithm to determine the optimal combination threshold of the maximum entropy multi-threshold segmentation method used in the kidney image processing unit; in the particle swarm algorithm, the particle swarm uses the following method to iteratively update the particle solution: V i (t+1)=ωV i (t)+c i1 (t)r1(Pbest i (t)-X i (t))+c i2 (t)r2(Gbest(t)-X i (t)) X i (t+1)=X i (t)+V i (t+1) Where, X i (t+1) and V i (t+1) respectively represent the solution and velocity of particle i at the (t+1)th iteration, X i (t) and V i (t) represent the solution and velocity of particle i at the tth iteration, ω represents the inertia weight factor, Pbest i (t) represents the historical optimal solution of particle i at the tth iteration, Gbest(t) represents the global optimal solution of the particle swarm at the tth iteration, r1 and r2 represent randomly generated random numbers between 0 and 1, and c i1 (t) represents the cognitive learning factor of particle i at the tth iteration, c i2 (t) represents the social learning factor of particle i at the tth iteration; Set the particle swarm to adjust the cognitive learning factor c of particle i at the tth iteration in the following way i1 (t) and social learning factor c i2 The value of (t): Assume Ω i (t) represents the local neighborhood of particle i at the tth iteration, and Ω i (t) is the solution to X i (t) is the center and R(t) is the radius. R(t) is the local neighborhood radius of the particle swarm at the tth iteration. The value of R(t) is set as: in, $R_i(t)$ represents the local neighborhood radius of particle $i$ at the $t$-th iteration, and represents the solution of the $l$-th nearest particle to the solution $X$ i (t) in the particle swarm at the $t$-th iteration, $c$ represents a given positive integer, and $c < N$, where $N$ represents the number of particles in the particle swarm; $\Omega$ i,best (t) represents the historical best neighborhood of particle $i$ at the $t$-th iteration, and $\Omega$ i,best (t) is a circular region centered at the solution $Pbest$ i (t) with a radius of $r(t)$, where $r(t)$ is the historical best neighborhood radius of the particle swarm at the $t$-th iteration, and where $r_i(t)$ represents the historical best neighborhood radius of particle $i$ at the $t$-th iteration, and In the formula, represents the solution of the i -th nearest particle to the solution $Pbest$ (t) in the particle swarm at the $t$-th iteration; Define the function q i (t) is used to detect the optimization uniformity of particle i at the tth iteration, and set q i The calculation formula of (t) is: In the formula, q i (t) represents the optimization uniformity of particle i at the tth iteration, J i (t) represents the local optimization uniformity of particle i at the tth iteration, S i (t) represents the historical optimization uniformity of particle i at the tth iteration, J i (t) and S i The values ​​of (t) are: Where, X j (t) represents the solution of particle j at the tth iteration, X b (t) represents the solution of particle b at the tth iteration, and X j (t)≠X b (t), X k (t) represents the solution of particle k at the tth iteration, X o (t) represents the solution of particle 0 at the tth iteration, and X k (t)≠X o (t), λ i，j (t) is the optimal judgment function between particles i and j at the tth iteration, let f i (t) and f j (t) represent the fitness function values ​​of particle i and particle j at the tth iteration respectively, then λ i，b (t) is the optimal judgment function between particle i and particle b at the tth iteration, let f b (t) represents the fitness function value of particle b at the tth iteration, then m i，best (t) represents the historical optimal neighborhood Ω i，best The number of particles in (t); When i (t) <H q1 (t) or q i (t)>H q2 (t), then c i1 (t) and c i2 The values ​​of (t) are adjusted to: When H q1 (t)≤q i (t)≤H q2 (t), then c i1 (t) and c i2 The values ​​of (t) are adjusted to: In the above formula, α and β are two constants used to control c i1 (t) and c i2 (t) range, H q1 (t) is the lower limit of the global optimization uniformity of the particle swarm at the tth iteration, and H q2 (t) is the upper limit of the global optimization uniformity of the particle swarm at the tth iteration, and in, represents the mean value of the optimization uniformity of the particles in the particle swarm at the tth iteration, and T max is the maximum number of iterations given.

2. The system for automatic detection of kidney disease based on machine learning according to claim 1, characterized in that: The renal image collection module collects data from the sources of site pathogenicity assessment, disease database, and renal images of in-hospital patients, and aggregates them into a SQL database for online management.

3. The system for automatic detection of kidney disease based on machine learning according to claim 1, characterized in that: The renal imaging update unit screens renal images with renal disease diagnostic labels, including calcifications, calyx and pelvic density, renal medulla, cysts, and cancer thrombi.

4. The system for automatic detection of kidney disease based on machine learning according to claim 1, characterized in that: The back propagation algorithm is used to further classify the kidney diseases identified by the kidney imaging intelligent diagnosis module to ensure that patients using the system can obtain diagnosis and treatment plans through self-examination of kidney images. The specific steps are as follows: First, build an M-layer neural network. For the output of the m+1th layer network, we have: a m+1 =f m+1 (W m+1 a m +b m+1 ),m=0,1,…,M-1 Among them, M is the number of network layers, a is the output of the neural network, that is, the category of kidney disease, and there is a 0 =p and a=a M , a m+1 is the output of the m+1th layer of the neural network, a m is the output of the mth layer of the neural network, f m+1 is the activation function of the m+1th layer of the neural network, W m+1 is the weight of the m+1th layer of the neural network, b m+1 is the bias of the m+1th layer of the neural network. Assuming that for each type of kidney disease feature input p, there is a corresponding kidney disease type output t, the following corresponding relationship is established: {p1, t1}, {p2, t2}, ..., {p q , t q }, ..., {p Q , t Q }, where p1 is the input of the first kidney image feature, t1 is the expected output corresponding to the first classification, p2 is the input of the second kidney image feature, t2 is the expected output corresponding to the second classification, and p q is the input of the qth kidney image feature, t q is the expected output corresponding to the qth classification, p Q is the input of the Qth kidney image feature, t Q is the expected output corresponding to the Qth classification, and the algorithm will adjust the network parameters to minimize the mean square error: F(x) = E[(ta) 2 ]=E[(ta) T (ta)], let e = ta represent the error between the expected output and the output result of the neural network, where F represents the mean square error, x represents the independent variable of F, E represents the mathematical expectation, (ta) T Denote the transpose of the error, and use Substitute F(x) to approximate the mean square error: Where k is the kth iteration, and the steepest descent method for approximate mean square error is: Where α is the learning rate, represents the weight of the i-th element to the j-th element in the k+1-th iteration of the m-th layer neural network, represents the weight of the i-th element to the j-th element in the k-th iteration of the m-th layer neural network, represents the bias of the i-th element at the k+1-th iteration of the m-th layer neural network, Represents the bias of the i-th element at the k-th iteration of the m-th layer neural network.

5. According to the system for automatic detection of kidney disease based on machine learning in claim 4, the chain rule is defined as: If f(n)=e^n and n=2w, then f(n(w))=e^{2w}, we have Then using the chain rule and Written as: For the net input n of the mth layer, we have: in, s represents sensitivity, represents the weight of the i-th element to the j-th element in the m-th layer of the neural network, s m-1 represents the sensitivity of the m-1th layer of the neural network, represents the net input of the i-th element in an m-layer neural network, represents the output of the jth element in the m-1 layer neural network, represents the bias of the i-th element in an m-layer neural network, so: If defined: but and Simplified to: Using the approximate steepest descent method, it can be expressed as: in, represents the bias of the i-th element in the m-layer neural network, Λ represents the steepest descent operator, then the matrix form is expressed as: W m (k+1)=w m (k)-As m (a m-1 ) T b m (k+1)=b m (k)-As m Among them, W m (k+1) represents w m The matrix form of (k+1) is the weight set after the k+1th iteration of the m-layer neural network. Here s m have: in, It represents the partial derivative of the approximate error with respect to the net input of the mth layer, It represents the partial derivative of the approximate error with respect to the first net input of the mth layer. It represents the partial derivative of the approximate error with respect to the second net input of the mth layer, represents the approximation error for the sth layer of the mth layer m Find the partial derivative of the net input and record the Jacobian matrix as: in, represents the partial derivative of the net input of the m+1th layer of the neural network to the net input of the mth layer of the neural network, represents the partial derivative of the first net input of the m+1th layer of the neural network to the first net input of the mth layer of the neural network, represents the partial derivative of the first net input of the m+1th layer of the neural network to the second net input of the mth layer of the neural network, It represents the first net input of the m+1th layer of neural network to the sth layer of the mth layer of neural network. m The partial derivative of the net input, represents the partial derivative of the second net input of the m+1th layer of the neural network to the first net input of the mth layer of the neural network, represents the partial derivative of the second net input of the m+1th layer of the neural network to the second net input of the mth layer of the neural network, It represents the second net input of the m+1th layer of neural network to the sth layer of the mth layer of neural network m The partial derivative of the net input, represents the sth layer of the m+1th neural network m+1 The partial derivative of the net input to the first net input of the m-th layer neural network, represents the sth layer of the m+1th neural network m+1 The partial derivative of the net input to the second net input of the m-th layer of the neural network, represents the sth layer of the m+1th neural network m+1 The net input to the sth layer of the mth neural network m The partial derivative of the net input, considering the i, j elements of the matrix: in: Then the Jacobian matrix is ​​written as: in: F m (n m ) indicates that the net input of the m-th layer neural network is n m The objective function when Indicates that the net input of the m-th layer neural network is The objective function when Indicates that the net input of the m-th layer neural network is The objective function when Indicates that the net input of the m-th layer neural network is The objective function when , the recursive relation of sensitivity is written by the chain rule in matrix form:

6. The system for automatic detection of kidney disease based on machine learning according to claim 5, characterized in that: The back-propagation algorithm and chain rule are used to obtain the optimal training weights, and the patient's disease information is input into the trained machine learning model to automatically detect whether the patient has calcifications, renal calyx and pelvis density, renal medulla, cysts and cancer thrombi.

7. The system for automatic detection of kidney disease based on machine learning according to claim 1, characterized in that: For the established machine learning model, automatic disease detection of kidney images that need to be identified is realized, and the trained kidney images with kidney disease diagnosis labels are saved in the kidney image database, and a model for automatic disease detection of kidney images that can be identified is generated so that it can be used directly next time without re-training. The kidney disease classification module further classifies the kidney diseases identified by the kidney image intelligent diagnosis module to ensure that patients using the system can obtain diagnosis and treatment plans through self-examination of kidney images.

Images