A method for measuring the differential skin impedance of acupoint areas during acupuncture and rapid imaging based on a neural network model

Abstract

The present invention discloses a method for measuring the differential skin impedance and rapid imaging of acupoint regions during acupuncture based on a neural network model. The steps are as follows: First, a corresponding mathematical model is established for the boundary value problem of skin impedance measurement in acupoint regions; then, according to the above-mentioned mathematical model of the boundary value problem, the forward problem is solved multiple times by the finite element method to establish a neural network training set; next, a neural network model is constructed, and the model parameters of the neural network are solved by the stochastic gradient descent method; finally, the measured data is substituted into the trained neural network model for rapid imaging. While maintaining rapid imaging, the present invention overcomes the deficiencies of low spatial resolution and poor imaging accuracy of the reconstructed image caused by the severe ill-conditioning of the inversion matrix in the linear image reconstruction algorithm, and realizes the rapid imaging of the dynamic changes in skin impedance of acupoint regions during acupuncture.

Description

Technical Field

[0001] The present invention belongs to the fields of traditional Chinese medicine acupoints and medical imaging, and particularly relates to a method for measuring the differential skin impedance of acupoint areas during acupuncture and rapid imaging based on a neural network model. Background Art

[0002] Acupuncture is an important part of traditional medicine. Among them, acupuncture manipulation is the key technology of acupuncture treatment. All kinds of acupuncture manipulations, acupuncture frequencies, and acupuncture intensities have an impact on the human body, directly determining the prognosis of diseases. The quantitative research on acupuncture manipulation has greatly promoted the development of acupuncture clinical practice.

[0003] The measurement of skin impedance in acupoint areas is the earliest scientific research direction of acupuncture. This aspect of research is expected to reveal the action laws and mechanisms of traditional acupuncture and quantify the treatment effects of acupuncture. Among them, the four-electrode resistance detector can measure relatively stable resistance values and has become the main method for detecting the electrical properties of biological tissues. However, the method of single-measurement electrode moving and scanning requires manual positioning, which not only has low detection efficiency but also is greatly affected by subjective factors, resulting in insufficient repeatability and reliability of the measurement results. With the in-depth study of the scientific research of acupuncture, it has been recognized that during acupuncture with different manipulations (such as retaining needles, reinforcing acupuncture, and reducing acupuncture), the electrical properties of acupoint points will change differently. This discovery is of great significance for exploring the action laws and treatment mechanisms of different acupuncture methods in traditional acupuncture. However, the existing single-acupoint electrical property measuring instrument can only obtain data of a single measurement point each time, which reflects the average apparent resistivity of the skin in the acupoint area, has no spatial resolution, and is generally a static measurement, making it difficult to track the dynamic changes of skin impedance in the acupoint area during acupuncture in real time.

[0004] In recent years, the measurement of acupoint impedance has developed towards multi-channel and visualization. Ye Xiaohong et al. developed a 64-channel array measurement system based on the four-electrode measurement method and displayed the measured impedance values on each measurement channel in the form of a grayscale image. This measurement method expands the detection of skin impedance from "point" to "surface", but it is still an expansion of the traditional four-electrode method in terms of quantity and does not involve an imaging method.

[0005] As a highly non-linear model, artificial neural network provides a new idea for solving inverse problems. Currently, common neural network models mostly rely on a data-driven approach when solving inverse problems. The network model learns a large number of input-output sample pairs, iteratively updates the network weight parameters, and finally obtains the function mapping relationship between the input and the output. This data-driven neural network inverse problem solving method has achieved good results in some practical problems. Summary of the Invention

[0006] The purpose of this method is to overcome the deficiencies of low spatial resolution and poor imaging accuracy of the reconstructed image caused by the severe ill-conditioning of the inversion matrix in the linear image reconstruction algorithm while maintaining fast imaging, and to achieve fast imaging of the dynamic changes in the skin impedance of the acupoint area during acupuncture.

[0007] The present invention proposes a method for measuring the differential skin impedance and fast imaging of the acupoint area during acupuncture based on a neural network model. The method for measuring the differential skin impedance and fast imaging of the acupoint area during acupuncture based on the neural network model can be used for measuring the differential changes in the skin impedance of the three-dimensional area of the acupoint during acupuncture, and through the trained neural network model for fast prediction, to achieve fast imaging of the dynamic changes in the skin impedance of the acupoint area during acupuncture.

[0008] To achieve the above object, the technical solution adopted by the present invention is:

[0009] A method for measuring the differential skin impedance and fast imaging of the acupoint area during acupuncture based on a neural network model, using the current applied to the treatment needle during acupuncture as the excitation current required for detection, taking the sensing needle inserted into the surface skin as the detection electrode array, the sensing needle is arranged with detection points at different height layers, the number of detection points arranged on each sensing needle is the same, and the detection points at the same height layer are kept on the same height plane during detection; the detection points are connected to a potential detector, and the voltage information of the skin at the detection points is obtained through the potential detector; the fast imaging method is characterized by including the following steps:

[0010] Step 1, establish a corresponding mathematical model for the boundary value problem of measuring the skin impedance of the acupoint area.

[0011] Step 2, solve the forward problem by the finite element method to establish a neural network training set.

[0012] Step 3, construct a neural network model.

[0013] Step 4, solve the model parameters of the neural network by the stochastic gradient descent method.

[0014] Step 5, substitute the measured data into the trained neural network model for fast imaging.

[0015] Further, the said Step 1 includes:

[0016] Assume the measured three-dimensional acupoint area V, boundary S, skin impedance ρ, during acupuncture, the excitation current I is applied to the treatment needle and injected from two points A and B on the treatment needle, and the potential distribution φ(r) in the area V satisfies the following Laplace equation:

[0017]

[0018] Among them, is the Nabla operator, δ is the Dirac function, ρ is the skin impedance, φ(r) is the potential distribution within the measured three-dimensional acupoint region V, I is the excitation current, r represents the three-dimensional space coordinate, r A and r B are the coordinates of points A and B where the excitation current is injected on the treatment needle respectively.

[0019] Since the skin is a continuous weakly conductive medium, the potential satisfies the continuity boundary condition. An infinite far boundary condition is constructed, and it is assumed that on the infinite far boundary S ∞ the potential is zero, or the normal component of the potential is zero, that is:

[0020]

[0021] Or

[0022]

[0023] Furthermore, the second step includes:

[0024] According to the mathematical model of the above boundary value problem, the forward problem is solved multiple times by the finite element method; it is assumed that the skin impedance within the measured two-dimensional cross-section is the background impedance ρ 0 which is uniformly distributed in the non-needle-inserted state. At this time, the voltage value measured by the detection points on the sensing needle is V 0 ; the detection points at the same height layer on the sensing needle form a detection cross-section. After the needle is inserted, as the insertion depth h of the treatment needle changes, different small perturbations are generated in the skin impedance. This perturbation is approximated as a circular anomaly with a uniform impedance distributed near the acupoint on each detection cross-section. The center coordinates of this anomaly are (m, n), the diameter is d, and the impedance is ρ. At this time, the voltage value measured by the detection points on the sensing needle is V, and the change value of the measured voltage relative to the background impedance is △V = V - V 0 ; by solving the forward problem multiple times by the finite element method, a data set composed of the center coordinate position (m, n) of the anomaly within the detection cross-section, the diameter d of the circle, the impedance value ρ of the anomaly, etc., and the corresponding detection cross-section height l, the insertion depth h of the treatment needle, the three-dimensional space coordinates (i, j, k) of the sensing needle, and the measured voltage change value △V is obtained as the training set of the neural network.

[0025] Furthermore, the third step includes:

[0026] The voltage change △V between the measurement electrodes and the relative change △ρ of the skin impedance in the acupoint region satisfy a non-linear relationship. Let this non-linear relationship be:

[0027] △V = f(ρ, r, d, h)

[0028] According to the universal approximation theorem, f can be solved by a non-linear neural network; the entire neural network is regarded as a composite function f(x; W, b), where W and b are the model parameters of the neural network, representing the connection weights and biases of all layers in the network respectively; the neural network model includes an input layer, two hidden layers and an output layer, and its input signal has 6 nodes, x 1 -x 6 They are respectively the height l of the detection cross-section to be inverted, the penetration depth h of the treatment needle, the three-dimensional spatial coordinates (i, j, k) of the sensing needle, and the change value △V of the voltage measured by the sensing needle during acupuncture. The output of the last layer of the neural network is the output y of this function, and the neural network output has 4 nodes, y 1 -y 4 They are respectively the coordinate positions (m, n) of the center of the abnormal body in the detection cross-section, the diameter d of the circle, and the impedance value ρ of the abnormal body.

[0029] Further, the step four includes:

[0030] Given the initial values of the model parameters W and b of the neural network, extract the input quantity X from the training set obtained in the step two. The data dimension k of X is the number of times of solving the forward problem in the step two. Each dimension of the input quantity contains 6 nodes, which are respectively the height l of the detection cross-section to be inverted, the penetration depth h of the treatment needle, the three-dimensional spatial coordinates (i, j, k) of the sensing needle, and the change value △V of the voltage measured by the sensing needle during acupuncture. The coordinate positions (m, n) of the center of the abnormal body in the detection cross-section, the diameter d of the circle, and the impedance value ρ of the abnormal body corresponding to each dimension of the input quantity in the step two are used as the correct solution t. After the input quantity X is input into the neural network, the coordinate positions (m', n') of the center of the abnormal body in the detection cross-section, the diameter d' of the circle, and the impedance value ρ' of the abnormal body calculated by the neural network model are used as the predicted value y of the network. By optimizing the distance between the predicted value y of the network and the correct solution t, that is, the loss function, the model parameters of the neural network can be solved;

[0031] Use the mean square error as the loss function to measure the distance between the network predicted value y and the correct solution t:

[0032]

[0033] where, y k represents the output of the neural network, t k represents the correct solution, and k represents the dimension of the data;

[0034] The neural network model parameter optimization algorithm adopts the stochastic gradient descent method. Using the gradient of the parameters as a clue, update and iterate the values of the network model parameters W and b along the gradient direction to complete the training of the neural network model.

[0035] Further, in the fifth step, according to the height l of the detection cross-section, the penetration depth h of the treatment needle, the coordinates (i, j, k) of each sensing needle, and the change value △V of the voltage at the detection point of the sensing needle measured during the actual acupuncture process, substituting them into the neural network model trained in the fourth step, the predicted values of the center coordinate position (m, n) of the abnormal body, the diameter d of the circle, and the impedance value ρ of the abnormal body in the detection cross-section can be quickly obtained from the output layer of the network model, realizing the rapid imaging of the dynamic change of the skin impedance in the acupoint area.

[0036] Beneficial effects:

[0037] Based on the principle of electrical impedance tomography of the skin, the biggest difference between the present invention and the traditional electrical impedance tomography method is that the present invention ingeniously uses the current applied to the treatment needle during the acupuncture process as the excitation current required for detection, uses the sensing needles inserted into the surface skin as the detection electrode array, and obtains the voltage information of the three-dimensional cortical tissue at different depths according to the detection points at different heights on the sensing needles. Through the differential measurement method and combined with the neural network model, the rapid imaging of the dynamic change of the skin impedance in the acupoint area during the acupuncture process is realized. Compared with the linear image reconstruction algorithms of electrical impedance tomography such as equipotential line filtered backprojection, while maintaining rapid imaging, the present invention overcomes the deficiencies of low spatial resolution and poor imaging accuracy of the reconstructed image caused by the severe ill-condition of the inversion matrix in the linear image reconstruction algorithm. Brief description of the drawings

[0038] Figure 1 Schematic diagram of the sensor electrode array of the present invention;

[0039] Figure 2 Neural network model structure constructed by the present invention;

[0040] Figure 3 Flow chart of neural network model parameter calculation of the present invention. Detailed implementation manners

[0041] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0042] Such as Figure 1As shown in the figure, the measurement method of acupoint impedance in the present invention utilizes the current applied to the treatment needle 1 during the acupuncture process as the excitation current required for detection, and uses the sensing needle 2 inserted into the surface skin as the detection electrode array. The sensing needle 2 is provided with detection points at different height layers, and the number of detection points arranged on each sensing needle 2 is the same. During detection, the detection points at the same height layer are kept on the same height plane. The detection points are connected to the potential detector 3, and the voltage information of the skin at the detection points is obtained through the potential detector 3. The advantages of this measurement method are as follows: (1) During the detection process, the current on the treatment needle is used as the excitation current, and there is no additional introduced other excitation current signals, which will not cause electrical signal interference to the acupuncture process; (2) The detection object is the change amount of skin impedance, and this differential measurement method is beneficial to reducing the influence of individual differences on the analysis of general laws.

[0043] Although this measurement method has its unique advantages, due to the small amount of effective data in a single measurement process, when using a linear approximation real-time imaging algorithm, the ill-conditioning of its inversion matrix is very serious, resulting in low spatial resolution and poor image accuracy of the reconstructed image.

[0044] To overcome the above defects, the present invention proposes a method for differential measurement and rapid imaging of skin impedance in the acupoint area during acupuncture based on a neural network model, so as to realize rapid imaging of the dynamic change of skin impedance in the acupoint area during acupuncture. The main steps are as follows:

[0045] Step 1, for the boundary value problem of skin impedance measurement in the acupoint area, establish a corresponding mathematical model;

[0046] Assume that the measured three-dimensional acupoint area is V, the boundary is S, the skin impedance is ρ, during acupuncture, the excitation current I is applied to the treatment needle and injected from two points A and B on the treatment needle. The potential distribution φ(r) in the area V satisfies the following Laplace equation:

[0047]

[0048] Among them, is the Nabla operator, δ is the Dirac function, ρ is the skin impedance, φ(r) is the potential distribution in the measured three-dimensional acupoint area V, I is the excitation current, r represents the three-dimensional space coordinate, r A and r B are the coordinates of points A and B where the excitation current is injected on the treatment needle respectively.

[0049] Since the skin is a continuous weakly conductive medium, the potential satisfies the continuity boundary condition. An infinite far boundary condition is constructed, and it is assumed that the potential is zero or the normal component of the potential is zero on the infinite far boundary S ∞ That is:

[0050]

[0051] or

[0052]

[0053] Step 2: Solve the forward problem by the finite element method to establish a neural network training set;

[0054] According to the mathematical model of the above boundary value problem, solve the forward problem by the finite element method multiple times; assume that the skin impedance in the measured two-dimensional cross-section is the background impedance ρ with uniform distribution under the non-needling state 0 , and the voltage value measured by the detection points on the sensing needle at this time is V 0 ; the detection points at the same height layer on the sensing needle form a detection cross-section. After inserting the needle, as the penetration depth h of the treatment needle changes, different small perturbations are generated in the skin impedance. This perturbation is approximated as a circular anomaly with uniform impedance distributed near the acupoint in each detection cross-section. The center coordinates of this anomaly are (m, n), the diameter is d, and the impedance is ρ. At this time, the voltage value measured by the detection points on the sensing needle is V, and the change value of the measured voltage △V = V - V 0 ; solve the forward problem by the finite element method multiple times to obtain a data set composed of the center coordinate position (m, n) of the anomaly in the detection cross-section, the diameter d of the circle, the impedance value ρ of the anomaly, etc., and the corresponding detection cross-section height l, the penetration depth h of the treatment needle, the three-dimensional space coordinates (i, j, k) of the sensing needle, and the measured voltage change value △V, as the training set of the neural network.

[0055] Step 3: Construct a neural network model;

[0056] The voltage change △V between the measurement electrodes and the relative change △ρ of the skin impedance in the acupoint area satisfy a non-linear relationship. Let this non-linear relationship be:

[0057] △V = f(ρ, r, d, h)

[0058] According to the universal approximation theorem, f can be solved by a non-linear neural network; the entire neural network is regarded as a composite function f(x; W, b), where W and b are the model parameters of the neural network, representing the connection weights and biases of all layers in the network respectively; the neural network model includes an input layer, two hidden layers and an output layer. Its input signal has 6 nodes, x 1 -x 6 are respectively the detection cross-section height l to be inverted, the penetration depth h of the treatment needle, the three-dimensional space coordinates (i, j, k) of the sensing needle, and the change value △V of the voltage measured by this sensing needle during acupuncture. The output of the last layer of the neural network is the output y of this function. The neural network output has 4 nodes, y 1 -y 4They are the coordinate positions (m, n) of the center of the anomaly in the detection cross-section, the diameter d of the circle, and the impedance value ρ of the anomaly, respectively.

[0059] Step 4: Solve the model parameters of the neural network by the stochastic gradient descent method.

[0060] Given the initial values of the model parameters W and b of the neural network, extract the input quantity X from the training set obtained in Step 2. The data dimension k of X is the number of times of solving the forward problem in Step 2. Each dimension of the input quantity contains 6 nodes, which are the height l of the detection cross-section to be inverted, the penetration depth h of the treatment needle, the three-dimensional spatial coordinates (i, j, k) of the sensing needle, and the change value △V of the voltage measured by the sensing needle during the acupuncture process. The coordinate position (m, n) of the center of the anomaly, the diameter d of the circle, and the impedance value ρ of the anomaly in the detection cross-section corresponding to each dimension of the input quantity in Step 2 are used as the correct solution t. After the input quantity X is input into the neural network, the coordinate position (m’, n’) of the center of the anomaly, the diameter d’ of the circle, and the impedance value ρ' of the anomaly obtained through the calculation of the neural network model are used as the predicted value y of the network. By optimizing the distance between the predicted value y of the network and the correct solution t, that is, the loss function, the model parameters of the neural network can be solved.

[0061] Use the mean square error as the loss function to measure the distance between the network predicted value y and the correct solution t:

[0062]

[0063] where y k represents the output of the neural network, t k represents the correct solution, and k represents the dimension of the data;

[0064] The neural network model parameter optimization algorithm uses the stochastic gradient descent method. Taking the gradient of the parameters as a clue, update and iterate the values of the network model parameters W and b along the gradient direction to complete the training of the neural network model.

[0065] Step 5: Substitute the measured data into the trained neural network model for fast imaging.

[0066] According to the height l of the detection cross-section, the penetration depth h of the treatment needle, the coordinates (i, j, k) of each sensing needle, and the change value △V of the voltage measured at the sensing needle detection point during the actual acupuncture process, substitute them into the neural network model trained in Step 4. Then, the predicted values of the parameters such as the coordinate position (m, n) of the center of the anomaly, the diameter d of the circle, and the impedance value ρ of the anomaly in the detection cross-section can be quickly obtained from the output layer of the network model, realizing the fast imaging of the dynamic change of the skin impedance in the acupoint area.

[0067] Those skilled in the art can easily understand that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for measuring the differential skin impedance and rapid imaging of acupoint regions during acupuncture based on a neural network model. Using the current applied to the treatment needle during acupuncture as the excitation current required for detection, and the sensing needle inserted into the surface skin as the detection electrode array. The sensing needle is arranged with detection points at different height layers, and the number of detection points arranged on each sensing needle is the same. During detection, the detection points at the same height layer are maintained on the same height plane. The detection points are connected to a potential detector, and the voltage information of the skin at the detection points is obtained through the potential detector. The rapid imaging method Characterized in that It includes the following steps: Step 1, for the boundary value problem of skin impedance measurement in the acupoint region, establish a corresponding mathematical model, including: Assume that the three-dimensional acupoint region V to be measured has a boundary S and a skin impedance of , during the acupuncture process, an excitation current I is applied to the treatment needle and injected from two points A and B on the treatment needle. The potential distribution in the region V satisfies the following Laplace equation: Among them, is the Nabla operator, is the Dirac function, is the skin impedance, is the potential distribution in the measured three-dimensional acupoint region V, I is the excitation current, r represents the three-dimensional space coordinate, r A and r B are the coordinates of point A and point B where the excitation current is injected on the treatment needle, respectively; Since the skin is a continuous weakly conductive medium and the potential satisfies the continuity boundary condition, an infinite far boundary condition is constructed. Let the potential be zero or the normal component of the potential be zero on the infinite far boundary, that is: ​ Or ； Step 2, solve the forward problem by the finite element method and establish a neural network training set, including: According to the mathematical model of the above boundary value problem, the forward problem is solved multiple times by the finite element method; it is assumed that the skin impedance in the measured two-dimensional cross-section is the background impedance with uniform distribution under the state of no needle insertion. , and the voltage value measured by the detection points on the sensing needle at this time is ; the detection points on the same height layer of the sensing needle form a detection cross-section; after the needle is inserted, as the penetration depth h of the treatment needle changes, different small perturbations are generated in the skin impedance. This perturbation is approximated as a circular anomaly with uniform impedance distributed near the acupoint in each detection cross-section. The center coordinates of this anomaly are (m, n), the diameter is d, and the impedance is , and the voltage value measured by the detection points on the sensing needle at this time is , and the change value of the measured voltage relative to the background impedance; the forward problem is solved multiple times by the finite element method to obtain the center coordinate position (m, n) of the anomaly in the detection cross-section, the diameter d of the circle, and the impedance value of the anomaly, the corresponding detection cross-section height l, the penetration depth h of the treatment needle, the three-dimensional space coordinates (i, j, k) of the sensing needle, and the change value of the measured voltage to form a data set as the training set of the neural network; Step 3, construct a neural network model, including: Measure the voltage change between the electrodes and the relative change in skin impedance in the acupoint area satisfy a non-linear relationship. Let this non-linear relationship be: ； According to the universal approximation theorem, it can be solved by a non - linear neural network; the entire neural network is regarded as a composite function , where W and b are the model parameters of the neural network, representing the connection weights and biases of all layers in the network respectively; the neural network model includes an input layer, two hidden layers and an output layer, and its input signal has 6 nodes, x 1 - x 6 are respectively the height l of the detection cross - section to be inverted, the penetration depth h of the treatment needle, the three - dimensional space coordinates (i, j, k) of the sensing needle, and the change value of the voltage measured by the sensing needle during acupuncture , the output of the last layer of the neural network is the output y of the function, and the neural network output has 4 nodes, y 1 -y 4 are respectively the coordinate positions (m, n) of the center of the abnormal body in the detection cross - section, the diameter d of the circle, and the impedance value of the abnormal body ; Step 4, solve the model parameters of the neural network by the stochastic gradient descent method; Step 5, substitute the measured data into the trained neural network model for rapid imaging.

2. A method for measuring the differential skin impedance and rapid imaging of acupoint regions during acupuncture based on a neural network model according to claim 1, Characterized in that The said step 4 includes: Given the initial values of the model parameters W and b of the neural network, extract the input quantity X from the training set obtained in the second step. The data dimension k of X is the number of times of solving the forward problem in the second step. The input quantity of each dimension contains 6 nodes, namely the height l of the detection cross-section to be inverted, the penetration depth h of the treatment needle, the three-dimensional spatial coordinates (i, j, k) of the sensing needle, and the change value of the voltage measured by the sensing needle during the acupuncture process ; the center coordinate position (m,n), the diameter d of the circle, and the impedance value of the abnormal body within the detection cross-section corresponding to the input quantity of each dimension in the second step are used as the correct solution t; after the input quantity X is input into the neural network, the center coordinate position (m’,n’), the diameter d’ of the circle, and the impedance value of the abnormal body within the detection cross-section calculated by the neural network model are used as the predicted value y of the network; by optimizing the distance between the predicted value y of the network and the correct solution t, that is, the loss function, the model parameters of the neural network can be solved; Using the mean square error as the loss function to measure the distance between the network predicted value y and the correct solution t: ； Among them, y k represents the output of the neural network, t k represents the correct solution, and k represents the dimension of the data; The neural network model parameter optimization algorithm adopts the stochastic gradient descent method, using the gradient of the parameters as a clue, and updating and iterating the network model parameter values W and b along the gradient direction to complete the training of the neural network model.

3. A method for measuring the differential skin impedance and rapid imaging of acupoint regions during acupuncture based on a neural network model according to claim 2, Characterized in that In the fifth step, according to the height l of the detection cross-section, the penetration depth h of the treatment needle, the coordinates (i, j, k) of each sensing needle, and the change value of the voltage at the detection point of the sensing needle measured during the actual acupuncture process , substitute them into the neural network model trained in the fourth step, that is, quickly obtain the predicted values of the center coordinates (m, n) of the abnormal body in the detection cross-section, the diameter d of the circle, and the impedance value of the abnormal body from the output layer of the network model , so as to realize the rapid imaging of the dynamic change of skin impedance in the acupoint area.

Images