A quantitative intravascular optical coherence tomography method and system

Abstract

The present invention relates to a method for quantitative intravascular optical coherence tomography, comprising the following steps: performing optical simulation based on the IVOCT forward imaging model to obtain the theoretical values of the backscattered signals on the vascular cross-section corresponding to different optical attenuation coefficients; using the distribution map of the theoretical values of the backscattered signals on the vascular cross-section as the input of the sample, and using the corresponding distribution map of the optical attenuation coefficients as the expected output of the sample to construct a data set for quantitative analysis of IVOCT; building a RIM network model; training the RIM network model using the data set to obtain a trained RIM network model; and performing quantitative analysis using the trained RIM network model. The present invention first constructs a data set based on the IVOCT forward imaging model, then trains the RIM network model based on the data set, and performs quantitative analysis based on the trained RIM network model. The present invention realizes quantitative intravascular optical coherence tomography based on the IVOCT forward imaging model and the RIM network model.

Description

Technical Field

[0001] The present invention relates to the technical field of medical imaging, and particularly to a method and system for quantitative intravascular optical coherence tomography. Background Art

[0002] Intravascular Optical Coherence Tomography (IVOCT) is currently the interventional vascular imaging technology with the highest resolution, which can clearly display the inner wall of blood vessels and plays an important role in evaluating the degree of coronary artery lesions, accurately classifying atherosclerotic plaques, identifying vulnerable plaques, guiding stent implantation, and evaluating the effect of interventional treatment.

[0003] IVOCT is a pure optical imaging method, and the imaging depth of light in highly scattering tissues can usually only reach 1 mm. Therefore, its tissue penetration ability is weak and the imaging depth is shallow. In addition, in the early stage of the occurrence of lesions, the light scattering characteristics of normal and diseased tissues are not very different. Only based on the tomographic images of the tissue structure, it is impossible to accurately distinguish deeper tissues such as lipids, calcified plaques, macrophages, and necrotic nuclei under the fibrous cap. A comparison mechanism of physiological information other than the morphological structure is required to obtain the tissue optical characteristic parameter - light attenuation coefficient with clinical diagnostic value, that is, quantitative IVOCT (quantitative intravascular optical coherence tomography). Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for quantitative intravascular optical coherence tomography to achieve quantitative intravascular optical coherence tomography.

[0005] To achieve the above purpose, the present invention provides the following solutions:

[0006] A method for quantitative intravascular optical coherence tomography includes the following steps:

[0007] Based on the IVOCT forward imaging model, perform optical simulation to obtain the theoretical values of the backscattered signals on the cross-section of the blood vessel corresponding to different light attenuation coefficients, and construct a light attenuation coefficient distribution map and a distribution map of the theoretical values of the backscattered signals on the cross-section of the blood vessel corresponding to the light attenuation coefficient distribution map;

[0008] Use the distribution map of the theoretical values of the backscattered signals on the cross-section of the blood vessel as the input of the sample, and use the corresponding light attenuation coefficient distribution map as the expected output of the sample to construct a data set for quantitative analysis of IVOCT;

[0009] Build a RIM network model;

[0010] Use the data set to train the RIM network model to obtain a trained RIM network model;

[0011] The measurement value distribution map of the backscattering signal on the cross-section of the blood vessel to be analyzed is input into the trained RIM network model for quantitative analysis, and the light attenuation coefficient distribution map corresponding to the measurement value distribution map of the backscattering signal on the cross-section of the blood vessel to be analyzed is obtained.

[0012] Optionally, the optical simulation is performed based on the IVOCT forward imaging model to obtain the theoretical values of the backscattering signals on the cross-section of the blood vessel corresponding to different light attenuation coefficients, specifically including:

[0013] The Monte Carlo method is used to simulate the transmission of incident light in tissues with different light attenuation coefficients, and the photon weights and radial depths of the incident light in tissues with different light attenuation coefficients are obtained; the photon weights are the photon weights collected from the surface of the tissue, and the radial depth is the transmission depth of the photon in the tissue;

[0014] According to the photon weights and transmission depths of the incident light in tissues with different light attenuation coefficients, the theoretical values of the backscattering signals on the cross-section of the blood vessel corresponding to different light attenuation coefficients are calculated.

[0015] Optionally, according to the photon weights and transmission depths of the incident light in tissues with different light attenuation coefficients, the calculation formula for the theoretical values of the backscattering signals on the cross-section of the blood vessel corresponding to different light attenuation coefficients is:

[0016]

[0017] where I(r) represents the theoretical value of the backscattering signal at position r on the cross-section of the blood vessel, I0 is a constant characterizing the imaging system characteristics, N ph is the total number of photons, |r| is the radial depth of position r, λ c is the central wavelength of the incident light, L i is the optical path length of the i-th photon transmitted in the tissue, w i is the weight of the i-th photon, l coh is the coherence length of the detection light source; Δλ is the full width at half maximum of the light source spectrum.

[0018] Optionally, the RIM network model includes a plurality of RIM network structure units connected in cascade.

[0019] Optionally, the RIM network structure unit includes a first convolutional layer, a first gated recurrent unit, a second convolutional layer, a second gated recurrent unit, and a third convolutional layer connected in sequence;

[0020] The sizes of the filter kernels of the first convolutional layer and the second convolutional layer are 3×3×M; where M represents the number of feature maps input in the current layer;

[0021] The size of the filter kernel of the third convolutional layer is 1×1×M;

[0022] Both the first gated recurrent unit and the second gated recurrent unit include an update gate and a reset gate, and the gate functions of the update gate and the reset gate are both sigmoid activation functions.

[0023] Optionally, training the RIM network model using the dataset to obtain a trained RIM network model specifically includes:

[0024] Initialize the network parameters of the RIM network model, set the time step t = 0, divide the training set A from the dataset, and divide the training set A into multiple mini-batch training sets;

[0025] Set the number of times epoch to traverse all mini-batch training sets in the training set A to 0;

[0026] Set the index j of the mini-batch training set in the training set A to 1;

[0027] Input the Q samples in the j-th mini-batch training set A j in the training set A into the RIM network model for forward propagation;

[0028] Use the loss function to calculate the loss value of the forward propagation; the loss function is:

[0029]

[0030] In the formula, W and b are the weight parameter set and the bias parameter set of the RIM network model respectively, L(W,b) is the loss function with respect to W and b, m is the total number of pixels of the Q samples in A j the total number of pixels of the Q samples in A, T is the total number of time steps for training the network, μ t,i is the light attenuation coefficient distribution map obtained after the i-th sample in A j has been trained for t time steps, μ i is the target image of the i-th sample in A j ;

[0031] According to the loss value, calculate the gradient of the output layer parameters of the RIM network model using the following formula;

[0032]

[0033]

[0034] In the formula, W C and b C are the weight parameter matrix and the bias parameter vector of the C-th layer network respectively, W C and b CC in it represents the total number of network layers in the RIM network model, and Y C-1 is the output of the (C-1)-th layer network in the forward propagation. "⊙" represents the Hadamard product of matrices, σ'() is the first derivative of the ReLU activation function, the superscript T represents matrix transpose, and Z C is the linear activation response generated by the C-th layer network in the forward propagation: Z C = W C Y C-1 + b C ;

[0035] Calculate the error vector of each layer network of the RIM network model according to the following formula;

[0036]

[0037] In the formula, δ c represents the error vector of the c-th layer network of the RIM network model, W c+1 is the weight parameter matrix of the (c + 1)-th layer network, and Z c is the linear activation response generated by the c-th layer network in the forward propagation;

[0038] Calculate the gradient of each layer network of the RIM network model according to the error vector of each layer network of the RIM network model and the loss value by using the following formula:

[0039]

[0040]

[0041] In the formula, c = C, C - 1, C - 2,..., 2, and Y c-1 is the output of the (c - 1)-th layer network in the forward propagation, W c and b c are the weight parameter matrix and bias parameter vector of the c-th layer network respectively;

[0042] Update the weight parameter matrix and bias parameter vector of each layer network of the RIM network model by using the adaptive moment estimation algorithm according to the gradient of each layer network of the RIM network model by using the following formula;

[0043]

[0044] In the formula, W′ c and b′ c are the updated weight parameter matrix and bias parameter vector of the c-th layer network respectively, α represents the learning rate, ε is a constant used to maintain the stability of numerical calculation, and are the corrected Momentum exponentially weighted moving averages, and It is the corrected RMSprop exponentially weighted average, let k←k + 1:

[0045]

[0046] In the formula, k is the corrected number of times for accumulating the gradients of network parameters (the initial value is 0), V dW and V db are the Momentum exponentially weighted averages of dW and db respectively, S dW and S db are the RMSprop exponentially weighted averages of dW and db respectively, β1 and β2 are the exponential decay rates of the first - order moment estimate and the second - order moment estimate respectively, dW is the gradient matrix of W c and db is the gradient vector of b c :

[0047]

[0048] Let the value of j increase by 1, determine whether the formula j < P + 1 holds, and obtain the first judgment result;

[0049] If the first judgment result is yes, then return to the step "Input the Q samples in the j - th mini - batch training set A in the training set A j into the RIM network model for forward propagation";

[0050] If the first judgment result is no, then let the value of epoch increase by 1, determine whether the formula epoch < epoch max holds, and obtain the second judgment result; epoch max is the preset maximum value of epoch;

[0051] If the second judgment result is yes, then return to the step "Let the index j of the mini - batch training set in the training set A be equal to 1";

[0052] If the second judgment result is no, then let the value of t increase by 1, determine whether the formula t < T holds, and obtain the third judgment result;

[0053] If the third judgment result is yes, then return to the step "Let the number of times epoch for traversing all mini - batch training sets in the training set A be equal to 0";

[0054] If the third judgment result is no, then output the RIM network model with updated parameters as the trained RIM network model.

[0055] A quantitative intravascular optical coherence tomography system, comprising:

[0056] An optical simulation module for performing optical simulation based on the IVOCT forward imaging model to obtain the theoretical values of the backscattered signals on the vascular cross-section corresponding to different optical attenuation coefficients, and constructing a distribution map of optical attenuation coefficients and a distribution map of the theoretical values of the backscattered signals on the vascular cross-section corresponding to the distribution map of optical attenuation coefficients;

[0057] A dataset construction module for taking the distribution map of the theoretical values of the backscattered signals on the vascular cross-section as the input of the sample and the corresponding distribution map of optical attenuation coefficients as the expected output of the sample to construct a dataset for quantitative analysis of IVOCT;

[0058] A RIM network model building module for building a RIM network model;

[0059] A RIM network model training module for training the RIM network model using the dataset to obtain a trained RIM network model;

[0060] A quantitative analysis module for inputting the distribution map of the measured values of the backscattered signals on the vascular cross-section to be analyzed into the trained RIM network model for quantitative analysis to obtain the distribution map of optical attenuation coefficients corresponding to the distribution map of the measured values of the backscattered signals on the vascular cross-section to be analyzed.

[0061] Optionally, the optical simulation module specifically includes:

[0062] An optical transmission simulation sub-module for simulating the transmission of incident light in tissues with different optical attenuation coefficients using the Monte Carlo method to obtain the photon weights and radial depths of the incident light in tissues with different optical attenuation coefficients; the photon weight is the photon weight collected from the surface of the tissue, and the radial depth is the transmission depth of the photon in the tissue;

[0063] A backscattered signal calculation sub-module on the vascular cross-section for calculating the theoretical values of the backscattered signals on the vascular cross-section corresponding to different optical attenuation coefficients according to the photon weights and transmission depths of the incident light in tissues with different optical attenuation coefficients.

[0064] Optionally, calculating the theoretical values of the backscattered signals on the vascular cross-section corresponding to different optical attenuation coefficients according to the photon weights and transmission depths of the incident light in tissues with different optical attenuation coefficients is:

[0065]

[0066] where I(r) represents the theoretical value of the backscattered signal at position r on the vascular cross-section, I0 is a constant characterizing the imaging system, N ph is the total number of photons, |r| is the radial depth of position r, λ c is the central wavelength of the incident light, L iis the optical path length of the i-th photon transmitted in the tissue, w i is the weight of the i-th photon, l coh is the coherence length of the detection light source; Δλ is the full width at half maximum of the light source spectrum.

[0067] Optionally, the RIM network model includes a plurality of RIM network structure units connected in cascade.

[0068] According to the specific embodiments provided by the present invention, the following technical effects are disclosed by the present invention:

[0069] The present invention discloses a method for quantitative intravascular optical coherence tomography, including the following steps: performing optical simulation based on the IVOCT forward imaging model to obtain the theoretical values of the backscattered signals on the vascular cross-section corresponding to different optical attenuation coefficients, constructing a distribution map of optical attenuation coefficients and a distribution map of the theoretical values of the backscattered signals on the vascular cross-section corresponding to the distribution map of optical attenuation coefficients; using the distribution map of the theoretical values of the backscattered signals on the vascular cross-section as the input of the sample, and using the corresponding distribution map of optical attenuation coefficients as the expected output of the sample to construct a data set for quantitative analysis of IVOCT; building a RIM network model; training the RIM network model using the data set to obtain a trained RIM network model; inputting the distribution map of the measured values of the backscattered signals on the vascular cross-section to be analyzed into the trained RIM network model for quantitative analysis to obtain the distribution map of optical attenuation coefficients corresponding to the distribution map of the measured values of the backscattered signals on the vascular cross-section to be analyzed. The present invention first constructs a data set based on the IVOCT forward imaging model, trains a RIM network model based on the data set, and performs quantitative analysis based on the trained RIM network model. The present invention realizes quantitative intravascular optical coherence tomography based on the IVOCT forward imaging model and the RIM network model. Description of the Drawings

[0070] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0071] Figure 1 is a flowchart of a method for quantitative intravascular optical coherence tomography provided by an embodiment of the present invention;

[0072] Figure 2 is a schematic diagram of a method for quantitative intravascular optical coherence tomography provided by an embodiment of the present invention;

[0073] Figure 3Schematic structural diagram of the RIM network model provided by an embodiment of the present invention;

[0074] Figure 4 Schematic structural diagram of the gated recurrent unit provided by an embodiment of the present invention;

[0075] Figure 5 Flowchart for building the RIM network model provided by an embodiment of the present invention;

[0076] Figure 6 Flowchart for training the RIM network model provided by an embodiment of the present invention. Detailed implementation manners

[0077] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0078] The purpose of the present invention is to provide a quantitative intravascular optical coherence tomography method and system to achieve quantitative intravascular optical coherence tomography, thereby improving the accuracy of tissue classification and calibration and providing a reliable basis for computer-aided diagnosis of coronary atherosclerosis.

[0079] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0080] As Figure 1 and 2 shown, the present invention provides a quantitative intravascular optical coherence tomography method, including the following steps:

[0081] Step 101, perform optical simulation based on the IVOCT forward imaging model to obtain the theoretical values of the backscattered signals on the vascular cross-section corresponding to different optical attenuation coefficients, and construct a distribution map of the optical attenuation coefficients and a distribution map of the theoretical values of the backscattered signals on the vascular cross-section corresponding to the distribution map of the optical attenuation coefficients.

[0082] Based on the IVOCT forward imaging model, optical simulation is carried out to obtain the theoretical values of the backscattering signals on the vascular cross-section corresponding to different optical attenuation coefficients, specifically including: using the Monte Carlo method to simulate the transmission of incident light in tissues with different optical attenuation coefficients, obtaining the photon weights and radial depths of the incident light in tissues with different optical attenuation coefficients; the photon weight is the photon weight collected from the surface of the tissue, and the radial depth is the transmission depth of photons in the tissue; according to the photon weights and transmission depths of the incident light in tissues with different optical attenuation coefficients, calculate the theoretical values of the backscattering signals on the vascular cross-section corresponding to different optical attenuation coefficients.

[0083] Exemplarily, the IVOCT forward imaging model includes three parts: generating a Gaussian beam, simulating the propagation of photons in a medium, and simulating the backscattering signal.

[0084] Step 201, generate a Gaussian beam.

[0085] First, calculate the initial emission point coordinates (x0, y0) of photons in the Gaussian beam:

[0086]

[0087] In the formula, w0 is the waist radius of the Gaussian beam, ξ1 and ξ2 are uniformly distributed and independent random numbers in the interval [0, 1], R ≤ 1 and R ≠ 0. x0 and y0 follow a Gaussian distribution with a mean of 0 and a variance of and

[0088]

[0089] The average propagation direction of a large number of photons emitted from this point is the same as the Poynting vector of the corresponding Gaussian beam.

[0090] Then, track the propagation trajectory of photons in the focusing system and calculate the optical distribution parameters on the back of the lens, including the optical path L:

[0091] L = l · n (3)

[0092] And the energy loss of photons, that is, the reflection loss R:

[0093]

[0094] And the absorption loss T:

[0095] T = 1 - e -ax (5)

[0096] Wherein, l and n are respectively the distance between two intersection points of a photon and a lens surface and the refractive index of the medium between these two points (i.e., the refractive index of the medium before the transmission interface), n' is the refractive index of the medium after the transmission interface, a is the absorption coefficient of the lens optical material, and x is the thickness of the lens.

[0097] Step 202: Use the Monte Carlo method to simulate the transmission of incident light in tissue and obtain the photon weights and path lengths collected from the medium surface.

[0098] The propagation process of light in tissue is as follows: Photons are incident along the direction perpendicular to the tissue surface, and after multiple scattering and absorption, some photons escape from the tissue and are finally received by the detector. The steps of using the Monte Carlo method to simulate the propagation trajectory of photons in a weakly absorbing and strongly scattering medium include: emitting and initializing photons, setting the photon sampling step size, absorption and scattering of photons, and ending of photons. Specifically as follows:

[0099] Initialize a single photon vertically incident on the tissue surface, including setting the initial position, direction cosine, and weight. The number of scattering times and the distances between scattering points when photons propagate in tissue are random. Therefore, a random number is used to sample the photon movement step size s (i.e., the free path length):

[0100]

[0101] Wherein, ξ is a random variable uniformly distributed in the interval [0, 1], and μ t is the light attenuation coefficient of the tissue (i.e., the extinction coefficient):

[0102] μ t = μ a + μ s (7)

[0103] Wherein, μ a and μ s are respectively the light absorption coefficient and light scattering coefficient of the tissue.

[0104] When photons are scattered in tissue, the transmission direction changes. Its azimuth angle is only related to the medium characteristics, while the deflection angle θ is related to scattering. The azimuth angle is uniformly distributed in the interval (0, 2π), and is directly sampled to obtain The deflection angle θ ∈ [0, π], and its cosine is:

[0105]

[0106] Wherein, g is the scattering anisotropy factor of the tissue.

[0107] The transmission depth r of photons in tissue is:

[0108]

[0109] where N s is the total number of scattering times of a single photon in the tissue, s j is the step length of the j-th photon movement, μ z,j is the cosine of the j-th scattering of the photon:

[0110]

[0111] After the photon is scattered in the medium, part of its energy is absorbed by the medium, resulting in a decrease in the photon weight. If the scattering position is not in the target tissue layer, the weight of the scattered photon is:

[0112]

[0113] where w is the initial weight of the photon and w′ is the weight of the scattered photon. If the scattering occurs in the target tissue layer, the weight of the scattered photon is:

[0114]

[0115] If the photon weight is less than the preset threshold (i.e., the photon weight when the scattered medium is completely absorbed) or the photon escapes from the tissue surface, the propagation of the photon terminates.

[0116] Step 203, calculate the theoretical value of the backscattering signal at each position on the blood vessel cross-section.

[0117] The theoretical value of the backscattering signal at position r on the blood vessel cross-section is:

[0118]

[0119] where I0 is a constant characterizing the imaging system, N ph is the total number of photons, |r| is the radial depth of position r, λ c is the central wavelength of the incident light, L i is the optical path length of the i-th photon transmitted in the medium (i.e., the product of the refractive index of the medium and the total step length of the photon movement), w i is the weight of the i-th photon, l coh is the coherence length of the detection light source:

[0120]

[0121] where Δλ is the full width at half maximum of the light source spectrum.

[0122] Step 102: Taking the theoretical value distribution map of backscattered signals on the cross section of the blood vessel as the input of the sample and the corresponding light attenuation coefficient distribution map as the expected output of the sample, a data set for IVOCT quantitative analysis is constructed.

[0123] Specifically, the theoretical value distribution map of the backscattered signal on the cross section of the blood vessel is obtained according to the IVOCT forward imaging model, which is used as the input of the sample, and the corresponding light attenuation coefficient distribution map is used as the expected output of the sample to construct a data set suitable for quantitative IVOCT. All samples in the data set are randomly shuffled and divided into training set A, validation set B, and test set C in a ratio of 6:2:2. The samples in the training set A are evenly divided into P groups, each of which is a small batch training set containing Q samples.

[0124] Step 103: Building a RIM network model.

[0125] like Figure 5 As shown, the specific steps of building a RIM network model, that is, designing a forward propagation RIM network model, include:

[0126] Step 301: Building a RIM network structure.

[0127] The RIM network structure constructed by the method of the present invention is as follows Figure 3 As shown in the figure, a structural unit consists of three convolutional layers and two gated recurrent units (GRU). The filter kernel size used in the convolution operation is 3×3×M (the first two layers) and 1×1×M (the third layer), where M is the number of feature maps input to the current layer. The convolution layer uses the same convolution with a step size of 2, where the initial number of filter kernels, i.e., the number of feature channels, is channels = 64, and the activation function uses the linear rectification function ReLU.

[0128] like Figure 4 As shown, GRU is updated by gate z t and reset gate r t Two gate functions σ (sigmoid activation function), tanh activation function, matrix Hadamard product "⊙", matrix addition And "1-" operation. By updating the hidden state of the network, the relevant information from the previous time step is passed. The specific process is:

[0129] First, according to the hidden state h transmitted from time step t-1 t-1 and the GRU input x at time step t t Get two gate states and get the gate signal r t and z t :

[0130]

[0131] Where W r and W z are the weight parameter sets of r t and z t respectively, and "[·,·]" represents the concatenation operation on two matrices.

[0132] Then, perform the Hadamard product operation on r t and h t-1 , concatenate the resulting result with x t , and then obtain the candidate hidden state at the current time step through the tanh activation function

[0133]

[0134] Where is 's weight parameter set.

[0135] Finally, z t after performing the "1-" operation (i.e., 1 - z t ) and z t are respectively subjected to the Hadamard product with h t-1 and , and the results are then subjected to matrix addition to obtain the hidden state h t at time step t:

[0136]

[0137] Step 302, initialize the network parameters.

[0138] The feature layer number c = 1, the number of convolutions conv = 0, the number of feature channels channels = 64, the time step t = 0, the maximum time step t max = T = 8, and the initial hidden state h 0,1 = h 0,2 = 0.

[0139] Step 303, calculate the likelihood gradient of the (t + 1)-th network structure unit input:

[0140]

[0141] Where is the likelihood gradient of the (t + 1)-th network structure unit input, i is the sample number in the first mini-batch training set A1 in the training set A, I i is the input backscattering signal distribution map of the i-th sample, H is the forward imaging operator (i.e., the process of obtaining the theoretical backscattering signal values at various places on the blood vessel cross-section from the light attenuation coefficient of the tissue), H *is the adjoint operator of H, μ t,i is the optical attenuation coefficient distribution map output after the i-th sample passes through the t-th time step:

[0142]

[0143] where μ t,i and μ t-1,i are the optical attenuation coefficient distribution maps output after the i-th sample passes through the t-th and t - 1-th time steps respectively, μ 0,i is the initial value of μ t,i Δμ t,i is the incremental image output after the i-th sample passes through the t-th time step:

[0144]

[0145] where is the network structure unit at the t-th time step with learning parameter φ, is the likelihood gradient input to the t-th network structure unit, h t is the hidden state of the network structure unit at the t-th time step (including h t,1 and h t,2 two components):

[0146]

[0147] where is the GRU at the t-th time step with learning parameter φ.

[0148] Step 304, input all Q samples in A1 into the network.

[0149] Step 305, simply concatenate and μ t,i along the channel dimension to generate the input of the first convolutional layer in the network structure unit at the t + 1-th time step, as the feature layer map c .

[0150] Step 306, use map c as the network input feature layer map at the t + 1-th time step map t .

[0151] Step 307, perform a same convolution operation with a stride of 2 on map c in parallel through multiple feature channels to generate multiple groups of linear activation responses, calculate the ReLU function values of each linear activation response, and obtain the feature layer map c after the convolution operation.

[0152] Step 308, let c ← c + 1 and conv ← conv + 1.

[0153] Step 309, the convolutional feature layer map c and the state h at the previous time step t,1 enter the first GRU, obtain two gated states, and get a new feature layer map c and a new state h t+1,1 , and let c←c + 1.

[0154] Step 3010, for map c perform a same convolution operation with a stride of 2 in parallel through multiple feature channels, generate multiple sets of linear activation responses, calculate the ReLU function values of each linear activation response, and obtain the convolutional feature layer map c .

[0155] Step 3011, let c←c + 1 and conv←conv + 1.

[0156] Step 3012, the convolutional feature layer map c and the state h at the previous time step t,2 pass through the second GRU, obtain two gated states, and get a new feature layer map c and a new state h t+1,2 , and let c←c + 1.

[0157] Step 3013, for map c perform a same convolution operation with a stride of 2 in parallel through multiple feature channels, generate multiple sets of linear activation responses, calculate the ReLU function values of each linear activation response, and obtain the convolutional feature layer map c .

[0158] Step 3014, respectively let c←c + 1, conv←conv + 1 and map c ←map c +map t , then calculate the ReLU function value of the linear activation response to obtain the convolutional feature layer.

[0159] Step 3015, let t←t + 1. If t < 8, return to Step 305, otherwise execute Step 3016.

[0160] Step 3016, output the predicted images of Q samples in A1. At this time, c = 41.

[0161] Step 104, use the dataset to train the RIM network model to obtain the trained RIM network model.

[0162] Such as Figure 6As shown, when the number of layers of the exemplary RIM network model is 41, the specific implementation steps of step 104 of the present invention are:

[0163] Step 401, initializing network parameters: setting time step t=0.

[0164] Step 402, set the number of epochs of traversing all small batch training sets in training set A to 0.

[0165] Step 403, set the index j of the small batch training set in A to 1.

[0166] Step 404: The jth group of small batch training set A in the training set A is j The Q samples in are input into the built RIM network for forward propagation.

[0167] Step 405, set the loss function:

[0168]

[0169] Where W and b are the weight parameter set and bias parameter set of the network, L(W,b) is the loss function about W and b, and m is A j The total number of pixels in Q samples, T is the total number of time steps for training the network, 8 for example, μ t,i Yes A j The light attenuation coefficient distribution of the i-th sample in after t time steps of training, μ i Yes A j The target image of the i-th sample in (i.e., the simulated light attenuation coefficient distribution diagram).

[0170] Step 406, calculate the gradient of the parameters of the network output layer (exemplarily when the number of layers is 41, the output layer is the 41st layer):

[0171]

[0172]

[0173] Where W 41 and b 41 are the weight parameter matrix and bias parameter vector of the 41st layer network, Y 40 is the output of the 40th layer of the network in the forward propagation (also the input of the 41st layer), "⊙" represents the Hadamard product of the matrix, σ'() is the first-order derivative of the ReLU activation function, the superscript "T" represents the matrix transpose, and Z 41 is the linear activation response produced by the 41st layer of the network in the forward propagation:

[0174] Z 41 =W 41 Y40 +b 41 (25)

[0175] Step 407: Calculate the error vectors of each layer of the network.

[0176] The error vector of the network output layer (i.e., the 41st layer) is:

[0177]

[0178] The error vectors of the 40th to 2nd layers of the network are obtained by recursion according to Equation (26):

[0179] δ c =((W c+1 ) T δ c+1 )⊙σ'(Z c ) (27)

[0180] where c = 40, 39, 38,..., 2, δ c is the error vector of the cth layer of the network, W c+1 is the weight parameter matrix of the (c + 1)th layer of the network, and Z c is the linear activation response generated by the cth layer of the network in forward propagation.

[0181] Step 408: Calculate the gradients of the parameters of each layer of the network:

[0182]

[0183]

[0184] where c = 41, 40, 39, 38,..., 2, Y c-1 is the output of the (c - 1)th layer of the network in forward propagation, and W c and b c are the weight parameter matrix and bias parameter vector of the cth layer of the network, respectively.

[0185] Step 409: Update the parameters W c and b c using the adaptive moment estimation (Adam) algorithm:

[0186]

[0187] where W c ' and b c ' are the updated weight parameter matrix and bias parameter vector of the cth layer of the network, respectively, α = 0.001 represents the learning rate, and ε = 10 -8 is a constant used to maintain the stability of numerical calculations. and is the corrected Momentum exponentially weighted average, and is the corrected RMSprop exponentially weighted average. Let k←k + 1:

[0188]

[0189] where k is the corrected number of times of network parameter gradient accumulation (initial value is 0), V dW and V db are the Momentum exponentially weighted averages of dW and db respectively, S dW and S db are the RMSprop exponentially weighted averages of dW and db respectively, β1 and β2 are the exponential decay rates of the first - order moment estimate and the second - order moment estimate respectively, dW is the gradient matrix of W c and db is the gradient vector of b c :

[0190]

[0191] Step 4010, let j←j + 1. If j < P + 1, then go to step 404; otherwise, let epoch←epoch + 1 and go to step 4011.

[0192] Step 4011, if epoch < epoch max , epoch max is the preset maximum value of epoch, then go to step 403; otherwise, stop traversing and execute step 4012.

[0193] Step 4012, let t←t + 1; if t < 8, then return to step 402; otherwise, go to step 4013.

[0194] Step 4013, determine the final network parameters to obtain the optimized RIM network model.

[0195] Step 105, input the measurement value distribution map of the backscattering signal on the cross - section of the blood vessel to be analyzed into the trained RIM network model for quantitative analysis, and obtain the light attenuation coefficient distribution map corresponding to the measurement value distribution map of the backscattering signal on the cross - section of the blood vessel to be analyzed.

[0196] Perform grayscale processing on the standard cross - section image (i.e., the transverse view) output by the imaging system, and obtain the measurement values of the backscattering signal at each point in the image plane according to the gray - scale values of the pixels in the gray - scale image:

[0197]

[0198] where \(g(r)\) is the 8-bit gray value at position \(r\) in the IVOCT gray-scale image (the value range is \([0, 255]\)), and \(I\) m (r) is the measured value of the backscattering signal at position \(r\). The distribution map of the measured values of the backscattering signal is input into the trained RIM network, and finally the distribution map of the optical attenuation coefficient is output.

[0199] The present invention also provides a quantitative intravascular optical coherence tomography system, including:

[0200] An optical simulation module, configured to perform optical simulation based on the IVOCT forward imaging model, obtain the theoretical values of the backscattering signals on the vascular cross-section corresponding to different optical attenuation coefficients, and construct a distribution map of the optical attenuation coefficient and a distribution map of the theoretical values of the backscattering signals on the vascular cross-section corresponding to the distribution map of the optical attenuation coefficient.

[0201] The optical simulation module specifically includes: a light transmission simulation sub-module, configured to simulate the transmission of incident light in tissues with different optical attenuation coefficients by using the Monte Carlo method, and obtain the photon weights and radial depths of the incident light in tissues with different optical attenuation coefficients; the photon weight is the photon weight collected from the surface of the tissue, and the radial depth is the transmission depth of the photon in the tissue; a backscattering signal calculation sub-module on the vascular cross-section, configured to calculate the theoretical values of the backscattering signals on the vascular cross-section corresponding to different optical attenuation coefficients according to the photon weights and transmission depths of the incident light in tissues with different optical attenuation coefficients.

[0202] Among them, calculating the theoretical values of the backscattering signals on the vascular cross-section corresponding to different optical attenuation coefficients according to the photon weights and transmission depths of the incident light in tissues with different optical attenuation coefficients is:

[0203]

[0204] where \(I(r)\) represents the theoretical value of the backscattering signal at position \(r\) on the vascular cross-section, \(I_0\) is a constant characterizing the imaging system characteristics, \(N\) ph is the total number of photons, \(|r|\) is the radial depth of position \(r\), \(\lambda\) c is the central wavelength of the incident light, \(L\) i is the optical path length of the \(i\)-th photon transmitted in the tissue, \(w\) i is the weight of the \(i\)-th photon, \(l\) coh is the coherence length of the detection light source; \(\Delta\lambda\) is the full width at half maximum of the light source spectrum.

[0205] A data set construction module, configured to use the distribution map of the theoretical values of the backscattering signals on the vascular cross-section as the input of the sample, and use the corresponding distribution map of the optical attenuation coefficient as the expected output of the sample, and construct a data set for IVOCT quantitative analysis.

[0206] The RIM network model building module is used to build a RIM network model.

[0207] The RIM network model includes a plurality of RIM network structure units connected in cascade.

[0208] The RIM network model training module is used to train the RIM network model by using the data set to obtain a trained RIM network model.

[0209] The quantitative analysis module is used to input the measured value distribution map of the backscattering signal on the cross-section of the blood vessel to be analyzed into the trained RIM network model for quantitative analysis, and obtain the light attenuation coefficient distribution map corresponding to the measured value distribution map of the backscattering signal on the cross-section of the blood vessel to be analyzed.

[0210] According to the specific embodiments provided by the present invention, the following technical effects are disclosed in the present invention:

[0211] First of all, the present invention establishes a forward imaging model of IVOCT, in which the Monte Carlo method (MC) is used to simulate the transmission process of incident light in tissues to obtain the theoretical value of the tissue backscattering signal; then, the theoretical value distribution map of the backscattering signal obtained by computer simulation is used as the input of the sample, and the corresponding light attenuation coefficient distribution map is used as the expected output of the sample to construct a data set suitable for quantitative IVOCT; secondly, a RIM network model is built and trained by using the above data set to optimize the network parameters; finally, according to the mapping relationship between the gray value of each pixel in the IVOCT gray-scale image and the backscattering signal, the measured value distribution map of the backscattering signal is obtained and input into the trained RIM network model, and finally the light attenuation coefficient distribution map on the cross-section of the blood vessel is output.

[0212] In the description of this specification, each embodiment is described in a progressive manner. The key points of each embodiment are the differences from other embodiments. The same and similar parts among the embodiments can be referred to each other.

[0213] In this article, specific examples are used to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A quantitative intravascular optical coherence tomography method, characterized in that, It includes the following steps: Based on the IVOCT forward imaging model, perform optical simulation to obtain the theoretical values of the backscattered signals on the vascular cross-section corresponding to different optical attenuation coefficients, and construct a distribution map of optical attenuation coefficients and a distribution map of the theoretical values of the backscattered signals on the vascular cross-section corresponding to the distribution map of optical attenuation coefficients; Use the distribution map of the theoretical values of the backscattered signals on the vascular cross-section as the input of the sample, and use the corresponding distribution map of optical attenuation coefficients as the expected output of the sample to construct a dataset for quantitative analysis of IVOCT; Build a RIM network model; Use the dataset to train the RIM network model to obtain a trained RIM network model; Input the distribution map of the measured values of the backscattered signals on the vascular cross-section to be analyzed into the trained RIM network model for quantitative analysis to obtain the distribution map of optical attenuation coefficients corresponding to the distribution map of the measured values of the backscattered signals on the vascular cross-section to be analyzed; The optical simulation based on the IVOCT forward imaging model to obtain the theoretical values of the backscattered signals on the vascular cross-section corresponding to different optical attenuation coefficients specifically includes: Adopt the Monte Carlo method to simulate the transmission of incident light in tissues with different optical attenuation coefficients to obtain the photon weights and radial depths of the incident light in tissues with different optical attenuation coefficients; the photon weights are the photon weights collected from the surface of the tissue, and the radial depth is the transmission depth of photons in the tissue; According to the photon weights and transmission depths of the incident light in tissues with different optical attenuation coefficients, calculate the theoretical values of the backscattered signals on the vascular cross-section corresponding to different optical attenuation coefficients as: Among them, I(r) represents the theoretical value of the backscattering signal at position r on the cross-section of the blood vessel, I0 is a constant characterizing the imaging system's characteristics, N ph is the total number of photons, |r| is the radial depth of position r, λ c is the central wavelength of the incident light, L i is the optical path length that the i-th photon travels in the tissue, w i is the weight of the i-th photon, l coh is the coherence length of the detection light source; Δλ is the full width at half maximum of the light source spectrum.

2. The quantitative intravascular optical coherence tomography method according to claim 1, characterized in that, The RIM network model includes multiple RIM network structure units connected in cascade.

3. The quantitative intravascular optical coherence tomography method according to claim 1, wherein The RIM network structure unit includes a first convolutional layer, a first gated recurrent unit, a second convolutional layer, a second gated recurrent unit, and a third convolutional layer connected in sequence; The sizes of the filter kernels of the first convolutional layer and the second convolutional layer are 3×3×M; where M represents the number of feature maps input to the current layer; The size of the filter kernel of the third convolutional layer is 1×1×M; Both the first gated recurrent unit and the second gated recurrent unit include an update gate and a reset gate, and the gate functions of the update gate and the reset gate are both sigmoid activation functions.

4. The quantitative intravascular optical coherence tomography method according to claim 1, characterized in that The training of the RIM network model using the dataset to obtain a trained RIM network model specifically includes: Initialize the network parameters of the RIM network model, set the time step t = 0, divide the training set A from the dataset, and divide the training set A into multiple mini-batch training sets; Set the number of times epoch for traversing all mini-batch training sets in the training set A to 0; Set the index j of the mini-batch training set in the training set A to 1; Input the Q samples in the j-th mini-batch training set A in the training set A into the RIM network model for forward propagation; j ​ Use the loss function to calculate the loss value of the forward propagation; the loss function is: Where, W and b are the weight parameter set and bias parameter set of the RIM network model respectively, L(W, b) is the loss function with respect to W and b, and m is the total number of pixels of Q samples in A j ; T is the total number of time steps for training the network, and μ t,i is the light attenuation coefficient distribution map obtained after training the i-th sample in A j for t time steps, and μ i is the target image of the i-th sample in A j ; According to the loss value, calculate the gradient of the output layer parameters of the RIM network model using the following formula; Where, W C and b C are the weight parameter matrix and the bias parameter vector of the C-th layer network respectively, and C in W C and b C represents the total number of layers in the RIM network model. Y C-1 is the output of the (C - 1)-th layer network in the forward propagation. "⊙" represents the Hadamard product of matrices. σ'() is the first derivative of the ReLU activation function. The superscript T represents the matrix transpose. Z C is the linear activation response generated by the C-th layer network in the forward propagation: Z C = W C Y C-1 + b C ; Calculate the error vector of each layer of the RIM network model using the following formula; where, δ c represents the error vector of the c-th layer network of the RIM network model, and W c+1 is the weight parameter matrix of the (c + 1)-th layer network, and Z c is the linear activation response generated by the c-th layer network in the forward propagation; According to the error vector of each layer of the RIM network model and the loss value, calculate the gradient of each layer of the RIM network model using the following formula: where \(c = C, C - 1, C - 2, \cdots, 2, Y\) c-1 is the output of the \((c - 1)\)-th layer network in the forward propagation, \(W\) c and \(b\) c are the weight parameter matrix and the bias parameter vector of the \(c\)-th layer network, respectively; According to the gradients of each layer of the RIM network model, the adaptive moment estimation algorithm is used to update the weight parameter matrix and bias parameter vector of each layer of the RIM network model using the following formula; Where, W c ' and b c ' are respectively the weight parameter matrix and the bias parameter vector of the updated c-th layer network, α represents the learning rate, ε is a constant used to maintain the stability of numerical calculations, and are the corrected Momentum exponentially weighted moving averages, and are the corrected RMSprop exponentially weighted moving averages. Let k ← k + 1: where k is the number of corrections for network parameter gradient accumulation, with an initial value of 0 for k, V dW and V db are the Momentum exponentially weighted moving averages of dW and db respectively, S dW and S db are the RMSprop exponentially weighted moving averages of dW and db respectively, β1 and β2 are the exponential decay rates of the first-order and second-order moment estimates respectively, dW is the gradient matrix of W c , and db is the gradient vector of b c : Let the value of j increase by 1, and determine whether the formula j < P + 1 holds to obtain the first judgment result; If the first judgment result is "yes", return to the step of "inputting the Q samples in the j-th mini-batch training set A in the training set A into the RIM network model for forward propagation"; j in the RIM network model for forward propagation"; If the first judgment result is negative, increase the value of epoch by 1 and judge whether the formula epoch < epoch max holds to obtain a second judgment result; epoch max is the pre-set maximum value of epoch; If the second judgment result indicates yes, return to the step "Let the index j of the mini-batch training set in the training set A be equal to 1"; If the second judgment result indicates no, let the value of t increase by 1, and determine whether the formula t < T holds to obtain the third judgment result; If the third judgment result indicates yes, return to the step "Let the number of epochs for traversing all mini-batch training sets in the training set A be equal to 0"; If the third judgment result indicates no, output the RIM network model with updated parameters as the trained RIM network model.

5. A quantitative intravascular optical coherence tomography system, characterized in that, Including: An optical simulation module for performing optical simulation based on the IVOCT forward imaging model to obtain the theoretical values of the backscattered signals on the vascular cross-section corresponding to different optical attenuation coefficients, and constructing an optical attenuation coefficient distribution map and a distribution map of the theoretical values of the backscattered signals on the vascular cross-section corresponding to the optical attenuation coefficient distribution map; A data set construction module for constructing a data set for IVOCT quantitative analysis by using the distribution map of the theoretical values of the backscattered signals on the vascular cross-section as the input of the sample and the corresponding optical attenuation coefficient distribution map as the expected output of the sample; A RIM network model building module for building a RIM network model; A RIM network model training module for training the RIM network model using the data set to obtain a trained RIM network model; A quantitative analysis module for inputting the measurement value distribution map of the backscattered signals on the vascular cross-section to be analyzed into the trained RIM network model for quantitative analysis to obtain the optical attenuation coefficient distribution map corresponding to the measurement value distribution map of the backscattered signals on the vascular cross-section to be analyzed; The optical simulation module specifically includes: An optical transmission simulation sub-module for simulating the transmission of incident light in tissues with different optical attenuation coefficients using the Monte Carlo method to obtain the photon weights and radial depths of the incident light in tissues with different optical attenuation coefficients; the photon weights are the photon weights collected from the surface of the tissue, and the radial depth is the transmission depth of the photon in the tissue; A backscattered signal calculation sub-module on the vascular cross-section for calculating the theoretical values of the backscattered signals on the vascular cross-section corresponding to different optical attenuation coefficients according to the photon weights and transmission depths of the incident light in tissues with different optical attenuation coefficients as: Among them, I(r) represents the theoretical value of the backscattering signal at position r on the cross-section of the blood vessel, I0 is a constant characterizing the imaging system's characteristics, N ph is the total number of photons, |r| is the radial depth of position r, λ c is the central wavelength of the incident light, L i is the optical path length that the i-th photon travels in the tissue, w i is the weight of the i-th photon, l coh is the coherence length of the detection light source; Δλ is the full width at half maximum of the light source spectrum.

6. The quantitative intravascular optical coherence tomography system according to claim 5, wherein The RIM network model includes a plurality of RIM network structure units connected in cascade.

Images