Diffusion optical tomography reconstruction method based on two-dimensional convolution physical information neural network

Abstract

The invention discloses a diffusion optical tomography reconstruction method based on a two-dimensional convolution physical information neural network, and belongs to the field of medical image processing. Comprising three modules: a data preprocessing module, a two-dimensional space physical information loss function module and a two-dimensional convolutional neural network reconstruction module. Wherein the data preprocessing module is used for mapping data measured by a detector into a two-dimensional data space through a specific rule, so that the data can be used as input data of the two-dimensional convolutional neural network reconstruction module; according to the method, the two-dimensional convolutional physical information neural network is creatively provided, the physical constraint of the diffusion equation is applied to training of the two-dimensional convolutional neural network, convergence of the neural network is constrained through an interpretable physical formula, and a high-quality optical parameter distribution diagram of the to-be-detected tissue body is directly reconstructed. Experimental results show that the method can realize accurate diffusion optical tomography reconstruction.

Description

Technical Field

[0001] The invention belongs to the field of medical image processing and relates to a diffusion optical tomography reconstruction method based on a convolutional physical information neural network. Background Art

[0002] At present, clinical detection of breast cancer mainly relies on modern medical imaging technology, such as breast X-ray photography, ultrasound detection, MRI and PET. After the abnormality is detected, pathological examination is often used to give the final diagnosis. However, the above detection methods all have shortcomings. Breast X-ray photography and PET have the risk of radiation-induced cancer, MRI detection is prone to misjudgment, and ultrasound detection is not deep enough. Diffuse optical tomography is an emerging tumor detection method. It has many advantages such as harmlessness, high cost performance, and real-time measurement. It is a breast tumor detection method with great potential.

[0003] Diffusion Optical Tomography (DOT) uses non-ionizing near-infrared light (wavelength 600nm-900nm) to illuminate human tissues, detects the emitted light of the organism, and combines image reconstruction technology and diffuse light measurement technology to obtain two-dimensional or three-dimensional tissue optical parameter distribution images, which can provide a reliable basis for cancer diagnosis. Compared with the several detection methods mentioned above, DOT technology has great advantages in equipment complexity, maintenance cost and maintenance difficulty, and has the hope of being popularized as an early screening technology for breast cancer. It has many advantages such as non-invasive, cost-effective, real-time measurement and non-ionizing radiation imaging.

[0004] At present, the physical models that describe the propagation of light in tissues mainly include the Monte Carlo model, the radiation transfer equation model and the diffusion equation model. Among them, the Monte Carlo method has a very slow calculation speed; the radiation transfer equation cannot obtain an analytical solution in the case of complex geometry. The diffusion equation is a simplified form of the radiation transfer equation and has a better simulation effect in low scattering areas. The diffusion equation is as follows:

[0005]

[0006] where κ(r) is the optical parameter distribution at position r in the tissue, μ a (r) is the absorption coefficient distribution at r in the tissue body, Φ(r) is the light intensity value at r in the tissue body, and q0(r) is the light source term at r in the tissue body.

[0007] Due to the irregularity of the tissue, the diffusion equation cannot be solved directly. It is necessary to introduce the finite element analysis method to transform the diffusion equation into a matrix equation under the finite element framework:

[0008] (K(κ)+Cμ a+(1 / 2A)F)Φ=q0#(2)

[0009] The open source software package NIRFAST provides a solver that completes the above-mentioned solution process, which can calculate the light emitted from the tissue. The process of reconstructing the optical parameter distribution map of the tissue is completely opposite to the above process. It is the process of inversely solving the optical parameter distribution after obtaining the emitted light of the tissue. However, there are problems with both the forward solution and the inverse solution. The forward solution cannot calculate the analytical solution; the inverse solution requires linear approximation and then amplifying the change in light flux caused by the optical parameters for solution. The disadvantage of this is that noise and errors are also amplified, affecting the final imaging quality. And because the propagation of light in the tissue is very random, there will be different optical parameter distributions corresponding to the same light intensity detection value on the surface of the tissue, which means that the inverse reconstruction process is an ill-posed problem.

[0010] Physics Informed Neural Network (PINN) is an application method of scientific machines in the traditional numerical field, that is, adding real physical constraints (such as physical equations) as restrictions to network training, so that the fitting results are more in line with physical laws. The principle is to minimize the loss function by training the neural network to approximate the solution of PED. The so-called loss function terms include the residual terms of the initial and boundary conditions, and the residual of the partial differential equation at the selected points (i.e., matching points) in the region. After the physical information neural network is trained, the value at the required time and space point can be obtained by direct reasoning. Compared with traditional neural networks, PINN directly uses physical equations as loss functions, combined with loss functions such as mean square error to guide the convergence of neural networks. It connects data with physical knowledge, and well makes up for the weakness of pure data-driven in the field of scientific machine learning (pure data-driven mode is suitable for processing applications with a large number of observations, but when the amount of data decreases, the generalization of the model will inevitably deteriorate. In other words, the inference results of the network may not conform to the actual physical situation, that is, it is unreliable). With the help of automatic differentiation functions in various deep learning frameworks, it becomes very simple to derive variables (find gradients), which is also a major advantage of using physical information neural networks.

[0011] In response to the ill-posedness of the inverse problem mentioned above and the unreliability and poor generalization of ordinary neural networks, the present invention proposes for the first time a diffuse optical tomography reconstruction method based on a two-dimensional convolutional physical information neural network, pioneeringly introduces physical information neural networks into diffuse optical tomography reconstruction, and realizes automatic gradient calculation for two-dimensional data. This solves the pain points of the above-mentioned traditional methods. Compared with general deep learning reconstruction algorithms, the present invention uses the physical constraint of the diffusion equation that describes the transmission process of light inside the tissue to guide the training of the neural network, making the neural network physically interpretable. Summary of the invention

[0012] The present invention designs a diffusion optical tomography reconstruction method based on convolutional physical information neural network, which includes three modules: data preprocessing module, two-dimensional space physical information loss function module and two-dimensional convolutional neural network reconstruction module. Among them, the function of the data preprocessing module is to map the data measured by the detector into a two-dimensional data space through specific rules, so that it can be used as input data of the two-dimensional convolutional neural network reconstruction module. Its specific implementation is as follows:

[0013] The input data is the light detection value of the tissue body φ boundary , the output is the two-dimensional data feature map data after data preprocessing feature :

[0014] data feature =f Mapping (f Norm (log(φ boundary )))

[0015] Among them, φ boundary is the one-dimensional near-infrared light signal collected on the surface of the tissue, f Norm represents the normalized calculation, f Mapping Represents the mathematical mapping of data from one dimension to two dimensions. feature is the obtained two-dimensional data feature map.

[0016] The function of the two-dimensional space physical information loss function module is to constrain the training of the two-dimensional convolutional neural network reconstruction module by calculating the diffusion equation. Its specific implementation is as follows:

[0017] A special first-order differential operator is used to calculate the known light intensity value Φ(r) at each position of the tissue and the optical parameter distribution f of the tissue reconstructed by the neural network. θ (φ boundary ) to find the gradient, and then bring the obtained result and other known quantities into the following formula to constrain the training of the neural network:

[0018]

[0019] where f θ represents a neural network, θ represents the network parameters to be trained, q0(r) is a known light source term, and Φ(r) is a known distribution of tissue light intensity values ​​(q0(r) and Φ(r) are both given by the open source software package NIRFAST in MATLAB). In the above formula:

[0020]

[0021] This is the rewritten diffusion equation. Ideally, the result of the above equation should be 0. The absorption coefficient μ at r in the tissue is a (r) can be reduced to the scattering coefficient μ from the known tissue volume s (r)' and the distribution of optical parameters of tissue predicted by neural network f θ (φ boundary ) is calculated, and the relationship is as follows:

[0022]

[0023] The function of the two-dimensional convolutional neural network reconstruction module is to extract the features of the two-dimensional data feature map and restore the optical parameter distribution map of the tissue to be tested. The network structure and specific implementation of the neural network are as follows:

[0024] The two-dimensional convolutional neural network reconstruction module consists of a dimension expansion module, a downsampling module and an upsampling module.

[0025] 1) The dimension expansion module uses 32 3×3 convolution kernels to expand the dimension of the input two-dimensional data feature map from 1 to 32 while ensuring that the size of the two-dimensional data feature map remains unchanged. Its purpose is to expand the total amount of available features of the data. The dimension expansion module can be expressed as follows:

[0026] data output =f2(f1(data input ,ω1,b1),ω2,b2)

[0027] Where f1 represents the two-dimensional convolution layer and batch normalization layer, ρ1, b1 are the weights and biases of the two-dimensional convolution layer; f2 represents the deep feature extraction layer and batch normalization layer, ω2, b2 are the weights and biases of the deep feature extraction layer;

[0028] 2) The downsampling module uses three identical downsampling blocks to extract high-dimensional features of the data. The network layers of the downsampling block are, in order: maximum pooling layer, downsampling two-dimensional convolution layer and batch normalization layer, deep feature extraction layer and batch normalization layer. The function of the maximum pooling layer is to reduce the length and width of the feature map to 1 / 2 of the original, and the function of the downsampling two-dimensional convolution layer is to expand the dimension of the feature map to twice the original without changing the size of the feature map; the functions of the deep feature extraction layer and batch normalization layer are consistent with those of the deep feature extraction layer and batch normalization layer in the dimensional expansion module, which will not be repeated here. After three downsampling modules, the dimension of the feature is 256. The downsampling module can be expressed by the following formula:

[0029] data output =f deepdown (f down (D(data input ),ω down , b down ),ω deepdown , b deepdown ) where D represents the downsampling layer; f down represents the downsampling 2D convolutional layer and batch normalization layer, ω down , b down are the weights and biases of the down-sampled 2D convolutional layer; f deepdown represents the deep feature extraction layer and batch normalization layer, ω deepdown , b deepdown The weights and biases of the deep feature extraction layer;

[0030] 3) The function of the upsampling module is to enlarge the size of the feature map to restore the characteristics of the optical parameter distribution of the tissue to be tested. It contains two upsampling blocks, of which the first two upsampling blocks have the same structure, and the network layers are in the following order: upsampling layer, upsampling two-dimensional volume layer and batch normalization layer, deep feature extraction layer and batch normalization layer. Among them, the upsampling layer uses the nearest neighbor difference to expand the length and width of the feature map to twice the original while ensuring that the dimension of the feature map remains unchanged; the function of the upsampling two-dimensional volume layer and batch normalization layer is to reduce the dimension of the feature map to 1 / 2 of the original while ensuring that the length and width of the feature map remain unchanged; and the role of the deep feature extraction layer and batch normalization layer is consistent with that of the deep feature extraction layer and batch normalization layer in the dimension expansion module, which will not be repeated here. The first and second upsampling blocks can be expressed by the following formula:

[0031] data output =f deepup (f up (U(data input ),ω up , b up ),ω deepup , bdeepup )

[0032] Among them, U represents the upsampling layer; f up represents the upsampling 2D convolutional layer and batch normalization layer, ω up , b up are the weights and biases of the upsampling 2D convolutional layer; f deepup represents the deep feature extraction layer and batch normalization layer, ω deepup , b deepup The weights and biases of the deep feature extraction layer;

[0033] The third upsampling block uses a different structure from the first two upsampling blocks, which consists of an upsampling layer and three upsampling two-dimensional convolutional layers. The upsampling layer uses the nearest neighbor difference to expand the length and width of the feature map to twice the original size while keeping the feature map dimension unchanged; the upsampling two-dimensional convolutional layer reduces the dimension of the feature map to 1 / 2 of the original size while keeping the length and width of the feature map unchanged. After passing through the above three upsampling blocks, the predicted optical parameter distribution map of the tissue to be tested can be obtained. The third upsampling layer can be expressed by the following formula:

[0034] data output =f up3 (f up2 (f up1 (U(data input ),ω up1 ,b up1 ),ω up2 , b up2 ),ω up3 , b up3 )

[0035] Among them, U represents the upsampling layer; f up1 represents the upsampling 2D convolutional layer and batch normalization layer, ω up1 ,b up1 are the weights and biases of the upsampling 2D convolutional layer; f up2 and f up3 represents the upsampling 2D convolutional layer, ω up2 ,b up2 ,ω up3 ,b up3 These are the weights and biases of the second and third upsampling 2D convolutional layers, respectively.

[0036] Compared with the prior art, the present invention designs a diffusion optical tomography reconstruction method based on convolutional physical information neural network. This method innovatively adopts the diffusion equation as the loss function of the neural network, combines the physical information neural network with the two-dimensional convolution, and can directly reconstruct a high-quality optical parameter distribution map of the inside of the tissue to be tested after inputting the preprocessed data. Compared with the traditional method, this method has higher quality and faster speed; compared with the reconstruction algorithm based on neural network, the method in this article has better generalization and physical interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a schematic diagram of the phantom, where the brighter circular area is the real abnormal area.

[0038] Figure 2 Schematic diagram of the encoder-decoder network model structure.

[0039] Figure 3 Model training and inference process of the diffusion optical tomography reconstruction method based on convolutional physical information neural network.

[0040] Figure 4 The optical parameter distribution map of the tissue to be measured reconstructed by the reconstruction method proposed in the present invention (with Figure 1 for the same tissue). DETAILED DESCRIPTION

[0041] In order to make the objectives, technical solutions and advantages of the present invention more clear, the following is a further detailed description of the implementation mode of the present invention in combination with specific embodiments and with reference to the accompanying drawings.

[0042] Step 1: Establish a mathematical model of light transmission in tissue. The radiative transfer equation (RTE, hereinafter referred to as the RTE equation) is a differential equation that describes the attenuation of electromagnetic wave radiation when it propagates in a medium. It can be derived from the law of conservation of energy. In short, RTE states that a light beam loses energy through divergence and extinction (including absorption and scattering of the principle light beam), and gains energy from the light source in the medium and the scattering of the directional light beam. For general tissue bodies, their optical properties such as refractive index, absorption coefficient and scattering coefficient of tissue, and scattering anisotropy are considered to be invariant to time, but vary with space. The RTE equation is as follows:

[0043]

[0044] where c is the speed of light in the tissue, determined by the relative refractive index; is the phase function, which represents the probability that light in the propagation direction is scattered around a solid angle. In most cases, the phase function depends only on the angle between the scattering direction and the incident direction; is the light source term. In actual situations, since the RTE equation contains both angle variables and control variables, and is an integral-differential equation, its calculation is extremely complex and requires a large amount of computing resources and time. In addition, when the shape of the phantom is not a regular geometric body, the RTE equation is often unable to calculate an analytical solution, and generally a numerical solution is used to give an accurate and quantitative distribution of the RTE equation solution. Therefore, in practical applications, the approximate equation of the RTE equation, the diffusion equation (DE, hereinafter referred to as the DE equation), is often used to simplify the calculation and reduce the consumption of computing resources. In low-scattering areas, the DE equation has a better simulation effect. In the present invention, the light transmission model used for diffuse optical tomography is the DE equation.

[0045] Step 2: Discrete numerical solution of the diffusion approximation equation.

[0046] Diffuse optical tomography (DOT) includes two processes:

[0047] 1) In the forward process, the light source emits near-infrared light to illuminate the biological tissue to be tested, and the diffusion equation physical model calculates the intensity of the emitted light at a given position on the surface of the tissue:

[0048]

[0049] 2) In the reverse process, when the intensity of the emitted light at a given position on the surface of the tissue is known, the distribution of the optical parameters inside the tissue is continuously approximated by the Tikhonov regularization method and the forward model, in order to obtain the optical parameter distribution of the tissue that best matches the known intensity of the emitted light:

[0050]

[0051] in, It is the detection value obtained by the detector arranged at a given position on the surface of the object to be measured; is the detector detection value obtained by forward model calculation of the predicted optical parameters; μ0 is the initial estimate of the optical parameters, which is generally obtained through a data calibration procedure, or the prior distribution of the existing optical parameters. When μ0 is not much different from the actual parameter distribution, this item can be ignored; NM is the total number of detectors deployed; NN is the total number of nodes after the phantom to be tested is discretized by finite element, and λ is the Tikhonov regularization parameter, which is defined as the ratio of the variance of the measured data to the variance of the optical properties:

[0052]

[0053] Step 3: Generation of training data set. The size of the numerical phantom used in the present invention is set to 40 mm × 40 mm, and is divided into 1681 finite element nodes and 3200 triangular mesh units by the finite element method. Figure 1 As shown. The rule for adding abnormal areas is: random position, diameter within the range of 4mm-7mm, and the contrast between the optical parameters and the background tissue is 5. The total number of light sources and detectors is 16, which are arranged equidistantly on each side of the phantom, with 4 on each side. The forward data of the phantom is calculated by the NIRFAST toolkit, and the detector detection light intensity value φ is saved boundary , the light intensity value Φ(r) of the entire phantom, the light source distribution q0(r) and the optical parameter distribution k(r) of the entire phantom are used as training. 10,000 sets of data are generated as training data and 1,000 sets of data are used as test data.

[0054] Step 4: Training of convolutional physical information neural network model.

[0055] In the model training stage, the detector detection light intensity value φ in the data boundary As the input of the network, the output is the predicted optical parameter distribution κ(r)' of the tissue to be tested. GTX3090 GPU 24G is used for training, the framework uses Pytorch2, the initial learning rate is set to 0.001, the regularization parameter is set to 0.001, and the training epoch is set to 1000.

[0056] 1): Input one-dimensional fluorescence signal φ boundary After mathematical mapping and normalization, the signal characteristic graph data is obtained feature .

[0057] 2): Signal feature graph data feature After the encoder-decoder network, the predicted optical parameter distribution κ(r)' of the tissue to be tested is obtained.

[0058] 3): Calculate the gradient of the network-predicted tissue optical parameter distribution κ(r)' and the light intensity value Φ(r) of the entire tissue to obtain and Then, other known quantities are introduced into equation (4) to constrain the training of the neural network and guide its convergence.

[0059] Step 5: Test of the diffusion optical tomography reconstruction method based on convolutional physical information neural network. In the model testing phase, the detector detects the light intensity value φ boundary As input (see Appendix Figure 3 The blue part) directly outputs the predicted distribution of the optical parameters of the tissue. Figure 4 See the reconstruction results of the abnormal area.

[0060] Experiments have shown that the method proposed in the present invention can effectively reconstruct the optical parameter distribution of a tissue and can physically explain it based on the diffusion equation theory.

Claims

1. A diffusion optical tomography reconstruction method based on a two-dimensional convolutional physical information neural network, characterized in that: Directly reconstruct the distribution of optical property parameters from the acquired signals through a two-dimensional convolutional physical information neural network; To directly reconstruct the optical characteristic parameter image from the collected signal, the collected signal is first normalized and then mapped to a two-dimensional data space, which can be expressed as follows: data feature =f Mapping (f Norm (log(φ boundary ))) Among them, φ boundary is the one-dimensional near-infrared light signal collected on the surface of the tissue, f Norm represents the normalized calculation, f Mapping Represents the mathematical mapping of data from one dimension to two dimensions. feature is the obtained two-dimensional data feature map; A two-dimensional convolutional physical information neural network is designed for reconstruction, and the loss function of the network is constructed using the diffusion approximation equation describing the propagation of light in tissues, which is expressed as follows: Among them, f θ represents a neural network, θ represents the network parameters that need to be trained, q0(r) is the light source term, and Φ(r) is the known light intensity distribution of the tissue; Design a two-dimensional convolutional physical information neural network for reconstruction, which includes a dimension expansion module, three downsampling modules and three upsampling modules; 1) Dimension expansion module, which is characterized by expanding the dimension of the feature map to increase the total amount of features available for learning, which can be expressed as: date output =f2(f1(data input ,ω1,b1),ω2,b2) (3) Where f1 represents the two-dimensional convolution layer and batch normalization layer, ω1, b1 are the weights and biases of the two-dimensional convolution layer; f2 represents the deep feature extraction layer and batch normalization layer, ω2, b2 are the weights and biases of the deep feature extraction layer; 2) Downsampling module, which reduces the size of the feature map to extract the high-dimensional features of the data, can be expressed as: data output =f deepdown (f down (D(data input ),oh down ,b down ),oh deepdown ,b deepdown ) (4) Among them, D represents the downsampling layer; f down represents the downsampling 2D convolutional layer and batch normalization layer, ω down ,b down are the weights and biases of the down-sampling 2D convolutional layer; f deepdown represents the deep feature extraction layer and batch normalization layer, ω deepdown ,b deepdown The weights and biases of the deep feature extraction layer; 3) The upsampling module includes two different upsampling blocks, wherein the first and second upsampling blocks are characterized by enlarging the size of the feature map while reducing the dimension of the feature map, which can be expressed as follows: data output =f deepup (f up (U(data input ),ω up ,b up ),ω deepup ,b deepup ) Among them, U represents the upsampling layer; f up represents the upsampling 2D convolutional layer and batch normalization layer, ω up ,b up are the weights and biases of the upsampling 2D convolutional layer; f deepup represents the deep feature extraction layer and batch normalization layer, ω deepup ,b deepup The weights and biases of the deep feature extraction layer; The third upsampling block restores the optical parameter distribution of the tissue to be tested based on the feature map, which can be expressed as follows: data output =f up3 (f up2 (f up1 (U(data input ),oh up1 ,b up1 ),oh up2 ,b up2 ),oh up3 ,b up3 ) Among them, U represents the upsampling layer; f up1 represents the upsampling 2D convolutional layer and batch normalization layer, ω up1 ,b up1 are the weights and biases of the upsampling 2D convolutional layer; f up2 and f up3 represents the upsampling 2D convolutional layer, ω up2 ,b up2 ,ω up3 ,b up3 These are the weights and biases of the second and third upsampling 2D convolutional layers, respectively.

2. The method for reconstruction of diffuse optical tomography based on two-dimensional convolutional physical information neural network according to claim 1, characterized in that: A mathematical model of light transmission in tissues is established; the radiation transfer equation RTE is a differential equation that describes the attenuation of electromagnetic radiation when it propagates in a medium; derived from the law of conservation of energy, the RTE equation is as follows: where c is the speed of light in the tissue, determined by the relative refractive index; is the phase function, which represents the probability that light in the propagation direction is scattered around a solid angle. In most cases, the phase function depends only on the angle between the scattering direction and the incident direction; is the light source term; the light transmission model used in diffuse optical tomography is the DE equation.

3. The method for reconstruction of diffuse optical tomography based on two-dimensional convolutional physical information neural network according to claim 2, characterized in that: Discrete numerical solution of diffusion approximation equation; Diffuse optical tomography, referred to as DOT technology, includes two processes: 1) In the forward process, the light source emits near-infrared light to illuminate the biological tissue to be tested, and the diffusion equation physical model calculates the intensity of the emitted light at a given position on the surface of the tissue: 2) In the reverse process, when the intensity of the emitted light at a given position on the surface of the tissue is known, the distribution of the optical parameters inside the tissue is continuously approximated by the Tikhonov regularization method and the forward model, in order to obtain the optical parameter distribution of the tissue that best matches the known intensity of the emitted light: in, It is the detection value obtained by the detector arranged at a given position on the surface of the object to be measured; is the detector detection value obtained by forward model calculation of the predicted optical parameters; μ0 is the initial estimate of the optical parameters; NM is the total number of detectors arranged; NN is the total number of nodes after finite element discretization of the phantom to be tested, and λ is the Tikhonov regularization parameter, which is defined as the ratio of the variance of the measured data to the variance of the optical properties:

4. The method for reconstruction of diffuse optical tomography based on two-dimensional convolutional physical information neural network according to claim 3, characterized in that: Generation of training data set; The size of the numerical phantom used was set to 40mm×40mm, and it was divided into 1681 finite element nodes and 3200 triangular mesh units by the finite element method; The rule for adding abnormal areas was as follows: random position, diameter within the range of 4mm-7mm, and the contrast between the optical parameters and the background tissue was 5; The total number of light sources and detectors was 16, which were arranged equidistantly on each side of the phantom, with 4 on each side; The forward data of the phantom was calculated by the NIRFAST toolkit, and the detector detection light intensity value φ was saved boundary , the light intensity value Φ(r) of the entire phantom, the light source distribution q0(r) and the optical parameter distribution κ(r) of the entire phantom are used as training; 10,000 sets of data are generated as training data and 1,000 sets of data are used as test data.

5. The method for reconstruction of diffuse optical tomography based on two-dimensional convolutional physical information neural network according to claim 4, characterized in that: Training of convolutional physical information neural network models; In the model training stage, the detector detection light intensity value φ in the data boundary As the input of the network, the output is the predicted optical parameter distribution κ(r)' of the tissue to be tested; GTX3090 GPU 24G is used for training, the framework uses Pytorch2, the initial learning rate is set to 0.001, the regularization parameter is set to 0.001, and the training epoch is set to 1000; 1): Input one-dimensional fluorescence signal φ boundary After mathematical mapping and normalization, the signal characteristic graph data is obtained feature ; 2): Signal feature graph data feature After the encoder-decoder network, the predicted optical parameter distribution κ(r)' of the tissue to be tested is obtained; 3): Calculate the gradient of the network-predicted tissue optical parameter distribution κ(r)' and the light intensity value Φ(r) of the entire tissue to obtain and Then, other known quantities are introduced into equation (4) to constrain the training of the neural network and guide its convergence.

6. The method for reconstruction of diffuse optical tomography based on two-dimensional convolutional physical information neural network according to claim 5, characterized in that: Test of the diffusion optical tomography reconstruction method based on convolutional physical information neural network; During the model testing phase, the detector detects the light intensity value φ boundary As input, the predicted distribution of optical parameters of the tissue body is directly output.

Images