Optical absorption coefficient resolving system and method based on photoacoustic imaging and non-training network

Abstract

The invention belongs to the field of biomedical photonics, photoacoustic imaging and deep learning, and provides an optical absorption coefficient resolving system and method based on photoacoustic imaging and a non-training network, and the method comprises the steps: taking physically-driven Monte Carlo simulation as an optical forward model, and embedding the optical forward model into a neural network frame which does not need to be trained in advance, direct and high-precision inversion from a single initial sound pressure distribution image to optical absorption coefficient distribution is realized through a self-constraint resolving process.

Description

Technical Field

[0001] This invention belongs to the fields of biomedical photonics, photoacoustic imaging, and deep learning, and in particular relates to a system and method for calculating the optical absorption coefficient based on photoacoustic imaging and an untrained network. Background Technology

[0002] The optical absorption coefficient is one of the core parameters characterizing the optical properties of biological tissues. It reflects the ability of biological tissues to absorb light energy of specific wavelengths and is closely related to blood content, blood oxygen saturation, and metabolic activity within the tissue. In the field of biomedical imaging, the accurate determination of the optical absorption coefficient is of great significance for achieving early tumor detection, blood oxygen metabolism monitoring, and tissue function assessment, and is also an important foundation for quantitative analysis of histopathology.

[0003] Existing methods for acquiring tissue optical parameters include optical coherence tomography (OCT) and diffuse optical tomography (DOT). These methods can achieve high optical resolution in shallow tissues; however, they are significantly affected by light scattering and have limited imaging depth, making it difficult to achieve both spatial resolution and imaging depth simultaneously for precise quantification in deep tissues. Photoacoustic imaging technology combines the high contrast of optical imaging with the high resolution and strong tissue penetration of ultrasound imaging, offering significant advantages in the structural and functional imaging of biological tissues. This technology uses short-pulse laser irradiation of tissue, causing instantaneous thermoelastic expansion in areas absorbing light energy, thereby exciting broadband ultrasonic signals. By detecting and reconstructing these acoustic signals, an initial sound pressure distribution image within the tissue can be obtained. The initial sound pressure distribution p0 is related to the tissue's optical absorption coefficient μ. a The luminous flux distribution Φ and the Grüneisen coefficient Γ satisfy the following relationship: p0=Γ·μ a ·Φ, where Γ is usually considered a constant. The process of solving for the optical absorption coefficient distribution from the initial sound pressure distribution is called quantitative photoacoustic imaging. This process belongs to the category of optical inversion and is a key step in realizing the quantification of photoacoustic imaging.

[0004] Currently, the methods for solving the optical absorption coefficient based on photoacoustic imaging technology mainly include the following three categories: (1) Iterative solution method: using a photon transport mathematical model (including the radiative transport equation and its simplified approximation equation, such as the diffusion equation) or a photon transport numerical model (including Monte Carlo simulation method based on stochastic statistics, finite element-finite difference method, etc.) as the forward model, the light flux distribution inside the tissue is calculated, and the predicted initial sound pressure distribution is obtained by combining the optical characteristic parameters. By continuously iterating and optimizing the light flux distribution and absorption coefficient distribution, the difference between the predicted initial sound pressure and the measured initial sound pressure is minimized, thereby gradually reversing the optical absorption coefficient distribution of the tissue. This type of method has clear physical interpretability and high quantitative accuracy, but it has problems such as slow convergence speed and poor nonlinear fitting ability. This method generally requires the pre-setting of boundary conditions. The more complex the boundary conditions, the more accurate the theoretical result, but the amount of calculation also increases significantly, often leading to slower iteration speed, decreased stability, or even non-convergence; and with the increase of the number of iterations, the image resolution may decrease. (2) Joint reconstruction-based methods: By introducing other modal imaging information (such as nuclear magnetic resonance imaging, optical coherence tomography, diffusion optical tomography, etc.) to assist in photoacoustic inversion, the accuracy of optical absorption coefficient solution can be improved. The introduction of multimodal imaging can compensate for the underdeterminacy problem of photoacoustic inversion to a certain extent, but the system construction is relatively complex, the image registration between modalities is difficult, the equipment cost is high, and it is difficult to achieve generalization and clinical application. (3) Deep learning-based methods: In recent years, deep learning has been widely used in optical absorption coefficient inversion tasks due to its powerful nonlinear fitting ability. Compared with traditional iterative methods, deep learning can directly establish the mapping relationship from sound pressure distribution to absorption coefficient distribution, avoiding complex boundary condition settings. However, this type of method usually relies on a large amount of labeled data for training, and has high requirements for the scale and quality of the dataset; when the distribution of test data and training set differs greatly, the prediction results often show large deviations. In addition, the neural network training process requires a lot of time and computing resources, and the training process may last for several hours to several days. Although existing research has attempted to reduce the dependence on labeled data through unsupervised or self-supervised learning, it is still necessary to build a certain amount of dataset and train the network. It has not yet truly gotten rid of data dependence, and thus it is difficult to guarantee universality across systems.

[0005] Existing methods for calculating optical absorption coefficients using photoacoustic imaging principles still have certain shortcomings in terms of computational accuracy and complexity, result convergence, system complexity, and data dependence. Therefore, there is an urgent need for a novel method for calculating optical absorption coefficients that possesses strong nonlinear fitting capabilities, low system complexity, and is independent of data dependence. This would improve the accuracy and versatility of absorption coefficient calculations, providing reliable data for biomedical applications such as tissue lesion detection, tumor detection and diagnosis, and blood oxygen distribution estimation. Ultimately, this would promote the widespread application of photoacoustic imaging technology in clinical diagnosis, treatment monitoring, and personalized medicine. Summary of the Invention

[0006] The purpose of this invention is to overcome the bottlenecks in computational efficiency, data dependence, and system generalization ability in the existing technology, and to provide a more efficient, universal, and practical optical absorption coefficient calculation scheme.

[0007] This invention provides a system and method for solving optical absorption coefficients based on photoacoustic imaging and untrained networks. Its core lies in embedding a physical-driven Monte Carlo simulation as an optical forward model into a neural network framework that does not require pre-training. Through a self-constrained solution process, it achieves direct and high-precision inversion from a single initial sound pressure distribution image to the optical absorption coefficient distribution.

[0008] This invention proposes an optical absorption coefficient calculation system based on photoacoustic imaging and an untrained network, comprising:

[0009] The data acquisition module acquires ultrasonic signals excited by short-pulse lasers and reconstructs the initial sound pressure distribution image p0 of the target body using acoustic reconstruction algorithms, including back-projection algorithms and delay summation algorithms.

[0010] The Monte Carlo simulation optical forward model construction module builds an optical Monte Carlo simulation optical forward model and initializes the optical forward model parameters. Specifically, it sets a matching light source model in the Monte Carlo simulation based on the actual light source configuration of the photoacoustic imaging system (such as beam shape and energy distribution). It also presets empirical values ​​for optical characteristic parameters that match the target biological tissue type, including the optical scattering coefficient μ. s Anisotropy factor g, refractive index n, etc.

[0011] The untrained neural network construction module builds an untrained neural network to learn the mapping relationship from a random noise vector to the distribution of the target optical absorption coefficient; initializes the parameters of the untrained neural network, sets the optimization hyperparameters of the untrained neural network, including the learning rate, maximum number of computations, and the convergence threshold of the loss function; and initializes a random array matrix as the fixed input of the neural network.

[0012] The self-constrained solution module inputs a random array matrix into an untrained neural network, and the network outputs a predicted value μ of the optical absorption coefficient distribution. a _pred; will change μ a The _pred and preset optical characteristic parameters are input into the forward optical model of Monte Carlo simulation to obtain the luminous flux distribution Φ inside the target body. Based on the luminous flux distribution, the predicted initial sound pressure distribution p0_pred is calculated. The loss value between the predicted initial sound pressure distribution p0_pred and the initial sound pressure distribution p0 obtained by the data acquisition module is calculated. An optimization algorithm is used to perform inverse optimization on the network, and the final output of the network is the optical absorption coefficient distribution μ. a _pred is the desired solution result.

[0013] The acceleration module was optimized by using the Fourier domain to speed up the solution process.

[0014] This invention further introduces an optimization strategy based on the Fourier domain to significantly accelerate the calculation process and improve the practicality and efficiency of the method.

[0015] The optimization and acceleration module divides the solution process into three stages: the early stage of computation, the middle stage of computation, and the late stage of computation.

[0016] In the early stages of the computation, the Monte Carlo simulation uses the first number of photons; in the middle stages, it uses the second number of photons; and in the later stages, it uses the third number of photons, with the third number of photons > the second number of photons > the first number of photons.

[0017] This invention proposes a method for calculating the optical absorption coefficient based on photoacoustic imaging and an untrained network, comprising the following steps:

[0018] Step S1: Acquire the ultrasonic signal excited by the short pulse laser, and reconstruct the initial sound pressure distribution image p0 of the target body using an acoustic reconstruction algorithm.

[0019] Step S2: Construct an optical Monte Carlo simulation forward optical model and initialize the parameters of the forward optical model; and preset empirical values ​​of optical characteristic parameters that match the target biological tissue type, including the optical scattering coefficient μ. s Anisotropy factor g, refractive index n.

[0020] Step S3: Construct an untrained neural network to learn the mapping relationship from the random noise vector to the distribution of the target optical absorption coefficient; initialize the parameters of the untrained neural network, and initialize a random array matrix as the fixed input of the neural network.

[0021] Step S4: Input the random array matrix into the untrained neural network, and the network outputs the predicted value μ of the optical absorption coefficient distribution. a_pred; will change μ a The _pred and preset optical characteristic parameters are input into the forward optical model of Monte Carlo simulation to obtain the luminous flux distribution Φ inside the target body. Based on the luminous flux distribution, the predicted initial sound pressure distribution p0_pred is calculated. The loss value between the predicted initial sound pressure distribution p0_pred and the initial sound pressure distribution p0 obtained by the data acquisition module is calculated. An optimization algorithm is used to perform inverse optimization on the network, and the final output of the network is the optical absorption coefficient distribution μ. a _pred is the desired solution result.

[0022] Step S5: The Fourier domain is used to accelerate the solution process in step S4.

[0023] In the early stages of the computation, the Monte Carlo simulation uses the first number of photons; in the middle stages, it uses the second number of photons; and in the later stages, it uses the third number of photons, with the third number of photons > the second number of photons > the first number of photons.

[0024] Compared with the prior art, the technical solution proposed in this invention brings the following significant beneficial effects:

[0025] 1. Achieving a balance between high precision and physical consistency: This invention incorporates Monte Carlo simulation—the gold standard physical model for studying light propagation in biological tissues—into the optimization process, ensuring that the solution process is strictly constrained by physical laws. The resulting optical absorption coefficient distribution not only has a precise mapping relationship with measured sound pressure data but also possesses clear physical meaning, guaranteeing the accuracy and interpretability of the quantitative results.

[0026] 2. It avoids the convergence problem of traditional iterative methods: By utilizing the powerful nonlinear mapping capability of deep neural networks, this method does not require manual pre-setting of complex boundary conditions during the inversion process, thereby fundamentally avoiding problems such as iterative oscillation, slow convergence speed or non-convergence, and resolution reduction caused by it, and ensuring the stability and convergence reliability of the solution process.

[0027] 3. Completely eliminate data dependence and improve practicality and universality: Due to the adoption of a network architecture that does not require pre-training, this method overcomes the stringent requirements of traditional deep learning for massive amounts of labeled data. It is particularly suitable for clinical photoacoustic imaging research where it is difficult to obtain real labeled data, which greatly reduces the threshold and cost of technology application.

[0028] 4. Excellent generalization ability and system robustness: The generalization ability of this method is reflected in the decoupling of its algorithmic framework from specific imaging systems, and the fact that no other modalities need to be introduced for compensation. Therefore, when facing different photoacoustic imaging systems, this method can exhibit strong cross-platform adaptability without the need to redesign the model or train the neural network for new systems.

[0029] 5. Significantly improved computational efficiency: By introducing an optimization strategy based on the Fourier domain, namely dynamically scheduling the computational resources of Monte Carlo simulation, a significant improvement in computational efficiency is achieved, enabling this high-precision quantitative method to be solved quickly even with lower hardware resources, thus possessing greater potential for clinical applications. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This invention presents an optical absorption coefficient calculation system based on photoacoustic imaging and an untrained network.

[0032] Figure 2 This is a diagram of the untrained neural network structure used in this invention;

[0033] Figure 3 This is a diagram of the non-trained neural network computation block structure of the present invention;

[0034] Figure 4 This is a schematic diagram of the self-constrained solution process of the present invention;

[0035] Figure 5 To explore the potential spectral characteristics of neural networks and Monte Carlo simulations in the Fourier domain;

[0036] Figure 6 This is a flowchart of the optical absorption coefficient calculation method based on photoacoustic imaging and untrained network proposed in this invention.

[0037] Figure 7 The diagram shows the results of the absorption coefficient calculation method of this invention. Detailed Implementation

[0038] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0039] like Figure 1 As shown, this invention proposes an optical absorption coefficient calculation system based on photoacoustic imaging and an untrained network, comprising:

[0040] The data acquisition module acquires ultrasonic signals excited by short-pulse lasers and reconstructs the initial sound pressure distribution image of the target body using acoustic reconstruction algorithms, including back-projection algorithms and delay summation algorithms.

[0041] The Monte Carlo simulation optical forward model construction module builds an optical Monte Carlo simulation optical forward model and initializes the optical forward model parameters. Specifically, it sets a matching light source model in the Monte Carlo simulation based on the actual light source configuration of the photoacoustic imaging system (such as beam shape and energy distribution). It also presets empirical values ​​for optical characteristic parameters that match the target biological tissue type, including the optical scattering coefficient μ. s Anisotropy factor g, refractive index n, etc.

[0042] The untrained neural network construction module builds an untrained neural network to learn the mapping relationship from a random noise vector to the distribution of the target optical absorption coefficient; initializes the parameters of the untrained neural network, sets the optimization hyperparameters of the untrained neural network, including the learning rate, maximum number of computations, and the convergence threshold of the loss function; and initializes a random array matrix as the fixed input of the neural network.

[0043] like Figure 2 As shown, the untrained neural network consists of k sequentially connected computational blocks; each computational block is composed of an upsampling layer, a convolutional activation layer, and a residual connection layer in sequence, as shown below. Figure 3 As shown in the diagram. The upsampling layer adjusts the spatial size of the feature map, progressively upsampling it to the same scale as the target optical absorption coefficient distribution image. The convolutional activation layer, composed of convolutional layers and activation functions, is responsible for nonlinear transformation and extraction of features. The residual connection layer adds the input of the current computation block to the output of the convolutional activation layer through skip connections, mitigating the gradient vanishing problem in deep networks and promoting network optimization.

[0044] The self-constrained solution module inputs a random array matrix into an untrained neural network, and the network outputs a predicted value μ of the optical absorption coefficient distribution. a _pred; will change μ a The _pred and preset optical characteristic parameters are input into the forward optical model of Monte Carlo simulation to obtain the luminous flux distribution Φ inside the target body. Based on the luminous flux distribution, the predicted initial sound pressure distribution p0_pred is calculated. The loss value between the predicted initial sound pressure distribution p0_pred and the initial sound pressure distribution p0 obtained by the data acquisition module is calculated. An optimization algorithm is used to perform inverse optimization on the network, and the final output of the network is the optical absorption coefficient distribution μ. a _pred is the desired solution result.

[0045] like Figure 4As shown, the self-constrained solution module first inputs a random array matrix into an untrained neural network, and the network outputs a predicted value μ of the optical absorption coefficient distribution. a _pred; will change μ a _pred and the preset optical scattering coefficient μ s Optical characteristic parameters such as anisotropy factor g and refractive index n are input into the forward optical model of Monte Carlo simulation to obtain the luminous flux distribution Φ inside the target body. According to the photoacoustic equation p0_pred=Γ·μ a _pred·Φ yields the predicted initial sound pressure distribution p0_pred. A loss function is then used to quantify the difference between the predicted initial sound pressure distribution p0_pred and the simulated or experimentally measured initial sound pressure distribution p0, i.e., to calculate the loss value.

[0046] Next, using the backpropagation algorithm, the gradient of the loss value with respect to all adjustable parameters of the neural network is calculated. The gradient descent optimization algorithm is then used to update the neural network parameters based on this gradient. After the parameter update, the network generates a new, optimized μ from the random data matrix again. a The above steps are repeated. This calculation process continues, constantly narrowing the difference between the predicted and measured sound pressure levels until the following conditions are met: the loss value is below a preset convergence threshold, or the number of calculations reaches a preset maximum. When the loop terminates, the final output optical absorption coefficient distribution μ of the network is obtained. a _pred is the desired solution result.

[0047] The acceleration module was optimized by using the Fourier domain to speed up the solution process.

[0048] The Monte Carlo simulation used in the above-mentioned solution method inherently involves a trade-off between computational accuracy and resource requirements. While the results become more reliable and accurate with an increased number of photons, the computational demands and time increase significantly. This invention further introduces an optimization strategy based on the Fourier domain to significantly accelerate the computation process and improve the practicality and efficiency of the method.

[0049] This strategy is based on the spectral characteristics shared by deep learning optimization processes and Monte Carlo statistical processes in the Fourier domain. Both prioritize fitting or presenting the low-frequency (macro) information of the data, and then gradually converge on the high-frequency (detail) information. Figure 5 The aforementioned spectral characteristics were visualized using a digital mouse model. Figure 5 (a) illustrates the potential frequency domain characteristics during neural network computation. In the initial stages of computation, the network output almost entirely reflects low-frequency information. As the number of computations increases, the neural network gradually fits high-frequency information, and the output reflects more detailed information. Similarly, Figure 5(b) demonstrates the potential spectral characteristics of the Monte Carlo simulation method. At low photon counts, the luminous flux distribution output by the Monte Carlo simulation primarily contains low-frequency information. As the photon count increases, the luminous flux distribution calculated by the Monte Carlo simulation reflects more high-frequency information. Based on this insight, the core solution method does not always employ a fixed high photon count for Monte Carlo simulations. Instead, an adaptive photon count scheduling scheme is implemented to avoid wasting computational resources by using excessive photons.

[0050] The optimization and acceleration module divides the solution process into three stages: the early stage of computation, the middle stage of computation, and the late stage of computation.

[0051] In the early stages of computation, when the network predictions are relatively coarse and the loss values ​​are large, Monte Carlo simulations use the first number of photons for calculation. Although the simulation results at this stage contain some statistical noise, they can reliably capture the low-frequency backbone information of the light flux distribution, which is sufficient to efficiently guide the network to make correct macroscopic structural adjustments and quickly reduce the loss value. Because low-photon-count simulations are faster, the computational efficiency at this stage is significantly improved.

[0052] In the middle stage of computation, as the calculation progresses, when the loss function decreases to a specific range or the computation enters the middle stage (e.g., the number of calculations reaches 10-20% or more of the preset maximum number of calculations), the macroscopic fitting of the network is basically complete. At this stage, the number of photons used in the Monte Carlo simulation is gradually increased, and a second number of photons is used for calculation. The higher-precision simulation can provide richer mid- and high-frequency information, guiding the network from contour fitting to detail optimization.

[0053] In the later stages of computation (e.g., when the number of calculations is above 80-90% of the preset maximum number of calculations), the loss function converges further. At this point, a high photon number Monte Carlo simulation is performed. This stage of high-precision optical forward computation provides the network with the most complete spectral information. The use of a third photon number ensures that it corrects for any remaining small errors, guaranteeing the accuracy and reliability of the solution results in the microstructure.

[0054] Among them, the third photon number > the second photon number > the first photon number.

[0055] Through the above dynamic optimization strategy, this invention avoids the huge computational burden caused by using high photon number simulation throughout the process, thereby improving computational efficiency and making the solution method consume less computational resources.

[0056] like Figure 6 As shown, this invention proposes a method for calculating the optical absorption coefficient based on photoacoustic imaging and an untrained network, comprising the following steps:

[0057] Step S1: Acquire the ultrasonic signal excited by the short pulse laser, and reconstruct the initial sound pressure distribution image of the target body using an acoustic reconstruction algorithm;

[0058] Specifically, acoustic reconstruction algorithms include back-projection algorithms, delay summation algorithms, etc.

[0059] Step S2: Construct an optical Monte Carlo simulation optical forward model and initialize the parameters of the optical forward model;

[0060] Specifically, based on the actual light source configuration of the photoacoustic imaging system, such as beam shape and energy distribution, a matching light source model is set in the Monte Carlo simulation; and empirical values ​​of optical characteristic parameters, including the optical scattering coefficient μ, are preset to match the target biological tissue type. s Anisotropy factor g, refractive index n, etc.

[0061] Step S3: Construct an untrained neural network to learn the mapping relationship from the random noise vector to the target optical absorption coefficient distribution; initialize the parameters of the untrained neural network, set the optimization hyperparameters of the untrained neural network, including the learning rate, maximum number of computations, and the convergence threshold of the loss function; and initialize a random array matrix as the fixed input of the neural network.

[0062] Step S4: Calculate the optical absorption coefficient distribution μ based on the Monte Carlo simulation optical forward model and the untrained neural network. a _pred;

[0063] The self-constrained solution module first inputs a random array matrix into an untrained neural network, and the network outputs a predicted value μ of the optical absorption coefficient distribution. a _pred; will change μ a _pred and the preset optical scattering coefficient μ s Optical characteristic parameters such as anisotropy factor g and refractive index n are input into the forward optical model of Monte Carlo simulation to obtain the luminous flux distribution Φ inside the target body. According to the photoacoustic equation p0_pred=Γ·μ a _pred·Φ yields the predicted initial sound pressure distribution p0_pred. A loss function is then used to quantify the difference between the predicted initial sound pressure distribution p0_pred and the simulated or experimentally measured initial sound pressure distribution p0, i.e., to calculate the loss value.

[0064] Next, using the backpropagation algorithm, the gradient of the loss value with respect to all adjustable parameters of the neural network is calculated. The gradient descent optimization algorithm is then used to update the neural network parameters based on this gradient. After the parameter update, the network generates a new, optimized μ from the random data matrix again. aThe above steps are repeated. This calculation process continues, constantly narrowing the difference between the predicted and measured sound pressure levels until the following conditions are met: the loss value is below a preset convergence threshold, or the number of calculations reaches a preset maximum. When the loop terminates, the final output optical absorption coefficient distribution μ of the network is obtained. a _pred is the desired solution result.

[0065] Step S5: The Fourier domain is used to accelerate the solution process in step S4.

[0066] In the early stages of the computation, the Monte Carlo simulation uses the first number of photons; in the middle stages, it uses the second number of photons; and in the later stages, it uses the third number of photons, with the third number of photons > the second number of photons > the first number of photons.

[0067] In a specific embodiment of this paper, the optical absorption coefficient of the reconstructed digital mouse model is taken as an example:

[0068] The first step is to obtain a digital mouse model containing optical absorption coefficient distribution data from an open-source dataset. In this embodiment, the abdominal region of the digital mouse model is selected, which includes soft tissues such as the liver, kidneys, pancreas, and skin. An initial sound pressure distribution image is obtained through simulation and used as input data for the solution method.

[0069] The second step involves configuring a light source model consistent with the simulation conditions in the Monte Carlo optical forward model (e.g., using a GPU-accelerated Monte Carlo simulation as the forward model), and setting initial empirical values ​​for the optical characteristic parameters of the digital mouse model, where the scattering coefficient μ... s The setting range is 1 to 10 mm. -1 The anisotropy factor g is set to 0.9, and the refractive index n is set to 1.37.

[0070] The third step, which can be performed simultaneously with the second step, involves setting the hyperparameters of the untrained neural network: an initial learning rate of 0.01, a maximum number of computations of 3000, and a loss function convergence threshold of 10. -8 Initialize a random array matrix (e.g., an array matrix of size 8×8×256) as the fixed input to the neural network.

[0071] The fourth step involves inputting the aforementioned array matrix into an untrained neural network to obtain the predicted distribution of absorption coefficients (e.g., the output absorption coefficient distribution data size is 256×256×1, and the actual physical size represented by each pixel is 0.1mm×0.1mm). This data, along with other preset optical characteristic parameters, is then input into a Monte Carlo simulation to calculate the luminous flux. During this process, a Fourier domain-based optimization strategy is executed, reducing the photon number from 10... 4Up to 10 7 Gradually increase the number of calculations. In the first 10% of the calculations (i.e., the first 300 calculations), use 10. 4 Monte Carlo simulations were performed using photon numbers; in the last 10% of the calculations (i.e., calculations 2700 to 3000), 10-1 was used. 7 Photon number; when the number of calculations is in the middle range (i.e., from the 300th to the 2700th calculation), the photon number increases from 10. 4 Gradually increase to 10 7 The number of photons added in each calculation is (10 7 -10 4 ) / (2700-300)≈4000. This means that a small number of photons are used in the initial stage of the calculation for rapid simulation to capture low-frequency information, and the number of photons is increased in the later stages of the calculation to optimize details. The initial sound pressure level is predicted using the light flux obtained above, and the loss is calculated by comparing it with the input initial sound pressure level. The network parameters are then updated through backpropagation.

[0072] Fifth, repeat the calculation process of step four until the loss value is lower than the preset threshold or the maximum number of calculations is reached; after the loop terminates, the final optical absorption coefficient distribution output by the network is taken as the final solution result of this embodiment.

[0073] The sixth step involves quantitative verification using evaluation metrics such as the mean squared error (MSE) and structural similarity index (SSIM) between the calculated results and the actual absorption coefficient distribution of the digital mouse model. Figure 7 As shown, Figure 7 (a) is the initial sound pressure distribution image. Figure 7 (b) is an image of the true optical absorption coefficient distribution. Figure 7 (c) is the optical absorption coefficient distribution image obtained by this solution method. Quantitative evaluation shows that the method shown in this embodiment can effectively reconstruct the optical absorption coefficient distribution in biological tissues.

[0074] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0075] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0076] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0077] In the embodiments provided by this invention, it should be understood that the disclosed apparatus / terminal devices and methods can be implemented in other ways. For example, the apparatus / terminal device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0078] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0079] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0080] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A system for calculating the optical absorption coefficient based on photoacoustic imaging and an untrained network, characterized in that, The solution system includes: The data acquisition module acquires ultrasonic signals excited by short-pulse lasers and reconstructs the initial sound pressure distribution image p0 of the target body using an acoustic reconstruction algorithm. The Monte Carlo simulation optical forward model construction module builds an optical Monte Carlo simulation optical forward model and initializes the optical forward model parameters; it also presets empirical values ​​of optical characteristic parameters that match the target biological tissue type, including the optical scattering coefficient μ. S Anisotropy factor g, refractive index n; The untrained neural network building module constructs an untrained neural network to learn the mapping relationship from a random noise vector to the distribution of the target optical absorption coefficient; it initializes the parameters of the untrained neural network and initializes a random array matrix as the fixed input of the neural network. The self-constrained solution module inputs a random array matrix into an untrained neural network, and the network outputs a predicted value μ of the optical absorption coefficient distribution. a _pred; will change μ a The _pred and preset optical characteristic parameters are input into the forward optical model of Monte Carlo simulation to obtain the luminous flux distribution Φ inside the target body. Based on the luminous flux distribution, the predicted initial sound pressure distribution p0_pred is calculated. The loss value between the predicted initial sound pressure distribution p0_pred and the initial sound pressure distribution p0 obtained by the data acquisition module is calculated. An optimization algorithm is used to perform inverse optimization on the network, and the final output of the network is the optical absorption coefficient distribution μ. a _pred is the desired solution result.

2. The optical absorption coefficient calculation system based on photoacoustic imaging and untrained network according to claim 1, the system further includes an optimization and acceleration module that uses the Fourier domain to accelerate the calculation process.

3. The optical absorption coefficient calculation system based on photoacoustic imaging and untrained network according to claim 2, wherein the optimization acceleration module divides the calculation process into a pre-calculation phase, a mid-calculation phase, and a post-calculation phase; In the early stages of the computation, the Monte Carlo simulation uses the first number of photons; in the middle stages, it uses the second number of photons; and in the later stages, it uses the third number of photons, with the third number of photons > the second number of photons > the first number of photons.

4. The optical absorption coefficient calculation system based on photoacoustic imaging and untrained network according to claim 1, wherein the initialization of the optical forward model parameters includes setting a matching light source model in Monte Carlo simulation according to the actual light source configuration of the photoacoustic imaging system.

5. The optical absorption coefficient calculation system based on photoacoustic imaging and untrained network according to claim 1, wherein the untrained neural network consists of k sequentially connected computational blocks; each computational block consists of an upsampling layer, a convolutional activation layer and a residual connection layer in sequence.

6. The optical absorption coefficient calculation system based on photoacoustic imaging and untrained network according to claim 5, wherein the residual connection layer adds the input of the current computation block to the output of the convolutional activation layer through skip connections.

7. The optical absorption coefficient calculation system based on photoacoustic imaging and an untrained network as described in claim 1, in the self-constrained calculation module, a loss function is used to quantify the difference between the predicted initial sound pressure distribution p0_pred and the initial sound pressure distribution p0 obtained by the data acquisition module, i.e., to calculate the loss value; using the error backpropagation algorithm, the gradient of the loss value with respect to all adjustable parameters of the neural network is calculated, and the gradient descent optimization algorithm is used to update the parameters of the neural network according to the gradient; after the parameter update, the network generates a new, optimized μ again from the random data matrix. a _pred, and repeat the above steps; the above calculation process continues, constantly narrowing the difference between the predicted sound pressure and the measured sound pressure, until the following conditions are met: the loss value is lower than the preset convergence threshold, or the number of calculations reaches the preset maximum number of calculations.

8. A method for calculating the optical absorption coefficient based on photoacoustic imaging and an untrained network, characterized in that, The solution method includes the following steps: Step S1: Acquire the ultrasonic signal excited by the short pulse laser, and reconstruct the initial sound pressure distribution image p0 of the target body using the acoustic reconstruction algorithm; Step S2: Construct an optical Monte Carlo simulation forward optical model and initialize the parameters of the forward optical model; and preset empirical values ​​of optical characteristic parameters that match the target biological tissue type, including the optical scattering coefficient μ. S Anisotropy factor g, refractive index n; Step S3: Construct an untrained neural network to learn the mapping relationship from the random noise vector to the target optical absorption coefficient distribution; initialize the parameters of the untrained neural network, and initialize a random array matrix as the fixed input of the neural network; Step S4: Input the random array matrix into the untrained neural network, and the network outputs the predicted value μ of the optical absorption coefficient distribution. a _pred; will change μ a The _pred and preset optical characteristic parameters are input into the forward optical model of Monte Carlo simulation to obtain the luminous flux distribution Φ inside the target body. Based on the luminous flux distribution, the predicted initial sound pressure distribution p0_pred is calculated. The loss value between the predicted initial sound pressure distribution p0_pred and the initial sound pressure distribution p0 obtained by the data acquisition module is calculated. An optimization algorithm is used to perform inverse optimization on the network, and the final output of the network is the optical absorption coefficient distribution μ. a _pred is the desired solution result; Step S5: The Fourier domain is used to accelerate the solution process in step S4.

9. The optical absorption coefficient calculation method based on photoacoustic imaging and untrained network according to claim 8, wherein in step S5, in the early stage of calculation, the Monte Carlo simulation uses a first number of photons; in the middle stage of calculation, the Monte Carlo simulation uses a second number of photons; and in the later stage of calculation, the Monte Carlo simulation uses a third number of photons, wherein the third number of photons > the second number of photons > the first number of photons.

10. The optical absorption coefficient calculation method based on photoacoustic imaging and untrained network according to claim 9, wherein the initialization of the optical forward model parameters includes setting a matching light source model in Monte Carlo simulation according to the actual light source configuration of the photoacoustic imaging system.

Images