Multi-Speed Photoacoustic Image Reconstruction Method and Device Based on Implicit Neural Representations

Abstract

This disclosure provides a method and apparatus for multi-velocity photoacoustic image reconstruction based on implicit neural representation, relating to the field of image processing technology. The method includes: adjusting the parameters of an implicit neural representation network. Adjusting the parameters involves: obtaining a first initial sound pressure distribution and a first sound velocity distribution based on spatial coordinates within an imaging region using the implicit neural representation network; obtaining a predicted detector signal based on the first initial sound pressure distribution and the first sound velocity distribution using a differentiable acoustic forward propagation model; minimizing the difference between the predicted detector signal and the detector measurement signal to update the parameters of the implicit neural representation network; iterating in this manner until a preset convergence condition is met; and using the parameter-adjusted implicit neural representation network to obtain a second initial sound pressure distribution based on spatial coordinates within the imaging region. This disclosure enables photoacoustic image reconstruction without requiring a sound velocity distribution map or paired ground truth.

Description

Technical Field

[0001] This disclosure relates to the field of image processing technology, and in particular to a method and apparatus for multisonic photoacoustic image reconstruction based on implicit neural representation. Background Technology

[0002] Photoacoustic imaging (PAI), as an emerging hybrid medical imaging modality, has shown significant application potential in biomedical fields such as brain functional imaging and early tumor detection due to its combination of the high contrast of optical imaging and the deep tissue penetration capability of ultrasound imaging. Its imaging principle lies in using pulsed laser irradiation of biological tissues, such as hemoglobin, where light absorbers absorb light energy and undergo instantaneous thermal expansion, generating ultrasonic signals (i.e., photoacoustic signals). These signals are received by a detector array, and reconstruction algorithms are used to reconstruct an image of the light absorption distribution within the tissue. However, the practical application of photoacoustic imaging technology, especially in complex scenarios such as transcranial imaging, faces a key technical bottleneck: the inherent acoustic heterogeneity of biological tissues. Different tissue types, such as fat, muscle, soft tissue, and skull, exhibit significant differences in acoustic properties such as sound velocity and density.

[0003] Current mainstream photoacoustic image reconstruction algorithms, such as Delay-and-Sum (DAS), Filtered Back Projection (FBP), and Time Reversal (TR), generally rely on a key assumption for computational simplicity: that sound waves propagate at a constant average speed of sound in a homogeneous medium. This simplified assumption is applicable to soft tissues with relatively homogeneous acoustic properties. However, in scenarios involving interfaces with high acoustic impedance differences, such as when sound waves pass through the skull, significant refraction, reflection, and scattering occur when the photoacoustic signal traverses multi-layered tissues with vastly different acoustic properties. This leads to significant deviations between the actual propagation path and propagation time of the sound waves and the calculation results under the homogeneous medium assumption. At this point, traditional reconstruction algorithms based on the constant speed of sound assumption will inevitably introduce severe image distortion. For example, the position, shape and size of target structures (such as blood vessels and lesions) in the reconstructed image will be significantly distorted, the spatial resolution and contrast of the image will be severely degraded, the recognition of fine structures such as microvascular vessels will be significantly reduced, and non-real structures caused by signal misfocusing, such as arc artifacts or strip artifacts, will also be generated. Therefore, a photoacoustic image reconstruction method that can effectively compensate for or correct the effects of acoustic heterogeneity is needed to improve the imaging quality of photoacoustic imaging in complex scenes. Summary of the Invention

[0004] In view of this, this disclosure provides a method and apparatus for multisonic photoacoustic image reconstruction based on implicit neural representation.

[0005] According to a first aspect of this disclosure, a method for multi-velocity photoacoustic image reconstruction based on implicit neural representation is provided, the method comprising:

[0006] The parameters of the implicit neural representation network are adjusted, which includes: obtaining a first initial sound pressure distribution and a first sound velocity distribution based on spatial coordinates within the imaging region using the implicit neural representation network; simulating the propagation process of the sound pressure wave excited by the first initial sound pressure distribution in the medium defined by the first sound velocity distribution using a differentiable acoustic forward propagation model to obtain a predicted detector signal; minimizing the difference between the predicted detector signal and the detector measurement signal to update the parameters of the implicit neural representation network; and iterating in this way until a preset convergence condition is met.

[0007] The second initial sound pressure distribution is obtained using a hidden neural representation network with adjusted parameters based on the spatial coordinates within the imaging region.

[0008] In some embodiments of the first aspect of this disclosure, adjusting the parameters of the implicit neural representation network further includes: obtaining a first density distribution while obtaining the first initial sound pressure distribution and the first sound velocity distribution; specifically, simulating the propagation process of the sound pressure wave excited by the first initial sound pressure distribution in the medium defined by the first sound velocity distribution using a differentiable acoustic forward propagation model to obtain the predicted detector signal involves simulating the propagation process of the sound pressure wave excited by the first initial sound pressure distribution in the medium defined by the first sound velocity distribution and the first density distribution using a differentiable acoustic forward propagation model to obtain the predicted detector signal.

[0009] In some embodiments of the first aspect of this disclosure, the method further includes: pre-training the implicit neural representation network before adjusting the parameters of the implicit neural representation network; wherein, pre-training the implicit neural representation network includes: using the DAS algorithm to initially reconstruct the detector measurement signal to obtain a third initial sound pressure distribution, using the third initial sound pressure distribution as the supervision target of the initial boost distribution, initializing the density distribution and sound velocity distribution to predetermined constants and constraining their physical consistency through the differentiable acoustic forward propagation model, and performing pre-training of the implicit neural representation network.

[0010] In some embodiments of the first aspect of this disclosure, minimizing the difference between the predicted detector signal and the detector measurement signal to update the parameters of the implicit neural representation network includes: calculating the mean square error between the predicted detector signal and the detector measurement signal, using the mean square error to perform backpropagation to obtain the gradient, and updating the parameters of the implicit neural representation network based on the gradient.

[0011] In some embodiments of the first aspect of this disclosure, the method further includes: obtaining a synthetic detector signal based on a second initial sound pressure distribution, a predetermined third sound velocity distribution, and a predetermined third density distribution using a differentiable acoustic forward model; and reconstructing the synthetic detector signal using a DAS algorithm to obtain a fourth initial sound pressure distribution.

[0012] In some embodiments of the first aspect of this disclosure, the method further includes: obtaining a second initial sound pressure distribution, a second sound velocity distribution, and / or a second density distribution based on spatial coordinates within the imaging region using an implicit neural representation network with adjusted parameters.

[0013] In some embodiments of the first aspect of this disclosure, the implicit neural representation network is a multilayer perceptron (MLP) with position encoding, and the differentiable acoustic forward propagation model is a j-Wave simulator implemented based on the JAX library.

[0014] According to a second aspect of this disclosure, a multi-velocity photoacoustic image reconstruction device based on implicit neural representation is provided, the device comprising:

[0015] A parameter fine-tuning unit is used to adjust the parameters of the implicit neural representation network. The adjustment of the parameters of the implicit neural representation network includes: obtaining a first initial sound pressure distribution and a first sound velocity distribution based on the spatial coordinates within the imaging region through the implicit neural representation network; simulating the propagation process of the sound pressure wave excited by the first initial sound pressure distribution in the non-uniform medium defined by the first sound velocity distribution using a differentiable acoustic forward propagation model to obtain a predicted detector signal; minimizing the difference between the predicted detector signal and the detector measurement signal to update the parameters of the implicit neural representation network; and iterating in this way until a preset convergence condition is met.

[0016] The reconstructed execution unit is used to obtain a second initial sound pressure distribution based on the spatial coordinates within the imaging region using a parameterized implicit neural representation network.

[0017] According to a third aspect of this disclosure, an electronic device is provided, comprising: one or more processors and a memory storing a program, the program including instructions that, when executed by the processor, cause the processor to perform the methods described above.

[0018] According to a fourth aspect of this disclosure, a computer-readable storage medium storing a program, the program including instructions that, when executed by one or more processors of a computing device, cause the computing device to perform the method described above.

[0019] As can be seen from the above technical solutions, the embodiments of this disclosure can achieve joint implicit estimation and compensation of the initial sound pressure distribution (i.e., photoacoustic image) and spatially varying acoustic parameters (sound velocity, density) through a single neural network without prior acquisition of sound velocity distribution maps or paired ground truth values. This significantly improves the reconstruction quality of photoacoustic images, including structural fidelity, detail clarity, and artifact suppression capabilities, especially the reconstruction quality of photoacoustic images in complex acoustic environments such as transcranial imaging. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart of a multisonic photoacoustic image reconstruction method based on implicit neural representation provided in this disclosure embodiment;

[0022] Figure 2 This is a schematic diagram of a single execution process for adjusting INR network parameters according to an embodiment of this disclosure;

[0023] Figure 3 A schematic diagram illustrating the process of transcranial photoacoustic imaging reconstruction using the method provided in this embodiment of the disclosure;

[0024] Figure 4 This is a comparison chart of the reconstruction results on a two-dimensional human brain numerical phantom using the method of the present disclosure, the actual initial sound pressure distribution, and the reconstruction results on a two-dimensional human brain numerical phantom using related techniques.

[0025] Figure 5 A schematic diagram of the structure of the multi-velocity photoacoustic image reconstruction device based on implicit neural representation provided in the embodiments of this disclosure;

[0026] Figure 6 A schematic structural block diagram of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0027] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0028] The terminology used in the embodiments of this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of this disclosure. The singular forms “a,” “the,” and “the” as used in the embodiments of this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0029] Depending on the context, words such as "if," "when," etc., used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrases "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0030] Among related technologies, the main technical solutions for reconstructing photoacoustic images include the following:

[0031] 1) Prior sound velocity map-based correction method: This method involves pre-obtaining the sound velocity distribution map of the imaging area using other imaging modalities such as ultrasound reflection in transmission ultrasound tomography (UTT), X-ray attenuation (CT), or magnetic resonance imaging (MRI), and then using this sound velocity distribution map to correct the reconstruction algorithm. The disadvantages of this method are that it requires additional hardware, increasing the system's hardware complexity and cost, and registration errors between multimodal images can affect the correction accuracy.

[0032] 2) Joint reconstruction of sound velocity and photoacoustic image: Both the sound velocity map and the initial sound pressure map are treated as unknowns and solved simultaneously within an iterative optimization framework. This method is highly nonlinear and nonconvex, easily gets trapped in local optima, is sensitive to initial guesses, and has a huge computational cost and difficulty in convergence.

[0033] 3) Supervised learning-based methods: These methods employ deep neural networks such as U-net to learn the mapping from low-quality images with sound speed distortion to high-quality ground truth images. However, this approach heavily relies on large-scale, high-quality paired training datasets (i.e., distorted images and their corresponding undistorted ground truth images). In real-world scenarios, obtaining such paired ground truth data is extremely difficult or even impossible, which limits its generalization ability and reliability in practical applications.

[0034] Therefore, there is an urgent need for a new method that can effectively handle the sound velocity heterogeneity problem and reconstruct high-quality photoacoustic images without relying on external sound velocity maps or paired ground truth data.

[0035] In view of this, embodiments of this disclosure provide the following method and apparatus for multi-velocity photoacoustic image reconstruction based on implicit neural representation. Without prior acquisition of sound velocity distribution maps or paired ground truth values, it effectively corrects image distortion and artifacts caused by sound velocity non-uniformity (especially in the skull). Through a single neural network, it achieves joint implicit estimation and compensation of the initial sound pressure distribution (i.e., photoacoustic image) and spatially varying acoustic parameters (sound velocity, density), significantly improving the reconstruction quality of photoacoustic images, including structural fidelity, detail clarity, and artifact suppression capabilities, especially the photoacoustic image reconstruction quality in complex acoustic environments such as transcranial imaging.

[0036] Figure 1 This diagram illustrates a flowchart of a multi-velocity photoacoustic image reconstruction method based on implicit neural representation provided in an embodiment of this disclosure. See also... Figure 1 The method of this disclosure embodiment may include the following steps:

[0037] Step 101: Adjust the parameters of the Implicit Neural Representation (INR) network. Adjusting the parameters of the INR network includes: obtaining the first initial sound pressure distribution and the first sound velocity distribution based on the spatial coordinates within the imaging region through the INR network; simulating the propagation process of the sound pressure wave excited by the first initial sound pressure distribution in the non-uniform medium defined by the first sound velocity distribution through a differentiable acoustic forward propagation model to obtain the predicted detector signal; minimizing the difference between the predicted detector signal and the detector measurement signal to update the parameters of the implicit neural representation network; iterating in this way until the preset convergence condition is met.

[0038] Step 102: Use the adjusted INR network to obtain the second initial sound pressure distribution based on the spatial coordinates within the imaging area.

[0039] The method of this disclosure, without prior acquisition of sound velocity distribution maps or paired ground truth supervision data, uses a single neural network to parameterize an unknown physical field and combines it with a differentiable physical model for end-to-end optimization, thereby achieving joint implicit estimation and compensation of the initial sound pressure distribution (i.e., photoacoustic image) and spatially varying acoustic parameters.

[0040] In this embodiment, the INR network can be a multilayer perceptron (MLP) with position coding. The input of the INR network is the spatial coordinates r(x,y) within the imaging region, and the output of the INR network is the values ​​of multiple physical fields corresponding to each spatial coordinate within the imaging region. These multiple physical fields can be the initial sound pressure distribution, the sound velocity distribution, and the density distribution, respectively. In one example, the multiple physical fields output by the INR network can be: the normalized initial sound pressure distribution p0(r) and the sound velocity distribution c(r). In another example, for complex scenarios such as transcranial imaging, these multiple physical fields can include: the initial sound pressure p0(r), the sound velocity c(r), and the density ρ(r).

[0041] The embodiments of this disclosure use a single INR network to share parameters to simultaneously represent the initial sound pressure field and the acoustic medium field, which establishes implicit constraints between the physical fields and enables a compact representation of the model parameters.

[0042] For example, the INR network can be constructed as a multilayer perceptron containing 8 layers, each with 512 neurons, and the activation function is ReLU. A position encoding module is placed before the input layer of this multilayer perceptron. This module can be used to map spatial coordinates r(x,y) to a high-dimensional feature space to enhance the representation of high-frequency details. The output layer of this multilayer perceptron can have 3 neurons, corresponding to the initial sound pressure distribution p0, sound velocity distribution c, and density distribution ρ, respectively. The outputs of these 3 neurons are constrained within a reasonable physical range using activation functions such as tanh and sigmoid, and linear transformations. This reasonable physical range can be preset; for example, the physical range of the sound velocity distribution can be set to [1450, 3500], in m / s.

[0043] In this embodiment of the disclosure, the differentiable acoustic forward propagation model can be, but is not limited to, a j-Wave simulator implemented based on the JAX library. Specifically, the differentiable acoustic forward propagation model can be a differentiable simulator Fsim, using j-Wave, with its initial settings consistent with the geometric position and time sampling rate of the detector used.

[0044] Figure 2 The diagram illustrates a single execution process of adjusting the INR network parameters in step 101. Figure 2 In the model, the forward propagation model is a differentiable acoustic forward propagation model, F θ This represents the INR network, where the coordinate r is the spatial coordinate within the imaging region, p0 represents the initial sound pressure distribution, c represents the sound velocity distribution, and P... pred P represents the predictor signal. means This represents the detector's measured signal, and Loss represents the difference between the predicted detector signal and the detector's measured signal. See also Figure 2 The spatial coordinates r within the imaging region are input into the INR network F. θ The physical field (p0, c) is generated, and then the predictive detector signal P is obtained through a differentiable forward model F. pred Predicting detector signal P pred With the detector measurement signal P meas The loss L is calculated, and the parameters θ of the INR network are updated via backpropagation. This closed-loop optimization process iterates multiple times until a predetermined convergence condition is reached. In specific applications, this convergence condition can be flexibly set. For example, the convergence condition could be: the number of iterations reaches a preset value; or, the loss value is lower than a preset threshold.

[0045] In some examples, step 101 can involve obtaining a first density distribution simultaneously with obtaining a first initial sound pressure distribution and a first sound velocity distribution. The propagation process of the sound pressure wave excited by the first initial sound pressure distribution within the medium defined by the first sound velocity distribution is simulated using a differentiable acoustic forward propagation model to obtain the predicted detector signal. Specifically, this includes simulating the propagation process of the sound pressure wave excited by the first initial sound pressure distribution within the medium defined by the first sound velocity distribution and the first density distribution using a differentiable acoustic forward propagation model to obtain the predicted detector signal. This approach is applicable to complex scenarios such as cranial imaging.

[0046] In step 101, the differentiable acoustic forward propagation model takes the physical field (p0(r), c(r), ρ(r)) or the physical field (p0(r), c(r)) generated by the INR network as input. Based on the acoustic wave equation, it numerically simulates the propagation process of the sound pressure wave excited by the initial sound pressure p0(r) in a non-uniform medium. The acoustic properties of the medium are defined by the sound velocity c(r) and density ρ(r). The output is the simulated time series signal at the detector array position, which is the predicted detector signal Ppred.

[0047] In step 101, without requiring any ground truth image, the parameters of the INR network are updated by minimizing the difference between the predicted detector signal and the detector measurement signal. That is, minimizing the difference between the predicted detector signal and the detector measurement signal to update the parameters of the implicit neural representation network includes: calculating the mean square error between the predicted detector signal and the detector measurement signal, using the mean square error for backpropagation to calculate the gradient, and updating the parameters of the implicit neural representation network based on the gradient.

[0048] Specifically, a first loss function is predefined to measure the difference between the predicted detector signal and the detector measurement signal. For example, the first loss function can be the mean square error (MSE) as shown in equation (1).

[0049] L(θ) = ||Ppred - P meas ||²(1)

[0050] Where L(θ) represents the loss value between the predicted detector signal and the detector measurement signal, P pred P represents the predictor signal. meas This indicates the signal measured by the detector.

[0051] Specifically, gradient descent optimization algorithms such as the Adam optimizer can be used to update the parameters θ of the INR network based on the first loss function. Here, since the entire process is end-to-end differentiable, the gradient at each step can be efficiently calculated through automatic differentiation.

[0052] As can be seen from the above, the embodiments of this disclosure only utilize the original measurement signal P. meas The optimization of the INR network can be achieved by using it as an optimization target, thus realizing unsupervised adjustment of the INR network parameters and eliminating the dependence on paired true data.

[0053] Further, prior to step 101, the method of this embodiment may also include: pre-training an INR network. Specifically, the pre-training INR network includes: using the DAS algorithm to initially reconstruct the detector measurement signal to obtain a third initial sound pressure distribution, using the third initial sound pressure distribution as the supervision target of the initial boost distribution, initializing the density distribution and sound velocity distribution to predetermined constants and constraining their physical consistency through a differentiable acoustic forward propagation model, and performing pre-training of an implicit neural representation network.

[0054] When pre-training the INR network, a second loss function can be used, which is the sum of the reconstruction loss and the physical consistency loss. The reconstruction loss can be the mean square error between the third initial sound pressure distribution and the predicted initial sound pressure distribution currently generated by the INR. The physical consistency loss can be the mean square error between the predicted detector signal and the detector measurement signal obtained by the differentiable acoustic forward propagation model with the predicted initial sound pressure distribution currently generated by the INR, the density distribution set to a predetermined constant, and the sound velocity distribution set to a predetermined constant as inputs.

[0055] Specifically, first use DAS to measure the detector signal P meas A coarse and distorted initial sound pressure distribution is obtained through preliminary reconstruction; this initial sound pressure distribution is the aforementioned third initial sound pressure distribution. This coarse and distorted initial sound pressure distribution is then used as the supervision target for the initial sound pressure distribution p0(r) predicted by the INR network. The sound velocity distribution c(r) and the density distribution ρ(r) are set as predetermined constants, and the INR network is pre-trained with a small number of iterations to obtain the initial INR network parameters. During this pre-training process, the spatial coordinates r within the imaging region are fixed.

[0056] Pre-training can provide a more reasonable starting point for the unsupervised optimization of step 101, which helps to accelerate the convergence of step 101 and avoid getting trapped in bad local optima.

[0057] In step 102, all coordinate points r of the imaging region are input into the INR network optimized in step 101 to obtain the second initial sound pressure distribution, which is the reconstructed photoacoustic image. Simultaneously, a second sound velocity distribution and a second density distribution can also be obtained through the INR network, which can serve as supplementary components for the reconstructed photoacoustic image.

[0058] Furthermore, step 102 may further include: obtaining a synthetic detector signal based on a second initial sound pressure distribution, a predetermined third sound velocity distribution, and a predetermined third density distribution using a differentiable acoustic forward model; and reconstructing the synthetic detector signal using the DAS algorithm to obtain a fourth initial sound pressure distribution. This enables post-processing of the reconstructed photoacoustic image, removing distortion and artifacts, and further improving the quality of the reconstructed photoacoustic image.

[0059] Specifically, the initial sound pressure field (i.e., the second initial sound pressure distribution) learned by the optimized INR network is extracted. The non-uniform acoustic parameters learned by the network (i.e., the second sound velocity distribution and the second density distribution) are ignored. An ideal, uniform virtual acoustic medium (i.e., the medium defined by the third sound velocity distribution and the third density distribution) is defined. Using the initial sound pressure field as the sound source, a forward simulation is performed again in the aforementioned ideal uniform medium using a differentiable acoustic forward model to generate a theoretically distortion-free synthetic detector signal. Finally, a standard reconstruction algorithm such as DAS is used to reconstruct the synthetic detector signal, obtaining a fourth initial sound pressure distribution. This fourth initial sound pressure distribution is the corrected, high-quality reconstructed photoacoustic image. The third sound velocity distribution and the third density distribution can be preset. For example, the third sound velocity distribution can be preset to 1540 m / s, and the third density distribution can be preset to 1000 kg / m³.

[0060] Therefore, by utilizing the learned sound source information and the ideal propagation model, the residual artifacts introduced by the complex acoustic propagation path are effectively "cleaned up", further improving the clarity and fidelity of the reconstructed photoacoustic image.

[0061] The method of this disclosure can reconstruct photoacoustic images using actual measured photoacoustic signals, without the need for supervised training with paired ground truth images. This significantly reduces the demand for data such as ground truth images while improving the quality of photoacoustic image reconstruction, making it directly applicable to imaging scenarios where obtaining ground truth images is difficult. Furthermore, this disclosure, through the algorithm's inherent estimation and compensation for sound velocity changes, eliminates the need for additional hardware to measure sound velocity maps, unlike methods such as UTT, thus maintaining the simplicity of the photoacoustic imaging system and reducing costs and potential registration errors.

[0062] The embodiments disclosed herein represent the entire physical field as the weights of a small neural network through an INR network, achieving high data compression. The model is compact and the representation is continuous, allowing images to be queried at any resolution, thus avoiding the limitations caused by pixel / voxel discretization.

[0063] Simulation results show that the photoacoustic images reconstructed by the method of this embodiment are significantly better than traditional reconstruction algorithms in terms of objective indicators such as peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). Especially in complex scenarios such as transcranial imaging, it can effectively remove geometric distortions and artifacts and clearly restore fine structures such as intracranial blood vessels.

[0064] Furthermore, embodiments of this disclosure can also provide multiphysics information, which, in addition to reconstructing high-quality photoacoustic images, can simultaneously output a sound velocity distribution map and a density distribution map consistent with the physical scene, providing additional valuable information for tissue characterization and quantitative analysis.

[0065] The following section uses a two-dimensional simulation of transcranial photoacoustic imaging as an example to describe the specific implementation process of the embodiments of this disclosure.

[0066] Prior to the reconstruction process, the following data preparation is performed to obtain measurement signals that include the effects of skull deformity, which is an example of the detector measurement signals mentioned earlier:

[0067] A two-dimensional numerical model of the human brain was constructed, incorporating structures such as the scalp, skull, cerebrospinal fluid, brain parenchyma, and blood vessels. Different optical and acoustic parameters were assigned to different tissues. Optical parameters included, but were not limited to, absorption coefficient and scattering coefficient, while acoustic parameters included sound velocity and density. The actual initial sound pressure distribution p generated within the two-dimensional numerical model of the human brain after pulsed laser irradiation was calculated using Monte Carlo optical simulation. 0GT The true initial sound pressure distribution p 0GT Using the acoustic parameters of a two-dimensional human brain numerical phantom as a sound source, and employing acoustic simulation tools such as k-Wave, sound wave propagation is simulated, and the measurement signal P received by the ring detector array, which includes skull distortion effects, is recorded. meas .

[0068] Figure 3 A schematic diagram illustrating the process of transcranial photoacoustic imaging reconstruction using the method provided in embodiments of this disclosure is shown. See also Figure 3 The transcranial photoacoustic imaging reconstruction process may include the following steps a1 to a3:

[0069] Step a1, initialization, to obtain the initial INR network F θ ;

[0070] For the measurement signal P meas Performing a standard DAS reconstruction, assuming a uniform sound velocity of 1540 m / s (i.e., the third sound velocity distribution mentioned earlier), yields a coarse initial sound pressure distribution p. 0DAS The rough initial sound pressure distribution p 0DAS This is an example of the aforementioned third initial sound pressure distribution. For the INR network F... θ Pre-training is performed, specifically: the spatial coordinates r(x,y) (r∈R) within the imaging region are fixed. 2 ), with a rough initial sound pressure distribution p 0DAS As the supervisory target for the initial sound pressure distribution p0, the sound velocity distribution c is set to a constant "1540", and the density distribution ρ is set to a constant "1000" (i.e., the third density distribution mentioned above). The Adam optimizer is used to perform about 1000 iterations to obtain the initialized INR network parameters.

[0071] Step a2, unsupervised optimization, yields the INR network F with adjusted parameters. θ* ;

[0072] Based on the initialized INR network parameters, unsupervised optimization of the INR network is performed, iterating 10,000 times or more to obtain the adjusted INR network parameters. Specifically, each iteration in the unsupervised optimization performs the following steps: Step b1, input the spatial coordinates r(x,y) within the imaging region into the INR network to obtain the predicted physical field (p 0pred , c pred ,ρ pred ), p 0pred c represents the first initial sound pressure distribution. pred Represents the first sound velocity distribution, ρ pred Indicate the first density distribution; step b2, input the predicted physics field into the forward model F sim That is, the j-Wave simulator (i.e., the differentiable acoustic forward propagation model) calculates the predicted detector signal P. pred Step b3, calculate the predictive detector signal P pred With the actual measured signal P measStep b4: Calculate the gradient of the mean squared error loss Loss with respect to the INR network parameters using automatic differentiation; Step b5: Update the INR network parameters based on the gradient using the Adam optimizer.

[0073] Step a3, post-processing, to obtain the fourth initial sound pressure distribution p final .

[0074] Specifically, from the INR network F after parameter adjustment θ* Extract the learned initial sound pressure field p0* (i.e., the second initial sound pressure distribution mentioned above), and call the forward model F. sim Using the initial sound pressure field p0* as the sound source in a homogeneous medium (i.e., c ideal =1540m / s,ρ ideal Under the condition of 1000 kg / m³, the synthesized distortion-free signal P was obtained by forward propagation. synth (That is, the synthesized detector signal mentioned above), for the synthesized distortion-free signal P synth The fourth initial sound pressure distribution p is obtained by performing a DAS reconstruction once. final The fourth initial sound pressure distribution p final This is the reconstructed photoacoustic image.

[0075] The results are evaluated as follows: the fourth initial sound pressure distribution p final Compared with the actual initial sound pressure distribution p 0GT Comparison and quantitative analysis show that the fourth initial sound pressure distribution p final The peak signal-to-noise ratio (PSNR) reached 33.42 dB, and the structural similarity index (SSIM) reached 0.939, both significantly higher than the results of DAS (17.57 dB, 0.629), TR (24.10 dB, 0.718), and supervised U-net (31.38 dB, 0.948). This demonstrates that the method provided in this disclosure has good effectiveness and superiority compared to related technologies.

[0076] Figure 4 The results of reconstructing a two-dimensional human brain numerical phantom using the method of embodiments of this disclosure are shown, along with the actual initial sound pressure distribution p. 0GT A comparison of the reconstruction results of two-dimensional human brain numerical phantoms using related technologies (DAS, TR, U-net). Figure 4From left to right: Reconstruction results of the Delay-and-Sum (DAS) algorithm on a 2D human brain numerical phantom; Reconstruction results of the Time Reversal Algorithm (TR) on a 2D human brain numerical phantom; Reconstruction results of a supervised U-shaped Convolutional Neural Network (U-Net) on a 2D human brain numerical phantom; Reconstruction results of the method of this embodiment on a 2D human brain numerical phantom; and the ideal ground truth (GT) image without cranial deformities (i.e., the true initial sound pressure distribution p). 0GT ).pass Figure 4 The superiority of the method of this disclosure embodiment over related technologies in terms of structural fidelity and artifact suppression can be seen intuitively.

[0077] Figure 5 A schematic diagram of a multisonic photoacoustic image reconstruction device based on implicit neural representation provided in an embodiment of this disclosure is shown. This device is applied to an electronic device 600. See also... Figure 5 The multisonic photoacoustic image reconstruction device 500 based on implicit neural representation may include:

[0078] The parameter fine-tuning unit 501 is used to adjust the parameters of the implicit neural representation network. The adjustment of the parameters of the implicit neural representation network includes: obtaining a first initial sound pressure distribution and a first sound velocity distribution based on the spatial coordinates within the imaging area through the implicit neural representation network; simulating the propagation process of the sound pressure wave excited by the first initial sound pressure distribution in the non-uniform medium defined by the first sound velocity distribution through a differentiable acoustic forward propagation model to obtain a predicted detector signal; minimizing the difference between the predicted detector signal and the detector measurement signal to update the parameters of the implicit neural representation network; and iterating in this way until a preset convergence condition is met.

[0079] The reconstruction execution unit 502 is used to obtain a second initial sound pressure distribution based on the spatial coordinates within the imaging region using an implicit neural representation network with adjusted parameters.

[0080] Furthermore, the multi-velocity photoacoustic image reconstruction device 500 based on implicit neural representation may also include: a pre-training unit 503 for pre-training an implicit neural representation network; wherein, the pre-training of the implicit neural representation network includes: using the DAS algorithm to perform preliminary reconstruction of the detector measurement signal to obtain a third initial sound pressure distribution, using the third initial sound pressure distribution as the supervision target of the initial boost distribution, initializing the density distribution and sound velocity distribution to predetermined constants and constraining their physical consistency through a differentiable acoustic forward propagation model, and performing pre-training of the implicit neural representation network.

[0081] Furthermore, the multi-velocity photoacoustic image reconstruction device 500 based on implicit neural representation may also include: a post-processing unit 504, used to obtain a synthetic detector signal based on a second initial sound pressure distribution, a predetermined third sound velocity distribution, and a predetermined third density distribution through a differentiable acoustic forward model, and to reconstruct the synthetic detector signal using a DAS algorithm to obtain a fourth initial sound pressure distribution.

[0082] Further technical details regarding the multisonic photoacoustic image reconstruction device 500 based on implicit neural representations can be found in the preceding section on multisonic photoacoustic image reconstruction methods based on implicit neural representations, and will not be repeated here. In specific applications, the multisonic photoacoustic image reconstruction device 500 based on implicit neural representations can be implemented by the electronic device 600 described below, or it can be implemented as software within the electronic device 600.

[0083] In addition, embodiments of this disclosure also provide a computer-readable storage medium storing a computer program thereon, the program including instructions that, when executed by one or more processors of a computing device, perform the steps of the aforementioned multisonic photoacoustic image reconstruction method based on implicit neural representation.

[0084] Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure is shown. See also... Figure 6 The electronic device 600 may include one or more processors 601, and a memory 602 storing one or more programs, which are executed by the one or more processors 601 to implement the method flow and / or program units corresponding to each unit in the apparatus shown in the above embodiments of this disclosure.

[0085] The various components are interconnected via different buses and can be mounted on a common motherboard or otherwise as required. Processor 601 can process instructions executed within the electronic device, including instructions stored in or on memory to display graphical information of a user interface on an external input / output device (such as a display device coupled to an interface). In other embodiments, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired.

[0086] Processor 601 may include one or more single-core or multi-core processors. Processor 601 may include any combination of general-purpose processors or special-purpose processors (such as graphics processors, application processors, baseband processors, etc.).

[0087] Memory 602 is the computer-readable storage medium provided in this disclosure, which can be used to store non-transitory software programs, non-transitory computer-executable programs, and units, such as those in the embodiments of this disclosure. Figure 1The program instructions / units corresponding to the multisonic photoacoustic image reconstruction method based on implicit neural representation are shown. The processor 601 executes non-transient software programs, instructions, and units stored in the memory 602, thereby performing operations such as those described in the above method embodiments. Figure 1 The program, instructions, and units corresponding to the multisonic photoacoustic image reconstruction method based on implicit neural representation are shown.

[0088] The electronic device 600 may further include an input device 603 and an output device 604. The processor 601, memory 602, input device 603, and output device 604 can be connected via a bus or other means. Figure 6 Taking the example of a connection between China and Israel via a bus.

[0089] The aforementioned programs (also known as software, software applications, or code) include the machine instructions of a programmable processor and can be implemented using object-oriented programming languages, assembly language, or machine language.

[0090] With the development of time and technology, the meaning of "medium" has become increasingly broad. The dissemination of computer programs is no longer limited to tangible media; they can also be downloaded directly from the network. Any combination of one or more computer-readable storage media can be used. Computer-readable storage media can be, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, device, or apparatus.

[0091] The technical solutions provided in this disclosure have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this disclosure. The descriptions of the embodiments above are only for the purpose of helping to understand the methods and core ideas of this disclosure. Furthermore, those skilled in the art will recognize that, based on the ideas of this disclosure, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this disclosure.

[0092] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Any modifications or equivalent substitutions made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for reconstructing multisonic photoacoustic images based on implicit neural representation, characterized in that, The method includes: The implicit neural representation network is pre-trained. The pre-training of the implicit neural representation network includes: using the DAS algorithm to initially reconstruct the detector measurement signal to obtain a third initial sound pressure distribution; using the third initial sound pressure distribution as the supervision target of the initial pressure rise distribution; initializing the density distribution and sound velocity distribution to predetermined constants and constraining their physical consistency through the differentiable acoustic forward propagation model; and performing pre-training of the implicit neural representation network. The parameters of the implicit neural representation network are adjusted, which includes: obtaining a first initial sound pressure distribution, a first sound velocity distribution, and a first density distribution based on spatial coordinates within the imaging region using the implicit neural representation network; simulating the propagation process of the sound pressure wave excited by the first initial sound pressure distribution in the medium defined by the first sound velocity distribution and the first density distribution using a differentiable acoustic forward propagation model to obtain the predicted detector signal; minimizing the difference between the predicted detector signal and the detector measurement signal to update the parameters of the implicit neural representation network; and iterating in this way until a preset convergence condition is met. The second initial sound pressure distribution is obtained using a hidden neural representation network with adjusted parameters based on the spatial coordinates within the imaging region.

2. The method according to claim 1, characterized in that, Minimizing the difference between the predicted detector signal and the detector measurement signal to update the parameters of the implicit neural representation network includes: calculating the mean square error between the predicted detector signal and the detector measurement signal, using the mean square error to perform backpropagation to obtain the gradient, and updating the parameters of the implicit neural representation network based on the gradient.

3. The method according to claim 1, characterized in that, The method further includes: The synthesized detector signal is obtained based on the second initial sound pressure distribution, the predetermined third sound velocity distribution, and the predetermined third density distribution using a differentiable acoustic forward model. The DAS algorithm is used to reconstruct the synthetic detector signal to obtain a fourth initial sound pressure distribution.

4. The method according to claim 1, characterized in that, Also includes: Using an implicit neural representation network with adjusted parameters, a second initial sound pressure distribution, a second sound velocity distribution, and / or a second density distribution are obtained based on the spatial coordinates within the imaging region.

5. The method according to claim 1, characterized in that, The implicit neural representation network is a multilayer perceptron (MLP) with position encoding, and the differentiable acoustic forward propagation model is a j-Wave simulator implemented based on the JAX library.

6. A multi-speed photoacoustic image reconstruction device based on implicit neural representation, characterized in that, The multi-speed photoacoustic image reconstruction device based on implicit neural representation includes: The pre-training unit is used to pre-train the implicit neural representation network. The pre-training of the implicit neural representation network includes: using the DAS algorithm to perform preliminary reconstruction of the detector measurement signal to obtain a third initial sound pressure distribution; using the third initial sound pressure distribution as the supervision target of the initial boost distribution; initializing the density distribution and sound velocity distribution to predetermined constants and constraining their physical consistency through the differentiable acoustic forward propagation model; and performing pre-training of the implicit neural representation network. A parameter fine-tuning unit is used to adjust the parameters of the implicit neural representation network. The adjustment of the implicit neural representation network parameters includes: obtaining a first initial sound pressure distribution, a first sound velocity distribution, and a first density distribution based on spatial coordinates within the imaging region using the implicit neural representation network; simulating the propagation process of the sound pressure wave excited by the first initial sound pressure distribution in the medium defined by the first sound velocity distribution and the first density distribution using a differentiable acoustic forward propagation model to obtain the predicted detector signal; minimizing the difference between the predicted detector signal and the detector measurement signal to update the parameters of the implicit neural representation network; and iterating in this manner until a preset convergence condition is met. The reconstructed execution unit is used to obtain a second initial sound pressure distribution based on the spatial coordinates within the imaging region using a parameterized implicit neural representation network.

7. An electronic device, characterized in that, include: A memory for storing one or more processors and programs, the programs comprising instructions that, when executed by the processor, cause the processor to perform the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a program, the program comprising instructions that, when executed by one or more processors of a computing device, cause the computing device to perform the method as claimed in any one of claims 1 to 5.

Images