Wavefront aberration correction and quantitative phase imaging method based on physical driven neural field

Abstract

This invention discloses an aberration correction and quantitative phase imaging method based on a physically driven neural field. The method constructs a physical forward model for differential phase-contrast imaging under partially coherent illumination conditions and implicitly represents the sample's complex transmission function and optical aberrations using a neural field. By jointly optimizing sample information and optical aberrations through a differentiable operator, it achieves an inversion solution from intensity information to the sample's phase distribution. This method overcomes the limitations of weak object approximation, achieving self-calibration of optical aberrations and high-precision phase recovery without additional calibration data or acquisition overhead. It is suitable for quantitative phase imaging of biological samples or micro / nanostructured samples.

Description

Technical Field

[0001] This invention belongs to the field of optical microscopy imaging technology, and in particular to a method for aberration correction and quantitative phase imaging based on a physically driven neural field. Background Technology

[0002] In fields such as biomedical microscopy, micro / nanostructure detection, and precision optical metrology, many analytes possess weak absorption, low contrast, or transparency characteristics, such as living cells, unstained tissue samples, and transparent functional materials. When light waves pass through such samples, their amplitude changes relatively little, with the main effect being phase delay caused by changes in optical path difference. However, conventional microscopy systems primarily record light intensity information and are not directly sensitive to phase information. Therefore, transparent or weakly absorbing samples exhibit low contrast under traditional bright-field microscopy conditions, making it difficult to obtain accurate structural information. Consequently, achieving quantitative measurement of sample phase information without the need for external labeling has become an important research direction in the field of microscopy.

[0003] Quantitative phase imaging (QPI) technology measures the phase change of light waves after passing through a sample to achieve quantitative analysis of the sample's refractive index distribution or thickness information, making it an important technique for label-free imaging. Differential phase-contrast microscopy (DPI) is a typical partially coherent QPI method. It reconstructs the sample's phase information by applying asymmetric illumination from different directions and acquiring multiple intensity images for differential calculation. This method is simple in structure, has high data acquisition efficiency, and is easily compatible with existing microscopy systems. However, existing DPI methods are typically based on weak object approximation and ideal aperture assumptions, linearizing the nonlinear imaging process. When the sample exhibits large phase changes or significant absorption effects, these approximation conditions are difficult to meet, leading to phase reconstruction errors. Furthermore, optical aberrations are unavoidable in practical microscopy systems. These aberrations alter the system's spatial frequency response characteristics and couple with the sample's phase information. Without accurate modeling and correction of these aberrations, the reconstruction resolution can easily decrease or the quantitative results can be distorted. In recent years, deep learning methods have been applied to phase retrieval, but most methods rely on large amounts of labeled data for supervised training, making it difficult to obtain the true phase values ​​under real imaging conditions. Meanwhile, purely data-driven methods lack explicit physical constraints, and their generalization ability and stability are difficult to guarantee when system or sample conditions change. Therefore, under partially coherent illumination conditions, how to achieve stable and accurate quantitative phase recovery while considering both sample absorption effects and system aberrations remains a pressing technical problem to be solved in this field. Summary of the Invention

[0004] The purpose of this invention is to address the deficiencies or shortcomings of existing technologies by providing a physically driven neural field-based aberration correction and quantitative phase imaging method. This method enables high-precision, high-fidelity microscopic imaging under complex imaging conditions, including large-phase samples, absorptive samples, and samples with optical aberrations. This invention achieves joint estimation of the sample's phase and absorption distributions by constructing a physically constrained nonlinear inversion framework, and adaptively corrects system aberrations, thereby realizing high-precision quantitative phase reconstruction without altering the structure of existing differential phase-contrast microscopy systems.

[0005] The technical solution for achieving the objective of this invention is: a method for aberration correction and quantitative phase imaging based on physically driven neural fields, the method comprising the following steps:

[0006] Step 1: Asymmetric partially coherent illumination is used to detect the sample under test, and raw light intensity images under multiple illumination conditions are acquired;

[0007] Step 2: Construct a neural field model to map the spatial coordinates within the imaging field of view into high-dimensional feature vectors;

[0008] Step 3: Process the high-dimensional feature vector using a deep learning network to output the complex transmission function of the sample under test and the optical aberration of the imaging system.

[0009] Step 4: Based on the partially coherent imaging model, the predicted light intensity images corresponding to multiple illumination conditions are calculated using the complex transmission function and optical aberrations.

[0010] Step 5: Construct a loss function using the residual between the original light intensity image and the predicted light intensity image, optimize the parameters of the deep learning network using the gradient descent algorithm, and simultaneously update the phase distribution, absorption distribution, and optical aberration parameters of the sample under test until the preset convergence condition is met.

[0011] Furthermore, the asymmetric partially coherent illumination described in step 1, based on the differential phase contrast principle, effectively converts the sample phase gradient into intensity contrast; specifically, it includes:

[0012] The sample to be tested is placed in a partially coherent illumination imaging system, with an adjustable spatial light field illumination device as the illumination source;

[0013] The illumination area is divided into multiple asymmetric illumination sub-regions, which are lit sequentially according to a preset illumination mode. A transmitted light intensity image is acquired under each illumination condition, thereby obtaining multiple original intensity images containing the phase gradient information encoding of the sample under test. Each illumination sub-region corresponds to angular illumination in different directions.

[0014] Furthermore, the asymmetric partial coherent illumination includes semi-circular illumination or semi-ring illumination, which is achieved by adjusting the adjustable spatial light field illumination device to a circular illumination area and dividing it into multiple asymmetric illumination sub-regions along different directions.

[0015] Furthermore, the adjustable spatial light field illumination device includes at least one of a digital micromirror device, a spatial light modulator, a liquid crystal display, or a programmable light-emitting diode.

[0016] Furthermore, step 2 specifically includes the following processes:

[0017] The spatial pixel coordinates within the imaging field of view are normalized, mapping the original pixel coordinates to the normalized range;

[0018] In the normalized coordinate space, the parameters of the neural field model are randomly initialized so that the neural field can implicitly model the complex transmission function of the sample under test; wherein, the complex transmission function includes the phase distribution and absorption distribution of the sample under test.

[0019] The normalized spatial coordinates are mapped using a spatial frequency position coding algorithm, transforming low-dimensional coordinates into high-dimensional coded features.

[0020] Furthermore, the spatial frequency position coding algorithm includes at least one of radial coding, hash grid coding, Fourier feature position coding, or spherical harmonic function coding.

[0021] Furthermore, for the radial coding spatial frequency position coding algorithm, the mapping matrix is ​​defined as:

[0022]

[0023] In the formula, High-dimensional encoded features are used to expand low-dimensional coordinates to a high-dimensional feature space. For normalized spatial coordinates, This represents the number of coding layers in radial coding. This is a rotation angle parameter matrix used to define the rotation angles corresponding to different coding layers.

[0024] Furthermore, the deep learning network in step 3 includes:

[0025] Sample neural networks are used to model the phase and absorption distributions of samples as continuous functions.

[0026] Aberration neural networks recover and represent aberrations using the coefficients of an optical aberration parameterization model.

[0027] Both the sample neural network and the aberration neural network employ a multilayer perceptron.

[0028] Furthermore, in the sample neural network, each hidden layer employs an activation function to introduce nonlinear characteristics, and the output layer uses a sigmoid normalization function to ensure that the output value is constrained within a bounded range; the sample neural network transforms the abstract output into the actual physical parameter distribution through a transformation matrix, specifically as follows:

[0029]

[0030] In the formula, and This represents the dual-channel output of the last layer of the multilayer perceptron. and These are used to represent the phase and absorption distribution of the sample, respectively. represents the transformation matrix, used to transform the abstract output of the multilayer perceptron into the actual physical parameter distribution; r represents the spatial coordinates of the sample under test;

[0031] In the aforementioned aberration neural network, the optical aberration parameterization model adopts the Zernike polynomial model. Based on the aberration distribution formed by the Zernike coefficients, the corresponding aperture function is expressed as:

[0032]

[0033] In the formula, For the first The Zernikic polynomial of the term, For aperture function, For the first The Zernike coefficient of the term, The imaginary unit, Zernike coefficient number of terms denoted as two-dimensional plane coordinates in Fourier space.

[0034] Further, step 4 specifically includes: based on a partially coherent imaging model, using the complex transmission function or its corresponding physical parameters as input, performing forward propagation calculations under various illumination conditions to obtain the corresponding predicted light intensity images; wherein, the partially coherent imaging model is expressed as:

[0035]

[0036] in, Image plane coordinates The normalized light intensity distribution detected at the location. For two-dimensional plane coordinates in Fourier space, For the illumination function, The Fourier transform of the complex transmission function of the sample under test in coordinate system... The value at that location, The complex conjugate of the Fourier transform of the complex transmission function of the sample under test in coordinate system... The value at that location, These represent the values ​​of the aperture function and its complex conjugate at the corresponding frequency domain coordinates. , It is the integral variable.

[0037] Further, the loss function in step 5 includes a data consistency term, a structural similarity index term, and a total variation regularization term, wherein the structural similarity index term is used to constrain the structural similarity between the predicted light intensity image and the original light intensity image, and the total variation regularization term is used to constrain the spatial smoothness of the sample physical parameters; the loss function Specifically, it is expressed as follows:

[0038]

[0039] In the formula, This represents the differential phase-contrast forward physical imaging model. To predict and measure the relationship between light intensity Norm, Represents the structural similarity index. The computation function representing total variation regularization, Indicates the absorption distribution of an object. Represents the phase distribution of an object. Represents the aperture function. The illumination function represents the illumination angle i. The total number of lighting angles, Indicates the first Light intensity diagram for each illumination angle and These are the weight parameters for the structural similarity index regularization term and the total variation regularization term, respectively.

[0040] Compared with the prior art, the significant advantages of this invention are:

[0041] (1) By embedding the physical forward model of partial coherent microscopy into the neural field optimization framework, the imaging process is modeled and solved under unified physical constraints, thereby directly handling the nonlinear relationship between intensity and phase under partial coherent imaging conditions. This avoids the problems of nonlinear distortion, phase underestimation and high sensitivity to system aberrations caused by the linearization process of traditional differential phase contrast imaging methods that rely on weak object approximation and ideal aperture assumption.

[0042] (2) While reconstructing the complex amplitude information of the sample, it can jointly estimate and adaptively correct unknown optical aberrations without the need for additional aberration calibration or system calibration, thereby improving the accuracy and stability of quantitative phase reconstruction.

[0043] (3) The present invention maintains the data acquisition method of standard difference phase contrast imaging. It does not require an additional number of acquisition frames or changes to the existing microscopic imaging system structure while achieving high-precision phase recovery. It is directly compatible with conventional bright field microscopes and adjustable spatial light field illumination devices, and has good engineering feasibility.

[0044] (4) In the process of neural field optimization with strict physical constraints, the present invention can achieve a spatial resolution close to the incoherent diffraction limit under partially coherent illumination conditions, and the physical information driven neural field solution framework constructed also provides a universal computational method for solving complex nonlinear inverse problems in the field of computational imaging.

[0045] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the overall process of a physical-driven neural field-based aberration correction and quantitative phase imaging method in one embodiment.

[0047] Figure 2 This is a schematic diagram of the algorithm flow of the method of the present invention in one embodiment.

[0048] Figure 3 This is a comparison of quantitative reconstruction of real biological samples using the method of the present invention and conventional methods in one embodiment; wherein the original acquired image resolution is 2690×1920, and intensity image acquisition is performed using a 10×0.4 objective lens. Figure 3 (a) shows the full-field phase distribution reconstruction of stained pathological sections (MSH3) using the method of the present invention; Figure 3 (b) in the figure represents the reconstruction of the full field of view absorption distribution of stained pathological sections by the method of the present invention; Figure 3 (c) in the figure represents the reconstruction results of long-term quantitative phase imaging of live HeLa cells using the method of the present invention. Continuous time-series acquisition was performed on HeLa cells in the full field of view, and the results include the phase distribution of the full field of view at the initial time and five hours after cell culture. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0050] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0051] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0052] In one embodiment, a method for aberration correction and quantitative phase imaging based on a physically driven neural field is provided. An adjustable spatial light field illumination device is used as a partially coherent illumination source. This device enables programmable control of the illumination mode to generate quasi-monochromatic plane wave illumination with different incident angles, thereby constructing multiple semi-circular asymmetric illumination conditions and acquiring corresponding intensity image data. After acquiring multiple differential illumination images, a two-dimensional continuous coordinate space for the sample is established based on the image size and normalized. Subsequently, radial or positional encoding is used to map the two-dimensional spatial coordinates to a high-dimensional feature space to enhance the ability to express high-frequency structural information. Based on this, two multilayer perceptron models are constructed: a sample neural network representing the sample's complex transmission function and an aberration neural network representing optical aberrations. The high-dimensional encoded features serve as the input to the sample neural network, while a set of trainable latent variable vectors are used as the input to the aberration neural network. Through multilayer nonlinear mapping, the inputs are converted into actual physical parameters to obtain the sample's phase distribution and absorption distribution, and Zernike polynomial coefficients characterizing optical aberrations are output, thus constructing a complex aperture function containing aberration modulation. Subsequently, a physical forward model is established based on partially coherent imaging theory. The sample complex transmission function, aberration parameters, and actual optical system parameters predicted by the network are substituted into the imaging propagation process to calculate the predicted light intensity distribution under different semicircular illumination conditions. The difference between the acquired original light intensity image and the predicted light intensity image is further calculated to construct a loss function. The parameters of the multilayer perceptron are iteratively optimized through gradient backpropagation to achieve joint updates of the sample phase distribution, absorption distribution, and Zernike aberration parameters of the imaging system until a preset convergence condition is met. Finally, a quantitative reconstruction result is obtained after aberration correction and effective decoupling of absorption and phase, achieving high-precision imaging of large-phase samples without introducing weak object approximations.

[0053] Combination Figure 1 and Figure 2 The method includes the following steps:

[0054] Step 1: Asymmetric partially coherent illumination is used to detect the sample under test, and raw light intensity images under multiple illumination conditions are acquired;

[0055] Step 2: Construct a neural field model to map the spatial coordinates within the imaging field of view into high-dimensional feature vectors;

[0056] Step 3: Process the high-dimensional feature vector using a deep learning network to output the complex transmission function of the sample under test and the optical aberration of the imaging system.

[0057] Step 4: Based on the partially coherent imaging model, the predicted light intensity images corresponding to multiple illumination conditions are calculated using the complex transmission function and optical aberrations.

[0058] Step 5: Construct a loss function using the residual between the original light intensity image and the predicted light intensity image, optimize the parameters of the deep learning network using the gradient descent algorithm, and simultaneously update the phase distribution, absorption distribution, and optical aberration parameters of the sample under test until the preset convergence condition is met.

[0059] Furthermore, in one embodiment, the asymmetric partially coherent illumination described in step 1, based on the differential phase contrast principle, effectively converts the sample phase gradient into intensity contrast to support aberration correction and absorption-phase decoupling in subsequent quantitative phase reconstruction; specifically including:

[0060] The sample to be tested is placed in a partially coherent illumination imaging system, with an adjustable spatial light field illumination device as the illumination source;

[0061] The illumination area is divided into multiple asymmetric illumination sub-regions, which are lit sequentially according to a preset illumination mode. A transmitted light intensity image is acquired under each illumination condition, thereby obtaining multiple original intensity images containing the phase gradient information encoding of the sample under test. Each illumination sub-region corresponds to angular illumination in different directions.

[0062] Preferably, in some embodiments, the asymmetric partial coherent illumination includes semi-circular illumination or semi-ring illumination, which is achieved by adjusting the adjustable spatial light field illumination device to a circular illumination area and dividing it into multiple asymmetric illumination sub-regions along different directions.

[0063] Preferably, in some embodiments, the adjustable spatial light field illumination device includes at least one of a digital micromirror device, a spatial light modulator, a liquid crystal display, or a programmable light-emitting diode.

[0064] Furthermore, in one embodiment, step 2 specifically includes the following process:

[0065] The spatial pixel coordinates within the imaging field of view are normalized, mapping the original pixel coordinates to the normalized range;

[0066] In the normalized coordinate space, the parameters of the neural field model are randomly initialized so that the neural field can implicitly model the complex transmission function of the sample under test; wherein, the complex transmission function includes the phase distribution and absorption distribution of the sample under test.

[0067] The normalized spatial coordinates are mapped using a spatial frequency position coding algorithm, transforming low-dimensional coordinates into high-dimensional coded features.

[0068] Preferably, in some embodiments, the spatial frequency location encoding algorithm includes, but is not limited to, radial encoding spatial frequency location encoding algorithms (such as hash grid encoding, such as Instant NGP, Fourier feature encoding, and spherical harmonic function encoding, etc.).

[0069] Here, for the radial coding spatial frequency position coding algorithm, the mapping matrix is ​​defined as:

[0070]

[0071] In the formula, High-dimensional encoded features are used to expand low-dimensional coordinates to a high-dimensional feature space. For normalized spatial coordinates, This represents the number of coding layers in radial coding. This is a rotation angle parameter matrix used to define the rotation angles corresponding to different coding layers.

[0072] Furthermore, in one embodiment, the deep learning network in step 3 includes:

[0073] Sample neural networks are used to model the phase and absorption distributions of samples as continuous functions.

[0074] Aberration neural networks recover and represent aberrations using the coefficients of an optical aberration parameterization model.

[0075] Both the sample neural network and the aberration neural network employ a multilayer perceptron.

[0076] Preferably, in some embodiments, the sample neural network employs activation functions to introduce nonlinear characteristics in each hidden layer, and a sigmoid normalization function in the output layer to ensure that the output value is constrained within a bounded range; the sample neural network transforms the abstract output into an actual physical parameter distribution through a transformation matrix, specifically as follows:

[0077]

[0078] In the formula, and This represents the dual-channel output of the last layer of the multilayer perceptron. and These are used to represent the phase and absorption distribution of the sample, respectively. represents the transformation matrix, used to transform the abstract output of the multilayer perceptron into the actual physical parameter distribution; r represents the spatial coordinates of the sample to be tested.

[0079] Here, in the sample neural network, the hidden state update of the (i+1)th layer is represented as:

[0080]

[0081] In the formula, , and Let represent the weight parameters, input, and activation function of the i-th layer multilayer perceptron, respectively. Represents the radial encoding matrix .

[0082] Preferably, in some embodiments, the aberrations are parameterized using the coefficients of several optical aberration parameterization models and expressed as a linear combination of corresponding Zernike polynomials. This reduces the constraints of the aberration recovery problem from a high-dimensional space to a low-dimensional parameter space, thereby narrowing the solution space and improving optimization efficiency. The mathematical expression for the forward propagation process of the aberration network is as follows:

[0083]

[0084] in, For activation function, For the weights of the aberration neural network, the neural network input... It is a set of random vectors. It is a linear transformation matrix. The output of the aberration neural network is the Zernike coefficient predicted by the network.

[0085] Preferably, in the aberration neural network, the optical aberration parameterization model adopts the Zernike polynomial model. Based on the aberration distribution formed by the Zernike coefficients, the corresponding aperture function is expressed as:

[0086]

[0087] In the formula, For the first The Zernikic polynomial of the term, For aperture function, The modulus of the aperture function. For the first The Zernike coefficient of the term, The imaginary unit, Zernike coefficient number of terms denoted as two-dimensional plane coordinates in Fourier space.

[0088] Here, the aperture function constructed based on the Zernike polynomial includes optical aberration components, which maps aberration information into the output light intensity signal, thereby realizing aberration recovery based on light intensity measurement.

[0089] Further, in one embodiment, step 4 specifically includes: based on a partially coherent imaging model, using the complex transmission function or its corresponding physical parameters as input, performing forward propagation calculations under various illumination conditions to obtain the corresponding predicted light intensity images; wherein, the partially coherent imaging model is expressed as:

[0090]

[0091] in, Image plane coordinates The normalized light intensity distribution detected at the location. For two-dimensional plane coordinates in Fourier space, For the illumination function, The Fourier transform of the complex transmission function of the sample under test in coordinate system... The value at that location, The complex conjugate of the Fourier transform of the complex transmission function of the sample under test in coordinate system... The value at that location, These represent the values ​​of the aperture function and its complex conjugate at the corresponding frequency domain coordinates. , The light intensity distribution predicted by the network is obtained through the physical forward imaging model, which is the integral variable.

[0092] Further, in one embodiment, in step 5, the residual between the acquired original light intensity image and the predicted light intensity image is calculated, and a loss function is constructed based on the residual. The loss function includes a data consistency term, a structural similarity index term, and a total variation regularization term, wherein the structural similarity index term is used to constrain the structural similarity between the predicted light intensity image and the original light intensity image, and the total variation regularization term is used to constrain the spatial smoothness of the sample's physical parameters; the loss function... Specifically, it is expressed as follows:

[0093]

[0094] In the formula, This represents the differential phase-contrast forward physical imaging model. To predict and measure the relationship between light intensity Norm, Represents the structural similarity index. The computation function representing total variation regularization, Indicates the absorption distribution of an object. Represents the phase distribution of an object. Represents the aperture function. The illumination function represents the illumination angle i. The total number of lighting angles, Indicates the first Light intensity diagram for each illumination angle and These are the weight parameters for the structural similarity index regularization term and the total variation regularization term, respectively.

[0095] Here, the parameters of the deep learning network are optimized using a gradient descent-based optimization algorithm, and the phase distribution, absorption distribution, and Zernike aberration parameters of the sample and the imaging system are updated synchronously until the preset convergence condition is met.

[0096] In one embodiment, an aberration correction and quantitative phase imaging system based on a physically driven neural field is provided, the system comprising:

[0097] The first module is used to: use asymmetric partially coherent illumination to detect the sample under test and acquire raw light intensity images under multiple illumination conditions;

[0098] The second module is used to: construct a neural field model and map the spatial coordinates within the imaging field of view into high-dimensional feature vectors;

[0099] The third module is used to: process the high-dimensional feature vector using a deep learning network, and output the complex transmission function of the sample under test and the optical aberration of the imaging system;

[0100] The fourth module is used to: calculate the predicted light intensity images corresponding to multiple illumination conditions based on the partially coherent imaging model, using the complex transmission function and optical aberrations;

[0101] The fifth module is used to: construct a loss function from the residual between the original light intensity image and the predicted light intensity image; optimize the parameters of the deep learning network through the gradient descent algorithm; and synchronously update the phase distribution, absorption distribution, and optical aberration parameters of the sample under test until the preset convergence condition is met.

[0102] Specific limitations regarding the physical-driven neural field-based aberration correction and quantitative phase imaging system can be found in the above description of the limitations of the physical-driven neural field-based aberration correction and quantitative phase imaging method, and will not be repeated here. Each module in the aforementioned physical-driven neural field-based aberration correction and quantitative phase imaging system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independent of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the corresponding operations of each module.

[0103] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements:

[0104] Step 1: Asymmetric partially coherent illumination is used to detect the sample under test, and raw light intensity images under multiple illumination conditions are acquired;

[0105] Step 2: Construct a neural field model to map the spatial coordinates within the imaging field of view into high-dimensional feature vectors;

[0106] Step 3: Process the high-dimensional feature vector using a deep learning network to output the complex transmission function of the sample under test and the optical aberration of the imaging system.

[0107] Step 4: Based on the partially coherent imaging model, the predicted light intensity images corresponding to multiple illumination conditions are calculated using the complex transmission function and optical aberrations.

[0108] Step 5: Construct a loss function using the residual between the original light intensity image and the predicted light intensity image, optimize the parameters of the deep learning network using the gradient descent algorithm, and simultaneously update the phase distribution, absorption distribution, and optical aberration parameters of the sample under test until the preset convergence condition is met.

[0109] For specific limitations on each step, please refer to the limitations on aberration correction and quantitative phase imaging methods based on physically driven neural fields mentioned above, which will not be repeated here.

[0110] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program being implemented when executed by a processor:

[0111] Step 1: Asymmetric partially coherent illumination is used to detect the sample under test, and raw light intensity images under multiple illumination conditions are acquired;

[0112] Step 2: Construct a neural field model to map the spatial coordinates within the imaging field of view into high-dimensional feature vectors;

[0113] Step 3: Process the high-dimensional feature vector using a deep learning network to output the complex transmission function of the sample under test and the optical aberration of the imaging system.

[0114] Step 4: Based on the partially coherent imaging model, the predicted light intensity images corresponding to multiple illumination conditions are calculated using the complex transmission function and optical aberrations.

[0115] Step 5: Construct a loss function using the residual between the original light intensity image and the predicted light intensity image, optimize the parameters of the deep learning network using the gradient descent algorithm, and simultaneously update the phase distribution, absorption distribution, and optical aberration parameters of the sample under test until the preset convergence condition is met.

[0116] For specific limitations on each step, please refer to the limitations on aberration correction and quantitative phase imaging methods based on physically driven neural fields mentioned above, which will not be repeated here.

[0117] As a specific example, the invention will be further verified and illustrated in one embodiment.

[0118] To verify the feasibility and stability of the method of this invention under complex sample conditions, quantitative phase reconstruction was first performed on fixed MSH3-stained pathological sections using the method of this invention. This sample contained both significant absorption effects and a large phase. The reconstruction results are as follows: Figure 3 (a) and Figure 3 As shown in (b), it can be seen that when performing quantitative phase imaging on large-phase samples with strong absorption, the present invention can stably reconstruct actual biological samples with a phase peak of about 7 radians and achieve effective decoupling of sample phase distribution and absorption distribution.

[0119] Furthermore, the method of this invention is also applied to the dynamic observation of live cells in complex biological environments. To verify its applicability under dynamic imaging conditions, this invention was used to perform continuous time-series quantitative phase imaging of live HeLa cells, and the reconstruction results are as follows: Figure 3 As shown in (c), it can be seen that the method of the present invention can maintain stable quantitative phase reconstruction results when the morphology of living cells changes over time. Figure 3 The right side of (c) in the figure shows the reconstruction results after 5 hours of cell culture, which is used to illustrate the stability of the present invention in the long-term observation process.

[0120] The above embodiments demonstrate that the method of the present invention can achieve joint reconstruction of sample phase distribution and absorption distribution under complex imaging conditions, and is applicable to quantitative phase imaging of static samples and dynamic live cell samples.

[0121] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention without departing from its spirit and scope should be included within the protection scope of the present invention.

Claims

1. A method for aberration correction and quantitative phase imaging based on physically driven neural fields, characterized in that, The method includes the following steps: Step 1: Asymmetric partially coherent illumination is used to detect the sample under test, and raw light intensity images under multiple illumination conditions are acquired; Step 2: Construct a neural field model to map the spatial coordinates within the imaging field of view into high-dimensional feature vectors; Step 3: Process the high-dimensional feature vector using a deep learning network to output the complex transmission function of the sample under test and the optical aberration of the imaging system. Step 4: Based on the partially coherent imaging model, the predicted light intensity images corresponding to multiple illumination conditions are calculated using the complex transmission function and optical aberrations. Step 5: Construct a loss function using the residual between the original light intensity image and the predicted light intensity image, optimize the parameters of the deep learning network using the gradient descent algorithm, and simultaneously update the phase distribution, absorption distribution, and optical aberration parameters of the sample under test until the preset convergence condition is met.

2. The aberration correction and quantitative phase imaging method based on physically driven neural fields according to claim 1, characterized in that, Step 1, the asymmetric partial coherent illumination, based on the differential phase contrast principle, effectively converts the sample phase gradient into intensity contrast; specifically, it includes: The sample to be tested is placed in a partially coherent illumination imaging system, with an adjustable spatial light field illumination device as the illumination source; The illumination area is divided into multiple asymmetric illumination sub-regions, which are then illuminated sequentially according to a preset illumination mode. A transmitted light intensity image is acquired under each illumination condition, thereby obtaining multiple raw intensity images containing the phase gradient information encoding of the sample under test. Each illumination sub-region corresponds to angular illumination from different directions.

3. The aberration correction and quantitative phase imaging method based on physically driven neural fields according to claim 2, characterized in that, The asymmetric partial coherent illumination includes semi-circular illumination or semi-ring illumination, which is achieved by adjusting the adjustable spatial light field illumination device to a circular illumination area and dividing it into multiple asymmetric illumination sub-regions along different directions.

4. The aberration correction and quantitative phase imaging method based on physically driven neural fields according to claim 2, characterized in that, Step 2 includes the following specific processes: The spatial pixel coordinates within the imaging field of view are normalized, mapping the original pixel coordinates to the normalized range; In the normalized coordinate space, the parameters of the neural field model are randomly initialized so that the neural field can implicitly model the complex transmission function of the sample under test; wherein, the complex transmission function includes the phase distribution and absorption distribution of the sample under test. The normalized spatial coordinates are mapped using a spatial frequency position coding algorithm, transforming low-dimensional coordinates into high-dimensional coded features.

5. The aberration correction and quantitative phase imaging method based on physically driven neural fields according to claim 4, characterized in that, The spatial frequency position coding algorithm includes at least one of radial coding, hash grid coding, Fourier feature position coding, or spherical harmonic function coding.

6. The aberration correction and quantitative phase imaging method based on physically driven neural fields according to claim 5, characterized in that, For the radial coding spatial frequency position coding algorithm, the mapping matrix is ​​defined as: In the formula, High-dimensional encoded features are used to expand low-dimensional coordinates to a high-dimensional feature space. For normalized spatial coordinates, This represents the number of coding layers in radial coding. This is a rotation angle parameter matrix used to define the rotation angles corresponding to different coding layers.

7. The aberration correction and quantitative phase imaging method based on physically driven neural fields according to claim 2, characterized in that, The deep learning network in step 3 includes: Sample neural networks are used to model the phase distribution and absorption distribution of the sample under test as continuous functions. Aberration neural networks recover and represent aberrations using the coefficients of an optical aberration parameterization model. Both the sample neural network and the aberration neural network employ a multilayer perceptron.

8. The aberration correction and quantitative phase imaging method based on physically driven neural fields according to claim 7, characterized in that, The sample neural network incorporates nonlinear characteristics in each hidden layer using activation functions, and employs a sigmoid normalization function in the output layer to ensure that the output value is constrained within a bounded range. The sample neural network transforms the abstract output into an actual physical parameter distribution using a transformation matrix, specifically as follows: In the formula, and This represents the dual-channel output of the last layer of the multilayer perceptron. and These are used to represent the phase and absorption distribution of the sample under test, respectively. represents the transformation matrix, used to transform the abstract output of the multilayer perceptron into the actual physical parameter distribution; r represents the spatial coordinates of the sample under test; In the aforementioned aberration neural network, the optical aberration parameterization model adopts the Zernike polynomial model. Based on the aberration distribution formed by the Zernike coefficients, the corresponding aperture function is expressed as: In the formula, For the first The Zernikic polynomial of the term, For aperture function, The modulus of the aperture function. For the first Zernike coefficient of the term, The imaginary unit, Zernike coefficient number of terms denoted as two-dimensional plane coordinates in Fourier space.

9. The aberration correction and quantitative phase imaging method based on physically driven neural fields according to claim 1, characterized in that, Step 4 specifically includes: based on a partially coherent imaging model, using the complex transmission function or its corresponding physical parameters as input, performing forward propagation calculations under various illumination conditions to obtain the corresponding predicted light intensity images; wherein, the partially coherent imaging model is expressed as: in, Image plane coordinates The normalized light intensity distribution detected at that location. For two-dimensional plane coordinates in Fourier space, For the illumination function, The Fourier transform of the complex transmission function of the sample under test in coordinate system... The value at that location, The complex conjugate of the Fourier transform of the complex transmission function of the sample under test in coordinate system... The value at that location, These represent the values ​​of the aperture function and its complex conjugate at the corresponding frequency domain coordinates. , It is the integral variable.

10. The aberration correction and quantitative phase imaging method based on physically driven neural fields according to claim 1, characterized in that, The loss function in step 5 includes a data consistency term, a structural similarity index term, and a total variation regularization term. The structural similarity index term constrains the structural similarity between the predicted and original light intensity images, and the total variation regularization term constrains the spatial smoothness of the sample's physical parameters. Specifically, it is expressed as follows: In the formula, This represents the differential phase-contrast forward physical imaging model. To predict and measure the relationship between light intensity Norm, Represents the structural similarity index. The computation function representing total variational regularization, Indicates the absorption distribution of an object. Represents the phase distribution of an object. Represents the aperture function. The illumination function represents the illumination angle i. The total number of lighting angles, Indicates the first Light intensity diagram for each illumination angle and These are the weight parameters for the structural similarity index regularization term and the total variation regularization term, respectively.

Images