AR-PAM image enhancement and noise reduction method based on Mean-reverting diffusion model

Abstract

The invention discloses an AR-PAM (Auto Register-Pulse Amplitude Modulation) image enhancement and noise reduction method based on a Mean-reverting diffusion model. According to the method, through a diffusion process of reverse SDE, a Mean-reverting diffusion model is utilized to convert a high-resolution image into a fixed Gaussian noise state, and an original high-resolution image is recovered through the reverse SDE process, so that resolution enhancement and noise suppression of the AR-PAM image are realized. Compared with a traditional method, the method has the advantages that the transverse resolution of the image can be remarkably improved on the premise that the AR-PAM imaging depth is not sacrificed, and noise is effectively restrained. Through verification of simulation data and in-vivo experiment data, the method provided by the invention shows an excellent enhancement effect in various degradation scenes, and especially under the conditions of high noise and low resolution, the peak signal-to-noise ratio and the structural similarity of the image can be remarkably improved, and the image quality is improved. Therefore, a new possibility is provided for application expansion of the photoacoustic microscopic imaging technology.

Description

Technical Field

[0001] The present invention relates to the technical field of biomedical imaging, and in particular to an AR-PAM image enhancement and noise reduction method based on a Mean-reverting diffusion model. Background Art

[0002] PAM (Photoacoustic Microscopy) is an advanced biomedical imaging technology that combines the high resolution of optical imaging with the deep penetration of ultrasound imaging. In recent years, PAM has been widely used in tumor detection, vascular morphology analysis, dermatology, and other biomedical research fields. According to the different imaging methods, PAM can be divided into OR-PAM (Optical-Resolution Photoacoustic Microscopy) and AR-PAM (Acoustic-Resolution Photoacoustic Microscopy). Among them, OR-PAM can achieve high lateral resolution imaging, but its imaging depth is limited by the scattering of light in biological tissues; on the contrary, AR-PAM has deeper tissue penetration ability, but its lateral resolution is low and is easily limited by acoustic diffraction, resulting in reduced image quality.

[0003] In practical applications, improving the lateral resolution of AR-PAM has always been a challenge. Since the acoustic resolution is limited by the diffraction effect, AR-PAM images usually show lateral blur and more noise, especially when dealing with small structures or low-contrast areas. Although existing image processing techniques, such as convolutional neural networks and model-based image reconstruction methods, have improved the image quality of AR-PAM to a certain extent, these methods often have problems such as over-smoothing, loss of details, and high computational complexity.

[0004] In recent years, diffusion model, as an emerging generative model, has shown significant advantages in the field of image processing. In particular, the mean-reverting diffusion model can effectively enhance the resolution of the image and remove noise by simulating the evolution of the image from high resolution to low resolution. However, traditional diffusion models usually require a large amount of iterative computing resources, which limits their application in real-time imaging and large-scale data processing.

[0005] In order to meet these challenges, this paper proposes an AR-PAM image enhancement and denoising method based on the mean-reverting diffusion model, aiming to improve the lateral resolution and overall quality of AR-PAM images while effectively reducing noise interference. By combining the forward and reverse diffusion processes of the diffusion model, this method can achieve high-resolution image restoration while maintaining computational efficiency. This innovative image processing method provides new ideas and tools for solving the resolution and noise problems in AR-PAM images, and has broad application potential and important clinical value. Summary of the invention

[0006] The purpose of the present invention is to solve the deficiencies in the prior art and provide an AR-PAM image enhancement and noise reduction method based on the mean-reverting diffusion model. The method aims to improve the resolution of the AR-PAM image and effectively suppress the noise that occurs during the imaging process, thereby solving the problem in the prior art that AR-PAM cannot be clearly imaged when the lateral resolution is low and the noise interference is severe.

[0007] In order to achieve the above technical objectives, the present invention adopts the following technical solutions.

[0008] An AR-PAM image enhancement and denoising method based on a Mean-reverting diffusion model, the method comprising the following steps: Step S1: In the training data generation step, image data is first obtained from a high-resolution OR-PAM image with a resolution ranging from 1 μm to 5 μm, and the OR-PAM image is converted into a low-resolution AR-PAM image with a resolution ranging from 20 μm to 100 μm by a convolution operation to simulate the imaging characteristics of the AR-PAM system; then, Gaussian noise of different intensities is added to the image data. , generate a noisy image dataset with different levels of degradation: , in, Represents the simulated AR-PAM image, is the original OR-PAM image, is the point spread function, is the convolution operation; Step S2: In the forward SDE step, the generated noisy image dataset is processed using the Mean-reverting diffusion model to gradually degrade the high-resolution image into a low-resolution image state with fixed Gaussian noise. This process is modeled by a stochastic differential equation: , in, Represents image data at timet status, is a fixed Gaussian noise mean state, represents the velocity parameter of the time-dependent mean reversion, is the diffusion coefficient, It is a Brownian motion process; Step S3, in the reverse SDE recovery step, the original high-resolution image is gradually restored from the noisy state by simulating the reverse SDE process. This process estimates the score function by training a noise network to guide the reverse generation of the image: , in, is the score function predicted by the noisy network.

[0009] Optionally, the specific process of step S1 is as follows: First, high-resolution OR-PAM images are obtained as the basic data source. OR-PAM images are tissue structure images captured with a high lateral resolution of less than 5 microns by using optical detection methods. Tissue structure images contain rich tissue details and characteristic information. In the process of converting the high-resolution OR-PAM image into the low-resolution AR-PAM image, it is assumed that the PSF of the acoustic resolution photoacoustic microscope system satisfies the Gaussian distribution, and its standard deviation is related to the focal spot size and frequency response of the imaging system. The high-resolution OR-PAM image is convolved with the PSF to generate a simulated low-resolution AR-PAM image. The mathematical model of the convolution operation is expressed as: , in, represents the convolution operation, represents the point spread function of the system; In order to simulate the noise interference that may occur in the real imaging process, Gaussian noise is added to the generated AR-PAM image. The mathematical expression of this process is: , in, n is Gaussian noise with zero mean.

[0010] Optionally, after step S3, a model optimization step is also included, and the specific process is as follows: Through multiple rounds of iterative optimization, the parameters of the model are further adjusted. , including the diffusion coefficient and mean reversion speed , ensuring that the noise removal and image restoration processes within each time step can achieve the best results; After each round of iteration, the denoising effect and image restoration quality of the model are evaluated by calculating the peak signal-to-noise ratio and structural similarity of the restored image. The performance of the model is gradually improved by continuously adjusting and optimizing the model parameters, making the image restoration effect in the reverse SDE process more stable and accurate. Eventually, the parameters of the model will be fixed to form a fully trained noise network and optimized model.

[0011] Optionally, after the model optimization step, an evaluation step is also included, and the specific process is as follows: The optimized noise network and optimization model are applied to a set of low-resolution AR-PAM image datasets collected actually. The forward SDE and reverse SDE recovery processes are performed in sequence by inputting the low-resolution AR-PAM image datasets collected actually into the optimized noise network and optimization model, thereby outputting enhanced high-resolution images. Multiple evaluation indicators are used for evaluation, including peak signal-to-noise ratio and structural similarity. The specific effect of the model in improving image quality is quantified by calculating the peak signal-to-noise ratio and structural similarity of the image before and after processing; Based on the evaluation results, the model parameters are further optimized or the algorithm structure is adjusted to improve the performance of the model and make it more suitable for various complex situations in practical applications.

[0012] On the other hand, the present invention also provides an application of an acoustic resolution photoacoustic microscope image enhancement and noise reduction method based on a mean-reverting diffusion model as described in the aforementioned embodiment in the field of image processing.

[0013] One or more technical solutions provided in the embodiments of the present invention have at least the following technical effects or advantages: 1. Efficient noise modeling and denoising capabilities: By adopting the forward SDE and reverse SDE restoration methods based on the mean-reverting diffusion model, it can effectively simulate the image degradation process and gradually restore high-resolution images, with significant denoising effects, especially in high-noise environments.

[0014] 2. Enhanced image quality: The present invention can effectively enhance image details and structures by simulating the noise degradation process and combining it with inverse SDE restoration, avoiding excessive smoothing of image details in traditional denoising methods and ensuring the realism and visibility of the image.

[0015] 3. Strong adaptability: By utilizing the imaging characteristics of the AR-PAM system and the training of the noise network, the present invention can flexibly process image data sets with different noise intensities and degradation degrees, and adapt to a variety of practical application scenarios, such as low-light environments, satellite images, medical images, etc.

[0016] 4. Improved accuracy and robustness of image processing: The present invention significantly improves the accuracy of the image restoration process by introducing the score function predicted by the noise network in the reverse process, making the resolution of the restored image closer to the original high-resolution image, and avoiding excessive blur or distortion that may be caused by traditional methods during the image denoising process.

[0017] 5. Unsupervised learning capability: This method performs unsupervised learning by training a noise network and does not rely on a large amount of manually annotated data, which makes the method more flexible and easy to expand in practical applications and suitable for various unknown noise types and image degradation scenarios.

[0018] 6. Improve the computational efficiency of image processing: By combining the forward SDE and reverse SDE recovery processes, the present invention reduces the computational complexity and improves the speed of image processing through an efficient training mechanism, making it suitable for application in large-scale image data processing tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Schematic diagram of forward and reverse propagation of a diffusion model in an embodiment of the present invention; Figure 2 A flowchart of a network model algorithm in an embodiment of the present invention; Figure 3 This is a diagram showing the improvement results of the AR-PAM image quality of point-shaped objects under different signal-to-noise ratios in an embodiment of the present invention; Figure 4 Result diagram of blood vessel image reconstruction under different FWHM in an embodiment of the present invention; Figure 5 Result diagrams of image enhancement and noise reduction at different SNRs in an embodiment of the present invention; Figure 6 This is a comparison chart of the model iteration of the method of the present invention and other algorithms in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to facilitate those of ordinary skill in the art to understand and implement the present invention, the following detailed description is given to each step of the method proposed by the present invention, and it should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by the appended claims of the present application.

[0021] Example The present invention provides an acoustic resolution photoacoustic microscope image enhancement and noise reduction method based on a mean-reverting diffusion model, comprising the following steps S1 to S3.

[0022] Step S1: In the training data generation step, image data is first obtained from high-resolution optical resolution photoacoustic microscopy (OR-PAM) images with a resolution range of 1 μm to 5 μm, and these high-resolution images are converted into low-resolution acoustic resolution photoacoustic microscopy (AR-PAM) images with a resolution range of 20 μm to 100 μm through convolution operations to simulate the imaging characteristics of the AR-PAM system. Then, by adding Gaussian noise of different intensities , generate a noisy image dataset with different levels of degradation: , in, Represents the simulated AR-PAM image, is the original OR-PAM image, is the point spread function (PSF), These data provide diverse support for subsequent model training.

[0023] Step S2: In the forward SDE step, the generated noisy image dataset is processed using the Mean-reverting diffusion model to gradually degrade the high-resolution image into a low-resolution image state with fixed Gaussian noise. This process is modeled by stochastic differential equations to describe the evolution of the image from the original data distribution to the noise distribution: , in, Represents image data at time t status, is a fixed Gaussian noise mean state, represents the velocity parameter of the time-dependent mean reversion, is the diffusion coefficient, It is a Brownian motion process. Through this diffusion process, the details in the image are gradually masked by noise, and finally reach a low-resolution noise state, preparing for the subsequent reverse SDE recovery process.

[0024] Step S3: In the reverse SDE recovery step, the original high-resolution image is gradually restored from the noisy state by simulating the reverse time stochastic differential equation process. This process estimates the score function by training a noise network to guide the reverse generation of the image: , in, is the score function predicted by the noise network. Through continuous denoising, the process gradually restores the image from a noisy state to a high-resolution state, achieving lateral resolution enhancement and noise suppression of the image.

[0025] In step S1, image data is first acquired from high-resolution optical resolution photoacoustic microscopy (OR-PAM) images. These high-resolution images have good lateral resolution but limited penetration depth. In order to simulate the imaging characteristics of acoustic resolution photoacoustic microscopy (AR-PAM) systems, these high-resolution OR-PAM images first need to be converted into low-resolution AR-PAM images.

[0026] In order to achieve the image degradation process, the OR-PAM image is degraded through the convolution operation to make it have the characteristics of the AR-PAM system. Specifically, it is assumed that the AR-PAM system can be approximated as a linear spatial translation invariant system, and its point spread function PSF is a Gaussian distribution. By convolving the OR-PAM image with the point spread function PSF, a low-resolution AR-PAM image is generated: , In the above formula, Represents the simulated AR-PAM image, represents the original OR-PAM image, Represents the point spread function PSF, which is generally a Gaussian distribution. is Gaussian noise, is the convolution operation; When generating AR-PAM images, the selection of the point spread function (PSF) is crucial. In an AR-PAM system, the focal area of ​​the ultrasonic transducer is usually Gaussian distributed with a mean and standard deviation If we assume that the OR-PAM image also has a Gaussian profile, its mean is , the standard deviation is , then their convolution result is still a mean of , standard deviation is Gaussian profile.

[0027] At this time, the standard deviation of the AR-PAM image The relationship between the PSF and the standard deviation of the OR-PAM image is: , The lateral resolution of photoacoustic microscopy images is usually measured by the half-width FWHM. Since the standard deviation of the Gaussian distribution is linearly related to its half-width FWHM, the expression of the half-width FWHM of the AR-PAM image can be further derived: , In the above formula, is the half width and height FWHM of the AR-PAM image, is the half width and height FWHM of the PSF, is the half width and height FWHM of the OR-PAM image.

[0028] In order to generate training data with different degradation levels, Gaussian noise of different intensities is added to the AR-PAM images generated above. The addition of noise simulates various interferences that may occur in the actual imaging process, thereby generating AR-PAM images with different signal-to-noise ratios (SNRs).

[0029] In step S2, the generated noisy image dataset is processed using the mean-reverting diffusion model, gradually degrading the high-resolution image into a low-resolution image state with fixed Gaussian noise. This process is modeled by stochastic differential equations, describing the evolution of the image from the original data distribution to the noise distribution: Specifically, the image data is in time t Status It evolves according to the following stochastic differential equation: , in, represents the state of the image data at time t, is a fixed Gaussian noise mean state, represents the velocity parameter of the time-dependent mean reversion, is the diffusion coefficient, is a Brownian motion process, which means that t Random perturbations on .

[0030] In order to make the above formula closed form, let ,in is the stationary variance, at a given time Any starting state , which can be expressed as:

[0031] in, Known, and pass the core is a mean and variance Gaussian distribution; assuming the starting state is , then at any time t the initial state is conditioned as follows:

[0032] In the initial stage of the diffusion process, the image data In a high-resolution state, the impact of noise is small. Over time, by adjusting and Due to the dynamic changes of the image, noise is gradually introduced into the image data, and the details in the image are gradually covered by the noise, resulting in a gradual decrease in the resolution of the image and gradually turning into a low-resolution state.

[0033] When the diffusion process is nearly complete, the image data Converge to the noise mean state , at this time, the image feature information has been masked by noise and the image resolution reaches the lowest. The image in this state can be regarded as a completely degraded low-resolution image, ready to enter the reverse SDE process for restoration and denoising.

[0034] In step S3, the original high-resolution image is gradually restored from the noisy state by simulating the process of the reverse temporal stochastic differential equation. This process is achieved by training a noise network to estimate the score function, thereby guiding the reverse generation and denoising of the image.

[0035] Inverse SDE is used to gradually restore the image from a noisy state to a high-resolution state. Image data In time t The state of evolves according to the following reverse SDE: , represents the speed parameter of time-dependent mean regression, which controls the speed at which image data recovers to a high-resolution state. is a fixed Gaussian noise mean state, is the diffusion coefficient, which indicates the noise intensity, is the score function predicted by the trained noisy network, is a reverse Brownian motion process, which means that in time t Random perturbations on .

[0036] The core of the reverse SDE part is to train a noise network To estimate the score function at any time , we can simulate the complete process of reverse SDE. However, in the face of complex degradation scenarios, directly learning the instantaneous noise within a given time may lead to instability during training. Therefore, a maximum likelihood objective method is adopted. For a given By finding The best path to stabilize the training process and further improve the quality of image generation can be maximized according to the following formula: : , According to the Bayes formula, we can get the training process shown below to minimize the negative log-likelihood: , in, Is The ideal state is reversed. Then, The optimal inverse solution of can be given by: .

[0037] After step S3, a model optimization step is also included. The model optimization step ensures the gradual improvement of image quality during the reverse SDE process by iteratively training the noise network and optimizing the model parameters. The specific process is as follows: First, the noise network is initially trained using a large-scale, high-resolution training dataset. The core of this process is to minimize the loss function between the predicted image and the real image to adjust the weights of the network so that it can effectively estimate the score function of the image data in a noisy state.

[0038] The expression of the loss function is:

[0039] In the above formula, is a positive weight, represents reverse SDE, is the mathematical expectation. After the loss function is trained, the obtained noise network can be used To predict the score at any time, the numerical method Euler-Maruyama is combined to iteratively obtain the score from the noise state In this way, the mean recovery SDE is well combined with the maximum likelihood objective of neural network training, which can not only easily train the noisy network more stably in various complex degradation scenarios, but also further improve the performance of model denoising.

[0040] After the initial training is completed, the model enters the optimization stage. In this stage, through multiple rounds of iterative optimization, the various parameters of the model are further adjusted, including the diffusion coefficient and mean reversion speed The optimization of these parameters ensures that the noise removal and image restoration processes within each time step can achieve the best results.

[0041] Specifically, the optimization process ensures the following relationship: , By adjusting and The dynamic changes of the image quality make the gradual elimination of noise match the restoration of image quality, avoiding the phenomenon of over-denoising or under-denoising.

[0042] After the model optimization step, an evaluation step is also included. After each round of iteration, quantitative indicators such as peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are used to evaluate the denoising effect and image restoration quality of the model. The evaluation of these indicators helps to gradually adjust and optimize the model parameters, improve the performance of the model, and make the image restoration effect in the reverse SDE process more stable and accurate.

[0043] Finally, after the optimization phase, the parameters of the model will be fixed to form a fully trained noise network and optimization model. At this point, the model has the ability to restore high-resolution images under different noise conditions, and can effectively improve the resolution and clarity of acoustic resolution photoacoustic microscopy images in practical applications.

[0044] In order to verify the effect of the present invention, a large number of segmented blood vessel data sets were generated by simulation in this embodiment. The entire data set consists of 4500 training sets and 500 test sets. All experimental procedures are implemented on the deep learning open source framework Pytorch. The sparse reconstruction effect is observed by comparing the label results with the results of reconstruction by the method of the present invention, and the effect is judged by comparing the quantitative indicators.

[0045] The experimental platform is configured with a GPU graphics card (GeForce RTX 2080 Ti).

[0046] The data verified in this experiment are the improvement of point-like object AR-PAM image quality under different signal-to-noise ratios, vascular image reconstruction under different FWHMs, and image enhancement and noise reduction under different SNRs.

[0047] like Figure 1 As shown, the forward IR-SDE first models the degradation process from the high-resolution OR-PAM image to the low-resolution AR-PAM image, and then restores the degraded mean state to the high-resolution image through the backward IR-SDE simulation.

[0048] Please also refer to Figure 1 and Figure 2 In the upper part of (a), the AR-PAM image and the OR-PAM image are input into the scoring network for training to solve the only uncertainty factor: the score function of the temporal data distribution. . The scoring network only needs to obtain arrive Noise level at each time point The score function of the data distribution at a specific time can be determined In this process, an improved scoring network is used to replace the traditional network as the backbone network for noise prediction. Compared with the traditional network, the scoring network improves the image restoration quality and computational efficiency by using fewer model parameters. Its simplified network structure is as follows: Figure 2 As shown in (b) in the figure. The innovative design of the scoring network is to adopt the concept of non-linear activation function, remove all traditional non-linear activation functions, and use "no activation gating" as a substitute. No activation gating divides the feature channel into two parts and multiplies them to generate output. This design not only reduces the computational complexity, but also greatly improves the performance of the network. For time embedding processing, the scoring network is optimized, and the improved modules are as follows Figure 2 As shown in (c) in the figure, MLP (Multilayer Perceptron) is added to the attention layer and the feedforward layer to generate channel-wise scaling and offset parameters. and This improvement enhances the model’s adaptability to temporal dynamics and provides more detailed control over attention mechanisms and feedforward operations, further improving the model’s performance and flexibility.

[0049] During the reconstruction phase, Figure 2 As shown in the lower part of (a), the noise in the image is gradually reduced by reversely simulating the diffusion path, and finally a high-resolution image is restored from the intermediate mean state. Specifically, the noisy version of the low-resolution image Start sampling and iteratively solve the inverse IR-SDE using the Euler method. The core idea of ​​the Euler method is to divide the continuous time interval into discrete small time steps. , at each time step, the noise network According to the current image status And the score function predicted by time information is used to calculate the next image state The Euler method approximates the random term by an incremental procedure ,in, Represents the increment of the reverse process, usually simulated as (z is a standard normal random variable with mean 0 and variance 1). In the inverse simulation process, it is very important to choose an appropriate time step, which directly affects the accuracy and stability of the approximation. In the execution of the inverse SDE, the time step t decreases from T to 0, and the corresponding subscript shows a decreasing trend.

[0050] like Figure 3 The following are the results of improving the quality of point-like object AR-PAM images under different signal-to-noise ratios (SNRs). Figure 3 (a) is a noise-free AR-PAM image. Figure 3 (b) to (d) are AR-PAM images with SNR of 35dB, 25dB, and 15dB, respectively. Figure 3 The FWHM of (a) to (d) is set to 55 μm. Figure 3(e) to (h) show the reverse iteration process, and the last column is the output image. Figure 3 (i) to (l) in show the same true value, Figure 3 (a1) to (d6) are Figure 3 The enlarged images of the white dotted line boxes 1 to 3 in (a) to (d) are shown in FIG. Figure 3 (e1) to (e3) in the figure show the same enlarged images corresponding to the true values; Figure 3 (m) to (o) in the figure show the different SNR levels. Figure 3 Signal distribution of white dotted line 1 to white dotted line 3 in (a1) to (d6); NPA represents normalized photoacoustic amplitude, SNR represents signal-to-noise ratio; In1, input noise-free image 1; In2, input image 2, signal-to-noise ratio is 35db; In3, input image 3, signal-to-noise ratio is 25db; In4, input image 4, signal-to-noise ratio is 15db; GT, true value; Out1, output noise-free image 1; Out2, output image 2, signal-to-noise ratio is 35db; Out3, output image 3, signal-to-noise ratio is 25db; Out4, output image 4, signal-to-noise ratio is 15db.

[0051] In the testing phase, the feasibility of the model is evaluated by simulating points with different noise levels and lateral spacing, e.g. Figure 3 shown. Figure 3 (a) shows the noise-free AR-PAM image, while Figure 3 (b) to (d) show images with signal-to-noise ratios (SNR) of 35dB, 25dB, and 15dB, respectively, and the half-width (FWHM) of all images is set to 55μm. Figure 3 (e) to (h) show the model pair Figure 3 The iterative process for the four input images in (a) to (d). Figure 3 (i) to (l) in the figure show the same true value (OR-PAM image). The reconstruction starts with a noisy version of the low-resolution image, and the outline of the points is still blurry at the 25th iteration. As the iterations progress, the noise in the image is gradually removed and the clarity of the points is significantly improved. Finally, after the 100th iteration, all points in the input image at different signal-to-noise ratio (SNR) levels are effectively reconstructed, and the output image is highly similar to the true value.

[0052] In order to further analyze the resolution limit of the enhanced image, points with lateral spacing of 70 μm, 55 μm, and 15 μm were selected for analysis, as shown in Figure 3 As shown in the white dashed boxes 1 to 3 in (a) to (d) and the white arrow in the true value. Figure 3 (a1) to (d6) show Figure 3(a) to (d) are enlarged images of selected points at different SNR levels. Figure 3 (e1) to (e3) show the corresponding true values. Figure 3 (m) to (o) show the different SNR signal-to-noise ratio levels. Figure 3 (a1) Figure 3 (a3) and Figure 3 The signal intensity distribution at the white dashed line in (a5) is shown in Figure 5. Figure 3 As shown in (m) in the figure, points with a spacing of 70 μm in the input image can be clearly distinguished, while Figure 3 As shown in (n) in , the points with a spacing of 55 μm are in the critical state of being distinguishable, which indicates that the lateral resolution limit of the input AR-PAM image is estimated to be about 55 μm. Figure 3 As shown in (o) in the figure, for dots with a spacing of 15 μm, the spacing is significantly lower than the resolution limit of the input image, resulting in signal overlap that cannot be distinguished. Figure 3 As shown in (m) to (o) in the figure, in the enhanced results of the model, points with spacings of 70, 55 μm, and 15 μm can be effectively distinguished at different SNR levels, indicating that the lateral resolution of the model-enhanced image exceeds 15 μm. In addition, the position and intensity of the points in the output image are highly consistent with the corresponding true values. These results show that the method proposed in the present invention can significantly improve the lateral resolution of AR-PAM images under different noise interferences.

[0053] Figure 4 The following are the results of vascular image reconstruction under different half-widths (FWHMs): Figure 4 (a) to (d) show the enhancement results of AR-PAM images with a signal-to-noise ratio (SNR) of 35 dB and a half-width (FWHM) of 35 μm, 55 μm, 75 μm, and 95 μm, respectively. Figure 4 (e) to (h) show the corresponding input images. Figure 4 (i) to (l) show the same true value (OR-PAM image), Figure 4 (a1) to (d3) and Figure 4 (e1) to (h3) are Figure 4 The enlarged images indicated by the white dotted line frame 1 and the white dotted line frame 2 in (a) to (l) are shown in FIG. Figure 4 (m) and Figure 4 The (n) in the figure shows the FWHM at different half-heights. Figure 4 (a1) to (d3) and Figure 4Signal distribution at the white dotted line position from (e1) to (h3); FWHM represents half-maximum width, NPA represents normalized photoacoustic amplitude; In1, input noise-free image 1; In2, input image 2, signal-to-noise ratio is 35db; In3, input image 3, signal-to-noise ratio is 25db; In4, input image 4, signal-to-noise ratio is 15db; GT, true value; Out1, output noise-free image 1; Out2, output image 2, signal-to-noise ratio is 35db; Out3, output image 3, signal-to-noise ratio is 25db; Out4, output image 4, signal-to-noise ratio is 15db.

[0054] According to step S1, AR-PAM images with half-width at half maximum FWHM of 35 μm, 55 μm, 75 μm and 95 μm are synthesized, covering the resolution range of most AR-PAM imaging systems. The signal-to-noise ratio (SNR) of all AR-PAM images is set to 35 dB to simulate the noise in a real imaging system. Figure 4 (a) to (d) show the output results of the model when the AR-PAM images with half-width at half maximum FWHM of 35μm, 55μm, 75μm and 95μm are used as input, respectively. Figure 4 (e) to (h) in Figure 3 show the corresponding input images. For reference, Figure 4 (i) to (l) in the figure show the same true value (OR-PAM image). As the FWHM increases, the topological structure of the AR-PAM image becomes blurred. Figure 4 The capillaries indicated by the two white arrows in (e) to (h) gradually become difficult to distinguish. Compared with the AR-PAM images, Figure 4 As shown in (a) to (d), the lateral resolution and clarity of the enhanced image are significantly improved. Figure 4 The capillaries indicated by the two white arrows in (a) to (d) can be well reconstructed by the model, especially in extremely blurred scenes with a half-width (FWHM) of 75 μm and 95 μm, the blood vessels can also be well reconstructed. Figure 4 (a1) to (d3) and Figure 4 (e1) to (h3) show Figure 4 The two magnified images indicated by the white dashed box 1 and the white dashed box 2 in (a) to (l) in the figure. At different FWHM levels, the capillaries that are too close to be distinguished in the input image are effectively resolved in the output image. Figure 4 (m) and Figure 4 (n) in the figure shows the input image, true value and enhanced image under different half-height width FWHM conditions. Figure 4The signal distribution at the positions of the white dotted box 1 and the white dotted box 2 in (a) to (l) in FIG. 1 . As the half-width FWHM increases, the blurring degree of the input image gradually increases, presenting a smoother signal curve. However, the signal distribution of the output image is highly consistent with the true value. These results show that the proposed method has superior performance in enhancing AR-PAM images at different blurring degrees.

[0055] Figure 5 The results of image enhancement and noise reduction under different signal-to-noise ratios (SNR) are shown. Figure 5 (a) to (d) show the enhancement results of noise-free AR-PAM images with SNR of 35dB, 25dB, and 15dB, respectively. The FWHM of all images is set to 55μm. Figure 5 (e) to (h) show the corresponding input images. Figure 5 (i) to (l) show the same true value (OR-PAM image), Figure 5 (a1) to (d3) and Figure 5 (e1) to (h3) are Figure 5 The enlarged images (close-up areas) indicated by the white dotted frame 1 and the white dotted frame 2 in (a) to (l) are shown in FIG. Figure 5 (m) and Figure 5 The (n) in the figure shows the difference between the SNR and the Figure 5 (a1) to (d3) and Figure 5 Signal distribution at the white dotted line position from (e1) to (h3), NPA represents normalized photoacoustic amplitude; SNR represents signal-to-noise ratio; In1, input noise-free image 1; In2, input image 2, signal-to-noise ratio is 35db; In3, input image 3, signal-to-noise ratio is 25db; In4, input image 4, signal-to-noise ratio is 15db; GT, true value; Out1, output noise-free image 1; Out2, output image 2, signal-to-noise ratio is 35db; Out3, output image 3, signal-to-noise ratio is 25db; Out4, output image 4, signal-to-noise ratio is 15db.

[0056] Since noise is inevitable in real imaging scenes, different levels of Gaussian noise are added to the AR-PAM images with the half-width at half maximum (FWHM) set to 55 μm to verify the reconstruction performance of the model under different noise levels. Figure 5 (a) to (d) show the enhancement results of AR-PAM images without noise and with SNR of 35dB, 25dB and 15dB respectively. Figure 5 (e) to (h) show the corresponding input images. Figure 5 (i) to (l) in show the same true value. Figure 5As shown by the white arrows in (e) to (h), as the noise level increases, the topological structure of the blood vessels in the AR-PAM image is gradually obscured, and some weak blood vessel signals are submerged by the noise. In contrast, the resolution and contrast of the output image are significantly improved, even if the blurred capillaries in the input image (such as Figure 5 As shown by the white arrows in (a) to (l), its structural features can also be clearly captured and analyzed. Figure 5 (a1) to (d3) and Figure 5 (e1) to (h3) show Figure 5 The two magnified images indicated by the white dotted box 1 and the white dotted box 2 in (a) to (l). Compared with the case where the vascular structure in the input image cannot be distinguished, the enhanced results of the model show clearer vascular contours and more details, and the boundaries between blood vessels are clear, which is highly consistent with the OR-PAM image of the same area. Figure 5 (m) and Figure 5 The (n) in the figure shows the input image, the true value and the enhanced image at different signal-to-noise ratio (SNR) levels. Figure 5 (a1) to (d3) and Figure 5 Signal distribution from the white dashed line in (e1) to (h3) in Figure 3. Under different degrees of noise interference, each sharp signal in the model output maintains good consistency with the true value, which proves the superior performance of the proposed method in enhancing AR-PAM image resolution and denoising.

[0057] Figure 6 The figure shows the comparison between the model iteration of the method of the present invention and other algorithms, where Figure 6 (a) and Figure 6 (b) in the figure shows the results of using the method proposed in the present invention and the RL deconvolution method (RL Deconv) at different numbers of iterations. Figure 6 (c) and Figure 6 (d) shows the change of peak signal-to-noise ratio PSNR and structural similarity SSIM with the number of iterations. Figure 6 (e) in is the input image, Figure 6 (f) to (i) in the figure show the results of the proposed method, RL deconvolution, CycleGAN and FDU-net respectively. Figure 6 (j) in the figure is the true value, GT represents the true value, NPA represents the normalized photoacoustic amplitude, and RL Deconv represents Richardson-Lucy deconvolution.

[0058] The performance of the proposed model was further evaluated using in vivo experimental data. Figure 6 (a) and Figure 6(b) in the figure shows the iterative reconstruction process of an AR-PAM image with a signal-to-noise ratio (SNR) of 35 dB and a half-width height (FWHM) of 55 μm using the method proposed in the present invention and the RL deconvolution method. Figure 6 As shown by the three white arrows in (b), for the result of RL deconvolution, at the 20th iteration, the image quality is improved and some capillaries become distinguishable. However, as the number of iterations increases, the improvement effect gradually weakens and the final image restoration degree is limited. For the method proposed in the present invention, reconstruction starts from a low-resolution noisy image. Figure 6 As shown in (a) in Figure 2, starting from the 60th iteration, the noise is gradually removed and the blood vessel contour becomes clearer. Figure 6 As shown by the three white arrows in (a), when the iterations exceed 80 times, most of the capillaries emerge from the noise, significantly improving the clarity and contrast of the image. Figure 6 (c) in Figure 6 and (d) in Figure 6 quantitatively show the changes of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) with the number of iterations. The RL deconvolution method slightly improves the structural similarity (SSIM) in the initial stage, but as the number of iterations increases, the structural similarity (SSIM) tends to stabilize or even decreases, and the peak signal-to-noise ratio (PSNR) also shows a similar downward trend. In the end, both remain at a low level, with the peak signal-to-noise ratio (PSNR) tending to 13.54 dB and the structural similarity (SSIM) tending to 0.59. For the method proposed in the present invention, when the number of iterations reaches Figure 6 (c) and Figure 6 Before the position indicated by the red arrow in (d), due to the presence of a lot of noise in the image, the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are lower than those of the RL method. However, when the number of iterations exceeds 80, the noise in the image is gradually removed, and the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) increase rapidly, significantly surpassing the results of RL deconvolution. At the 100th iteration, the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) reach 31.96 dB and 0.91, respectively. It should be noted that Figure 6 (c) and Figure 6 The two performance curves of the proposed method in (d) rise very slowly in the initial stage and are not completely stable in the last few iterations. This is due to the order in which the model processes the two types of degradation. The mean reversion diffusion model tends to prioritize the main degradation (i.e., deblurring) and only performs Gaussian denoising in the last few iterations. Therefore, the performance curve rises slowly in the early stage (the first 60 iterations). Figure 6As can be seen in (a) in the figure, during the iterative reconstruction of the blood vessels, most of the blur is removed in the intermediate time steps (the first 60 iterations). However, due to the presence of Gaussian noise, the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) improve slowly. In the last few iterations, the model begins to focus on denoising, resulting in a rapid improvement in image quality. This is reflected in the performance curve rising sharply and fully converging at 100 iterations. In summary, these results show that the proposed method achieves significantly superior performance in enhancing the resolution of AR-PAM images compared to traditional methods.

[0059] Figure 6 (e) shows the input AR-PAM image with a lateral resolution of 55μm and a signal-to-noise ratio (SNR) of 35dB. Figure 6 (f) to (i) in the figure show the enhancement results of AR-PAM images using the proposed method, RL deconvolution, CycleGAN and FDU-net respectively. The real value is as follows Figure 6 As shown in (j) in the figure. RL deconvolution has limited improvement in lateral resolution, and most areas are still blurred. CycleGAN and FDU-net improve the lateral resolution of AR-PAM to a certain extent, but due to the loss of high-frequency information, details are missing and there is a certain difference from the true value. The method proposed in the present invention significantly enhances the lateral resolution and image quality of AR-PAM, retaining most of the details consistent with the true value. Quantitative analysis shows that the method proposed in the present invention improves the peak signal-to-noise ratio PSNR and structural similarity SSIM of AR-PAM to 31.96dB and 0.91, respectively, which are 18.42dB and 0.32 higher than RL deconvolution, 2.5dB and 0.03 higher than CycleGAN, and 3.3dB and 0.17 higher than FDU-net. In summary, the method proposed in the present invention not only effectively improves the lateral resolution and image quality, but also retains most of the vascular details, achieving the best image quality and enhancement performance consistent with the true value.

[0060] The results shown in Tables 1 to 3 are quantitative results of the method of the present invention for AR-PAM images of point-like objects at different signal-to-noise ratios SNR, AR-PAM images with different half-width and half-height FWHM, and structural similarity SSIM and peak signal-to-noise ratio PSNR of images at different signal-to-noise ratios SNR.

[0061] Table 1 Figure 3 PSNR and SSIM values ​​of each image in

[0062] In order to quantitatively evaluate the performance of the model, the mean and standard deviation of the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are calculated based on 50 test images, as shown in Table 1. Under the conditions of no noise and when the SNR is 35dB and 25dB respectively, the peak signal-to-noise ratio (PSNR) of the enhanced image exceeds 45dB, which is about 91%, 92% and 102% (in dB) higher than the respective input images. As the SNR decreases from 35dB to 15dB, each time by 10dB, the structural similarity (SSIM) of the output image increases by 0.68 times, 9.5 times and 309 times compared with the input image. The quantitative data show that the method proposed in the present invention can effectively improve the lateral resolution and image quality of AR-PAM images at different SNR levels, and can achieve significant results even under the extreme condition of SNR of 15dB.

[0063] Table 2 Figure 4 PSNR and SSIM values ​​of each image in

[0064] In order to quantitatively evaluate the performance of the model, the mean and standard deviation of the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of the 37 test images are calculated, as shown in Table 2. When the half-width FWHM is 35μm, the peak signal-to-noise ratio (PSNR) of the enhanced image is improved by 9.83dB compared with the input image, and the structural similarity (SSIM) is improved by 0.29. As the half-width FWHM gradually increases from 55μm to 95μm, with an increase of 20μm each time, the peak signal-to-noise ratio (PSNR) of the output image is improved by about 56%, 63%, and 71% (in dB) compared with the input image, respectively. Similarly, the structural similarity (SSIM) of the output image is improved by 1.00 times, 1.48 times, and 1.70 times compared with the input image, respectively. The quantitative data show that the model shows strong applicability and stability in AR-PAM images of different lateral resolutions, and the enhancement effect of the model is more significant as the lateral resolution deteriorates.

[0065] Table 3 Figure 5 PSNR and SSIM values ​​of each image in

[0066] In order to quantitatively evaluate the performance of the model, a quantitative analysis was performed based on 37 test images, and the results are shown in Table 3. Under noise-free conditions, the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) between the output image and the true value are 28.19 dB and 0.84, respectively, while the peak signal-to-noise ratio (PSNR) of the input image is only 16.35 dB, and the structural similarity (SSIM) is only 0.41. The peak signal-to-noise ratio (PSNR) of the output image is increased by about 72% in dB, and the structural similarity (SSIM) is increased by 2.05 times. For other noise levels of 35, 25, and 15 dB SNR, compared with the input image, the peak signal-to-noise ratio (PSNR) of the output image is increased by about 56%, 51%, and 66% in dB, respectively, and the structural similarity (SSIM) is increased by 1.00, 1.96, and 4.80 times, respectively. The quantitative data show that the method proposed in the present invention can effectively enhance the resolution of AR-PAM images under different noise interferences, and the output results have good consistency with the true values.

[0067] The present invention also provides an application of an acoustic resolution photoacoustic microscope image enhancement and denoising method based on a mean-reverting diffusion model in the aforementioned embodiment in the field of image processing, which is specifically applied to actual scenarios such as medical image processing and satellite image denoising that require improved image quality, and can denoise images in low-light and high-noise environments.

[0068] It should be understood that the above description of the preferred embodiment is relatively detailed and cannot be regarded as limiting the scope of patent protection of the present invention. Under the enlightenment of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested for the present invention shall be based on the attached claims.

Claims

1. An AR-PAM image enhancement and denoising method based on Mean-reverting diffusion model, characterized in that: The method comprises the following steps: Step S1: In the training data generation step, image data is first obtained from a high-resolution OR-PAM image with a resolution ranging from 1 μm to 5 μm, and the OR-PAM image is converted into a low-resolution AR-PAM image with a resolution ranging from 20 μm to 100 μm by a convolution operation to simulate the imaging characteristics of the AR-PAM system; then, Gaussian noise of different intensities is added to the image data. , generate a noisy image dataset with different levels of degradation: ， in, Represents the simulated AR-PAM image, is the original OR-PAM image, is the point spread function, is the convolution operation; Step S2: In the forward SDE step, the generated noisy image dataset is processed using the Mean-reverting diffusion model to gradually degrade the high-resolution image into a low-resolution image state with fixed Gaussian noise. This process is modeled by a stochastic differential equation: ， in, Represents image data at time t status, is a fixed Gaussian noise mean state, represents the velocity parameter of the time-dependent mean reversion, is the diffusion coefficient, It is a Brownian motion process; Step S3, in the reverse SDE recovery step, the original high-resolution image is gradually restored from the noisy state by simulating the reverse SDE process. This process estimates the score function by training a noise network to guide the reverse generation of the image: ， in, is the score function predicted by the noisy network.

2. The AR-PAM image enhancement and denoising method based on the Mean-reverting diffusion model according to claim 1, characterized in that: The specific process of step S1 is as follows: First, high-resolution OR-PAM images are obtained as the basic data source. OR-PAM images are tissue structure images captured with a high lateral resolution of less than 5 microns by using optical detection methods. Tissue structure images contain rich tissue details and characteristic information. In the process of converting the high-resolution OR-PAM image into the low-resolution AR-PAM image, it is assumed that the PSF of the acoustic resolution photoacoustic microscope system satisfies the Gaussian distribution, and its standard deviation is related to the focal spot size and frequency response of the imaging system. The high-resolution OR-PAM image is convolved with the PSF to generate a simulated low-resolution AR-PAM image. The mathematical model of the convolution operation is expressed as: ， in, represents the convolution operation, represents the point spread function of the system; In order to simulate the noise interference that may occur in the real imaging process, Gaussian noise is added to the generated AR-PAM image. The mathematical expression of this process is: ， in, n is Gaussian noise with zero mean.

3. The AR-PAM image enhancement and denoising method based on the Mean-reverting diffusion model according to claim 1, characterized in that: After step S3, a model optimization step is also included, and the specific process is as follows: Through multiple rounds of iterative optimization, the parameters of the model are further adjusted. , including the diffusion coefficient and mean reversion speed , ensuring that the noise removal and image restoration processes within each time step can achieve the best results; After each round of iteration, the denoising effect and image restoration quality of the model are evaluated by calculating the peak signal-to-noise ratio and structural similarity of the restored image. The performance of the model is gradually improved by continuously adjusting and optimizing the model parameters, making the image restoration effect in the reverse SDE process more stable and accurate. Eventually, the parameters of the model will be fixed to form a fully trained noise network and optimized model.

4. The AR-PAM image enhancement and denoising method based on the Mean-reverting diffusion model according to claim 3, characterized in that: After the model optimization step, an evaluation step is also included, and the specific process is as follows: The optimized noise network and optimization model are applied to a set of low-resolution AR-PAM image datasets collected actually. The forward SDE and reverse SDE recovery processes are performed in sequence by inputting the low-resolution AR-PAM image datasets collected actually into the optimized noise network and optimization model, thereby outputting enhanced high-resolution images. Multiple evaluation indicators are used for evaluation, including peak signal-to-noise ratio and structural similarity. The specific effect of the model in improving image quality is quantified by calculating the peak signal-to-noise ratio and structural similarity of the image before and after processing; Based on the evaluation results, the model parameters are further optimized or the algorithm structure is adjusted to improve the performance of the model and make it more suitable for various complex situations in practical applications.

5. An application of an acoustic resolution photoacoustic microscopy image enhancement and noise reduction method based on a mean-reverting diffusion model as described in any one of claims 1 to 4 in the field of image processing.

Images