A PT-OCT high signal-to-noise ratio fast imaging method

Abstract

This invention proposes a high signal-to-noise ratio (SNR) fast imaging method for PT-OCT, comprising: S10: sampling images at multiple different locations within the same region to construct a training set, wherein each sample in the training set includes a short-sampled PT-OCT image and a long-sampled PT-OCT image at the same location; S20: inputting the training set into a Pix2Pix generative adversarial network for training to obtain a trained generative adversarial network, wherein the Pix2Pix generative adversarial network includes a generator, a discriminator, and a trainer, the generator being a U-Net architecture, and embedding multiple cascaded residual modules in the bottleneck layer; S30: inputting the short-sampled PT-OCT image of the region to be detected into the trained generative adversarial network to obtain a high SNR reconstructed photothermal OCT image, which enables PT-OCT to obtain high SNR imaging results close to those of long-sampled images while significantly reducing the number of M-scan sampling points and the amount of data.

Description

Technical Field

[0001] This invention relates to the fields of optical detection and biomedical functional imaging, and in particular to a PT-OCT high signal-to-noise ratio fast imaging method. Background Technology

[0002] Photothermal optical coherence tomography (PT-OCT) induces a photothermal effect in the absorber of the measured object through pump light, causing minute changes in the local refractive index and optical path. This allows for the acquisition of more sensitive functional contrast information on absorption distribution, building upon OCT tomography. Since biological samples have limited pump energy tolerance, the effective photothermal signal of PT-OCT is typically weak and susceptible to system noise and environmental disturbances. Therefore, existing techniques often employ lock-in amplification, synchronously accumulating and averaging the phase or correlation signals over multiple modulation cycles to improve the signal-to-noise ratio. However, such methods are highly dependent on the number of sampling points N in M-scan mode, i.e., the number of times the same spatial point is repeatedly sampled. When the system sampling rate fs is fixed, as the number of sampling points N increases, the frequency resolution of the Fourier transform Δf = fs / N increases. Furthermore, the total energy for white noise is approximately constant in the frequency domain, making the average noise power allocated to each frequency resolution unit inversely proportional to N. Simultaneously, the spectral energy of the narrowband photothermal signal at the modulation frequency f0 becomes more concentrated with increasing sampling duration, resulting in a linear signal-to-noise ratio that increases proportionally with N. When the number of sampling points increases from N=50 to N=1000, the signal-to-noise ratio can be improved by more than ten times. However, the cost is a significant increase in the single imaging acquisition time and a significant increase in the amount of data. This means that the existing technology has the following technical contradictions: reducing the number of sampling points can achieve faster imaging speed, but it will introduce stronger random noise and stripe artifacts, making it difficult to identify weak photothermal textures and tissue boundaries. Increasing the number of sampling points can improve the signal-to-noise ratio, but it will significantly reduce the imaging speed and increase the risk of motion artifacts. This makes it difficult for the existing PT-OCT to meet the needs of real-time diagnosis and dynamic process monitoring.

[0003] Therefore, in order to meet the need for an image reconstruction method that can recover the image quality of long-sampling under short-sampling conditions, thereby simultaneously achieving high signal-to-noise ratio and fast imaging in PT-OCT. Summary of the Invention

[0004] To address the above issues, this invention proposes a high signal-to-noise ratio (SNR) rapid imaging method using PT-OCT, which enables PT-OCT to achieve near-long-sampling-ratio imaging results while significantly reducing the number of M-scan sampling points and data volume. To achieve the above objective, this invention provides the following technical solution: On one hand, this invention provides a high SNR rapid imaging method using PT-OCT, comprising:

[0005] S10: Sample images from multiple different locations within the same region to construct a training set. Each sample in the training set includes a short-sampled PT-OCT image and a long-sampled PT-OCT image from the same location. S20: Input the training set into a Pix2Pix generative adversarial network for training to obtain the trained generative adversarial network. The Pix2Pix generative adversarial network includes a generator, a discriminator, and a trainer. The generator is a U-Net architecture and multiple residual modules are embedded in series in the bottleneck layer. S30: Input the short-sampled PT-OCT image of the region to be detected into the trained generative adversarial network to obtain a high signal-to-noise ratio reconstructed photothermal OCT image.

[0006] This invention reconstructs images of near-long-sampling (high signal-to-noise ratio) quality from short-sampling (low signal-to-noise ratio) data using a deep learning model, thereby achieving rapid PT-OCT imaging while significantly reducing the number of M-scan sampling points and data volume. Furthermore, the invention simultaneously improves the generator network structure by embedding cascaded residual modules into the bottleneck layer of the U-Net generator, enhancing deep feature representation and effectively solving the gradient vanishing problem in deep networks. This enables the model to recover more refined and realistic biological structural information from short-sampling images with noise interference, significantly improving the generative adversarial network's ability to recover weak photothermal textures and tissue boundaries, and increasing the signal-to-noise ratio of the reconstructed PT-OCT images. Thus, it achieves high signal-to-noise ratio and rapid imaging in PT-OCT.

[0007] Furthermore, the training process of the generative adversarial network includes: The generator sequentially encodes, performs multiple residual forward propagations, and decodes the short-sampled PT-OCT images in the training set to obtain the reconstructed PT-OCT images. The discriminator performs discrimination processing on the reconstructed PT-OCT image and its corresponding short-sampled PT-OCT image to obtain a first discrimination result; at the same time, it performs discrimination processing on the corresponding short-sampled PT-OCT image and its corresponding long-sampled PT-OCT image to obtain a second discrimination result. The trainer calculates the generator adversarial loss and the discriminant adversarial loss based on the first and second discrimination results; simultaneously, it calculates the pixel-level loss based on the pixel values ​​of the reconstructed PT-OCT image and its corresponding long-sampled PT-OCT image; then, it optimizes the parameters of the generator by minimizing the composite loss function composed of the adversarial loss and the pixel-level loss; and it optimizes the parameters of the discriminator by minimizing the discriminant adversarial loss. Repeat the above process until the model converges or the number of iterations exceeds a preset threshold, thus completing the training.

[0008] Furthermore, while the discriminator performs discrimination processing based on the reconstructed PT-OCT image and the corresponding short-sampled PT-OCT image, it also extracts intermediate features to obtain a first intermediate feature group. While performing discrimination processing based on the reconstructed PT-OCT image and the corresponding short-sampled PT-OCT image, intermediate features are extracted to obtain a second intermediate feature group. The trainer calculates the feature matching loss based on the first intermediate feature group and the second intermediate feature group; then it minimizes the composite loss function consisting of adversarial loss, pixel-level loss and feature matching loss to optimize the parameters of the generator; and minimizes the discriminative adversarial loss to optimize the parameters of the discriminator.

[0009] Furthermore, the pixel-level loss is calculated using the following formula:

[0010] in, H, W Image size, To reconstruct PT-OCT images ( The pixel value at the pixel point. For the corresponding long-sampled PT-OCT image ( The pixel value at the pixel point.

[0011] Furthermore, the feature matching loss is calculated using the following formula:

[0012] Where N is the total number of intermediate layers. For the first intermediate feature group in the th The output of the intermediate layer, For the second intermediate feature group in the th Output of the intermediate layer.

[0013] Furthermore, after step S10, preprocessing is also included, which involves normalizing the image intensity using the following formula:

[0014] Wherein, P is the original grayscale value. The maximum pixel value in the image sequence. This represents the minimum pixel value in the image sequence.

[0015] Furthermore, the discriminator is a PatchGAN discriminator.

[0016] On the other hand, the present invention also provides a generative adversarial network for high signal-to-noise ratio fast imaging in PT-OCT, including a generator, a discriminator and a trainer. The generator is a U-Net architecture and multiple residual modules are embedded in series in the bottleneck layer.

[0017] Furthermore, the generator is used to sequentially encode, perform multiple residual forward propagation, and decode the short-sampled PT-OCT images in the training set to obtain reconstructed PT-OCT images; The discriminator is used to perform discrimination processing on the reconstructed PT-OCT image and its corresponding short-sampled PT-OCT image to obtain a first discrimination result; at the same time, it performs discrimination processing on the corresponding short-sampled PT-OCT image and its corresponding long-sampled PT-OCT image to obtain a second discrimination result. The trainer includes a generation discriminative adversarial loss calculation subunit, a pixel loss calculation subunit, and an iterative optimization subunit; Generate discriminative adversarial loss calculation subunit: used to calculate the generative adversarial loss and the discriminative adversarial loss based on the first discrimination result and the second discrimination result; Pixel loss calculation subunit: used to calculate pixel-level loss based on the pixel values ​​of the reconstructed PT-OCT image and its corresponding long-sampled PT-OCT image; Iterative optimization subunit: used to minimize the composite loss function consisting of adversarial loss and pixel-level loss to optimize the parameters of the generator; and to minimize the discriminative adversarial loss to optimize the parameters of the discriminator.

[0018] Furthermore, the discriminator includes an intermediate feature extraction subunit, which is used to perform intermediate feature extraction while performing discrimination processing based on the reconstructed PT-OCT image and the corresponding short-sampled PT-OCT image, or the corresponding long-sampled PT-OCT image, to obtain a first intermediate feature group or a second intermediate feature group. The trainer also includes: Feature matching loss subunit: used to calculate the feature matching loss based on the first intermediate feature group and the second intermediate feature group.

[0019] In summary, this invention proposes a high signal-to-noise ratio fast imaging method for PT-OCT, which has the following advantages compared to existing technologies: 1. This invention provides a high signal-to-noise ratio (SNR) rapid imaging method for PT-PCT, resolving the technical contradiction of traditional PT-OCT in balancing imaging speed and SNR. By using a deep learning model to reconstruct images with near-long-sampling (high SNR) quality from short-sampling (low SNR) data, high SNR and rapid imaging are achieved in PT-OCT while significantly reducing the number of M-scan sampling points and data volume.

[0020] 2. This invention significantly improves the ability to recover weak photothermal textures and tissue boundaries by modifying the generator network structure. By embedding cascaded residual modules into the bottleneck layer of the U-Net generator, deep feature representation is enhanced, effectively solving the gradient vanishing problem in deep networks. This allows the model to recover more refined and realistic biological structural information from short-sampled images with noise interference.

[0021] 3. The composite loss function used in this invention incorporates feature matching loss, which effectively optimizes the structural fidelity and high-frequency texture realism of the reconstructed image. This loss forces the generated image to maintain consistency with the real image in the feature space of multiple intermediate layers of the discriminator, thereby guiding the generator to learn a more refined high-frequency detail recovery capability. This avoids the over-smoothing problem that may be caused by pixel loss, ensuring that the reconstructed image has both structural integrity and sharp edge texture.

[0022] 4. The method proposed in this invention significantly shortens the imaging acquisition time and improves image quality, expanding the application prospects of PT-OCT in the biomedical field. Its high signal-to-noise ratio and rapid imaging capabilities enable it to meet the needs of real-time diagnosis and dynamic process monitoring, providing more efficient and reliable technical support for disease diagnosis and treatment evaluation.

[0023] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0024] Figure 1 A structural diagram of an existing photothermal-optical coherence tomography system; Figure 2 This is a flowchart of the PT-OCT high signal-to-noise ratio fast imaging method of the present invention; Figure 3 This is an exemplary structural block diagram of a generative adversarial network applied to a PT-OCT high signal-to-noise ratio fast imaging system according to the present invention; Figure 4 for Figure 3 The flowchart shown illustrates the execution process of each component in the training process of a Generative Adversarial Network (GAN). Figure 5 This is a comparison of the frontal projection of black and white hair samples under short sampling, network reconstruction, and long sampling conditions in the example. Detailed Implementation

[0025] Existing photothermal optical coherence tomography is based on the following Figure 1 The illustrated photothermal optical coherence tomography system consists of a time-modulated 532 nm pump source and a spectral domain OCT (SD-OCT) system. The SD-OCT probe source is a superluminescent diode with a center wavelength of 1310 nm and a bandwidth of 55 nm. The emitted coherent light is split into a reference arm and a sample arm at a 50:50 ratio via a 2×2 fiber coupler. In the sample arm, a 532 nm modulated continuous laser serves as the pump source, coaxially illuminating the sample surface with the low-coherence probe light through a dichroic mirror, inducing a local photothermal response. The backscattered light from the sample interferes with the reference light reflected from the reference mirror at the coupler, and is received by a spectrometer and recorded by a linear CCD camera. The acquisition rate of the linear CCD camera is set to 46.816 kHz. Under this physical architecture, the signal-to-noise ratio improvement of the photothermal signal is highly dependent on the number of sampling points N in M-scan mode. From the frequency domain perspective of signal processing, theoretically, SNR increases positively with the increase of the number of sampling points N, and its variation follows the following pattern: This means that when the number of sampling points increases from 50 to 1000, although the signal-to-noise ratio can be improved by tens of times, the acquisition time increases from a few minutes to nearly two hours, and the amount of raw data surges from 23.8 GB to 476.0 GB, which seriously hinders the need for rapid diagnosis.

[0026] To address the aforementioned contradictions, this invention attempts to use generative adversarial networks (GANs) to generate high-noise samples from low-signal-to-noise ratio (SNR) samples. GANs can generate long-sampling PT-OCT images of the corresponding regions based on short-sampling input PT-OCT, thus improving imaging speed. However, ordinary U-Nets are insufficient at the bottleneck layer (i.e., the layer with the lowest feature map resolution and largest receptive field) for capturing weak photothermal textures dominated by noise, thus failing to meet the high SNR requirement. Therefore, this invention innovatively embeds three cascaded residual blocks into the generator bottleneck, resulting in an improved Pix2Pix model. Based on this model, a high SNR fast PT-OCT imaging method is proposed. (See reference...) Figure 2 It includes: S10: Sample images from multiple different locations within the same region to construct a training set. Each sample in the training set includes a short-sampled PT-OCT image and a long-sampled PT-OCT image from the same location.

[0027] PT-OCT images are produced by Figure 1The optical coherence tomography (OCT) system shown acquires data. Based on the set number of scans M, short-sampling and long-sampling OCT data are acquired for the same imaging object under the same imaging system and spatial location conditions, respectively, to obtain a training sample. Multiple training samples are used to construct a training set. Preferably, after step S10, preprocessing is further included, normalizing the pixel intensity of the short-sampling image and the long-sampling image using the following formula:

[0028] Wherein, P is the original grayscale value. The maximum pixel value in the image sequence. The minimum pixel value of the image sequence is used as the training set for the processed paired samples.

[0029] Then, step S20 is executed: the training set is input into a Pix2Pix generative adversarial network for training to obtain a trained generative adversarial network. The Pix2Pix generative adversarial network includes a generator, a discriminator, and a trainer. The generator is a U-Net architecture and multiple residual modules are embedded in series in the bottleneck layer.

[0030] Please see Figure 3 and Figure 4 The Pix2Pix generative adversarial network includes a generator, a discriminator, and a trainer.

[0031] The generator is used to sequentially encode, perform multiple residual forward propagation, and decode the short-sampled PT-OCT images in the training set to obtain the reconstructed PT-OCT images.

[0032] The existing Pix2Pix model's generator G uses U-Net as its basic skeleton. The encoder path is responsible for extracting multi-scale abstract features from the low signal-to-noise ratio input image, and the decoder path gradually recovers pixel-level semantic mapping through convolution and upsampling. However, it is insufficient for capturing weak light and heat textures dominated by noise. Therefore, each residual module consists of two 3x3 convolutional layers and a ReLU activation function, introducing identity mapping shortcut connections. Let the input features be... The forward propagation of the residual module is represented as Through this design, even in the backpropagation stage, the residual function... gradient Approaching zero, identity term The total gradient can still be guaranteed. By propagating forward at least at a unit intensity, the gradient vanishing problem in deep networks when dealing with complex noise in PT-OCT is fundamentally solved, enabling the bottleneck layer to progressively optimize features and allowing the model to recover more detailed and realistic biological structural information from short-sampled images with noise interference.

[0033] The discriminator is used to perform discrimination processing on the reconstructed PT-OCT image and its corresponding short-sampled PT-OCT image to obtain a first discrimination result; at the same time, it performs discrimination processing on the corresponding short-sampled PT-OCT image and its corresponding long-sampled PT-OCT image to obtain a second discrimination result.

[0034] The goal of the discriminator is to maximize the probability of classifying real image pairs as real and generated image pairs as fake. Therefore, during training, it is necessary to simultaneously classify the reconstructed PT-OCT image with the corresponding short-sampled PT-OCT image and the reconstructed PT-OCT image with the corresponding long-sampled PT-OCT image in order to calculate the loss of the discriminator D.

[0035] The trainer includes a generation discriminative adversarial loss calculation subunit, a pixel loss calculation subunit, and an iterative optimization subunit; Generate discriminative adversarial loss calculation subunit: used to calculate the generative adversarial loss and the discriminative adversarial loss based on the first discrimination result and the second discrimination result; Pixel loss calculation subunit: used to calculate pixel-level loss based on the pixel values ​​of the reconstructed PT-OCT image and its corresponding long-sampled PT-OCT image; Iterative optimization subunit: used to minimize the composite loss function consisting of adversarial loss and pixel-level loss to optimize the parameters of the generator; and to minimize the discriminative adversarial loss to optimize the parameters of the discriminator.

[0036] Generate discriminative adversarial loss calculation subunit: used to calculate the generative adversarial loss and the discriminative adversarial loss based on the first discrimination result and the second discrimination result; The optimization objective of standard GANs is to drive the generator to produce images that are indistinguishable from real images at the local patch level through a minimax game between the generator and the discriminator. The adversarial loss of the generator is defined as:

[0037] in, For the input short-sampled, low signal-to-noise ratio image, This corresponds to a fully sampled, high signal-to-noise ratio real image. The reconstructed output of the generator. This is the discriminator for PatchGAN. During training, the discriminator... The generator attempts to maximize the objective (i.e., correctly distinguish between true and false), It then attempts to minimize this objective (i.e., deceive the discriminator), and this adversarial mechanism provides the generator with gradient feedback on the realism of local textures.

[0038] The adversarial loss of the discriminator is defined as:

[0039] Among them, among them, For the input short-sampled, low signal-to-noise ratio image, This corresponds to a fully sampled, high signal-to-noise ratio real image. The reconstructed output of the generator. This is the discriminator for PatchGAN.

[0040] The pixel loss calculation subunit 303.B is used to perform step S303.B: calculate the pixel-level loss based on the point-by-point pixel values ​​of the reconstructed PT-OCT image and the corresponding long-sampled PT-OCT image.

[0041] The pixel-level loss is calculated using the following formula:

[0042] Where H and W are the image dimensions, To reconstruct PT-OCT images ( The pixel value at the pixel point. For the corresponding long-sampled PT-OCT image ( The pixel value at the pixel point.

[0043] Finally, the composite loss function of the generator's total loss is obtained as follows:

[0044] in, These are the pixel weight coefficients. The generative adversarial loss is recovered using PatchGAN. The real speckle statistics and edge gradients within the local receptive field, along with pixel-level loss, directly constrain the point-by-point differences between the generated image and the real image in pixel space. This ensures that the reconstruction result maintains a high degree of consistency with the real image in terms of overall structure, contrast, and grayscale distribution, thereby improving the quality of the generated and reconstructed image.

[0045] During training, the generator and discriminator are optimized alternately. That is, in each iteration, the generator parameters are first fixed, and the optimization is achieved by minimizing... Update the discriminator; then fix the discriminator parameters and minimize... The generator is updated. Through the collaborative optimization of this composite loss function, the improved Pix2Pix model is able to simultaneously maintain global fidelity, local realism, and multi-scale semantic consistency, laying the algorithmic foundation for high signal-to-noise ratio fast imaging in photothermal OCT.

[0046] In another preferred embodiment, the inventors found that although the aforementioned pixel-level loss can ensure that the reconstruction result maintains a high degree of consistency with the real image in terms of overall structure, contrast, and grayscale distribution, this constrained point-by-point difference reconstruction method leads to the loss of high-frequency details. To compensate for the inadequacy of pixel loss in high-frequency detail recovery and to enable the reconstructed image to have both structural integrity and sharp edge texture, a feature matching loss sub-unit is introduced in the trainer. During the training process of this embodiment, the discriminator enables the intermediate feature extraction function during initialization, causing it to return an intermediate layer feature list. That is, the discriminator, after returning to the first discrimination, will also return the first intermediate feature set and the second intermediate feature set. The feature matching loss is then calculated using the following formula:

[0047] Where N is the total number of intermediate layers. For the first intermediate feature group in the th The output of the intermediate layer, For the second intermediate feature group in the th Output of the intermediate layer.

[0048] At this point, the composite loss function of the generator's total loss is:

[0049] in, These are pixel weight coefficients. In this embodiment of the invention, the feature matching weight coefficients are used. L1 loss is used with a large weight during training, while adversarial loss is recovered using PatchGAN. The system incorporates real speckle statistics and edge gradients within the local receptive field; feature matching loss is introduced with appropriate weights, and multi-scale feature constraints compensate for the shortcomings of L1 loss in high-frequency detail recovery, so that the reconstructed image has both structural integrity and sharp edge texture.

[0050] S30: Input the short-sampled PT-OCT image of the region to be detected into the trained generative adversarial network to obtain a high signal-to-noise ratio reconstructed photothermal OCT image.

[0051] Based on the trained Pix2Pix generative adversarial network, experiments were conducted on frontal projection images of black and white hair samples under short sampling, network reconstruction, and long sampling conditions. The results are attached. Figure 5 As shown, in the 50-scan image, the hair boundaries are blurred and the signal is extremely discontinuous due to speckle noise interference. After the improved network reconstruction, the background ghosting is greatly suppressed, and not only is the hair shape clearly visible, but the excellent fine thermal accumulation response texture is also restored, which is highly consistent with the true value of 500-scan.

[0052] In summary, this invention proposes a high signal-to-noise ratio fast imaging method for PT-OCT, which has the following advantages compared to existing technologies: 1. This invention provides a high signal-to-noise ratio (SNR) rapid imaging method for PT-PCT, resolving the technical contradiction of traditional PT-OCT in balancing imaging speed and SNR. By using a deep learning model to reconstruct images with near-long-sampling (high SNR) quality from short-sampling (low SNR) data, high SNR and rapid imaging are achieved in PT-OCT while significantly reducing the number of M-scan sampling points and data volume.

[0053] 2. This invention significantly improves the ability to recover weak photothermal textures and tissue boundaries by modifying the generator network structure. By embedding cascaded residual modules into the bottleneck layer of the U-Net generator, deep feature representation is enhanced, effectively solving the gradient vanishing problem in deep networks. This allows the model to recover more refined and realistic biological structural information from short-sampled images with noise interference.

[0054] 3. The composite loss function used in this invention incorporates feature matching loss, which effectively optimizes the structural fidelity and high-frequency texture realism of the reconstructed image. This loss forces the generated image to maintain consistency with the real image in the feature space of multiple intermediate layers of the discriminator, thereby guiding the generator to learn a more refined high-frequency detail recovery capability. This avoids the over-smoothing problem that may be caused by pixel loss, ensuring that the reconstructed image has both structural integrity and sharp edge texture.

[0055] 4. The method proposed in this invention significantly shortens the imaging acquisition time and improves image quality, expanding the application prospects of PT-OCT in the biomedical field. Its high signal-to-noise ratio and rapid imaging capabilities enable it to meet the needs of real-time diagnosis and dynamic process monitoring, providing more efficient and reliable technical support for disease diagnosis and treatment evaluation.

[0056] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and the present invention also intends to include these modifications and variations.

Claims

1. A high signal-to-noise ratio fast imaging method using PT-OCT, characterized in that, include: S10: Sample images from multiple different locations within the same region to construct a training set. Each sample in the training set includes a short-sampled PT-OCT image and a long-sampled PT-OCT image from the same location. S20: Input the training set into a Pix2Pix generative adversarial network for training to obtain the trained generative adversarial network. The Pix2Pix generative adversarial network includes a generator, a discriminator, and a trainer. The generator is a U-Net architecture and multiple residual modules are embedded in series in the bottleneck layer. S30: Input the short-sampled PT-OCT image of the region to be detected into the trained generative adversarial network to obtain a high signal-to-noise ratio reconstructed photothermal OCT image.

2. The PT-OCT high signal-to-noise ratio fast imaging method according to claim 1, characterized in that, The training process of the generative adversarial network includes: The generator sequentially encodes, performs multiple residual forward propagations, and decodes the short-sampled PT-OCT images in the training set to obtain the reconstructed PT-OCT images. The discriminator performs discrimination processing on the reconstructed PT-OCT image and its corresponding short-sampled PT-OCT image to obtain a first discrimination result; at the same time, it performs discrimination processing on the corresponding short-sampled PT-OCT image and its corresponding long-sampled PT-OCT image to obtain a second discrimination result. The trainer calculates the generator adversarial loss and the discriminant adversarial loss based on the first and second discrimination results; simultaneously, it calculates the pixel-level loss based on the pixel values ​​of the reconstructed PT-OCT image and its corresponding long-sampled PT-OCT image; then, it optimizes the parameters of the generator by minimizing the composite loss function composed of the adversarial loss and the pixel-level loss; and it optimizes the parameters of the discriminator by minimizing the discriminant adversarial loss. Repeat the above process until the model converges or the number of iterations exceeds a preset threshold, thus completing the training.

3. The PT-OCT high signal-to-noise ratio fast imaging method according to claim 2, characterized in that, While performing discrimination processing based on the reconstructed PT-OCT image and the corresponding short-sampled PT-OCT image, the discriminator extracts intermediate features to obtain the first intermediate feature group. While performing discrimination processing based on the reconstructed PT-OCT image and the corresponding short-sampled PT-OCT image, intermediate features are extracted to obtain a second intermediate feature group. The trainer calculates the feature matching loss based on the first intermediate feature group and the second intermediate feature group; then it minimizes the composite loss function consisting of adversarial loss, pixel-level loss and feature matching loss to optimize the parameters of the generator; and minimizes the discriminative adversarial loss to optimize the parameters of the discriminator.

4. The PT-OCT high signal-to-noise ratio fast imaging method according to claim 3, characterized in that, The pixel-level loss is calculated using the following formula: Where H and W are the image dimensions, To reconstruct PT-OCT images ( The pixel value at the pixel point. For the corresponding long-sampled PT-OCT image ( The pixel value at the pixel point.

5. The PT-OCT high signal-to-noise ratio fast imaging method according to claim 4, characterized in that, The feature matching loss is calculated using the following formula: Where N is the total number of intermediate layers. For the first intermediate feature group in the th The output of the intermediate layer, For the second intermediate feature group in the th Output of the intermediate layer.

6. The PT-OCT high signal-to-noise ratio fast imaging method according to any one of claims 1-5, characterized in that, Following step S10, preprocessing is also included, which involves normalizing the image intensity using the following formula: Wherein, P is the original grayscale value. The maximum pixel value in the image sequence. This represents the minimum pixel value in the image sequence.

7. The PT-OCT high signal-to-noise ratio fast imaging method according to claim 6, characterized in that, The discriminator is a PatchGAN discriminator.

8. A generative adversarial network for high signal-to-noise ratio fast imaging in PT-OCT, comprising a generator, a discriminator, and a trainer, characterized in that, The generator uses a U-Net architecture and embeds multiple residual modules in series at the bottleneck layer.

9. The generative adversarial network according to claim 8, characterized in that, The generator is used to sequentially encode, perform multiple residual forward propagation, and decode the short-sampled PT-OCT images in the training set to obtain the reconstructed PT-OCT images; The discriminator is used to perform discrimination processing on the reconstructed PT-OCT image and its corresponding short-sampled PT-OCT image to obtain a first discrimination result; at the same time, it performs discrimination processing on the corresponding short-sampled PT-OCT image and its corresponding long-sampled PT-OCT image to obtain a second discrimination result. The trainer includes a generation discriminative adversarial loss calculation subunit, a pixel loss calculation subunit, and an iterative optimization subunit; Generate discriminative adversarial loss calculation subunit: used to calculate the generative adversarial loss and the discriminative adversarial loss based on the first discrimination result and the second discrimination result; Pixel loss calculation subunit: used to calculate pixel-level loss based on the pixel values ​​of the reconstructed PT-OCT image and its corresponding long-sampled PT-OCT image; Iterative optimization subunit: used to minimize the composite loss function consisting of adversarial loss and pixel-level loss to optimize the parameters of the generator; and to minimize the discriminative adversarial loss to optimize the parameters of the discriminator.

10. The generative adversarial network according to claim 9, characterized in that, The discriminator includes an intermediate feature extraction subunit, which is used to perform intermediate feature extraction while performing discrimination processing based on the reconstructed PT-OCT image and the corresponding short-sampled PT-OCT image, or the corresponding long-sampled PT-OCT image, to obtain a first intermediate feature group or a second intermediate feature group. The trainer also includes: Feature matching loss subunit: used to calculate the feature matching loss based on the first intermediate feature group and the second intermediate feature group.

Images