A low-cost universal virtual phase contrast method and device based on a cylindrical lens
By incorporating cylindrical lenses into a bright-field microscope and using conditional generative adversarial neural networks, low-cost virtual phase-contrast images are generated, overcoming the difficulty of resolving hierarchical structures and textures when observing living cells with ordinary microscopes. This achieves similar effects to standard phase-contrast microscopes and expands the scope of applications.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- FUDAN UNIVERSITY
- Filing Date
- 2024-12-30
- Publication Date
- 2026-07-07
AI Technical Summary
Ordinary bright-field microscopes are unable to accurately distinguish the layered structure and texture of samples without damaging living cells, and specialized phase-contrast microscopes are expensive, limiting their widespread application.
By adding cylindrical lenses to a conventional bright-field microscope and combining them with a conditional generative adversarial neural network with confidence negative feedback, low-cost virtual phase-contrast images are generated to simulate the standard phase-contrast effect.
It enables low-cost virtual phase-contrast imaging, improves cell segmentation and recognition, and expands its application potential in fields such as biomedicine and materials science.
Smart Images
Figure CN119960160B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of optics and image processing technology, and particularly relates to the field of microscopic imaging technology, specifically to a low-cost, universal virtual phase contrast method and device based on cylindrical lenses. Background Technology
[0002] Ordinary bright-field microscopes often struggle to accurately distinguish the layered structure and texture of unstained, transparent samples, making sample identification difficult and posing challenges for researchers in biology, medicine, and other fields. Pretreatment methods such as staining offer a solution, but these methods can damage cells or tissues. Furthermore, staining is often applied to dead cells, making it less than optimal for studies requiring observation of living cells.
[0003] The invention of phase-contrast microscopy provides a relatively ideal solution to the above problems. Its principle is based on utilizing the phase changes caused by variations in density and thickness of transparent samples. These phase changes are converted into changes in light intensity in the image, allowing researchers to better identify the hierarchical structure, texture, and even internal structural changes of the sample through light intensity contrast. Standard phase-contrast microscope images have sharp, clear edges, high contrast, and a strong overall three-dimensional effect, facilitating rapid and accurate determination of cell structure and possessing high practical application value.
[0004] To achieve phase-contrast imaging, a dedicated annular aperture and matching objectives are required, which increases the cost of the optical system and limits its wider application. Therefore, improving the illumination path with a single cylindrical lens and combining it with a conditional generative adversarial neural network using confidence negative feedback to achieve a low-cost virtual phase-contrast effect can significantly reduce costs and has high application value. This allows researchers to achieve phase-contrast imaging observation at low cost without altering the structure of existing bright-field microscopes. Summary of the Invention
[0005] The purpose of this invention is to provide a low-cost, universal virtual phase contrast method and apparatus. Specifically, by adding a cylindrical lens to the illumination optical path and employing a conditional generative adversarial neural network with confidence negative feedback, the problem of not being able to observe object details under a conventional bright-field microscope at low cost is solved, providing a solution for realizing low-cost, universal virtual phase contrast.
[0006] The specific technical solution for implementing this invention is as follows.
[0007] This invention provides a low-cost, universal virtual phase contrast method based on cylindrical lenses, comprising the following steps:
[0008] By adding a cylindrical lens to a conventional bright-field microscope with a Köhler illumination structure, asymmetric illumination is achieved from the illumination source, resulting in a bright-field image. The resulting image with higher light intensity carries more phase information.
[0009] The obtained bright-field image is used to generate a virtual contrast image through a conditional generative adversarial neural network under a trained credibility negative feedback, thereby achieving an effect equivalent to a standard contrast image at low cost.
[0010] In this invention, a cylindrical lens is added between the illumination system of the microscope system and the stage on which the sample is placed. The focal length of the cylindrical lens remains constant, and the distance between the cylindrical lens and the sample remains constant after the network training is completed.
[0011] In this invention, a conditional generative adversarial neural network based on credibility negative feedback is used, where "condition" refers to a supervised network training method guided by truth values; the overall network includes a virtual contrast image generator, a virtual bright field image generator, and a discriminator.
[0012] A virtual phase-contrast image generator is used to generate virtual phase-contrast images from bright-field images containing cylindrical lenses. It learns to extract and enhance phase information from the bright-field image, thereby generating a virtual image that closely resembles the effect of a standard phase-contrast image. The virtual phase-contrast image generator is based on a UNet architecture. The input and output of the virtual phase-contrast image generator are grayscale images of the same size.
[0013] A virtual bright-field image generator is used to generate a virtual bright-field image from a virtual phase-contrast image. It recovers the corresponding virtual bright-field image from the generated virtual phase-contrast image, helping the network to verify whether the generated phase-contrast image is reasonable. The virtual bright-field image generator has the same structure as the virtual phase-contrast image generator. The input and output of the virtual bright-field image generator are grayscale images of the same size.
[0014] The discriminator compares the generated virtual phase-contrast image with the standard phase-contrast image and determines the difference between them. The discriminator input is a 2-channel image, with the two images being a bright-field image containing a cylindrical lens and a virtual phase-contrast image, and a bright-field image containing a cylindrical lens and a standard phase-contrast image. The first pair of data is considered false, and the loss is fed back to the virtual phase-contrast image generator. The second pair of data is considered true, and the loss is fed back to the discriminator.
[0015] In this invention, the virtual phase-contrast image generator contains eight encoders, seven decoders, and one output layer. The encoders include a two-dimensional convolutional layer, a batch normalization layer, and a LeakyReLU activation layer. The decoders include a two-dimensional convolutional layer, an upsampling layer, a batch normalization layer, and a LeakyReLU activation layer. The output layer includes a two-dimensional convolutional layer, an upsampling layer, a batch normalization layer, a LeakyReLU activation layer, and a Tanh activation layer. The encoders and decoders are connected in a jumper configuration to ensure the reliability of the original data.
[0016] In this invention, the discriminator consists of multiple two-dimensional convolutional layers, batch normalization layers, and LeakyReLU activation layers stacked together in sequence, and finally outputs directly without using an activation function.
[0017] In this invention, credibility negative feedback It compares a virtual brightfield image generated by a virtual brightfield image generator with a real brightfield image, that is:
[0018]
[0019] Where: SSIM is the structural similarity evaluation function, B is the bright-field image containing the cylindrical lens, and P... g B is a virtual brightfield image generated from a virtual phase-contrast image using a virtual brightfield image generator;
[0020] The credibility of the content in the virtual contrast image is obtained through the negative feedback expression, and the loss is fed back to the virtual contrast image generator to ensure the authenticity of the virtual contrast image content.
[0021] In this invention, the loss function of the virtual contrast image generator is obtained by weighting multiple factors, including direct loss, gradient and total variational loss, frequency domain loss, bright field loss, and adversarial loss, wherein:
[0022] The direct loss measure is the difference between the virtual phase-contrast image and the standard phase-contrast image, expressed as:
[0023]
[0024] Where SSIM is the structural similarity evaluation function, L1 is the L1 loss, and P is the standard phase-contrast image. g For virtual phase-contrast images;
[0025] The gradient and total variational loss measure the smoothness and sharpness of the virtual contrast image, expressed as follows:
[0026]
[0027] Gradient loss It is the sum of Sobel gradients for the entire image, and the total variational loss. It is the sum of the differences between each pixel;
[0028] The frequency domain loss measure is the difference between the virtual phase-contrast image and the standard phase-contrast image in the spectral domain, expressed as:
[0029]
[0030] Frequency domain loss It represents the sum of the differences between each pixel in the frequency spectrum domain;
[0031] Brightfield loss measures the credibility of virtual contrast content, i.e., the degree of artifacts, and is expressed as:
[0032]
[0033] Where SSIM is the structural similarity evaluation function, B is the bright-field image containing cylindrical lenses, and P... g B is a virtual brightfield image generated from a virtual phase-contrast image using a virtual brightfield image generator;
[0034] The adversarial loss enhances the generation capability of the virtual contrast image generator by leveraging the discriminator's ability to detect false positives. The expression is:
[0035]
[0036] Combating losses This represents the difference between the true image pair and 1.
[0037] The present invention also provides a low-cost, universal virtual phase contrast device based on a cylindrical lens, which is used to implement the above-mentioned virtual phase contrast method. The device includes a cylindrical lens disposed between the illumination system of the microscope system and the stage on which the sample is placed.
[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0039] The low-cost, universal virtual phase contrast method and device based on cylindrical lenses of the present invention can not only achieve effects similar to those of standard phase contrast microscopes at low cost, but also outperform standard phase contrast microscopes in certain applications (cell segmentation, cell identification, etc.), expanding its application potential in fields such as biomedicine and materials science. Attached Figure Description
[0040] Figure 1 This is a diagram of the improved illumination optical path structure based on a cylindrical lens. The microscope illumination system, microscope objectives, and camera all use a standard bright-field structure without any adjustments. The illumination system employs Köhler illumination, emitting parallel illumination light. The cylindrical lens is placed after the illumination system and in front of the sample, causing the emitted parallel illumination light to converge in one direction.
[0041] Figure 2 The structure diagram of the post-processing algorithm for a conditional generative adversarial neural network with negative feedback for credibility.
[0042] Figure 3 This is the bright-field image with cylindrical lens input in Example 1.
[0043] Figure 4 This is the virtual contrast image output in Example 1.
[0044] Figure 5 This is a comparison of the cell segmentation and recognition results between the bright-field image and the output virtual phase-contrast image in Example 1. Detailed Implementation
[0045] The present invention will now be described in further detail with reference to the accompanying drawings.
[0046] This invention proposes an improved illumination optical path based on cylindrical lenses and a post-processing algorithm for conditional generative adversarial neural networks with credibility negative feedback.
[0047] In this invention, an improved illumination path based on a cylindrical lens is incorporated into a conventional bright-field microscope with a Köhler illumination structure. In this configuration, the focal length of the cylindrical lens and its distance from the sample remain fixed. This design ensures that the light illuminating the sample produces a focusing effect in a specific direction, thereby achieving asymmetric illumination of the sample. The cylindrical lens modifies the light emitted from the light source, causing the light vector illuminating the sample surface to differ along the x and y axes. Traditional bright-field microscope illumination is generally symmetrical, with the light vector distribution being identical in all directions. However, through the action of the cylindrical lens, the light beam acquires a horizontal light vector. In traditional bright-field microscopy, the intensity image typically reflects primarily the transmitted or reflected intensity of the sample, while phase information is relatively hidden and difficult to extract directly from the intensity image. However, with asymmetric illumination, the directional change of the light vector causes the light to undergo varying degrees of phase change as it passes through the sample. In particular, different parts of the sample are illuminated from different directions, resulting in the phase information being encoded into the intensity variation in the image. Therefore, cylindrical lenses can provide a different illumination pattern than traditional illumination during microscope illumination, thereby enabling the phase information of the sample to be effectively encoded into the brightfield image.
[0048] In this invention, the post-processing algorithm for a conditional generative adversarial neural network with credibility negative feedback consists of a dataset comprised of two parts: bright-field images containing cylindrical lenses and corresponding standard phase-contrast images (in which case the illumination path does not contain cylindrical lenses, and a ring aperture and phase-contrast objective are added to the microscope system). To ensure the effectiveness of training and the accuracy of the model, all images in the dataset possess the following characteristics: First, the magnification of all images is fixed, meaning all bright-field images and standard phase-contrast images are captured at the same magnification, ensuring a one-to-one correspondence between details and structures in the images; second, the orientation of the cylindrical lenses remains fixed during training, meaning the installation angle and position of the cylindrical lenses do not change, ensuring that each generated bright-field image has the same illumination characteristics and light intensity distribution, avoiding the impact of changes in illumination conditions on the training effect. During training, these two types of images are paired one-to-one as input and target images, respectively. Specifically, the bright-field images containing cylindrical lenses are used as input images, provided to the generative model for training. The standard phase-contrast images, corresponding one-to-one with the input bright-field images, are provided to the neural network as training targets. The goal of the network is to generate a virtual phase-contrast image that is as close as possible to a standard phase-contrast image by processing the input bright-field image containing cylindrical lenses, thereby simulating the effect of a phase-contrast microscope.
[0049] The following are specific examples.
[0050] Example 1
[0051] This invention provides a low-cost, universal virtual phase contrast method based on cylindrical lenses. It aims to improve upon conventional bright-field microscopy using a single cylindrical lens and employs a conditional generative adversarial neural network (GAN) for data processing to achieve a virtual phase contrast effect essentially equivalent to that of a standard phase contrast microscope. In terms of hardware, a standard bright-field microscope with Köhler illumination is used. A cylindrical lens (with a fixed distance between the lens and the sample) is added to both the illumination system and the sample, creating asymmetric illumination and introducing phase information into the intensity image. Algorithmically, a dataset is established by acquiring bright-field images (with cylindrical lenses positioned in a specific orientation) of various unstained transparent biological samples and standard phase contrast images. A conditional generative adversarial neural network with confidence negative feedback is then trained in a supervised manner using the established dataset. After training, the network can generate corresponding virtual phase contrast images by inputting a bright-field image with a cylindrical lens.
[0052] Combination Figure 1This paper proposes an improved illumination path based on cylindrical lenses, incorporating them into a conventional bright-field microscope with a Köhler illumination structure. The focal length of the cylindrical lens and its distance from the sample remain fixed, ensuring that the light illuminating the sample is focused in a specific direction, achieving asymmetric illumination. By modifying the light emitted from the light source, the cylindrical lens creates differences in the light vector illuminating the sample surface along the x and y axes, breaking the rotationally symmetric illumination pattern of traditional bright-field microscopes, particularly in the horizontal direction. Traditional bright-field microscope intensity images primarily reflect the transmitted or reflected intensity of the sample, making phase information difficult to extract. However, through asymmetric illumination, the directional change of the light vector causes varying degrees of phase change as the light passes through the sample, especially since different parts of the sample are illuminated from different directions, resulting in effective encoding of phase information into the intensity image. This illumination mode provides an effective way to encode the sample's phase information, overcoming the limitation of difficult phase information extraction from traditional intensity images.
[0053] Combination Figure 2 This paper describes a post-processing algorithm for a conditional generative adversarial neural network with credibility negative feedback. The dataset consists of two parts: bright-field images containing cylindrical lenses and corresponding standard phase-contrast images. All images have a fixed magnification and the cylindrical lens orientation is kept consistent to ensure training consistency. During training, the bright-field images containing cylindrical lenses are used as input, and the standard phase-contrast images are used as target images for one-to-one pairing training. The aim is to generate virtual phase-contrast images that closely resemble the standard images, simulating the effect of a phase-contrast microscope.
[0054] Combination Figure 2 The network architecture comprises three main components: a virtual phase-contrast image generator, a virtual bright-field image generator, and a discriminator. The virtual phase-contrast image generator produces a virtual phase-contrast image from a bright-field image containing a cylindrical lens. The virtual bright-field image generator then recovers the bright-field image from the virtual phase-contrast image. The discriminator compares the generated virtual phase-contrast image with a standard phase-contrast image, calculates and outputs an adversarial loss, and drives the network to optimize the quality of the generated images. Through this adversarial training, the generator gradually improves its ability to generate virtual phase-contrast images.
[0055] Specifically, the virtual phase-contrast image generator's task is to generate virtual phase-contrast images from bright-field images containing cylindrical lenses. It learns to extract and enhance phase information from the bright-field image, thus generating a virtual image that approximates the effect of a standard phase-contrast image. The virtual phase-contrast image generator adopts the UNet architecture, containing eight encoder layers, seven decoder layers, and one output layer. The encoder consists of 2D convolutional layers, batch normalization layers, and LeakyReLU activation layers. The 2D convolutional layers use 4×4 kernels with a stride of 2. The decoder consists of 2D convolutional layers, upsampling layers, batch normalization layers, and LeakyReLU activation layers. The 2D convolutional layers also use 4×4 kernels with a stride of 2, and the upsampling layers have a magnification of 4x and use linear interpolation. The output layer contains 2D convolutional layers, upsampling layers, batch normalization layers, LeakyReLU activation layers, and Tanh activation layers. The 2D convolutional layers use 4×4 kernels with a stride of 2, and the upsampling layers also have a magnification of 4x and use linear interpolation. The encoder and decoder are connected via jumpers to ensure the integrity of the original data. The number of channels is progressively increased from 1 to 512 in the encoder, and then reduced back to 1 in the decoder. The generator takes a 512×512 grayscale image as input and outputs a grayscale image of the same size. The virtual bright-field image generator performs the reverse operation, recovering the corresponding virtual bright-field image from the generated virtual contrast image, helping the network verify the validity of the generated contrast image. The structure of the virtual bright-field image is the same as that of the virtual contrast generator, which accelerates the training process. The discriminator compares the generated virtual contrast image with the standard contrast image and determines the difference between them. The discriminator consists of multiple 2D convolutional layers, batch normalization layers, and LeakyReLU activation layers stacked sequentially, ultimately outputting the result directly without using an activation function. The 2D convolutional layers use 4×4 kernels with a stride of 2, progressively increasing the number of channels from 2 to 512, and finally compressing it back to 1 at the output. The discriminator's input consists of a 512×512 pixel image with two channels. One pair of images is a bright-field image with a cylindrical lens and a virtual phase-contrast image; the other pair is a bright-field image with a cylindrical lens and a standard phase-contrast image. The first pair (virtual phase-contrast image) is considered pseudo-data, and its loss is fed back to the virtual phase-contrast image generator. The second pair (standard phase-contrast image) is considered real data, and its loss is fed back to the discriminator. Based on this, the discriminator calculates and outputs the corresponding adversarial loss, guiding the network to optimize the virtual phase-contrast image generator and the virtual bright-field image generator, thereby improving the quality and realism of the generated images. Throughout this process, the parameters of each component are continuously adjusted through this adversarial training mechanism, enabling the generator to generate increasingly accurate and high-quality virtual phase-contrast images.
[0056] The credibility negative feedback mechanism evaluates the credibility of the generated content by comparing the virtual bright-field image generated by the virtual bright-field image generator with the real bright-field image. A large difference indicates that the virtual contrast image may contain artifacts, and the network adjusts its generation strategy based on this information to reduce artifacts and optimize image quality. The loss function is optimized through multiple weighted loss terms, including direct loss, total variational and gradient loss, frequency domain loss, bright-field loss, and adversarial loss, comprehensively measuring image quality, smoothness, sharpness, spectral features, and physical plausibility, thereby generating higher-quality virtual contrast images.
[0057] Furthermore, regarding the credibility negative feedback mechanism, specifically, the virtual bright-field image generator first generates a corresponding bright-field image based on the generated virtual contrast image, and then compares this virtual bright-field image with the real bright-field image. By calculating the difference between the two, the credibility of the content contained in the virtual contrast image can be obtained. If the difference is large, it means that there may be artifacts or unreal parts in the generated virtual contrast image; if the difference is small, it means that the generated image is more realistic and credible. Based on this credibility information, the loss function will provide negative feedback to the virtual contrast image generator, prompting it to adjust its generation strategy, reduce the occurrence of artifacts, and optimize the quality of the contrast image. Through this mechanism, the network can continuously improve the generation process, ensuring that the content of the generated virtual contrast image is realistic and accurate, thereby improving the overall quality of the virtual contrast image and its similarity to the real contrast image.
[0058] Furthermore, the loss function of the virtual contrast image generator is optimized through multiple weighted loss terms to comprehensively measure the quality of the generated image. The direct loss measures the pixel-level difference between the virtual contrast image and the standard contrast image, ensuring that the generated image is as consistent as possible with the real image. The expression is as follows: Where SSIM is the structural similarity evaluation function, L1 is the L1 loss, and P is the standard phase-contrast image. g This is a virtual contrast image. The total variation and gradient loss measure the smoothness and sharpness of the image, respectively, to avoid generating images that are too blurry or too sharp, thus optimizing visual quality. Their expressions are: as well as The gradient loss is the sum of the Sobel gradients of the entire image, and the total variational loss is the sum of the differences pixel by pixel. The frequency domain loss compares the differences between the virtual and standard contrast images in the spectral domain to ensure that the generated image matches the features of the real image at the spectral level; its expression is: This represents the sum of differences between pixels in the spectral domain. Brightfield loss, on the other hand, focuses on the reliability of the brightfield information in the generated image, reducing artifacts and ensuring the physical plausibility of the image. Its expression is: Where SSIM is the structural similarity evaluation function, B is the bright-field image containing cylindrical lenses, and P... g B represents the virtual brightfield image generated by a virtual brightfield image generator from a virtual phase-contrast image. The adversarial loss, applied by a discriminator, discriminates the generated image, promoting generator optimization and enhancing its ability to generate realistic images. The expression is as follows: This represents the difference between the real image pair and 1. By weighting these loss terms, the virtual contrast image generator can comprehensively optimize multiple aspects of the image, such as detail, smoothness, and spectral features, thereby generating higher quality and more realistic virtual contrast images.
[0059] In specific implementation example 1, the input sample is as follows: Figure 3 The image shown is a bright-field image containing a cylindrical lens, a grayscale image of size 512×512. By inputting the input image into a trained virtual phase-contrast generation network, images like this can be generated. Figure 4 The virtual phase-contrast image shown.
[0060] The conditional generative adversarial neural network with credibility negative feedback was trained using the Adam optimizer, accumulating gradients for 5 epochs before updating. The learning rate was set to 0.0001 and decreased by 10% every 10 epochs, for a total of 300 epochs. The generated virtual contrast images have better image quality, and cells are more easily distinguishable in detail, such as... Figure 5 As shown, it has better application effects compared to bright field images.
Claims
1. A low-cost, universal virtual phase contrast method based on cylindrical lenses, characterized in that, Includes the following steps: By adding a cylindrical lens to a conventional bright-field microscope with a Köhler illumination structure, asymmetric illumination is achieved from the illumination source, resulting in a bright-field image; the obtained light intensity image carries more phase information. The obtained bright-field image is used to generate a virtual contrast image through a conditional generative adversarial neural network under trained credibility negative feedback, thereby achieving an effect equivalent to a standard contrast image at low cost; where: A cylindrical lens is added between the illumination system of the microscope system and the stage on which the sample is placed. The focal length of the cylindrical lens remains constant, and the distance between the cylindrical lens and the sample remains constant after the network training is completed. Conditional generative adversarial neural networks based on credibility negative feedback, where "condition" refers to a supervised network training method guided by truth values; the overall network includes a virtual contrast image generator, a virtual bright field image generator, and a discriminator. A virtual phase-contrast image generator is used to generate virtual phase-contrast images from bright-field images containing cylindrical lenses. It learns to extract and enhance phase information from the bright-field image, thereby generating a virtual image that closely resembles the effect of a standard phase-contrast image. The virtual phase-contrast image generator is based on a UNet architecture. The input and output of the virtual phase-contrast image generator are grayscale images of the same size. A virtual bright-field image generator is used to generate a virtual bright-field image from a virtual phase-contrast image. It recovers the corresponding virtual bright-field image from the generated virtual phase-contrast image, helping the network to verify whether the generated phase-contrast image is reasonable. The virtual bright-field image generator has the same structure as the virtual phase-contrast image generator. The input and output of the virtual bright-field image generator are grayscale images of the same size. The discriminator compares the generated virtual phase-contrast image with the standard phase-contrast image and determines the difference between them. The discriminator input is a 2-channel image, with the two images being a bright-field image containing a cylindrical lens and a virtual phase-contrast image, and a bright-field image containing a cylindrical lens and a standard phase-contrast image. The first pair of data is considered false, and the loss is fed back to the virtual phase-contrast image generator. The second pair of data is considered true, and the loss is fed back to the discriminator.
2. The low-cost, universal virtual phase contrast method based on cylindrical lenses according to claim 1, characterized in that, The virtual phase-contrast image generator contains eight encoders, seven decoders, and one output layer. The encoders include a 2D convolutional layer, a batch normalization layer, and a LeakyReLU activation layer. The decoders include a 2D convolutional layer, an upsampling layer, a batch normalization layer, and a LeakyReLU activation layer. The output layer includes a 2D convolutional layer, an upsampling layer, a batch normalization layer, a LeakyReLU activation layer, and a Tanh activation layer. The encoders and decoders are connected in a skip connection manner to ensure the reliability of the original data.
3. The low-cost, universal virtual phase contrast method based on cylindrical lenses according to claim 1, characterized in that, The discriminator consists of multiple two-dimensional convolutional layers, batch normalization layers, and LeakyReLU activation layers stacked together in sequence, and finally outputs directly without using an activation function.
4. The low-cost, universal virtual phase contrast method based on cylindrical lenses according to claim 1, characterized in that, Credibility negative feedback compares a virtual bright-field image generated by a virtual bright-field image generator with a real bright-field image, that is: ， in: For structural similarity evaluation functions, This is a bright-field image containing cylindrical lenses. A virtual brightfield image generated from a virtual phase-contrast image using a virtual brightfield image generator; The credibility of the content in the virtual contrast image is obtained through the negative feedback expression, and the loss is fed back to the virtual contrast image generator to ensure the authenticity of the virtual contrast image content.
5. The low-cost, universal virtual phase contrast method based on cylindrical lenses according to claim 1, characterized in that, The loss function of the virtual contrast image generator is obtained by weighting multiple factors, including direct loss, gradient and total variational loss, frequency domain loss, bright field loss, and adversarial loss, where: The direct loss measure is the difference between the virtual phase-contrast image and the standard phase-contrast image, expressed as: ， in For structural similarity evaluation functions, for loss, For standard phase-contrast images, For virtual phase-contrast images; The gradient and total variational loss measure the smoothness and sharpness of a virtual contrast image, expressed as follows: ， , Gradient loss It is the sum of Sobel gradients for the entire image, and the total variational loss. It is the sum of the differences between each pixel; The frequency domain loss measure is the difference between the virtual phase-contrast image and the standard phase-contrast image in the spectral domain, expressed as: ， Frequency domain loss It represents the sum of the differences between each pixel in the frequency spectrum domain; Brightfield loss measures the credibility of virtual contrast content, i.e., the degree of artifacts, and is expressed as: , in For structural similarity evaluation functions, This is a bright-field image containing cylindrical lenses. A virtual brightfield image generated from a virtual phase-contrast image using a virtual brightfield image generator; The adversarial loss enhances the generation capability of the virtual contrast image generator by leveraging the discriminator's ability to detect false positives. The expression is: ， Combating losses This represents the difference between the true image pair and 1.
6. A low-cost, universal virtual phase contrast device based on a cylindrical lens, characterized in that, It is used to implement the low-cost, general-purpose virtual phase contrast method of any one of claims 1-5, which includes a cylindrical lens disposed between the illumination system of the microscope system and the stage on which the sample is placed.