Mark-free single-cell two-dimensional light scattering imaging modal amplification method and system

Abstract

The invention belongs to the technical field of single cell imaging and analysis, and provides an unmarked single cell two-dimensional light scattering imaging modal amplification method and system. The method comprises the following steps: acquiring a two-dimensional light scattering image of an unmarked single cell, and preprocessing the two-dimensional light scattering image; according to the proposed fusion condition adversarial loss LcGAN, L1 distance loss LL1 and a new imaging modal amplification model of perception loss # imgabs0 # based on a VGG network, generating a corresponding single-cell bright field image and a fluorescence image by using the processed light scattering image, and realizing amplification of a single-cell multi-modal image; multi-modal fusion and classification identification are carried out on the single cell two-dimensional light scattering image, the corresponding bright field image and the fluorescence image, so that the cancer single cell analysis accuracy is improved; the method can generate the corresponding single-cell bright field image and fluorescence image according to the unmarked single-cell two-dimensional light scattering image, is especially suitable for diagnosis and classification of cancer cells, and facilitates promotion of biomedical research and clinical application.

Description

Technical Field

[0001] The present invention relates to the technical field of single-cell imaging and analysis, and particularly to a method and system for amplifying a label-free single-cell two-dimensional light scattering imaging modality. Background Art

[0002] Cells are the basic units that make up organisms, and single-cell imaging plays a crucial role in drug development, disease diagnosis, and personalized medicine. Most traditional single-cell imaging techniques rely on labeling methods such as fluorescence labeling and radioactive labeling. Although these methods can provide high-sensitivity and high-specificity detection capabilities, they usually require complex labeling steps for cells, increasing the complexity and cost of experiments. In addition, fluorescence labeling may have problems such as photobleaching, low labeling efficiency, and potential toxicity to cells.

[0003] Label-free single-cell imaging technology can obtain cell information without labeling, avoiding the above problems and having significant advantages in non-invasive analysis. Light scattering imaging technology obtains the morphological and internal structure information of cells by detecting the interaction between cells and light, and is a label-free and non-invasive cell imaging method. Light scattering imaging has the advantages of high sensitivity and high resolution, can monitor the cell state in real time without damaging cells, and is suitable for single-cell analysis. Compared with traditional bright-field imaging, light scattering imaging can capture more information about the internal structure of cells, but the complexity of its data is relatively high and image analysis is difficult.

[0004] To overcome the limitations of single-modal imaging, multi-modal imaging acquisition and fusion have become an important research direction in single-cell analysis. Multi-modal images can provide more comprehensive cell information, including cell morphology, internal structure, functional state, and metabolic activity, which helps to reveal the characteristics and dynamic changes of cells from multiple dimensions. However, the practical application of single-cell multi-modal imaging technology faces significant challenges. The acquisition of traditional multi-modal images depends on different imaging modules such as fluorescence detection and light scattering detection, which results in complex device structures, high costs, high requirements for the experimental environment, and cumbersome operation procedures, restricting the popularization and application of the technology. At the same time, the independently acquired single-cell multi-modal images need to be aligned in space and time, increasing the complexity of data fusion and analysis. Therefore, it is particularly important to innovate and develop methods for providing single-cell multi-modal imaging information. Summary of the Invention

[0005] To address the deficiencies of the prior art, the present invention provides a method and system for amplifying the two-dimensional light scattering imaging modality of label-free single cells. By using advanced artificial intelligence technology, it can generate high-quality bright-field images and fluorescence images of the same label-free single cell based on two-dimensional light scattering images, breaking through the limitations of traditional light scattering imaging, obtaining more accurate and multi-dimensional cell information, simplifying the design and operation process of experimental equipment, and avoiding the complexity and high cost of multi-modal imaging devices; the obtained multi-modal images enable the diagnosis and classification of cancer cells, and are expected to promote biomedical research and clinical applications.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] In the first aspect, the present invention provides a method for amplifying the two-dimensional light scattering imaging modality of label-free single cells.

[0008] A method for amplifying the two-dimensional light scattering imaging modality of label-free single cells includes the following processes:

[0009] Obtain a two-dimensional light scattering image of a single cell without labels, and preprocess the two-dimensional light scattering image;

[0010] Based on the preprocessed two-dimensional light scattering image and a pre-trained deep learning model, obtain a bright-field image and a fluorescence image corresponding to the two-dimensional light scattering image;

[0011] Post-process the two-dimensional light scattering image, the corresponding bright-field image, and the fluorescence image to obtain a multi-modal amplification result for cell classification according to a pre-trained classification model.

[0012] As a further limitation of the first aspect of the present invention, preprocessing the two-dimensional light scattering image includes: denoising, cropping, and normalization.

[0013] As a further limitation of the first aspect of the present invention, the pre-trained deep learning model adopts a generative adversarial network model, and the generative adversarial network model includes: a generator and a discriminator; the generator adopts a U-net architecture, including an encoder, a decoder, and skip connections;

[0014] The encoder is used to extract the features of the preprocessed two-dimensional light scattering image layer by layer, the decoder is used to restore the image information layer by layer, the encoder and the decoder are connected by convolutional layers, and skip connections are established between the corresponding layers of the encoder and the decoder to directly transfer the low-level features to the corresponding layers of the decoder, retaining the detailed information in the input image;

[0015] The discriminator is a convolutional neural network, which includes a convolutional layer, a batch normalization layer, and a LeakyReLU activation layer. The discriminator is used to classify the images output by the generator and determine whether the images output by the generator are real or generated.

[0016] As a further limitation of the first aspect of the present invention, the loss function L overall (G, D) of the pre-trained deep learning model includes a conditional adversarial loss L cGAN , an L1 distance loss L L1 and a perceptual loss based on the VGG network By setting different weight coefficients α, β, and γ to adjust the contribution degree of each loss, the generation network can be adaptively adjusted according to different task requirements and data characteristics.

[0017] As a further limitation of the first aspect of the present invention, the conditional adversarial loss is: L cGAN (G, D) = E x,y [logD(x, y)] + E x [log(1 - D(x, G(x)))], where x is the input image, y is the corresponding real output image, G(x) is the generated image, D(x, y) is the pair of real images in the discriminator, and D(x, G(x)) is the pair of generated images;

[0018] The L1 distance loss is: L L1 (G) = E x,y [|y - G(x)|];

[0019] The VGG perceptual loss is: where VGGl(·) represents the feature map extracted from the l-th layer of the VGG-19 network, Z is the set of layers selected for calculating the perceptual loss, W l and H l represent the width and height of the feature map of the l-th layer, respectively.

[0020] In the second aspect, the present invention provides a system for amplifying the two-dimensional light scattering imaging modality of unlabeled single cells.

[0021] A system for amplifying the two-dimensional light scattering imaging modality of unlabeled single cells includes:

[0022] A light source excitation unit: configured with a laser light source and a beam shaping component for providing an excitation beam for single cells;

[0023] A data imaging unit: including a detection chip and an image acquisition device for capturing light scattering signals and converting them into digital images;

[0024] Pretreatment unit: used to crop and filter the collected images, extract the region of interest of single-cell light scattering, and improve the image quality;

[0025] Modal amplification unit: includes an amplification model, used to amplify the two-dimensional light scattering image into a multi-modal image;

[0026] Result evaluation unit: based on a deep learning model, classify or diagnose and analyze the multi-modal image.

[0027] As a further limitation of the second aspect of the present invention, the modal amplification unit adopts an amplification model based on a convolutional neural network (CNN) to generate a multi-modal image with more information.

[0028] As a further limitation of the second aspect of the present invention, in the post-processing unit, the pre-trained deep learning model includes a loss function L overall :

[0029]

[0030] wherein, L cGAN (G,D) is the conditional adversarial loss, L L1 (G) is the L1 distance loss, is the VGG perceptual loss, and α, β, and γ are weight coefficients.

[0031] As a further limitation of the second aspect of the present invention, the result evaluation unit combines deep learning and traditional classification methods for cell type recognition, health status assessment, and disease diagnosis.

[0032] As a further limitation of the second aspect of the present invention, a light source excitation unit is used to excite the sample; a light scattering image is captured by a data imaging unit; the pretreatment unit is used to crop and filter the image; the modal amplification unit is applied to generate a multi-modal image; and the result evaluation unit is used for analysis to output a diagnostic result.

[0033] Compared with the prior art, the beneficial effects of the present invention are:

[0034] 1. The present invention innovatively proposes a label-free single-cell two-dimensional light scattering imaging modal amplification strategy, which can generate high-quality bright-field images and fluorescence images according to the two-dimensional light scattering image. Utilizing advanced artificial intelligence technology, it breaks through the limitations of light scattering imaging, obtains more accurate and multi-dimensional cell information, is particularly suitable for the diagnosis and classification of cancer cells, and helps to promote biomedical research and clinical applications.

[0035] 2. By combining adversarial loss, L1 distance loss, and VGG perceptual loss, the present invention proposes a brand-new network optimization method, which not only improves the quality and realism of generated images, but also achieves important breakthroughs in multi-level feature extraction, image structure preservation, and detail presentation. This method can be widely applied to various image processing tasks such as image generation, image restoration, and style transfer, and has broad application prospects.

[0036] 3. The present invention adopts label-free detection: without the need for fluorescent dyes or other chemical markers, it avoids potential damage and interference to cells during the labeling process, maintains the natural state of cells, and is suitable for real-time dynamic monitoring of living cells. Since there is no need to perform physical or chemical treatment on cells, the detection process is non-invasive to cells and is suitable for long-term cell monitoring and dynamic research.

[0037] 4. The present invention has the ability to obtain rich cell information: by using two-dimensional light scattering technology, it can efficiently capture cell information. Through modal amplification technology, it can extract various physical and biological characteristic information from single cells, including but not limited to cell size, morphology, internal structure, etc., providing comprehensive cell characterization.

[0038] 5. The present invention enables rapid automated analysis: by combining advanced image processing and machine learning algorithms, it realizes automatic analysis and classification of single-cell images, significantly improving the analysis efficiency and reducing human intervention and errors.

[0039] 6. The present invention has broad application potential: it is applicable to various types of cell analysis, including but not limited to cancer cell detection, stem cell research, immune cell analysis, etc., and has broad application prospects and promotion value.

[0040] 7. The present invention has the advantages of low cost and high benefit: compared with traditional labeling detection methods, it eliminates expensive fluorescent dyes and complex sample preparation processes, reduces experimental costs, and at the same time improves detection efficiency and experimental throughput.

[0041] 8. The present invention has strong compatibility: it can be compatible with existing microscopes and imaging devices, is easy to integrate and implement, and does not require large-scale transformation of existing laboratory equipment. The implementation of the present invention will contribute to the development of single-cell research, especially in the fields of cancer research, drug screening, cell differentiation, etc., providing new technical means and research tools. The advantages of additional aspects of the present invention will be partially given in the following description, partially will become obvious from the following description, or will be understood through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The specification drawings forming a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention.

[0043] Figure 1 Schematic diagram of the method for amplifying the label-free single-cell two-dimensional light scattering imaging modality provided in Embodiment 1 of the present invention;

[0044] Figure 2 Verification result diagram of the simulation model modality amplification provided in Embodiment 1 of the present invention;

[0045] Figure 3 Modality amplification result diagram of small balls of different sizes provided in Embodiment 1 of the present invention;

[0046] Figure 4 Schematic diagram of the cell modality amplification result provided in Embodiment 1 of the present invention, showing the generated bright field and fluorescence (FL) images;

[0047] Figure 5 Radar chart for comparing the classification performance indexes of different modalities of cervical cancer cells provided in Embodiment 1 of the present invention;

[0048] Figure 6 Schematic diagram of a single-cell label-free two-dimensional light scattering modality amplification system provided in Embodiment 2 of the present invention;

[0049] Wherein: 1 - laser light source; 2 - filter; 3 - cylindrical lens; 4 - sample chip; 5 - displacement stage; 6 - imaging objective lens; 7 - tube lens; 8 - CMOS detector; 9 - modality amplification and analysis module. Detailed implementation manners

[0050] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0051] It should be noted that the following detailed descriptions are all exemplary and are intended to provide further descriptions of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0052] In the case of no conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0053] Embodiment 1:

[0054] This implementation method proposes a method for amplifying the two-dimensional light scattering imaging modality of label-free single cells, aiming to expand the single-mode light scattering image into a multi-modal image through deep learning technology, so as to realize the acquisition and analysis of multi-modal information of label-free single cells. This method can generate high-quality bright-field and fluorescence images, which are particularly suitable for the diagnosis and classification of cancer cells, and contribute to promoting biomedical research and clinical applications; the present invention verifies the feasibility of expanding the label-free light scattering image into a multi-modal image and its potential in cell classification. The bright-field and fluorescence images generated by the deep learning model have basic feature consistency with the images obtained by traditional microscopes and can be effectively used for the feature analysis and classification of cells; the experimental results show that the classification accuracy of cervical cancer cells is significantly improved by using the expanded multi-modal images.

[0055] First, the technical terms and related concepts involved in this processing solution are briefly introduced as follows:

[0056] Light scattering image: A two-dimensional light scattering image of a label-free single cell is acquired by laser excitation and forward defocusing of the objective lens.

[0057] Bright-field image: Bright-field is an imaging mode of modern microscopes. If only the transmitted beam passes through the objective aperture for imaging, the resulting image is a bright-field image.

[0058] Fluorescence image: After the cell sample is stained and fixed with DAPI, under the irradiation of a halogen lamp, a cell fluorescence image in which the cell nucleus is specifically represented as blue is obtained through an excitation filter (excitation wavelength 325 - 369 nm) and an emission filter (emission wavelength 430 - 490 nm).

[0059] Gaussian filtering denoising: Based on the Gaussian function, the image is smoothed to reduce noise while trying to preserve the edges and other important features of the image.

[0060] Specifically, this implementation method first acquires a two-dimensional light scattering image. The specific acquisition device for the two-dimensional light scattering image includes:

[0061] Light source: Usually a laser is used to provide a stable and intensity-controllable light source, which is suitable for generating high-quality light scattering images.

[0062] Microscope system: Equipped with an objective lens with a high numerical aperture to ensure high-resolution imaging effects.

[0063] Imaging detector: Usually a high-sensitivity CCD or CMOS camera is used to capture the light scattering image to ensure that the scattering information of the tiny structures inside the cell can be captured.

[0064] Control and acquisition system: Used to control the switch of the light source and adjust the light intensity of different specifications of filters, adjust the focal length of the microscope, and perform image acquisition and storage.

[0065] The label-free single-cell two-dimensional light scattering imaging modality amplification method described in this implementation mode uses an optical system as follows Figure 1 shown, including: a laser light source 1, a filter 2, a cylindrical lens 3, a sample chip 4, a displacement stage 5, an imaging objective lens 6, a tube lens 7, a CMOS detector 8, and a modality amplification and analysis module 9. The modality amplification method includes the following processes:

[0066] S1: Image preprocessing. Specifically, preprocessing the acquired light scattering image is an important step to improve the accuracy of subsequent analysis, including:

[0067] Denoising: Commonly used methods such as Gaussian filtering are used to smooth the image to remove noise during the imaging process;

[0068] Cropping: Crop the irrelevant background part in the original image and only retain the area containing cells;

[0069] Normalization: Perform gray-scale normalization on the image so that the gray-scale values of the image are within a standardized range, improving the comparability between different images.

[0070] S2: Modality amplification. Specifically, build and train a deep learning model for processing light scattering images and generating artificial multi-modal images, including the following processes:

[0071] S2.1: Construction of the deep learning model. Specifically, it includes:

[0072] Model architecture: Select a deep learning model constructed with multiple convolutional layers, pooling layers, and fully connected layers, etc.;

[0073] Data preparation: Collect a large number of light scattering images and their corresponding bright-field and fluorescence images as the training set and validation set, and collect cell data of different sizes to ensure data diversity and enhance the generalization ability of the model;

[0074] Model training: Use the prepared data set to train the model. During the training process, appropriate loss functions and optimization algorithms are used to minimize the error between the prediction result and the real image.

[0075] In this implementation mode, optionally, the core framework of the deep learning model for light scattering image modality amplification is a generative adversarial network (GAN), which includes two main components: a generator and a discriminator.

[0076] The generator adopts the U-net architecture, which includes an encoder, a decoder, and skip connections. The skip connections allow low-level information, which is often shared between the input and output images, to bypass the bottleneck of the encoder-decoder network, thereby retaining more detailed information. The advantage of this design is that it can generate high-quality artificial images more effectively, especially when dealing with image conversion tasks, such as generating corresponding artificial bright-field or fluorescence images from two-dimensional light scattering images.

[0077] Encoder: Composed of multiple convolutional layers, gradually reducing the spatial dimension and extracting the feature information of the image. Decoder: Gradually restores the spatial dimension of the image through transposed convolution or upsampling layers to generate a high-resolution output image. Skip Connections: Establish direct connections between the corresponding layers of the encoder and the decoder, directly passing the low-level features to the corresponding layers of the decoder, retaining the detailed information in the input image, thereby generating higher-quality images.

[0078] The discriminator is a convolutional neural network (CNN), composed of convolutional layers, batch normalization (BatchNorm) layers, and LeakyReLU activation layers. Its main function is to classify the generated images and determine whether they are real or generated. Convolutional Layers: Extract the features of the image, enhancing the discriminator's ability to capture image features through layer-by-layer convolution operations; BatchNorm Layers: Used to accelerate the training process and improve the network stability; LeakyReLU Layers: Introduce non-linearity to improve the network's expressive ability and the ability to learn complex features.

[0079] In the present invention, the input image is an unlabeled 2D light scattering image, the output images are the corresponding artificial bright-field image and fluorescence image, and the discriminator is used to classify whether these artificial images are real or generated.

[0080] In this implementation, combining multiple loss functions can provide more diverse training signals, improve the stability and convergence speed of training, and optimize the generator in multiple aspects to generate better image quality. Specifically, it includes:

[0081] The traditional Pix2Pix framework uses a conditional adversarial generative network (cGAN) as the core for jointly optimizing the generator and the discriminator. The conditional adversarial loss L cGAN is:

[0082] L cGAN (G,D) = E x,y[logD(x,y)]+E x [log(1 - D(x,G(x)))](1);

[0083] Where x is the input image, y is the corresponding real output image, G(x) is the generated image, D(x,y) is the real image pair in the discriminator, and D(x, G(x)) is the generated image pair (fake image pair).

[0084] Conditional adversarial loss L cGAN Encourages the generator to generate outputs indistinguishable from real images, thereby enhancing the realism of the generated images; however, the standalone adversarial loss has certain limitations in capturing high-level semantic information of images. Therefore, the present invention further improves the visual quality of the generated images by combining other loss functions. L1 distance loss L L1 Measures the absolute difference between the generated image and the real image at the pixel level, including:

[0085] L L1 (G)=E x,y [|y - G(x)|](2);

[0086] The L1 distance loss ensures that the generated image is closer to the real image in terms of visual effect, especially having an advantage in retaining details. Although the L1 loss optimizes the pixel-level errors, it is still insufficient in terms of the overall structure and semantic understanding of the image. Therefore, using the L1 loss alone may result in unnatural visual perception of the generated image.

[0087] To further improve the perceptual quality of the generated images, the present invention introduces a perceptual loss based on the VGG-19 network. This perceptual loss helps the generator capture more semantic details and structural information related to human vision through high-level feature matching. The perceptual loss calculates the squared difference between the feature maps of the generated image and the real image on the selected layers of the VGG network. This innovation enables the generated image to not only be close to the real image at the pixel level but also be consistent at the semantic level, significantly enhancing the realism and detail of the image.

[0088]

[0089] Where VGG_l represents the feature map extracted from the l-th layer of the VGG-19 network, Z is the set of layers selected for calculating the perceptual loss, W l dimension and H l respectively represent the width and height of the feature map of the l-th layer.

[0090] This implementation method conducts a comprehensive loss design for the global optimization objective function. By weighted combination of the conditional adversarial loss, L1 distance loss, and VGG perceptual loss, a unified optimization objective function L is formed overall:

[0091]

[0092] During the optimization process, this function can balance the role of each loss, so as to generate higher-quality and more realistic images while ensuring the accuracy of image content. By adjusting the hyperparameters α, β, and γ, the present invention can flexibly control the relative importance of each loss term, and then optimize various indicators of the generated image.

[0093] In summary, by combining adversarial loss, L1 distance loss, and VGG perceptual loss, the present invention proposes a brand-new network optimization method, which not only improves the quality and realism of the generated images, but also achieves important breakthroughs in multi-level feature extraction, image structure retention, and detail presentation. This method can be widely applied to various image processing tasks such as image generation, image restoration, and style transfer, and has broad application prospects.

[0094] The present invention adopts an efficient multi-level feature extraction and fusion strategy. By selecting the conv2_2, conv3_3, and conv4_2 layers in the VGG-19 network as the basis for feature extraction, the present invention can integrate low-level texture information and high-level semantic information. These layers provide a balanced capture of detail textures and global structures, making the generated images more delicate and realistic in visual perception. In addition, the loss function design of the present invention takes into account the feature map differences of each layer, enabling the model to be optimized at multiple levels, so as to better handle different types of image generation tasks.

[0095] The deep learning model of this implementation mode is specifically designed for the needs of LS imaging meCytometry. The generated artificial bright-field and fluorescence images can be used for further cell analysis and research. These operation steps enable the deep learning architecture in the present invention to more efficiently process LS imaging data and generate high-quality artificial images, providing a powerful tool for research in related fields.

[0096] S2.2: Modal amplification. Specifically, through the trained deep learning model, the input light scattering image is converted into high-quality bright-field and fluorescence images. The specific process is as follows:

[0097] Input processing: Input the preprocessed light scattering image into the deep learning model;

[0098] Image generation: The model outputs the corresponding multi-modal images, namely bright-field images and fluorescence images, according to the input light scattering image;

[0099] Result optimization: Post-process the generated multi-modal images, such as image enhancement, contrast adjustment, etc., to ensure that the quality of the output images reaches the best state.

[0100] S3: Classification evaluation. Using the post-processed multimodal images, classify the cells through deep learning algorithms to verify the application effect of the label-free modality amplification method in cell classification;

[0101] Performance evaluation: Use common classification performance metrics, such as accuracy, sensitivity, specificity, F1-score, etc., to evaluate the classification effect of the model.

[0102] More specifically, the present implementation provides the following examples:

[0103] Example 1: Simulation model two-dimensional light scattering modality amplification.

[0104] In the field of light scattering modeling, especially for microscopic particles such as cells, numerical simulation has become a crucial tool. In order to verify the effectiveness and reliability of the label-free single-cell two-dimensional light scattering imaging modality amplification method and system proposed in the present invention, the present invention has conducted detailed simulation experiments.

[0105] Mie theory provides a powerful mathematical framework for describing the scattering process of light in spherical particles. The present invention proposes a two-dimensional light scattering simulation algorithm based on Mie theory, which greatly reduces the computational cost and makes it easier to understand the two-dimensional light scattering pattern.

[0106] In this example, the following parameters are set for the simulation of cells:

[0107] Organelle simulation: Organelles such as mitochondria are simulated as spheres with a radius of 0.25 μm, a total of 300, and the nucleus is simulated as a sphere with a radius of 2 μm;

[0108] Cell simulation: The cell components are randomly distributed inside a spherical cell with a diameter of 10 μm and a refractive index of 1.42, and the refractive index of the surrounding medium is 1.334;

[0109] Light source setting: The incident light is a laser beam with a wavelength of 532 nm, and the scattered signals are collected within an angular range of approximately 63° to 117°;

[0110] In total, 1000 groups of cell images and their corresponding two-dimensional light scattering patterns are simulated in this example, and they are randomly divided into a training set and a test set at a ratio of 9:1. Specifically, 900 pairs of two-dimensional light scattering images and cell images are used to train the deep learning model, and the remaining 100 pairs are used to test the performance of the model.

[0111] After training, the model is applied to the test set, and artificial cell images corresponding to the single-cell two-dimensional light scattering images are successfully generated. As Figure 2As shown, representative simulated cell images, corresponding two-dimensional light scattering images, and generated artificial cell images are presented. It can be seen that the generated artificial images are highly visually consistent with the real images and can clearly display the internal structure and characteristics of the cells. By comparing the average curves of the intensities of the middle five pixels scanned, the curves are basically the same, demonstrating the performance of the invention.

[0112] Example 2: Modal amplification applications of microspheres of different sizes.

[0113] Selection of microspheres and image acquisition. In the experiment, two different-sized microspheres were selected: microspheres with diameters of 3.87 μm and 4.19 μm. For each type of microsphere, 600 groups of two-dimensional light scattering (2D LS) images and corresponding experimental bright-field images were acquired, for a total of 1200 groups of images.

[0114] Randomly divided into a training set and a test set in a ratio of 4:1, a convolutional neural network (CNN) was used to train the acquired two-dimensional light scattering images with the goal of generating corresponding high-quality artificial bright-field images. During the training process, the input was the two-dimensional light scattering image, and the output was the real bright-field image. After training was completed, the model was applied to the test set to generate artificial bright-field images corresponding to the two-dimensional light scattering images in the test set. Figure 3 Shows representative experimental bright-field images, two-dimensional light scattering images, and generated artificial bright-field images.

[0115] The focus of the analysis is on the consistency between the artificial bright-field image and the experimental bright-field image. The specific method is to extract the average values of the central five rows of pixels of the artificial bright-field image and its corresponding experimental bright-field image and compare their intensity distributions, as Figure 3 shown. The valleys and peaks of the scanned distributions are basically the same, indicating that the model can accurately reproduce the structural characteristics of the microspheres.

[0116] Further analysis shows that the artificial images retain the key optical characteristics of the experimental bright-field images. For example, the brightness uniformity and edge sharpness in the artificial images are very close to the experimental data, which indicates that the deep learning model effectively learns the mapping relationship between the two-dimensional light scattering pattern and the bright-field image and maintains the fidelity of the key characteristics.

[0117] The experimental results also show that the network can handle changes in microsphere size without significantly reducing accuracy, highlighting the robustness of the model. Although there are minor differences in the intensity values observed at some points, these differences may be attributed to experimental noise and minor deviations in microsphere positioning, but these do not reduce the overall performance of the system.

[0118] Example 3:: Modal amplification applications of cells of different sizes.

[0119] In the experiment, three different types of cells were selected: H8 cells (cervical cancer cell line), HeLa cells (cervical cancer cell line), and K562 cells (leukemia cell line). 1000 sets of multimodal images were collected for each type of cell, and a total of 3000 sets of images were collected. These images included 2D LS images and the corresponding bright-field images and fluorescence images.

[0120] The 3000 sets of experimental data were randomly divided into a training set and a test set at a ratio of 9:1. Specifically, 2700 sets of data were used to train the deep learning model, and 300 sets of data were used to test the performance of the model.

[0121] The collected 2D LS images were trained using a convolutional neural network (CNN) with the goal of generating corresponding high-quality artificial bright-field images. During the training process, the input was the two-dimensional light scattering image, and the output was the real bright-field image or fluorescence image. After training, the model was applied to the test set to generate corresponding data. Figure 4 Representative 2D light scattering images, generated artificial bright-field images, and generated artificial fluorescence images are shown.

[0122] Example 4: Modal amplification of cervical cancer cells to achieve improved analysis and recognition performance.

[0123] 1000 light scattering images, bright-field images, and fluorescence images of H8 cells and HeLa cells were used. All images were randomly divided into a training set and a test set at a ratio of 4:1 and then input into ResNet-50 in the MATLAB Deep Network Designer for classification. The experiment also included generating corresponding multimodal images from the 2D LS images of H8 cells and HeLa cells, and these newly generated images were also randomly divided into a training set and a test set at a ratio of 4:1.

[0124] To evaluate the performance of the classification model, the present invention selected common classification metrics, including accuracy, sensitivity, specificity, F1_score, and AUC (Area Under Curve). The classification results of different cell morphology images are shown in Table 1. From Figure 5 The radar chart in it can more intuitively show the comparison of the classification performance of different modal cell images. The joint classification of the multimodal images generated by the present invention shows more excellent performance, achieving label-free classification of cervical cancer cells. Compared with the classification accuracy of a single modality, the images generated by the system show a significant improvement in classification accuracy. It is worth noting that compared with the classification accuracy of a single light scattering modality image, the improvement in classification performance is close to 20%.

[0125] Table 1: Comparison of the classification performance of different modal cervical cancer cells.

[0126] Cell modal type Accuracy Sensitivity Specificity F1_score AUC Two-dimensional light scattering modal classification result 73.50％ 72.50％ 74.50％ 0.7323 0.8075 Bright field modal classification result 69.00％ 64.50％ 73.50％ 0.6754 0.7581 Fluorescence modal classification result 64.75％ 71.00％ 58.50％ 0.6682 0.6875 Multi-modal classification result after amplification 92.75％ 91.00％ 94.50％ 0.9262 0.9803

[0127] In summary, the present invention improves the accuracy and efficiency of cell classification and analysis by expanding the single-mode light scattering image into a multi-modal image. Especially, its application in cancer cell diagnosis and classification has important value. The present invention uses a deep learning model to realize multi-modal image generation, providing a new method for label-free single-cell analysis and demonstrating broad application prospects.

[0128] Example 2:

[0129] As Figure 6 shown, this implementation provides a single-cell label-free two-dimensional light scattering modality amplification system, including:

[0130] (1) Light source excitation unit

[0131] Laser light source: After starting the system, the laser light source emits a stable and highly directional laser beam for irradiating the single-cell sample;

[0132] Beam shaping: Through a beam shaping device (such as a cylindrical lens, beam expander or fiber optic coupler), the laser beam is adjusted into a shape suitable for single-cell light scattering detection (such as a flat top or focused light spot) to ensure uniform irradiation of the cell surface.

[0133] (2) Data imaging unit

[0134] Detection chip: The excitation light passes through a detection chip composed of a double-layer glass slide and a coverslip. The sample in the sample pool of the detection chip is excited by the excitation light and emits scattered light.

[0135] Image acquisition: The light scattering signal emitted by the detection chip is converted into a digital image through an image acquisition device and saved as a light scattering image, providing raw data for subsequent processing.

[0136] (3) Preprocessing unit

[0137] Picture cropping: The system automatically or manually locates the region of interest (ROI) of light scattering, crops off the redundant background information, and focuses on the single-cell light scattering signal;

[0138] Image filtering: The image is processed through Gaussian filtering, median filtering or other denoising algorithms to eliminate the noise that may be generated during the acquisition process, and at the same time enhance the edges and contrast of the signal.

[0139] (4) Modality amplification unit

[0140] Amplification model: The modality amplification model based on deep learning amplifies the input light scattering image to generate artificial multi-modal images (such as bright field images and fluorescence images);

[0141] Multimodal image: The multimodal image after modality amplification contains more optical information, enhancing the dimension and accuracy of single-cell analysis.

[0142] (5) Result evaluation unit

[0143] Deep learning: Use a pre-trained deep learning model to perform feature analysis on the amplified multimodal image to identify cell types;

[0144] Diagnostic recognition: According to the results of model analysis, output single-cell classification, health status assessment or other diagnostic information to form a final analysis report.

[0145] It can be understood that the above-mentioned various modules can be separately or all combined into one or several other units to form, or a certain one (or some) of the units can be further split into multiple smaller units in terms of function to form, which can achieve the same operation without affecting the realization of the technical effects of the embodiments of this application. The above units are divided based on logical functions. In actual applications, the function of one unit can also be realized by multiple units, or the functions of multiple units can be realized by one unit. In other embodiments of this application, the single-cell label-free two-dimensional light scattering modality amplification system can also include other units. In actual applications, these functions can also be assisted by other units and can be realized by the cooperation of multiple units.

[0146] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A label-free single-cell two-dimensional light scattering imaging modality amplification method, characterized in that: The process includes: Acquiring a label-free two-dimensional light scattering image of a single cell, and preprocessing the two-dimensional light scattering image; Obtaining a bright field image and a fluorescence image corresponding to the two-dimensional light scattering image according to the preprocessed two-dimensional light scattering image and the pre-trained deep learning model; The two-dimensional light scattering image and the corresponding bright field image and fluorescence image are post-processed to obtain a multimodal amplification result for cell classification according to a pre-trained classification model.

2. The single cell label-free two-dimensional light scattering modality amplification method according to claim 1, characterized in that: The two-dimensional light scattering image is preprocessed, including: denoising, cropping and normalization.

3. The single cell label-free two-dimensional light scattering modality amplification method according to claim 1, characterized in that: The pre-trained deep learning model adopts a generative adversarial network model, which includes: a generator and a discriminator; the generator adopts a U-net architecture, including an encoder, a decoder and a jump connection; The encoder is used to extract the features of the preprocessed two-dimensional light scattering image layer by layer, and the decoder is used to restore the image information layer by layer. The encoder and the decoder are connected through a convolution layer, and a skip connection is established between the corresponding layers of the encoder and the decoder to directly transfer the low-level features to the corresponding layers of the decoder, thereby retaining the detail information in the input image; The discriminator is a convolutional neural network, which includes a convolutional layer, a batch normalization layer and a LeakyReLU activation layer. The discriminator is used to classify the image output by the generator and determine whether the image output by the generator is real or generated.

4. The single cell label-free two-dimensional light scattering modality amplification method according to any one of claims 1 to 3, characterized in that: The pre-trained deep learning model includes the loss function L overall : Among them, L cGAN (G, D) is the conditional adversarial loss, L L1 (G) is the L1 distance loss, is the VGG perceptual loss, α, β and γ are weight coefficients, and the contribution of each loss is adjusted by setting different weight coefficients α, β and γ so that the generated network can be adaptively adjusted according to different task requirements and data characteristics.

5. The single cell label-free two-dimensional light scattering modality amplification method according to claim 4, characterized in that: The conditional adversarial loss is: L cGAN (G,D)=E x,y [logD(x,y)]+E x [log(1-D(x,G(x)))], where x is the input image, y is the corresponding true output image, G(x) is the generated image, D(x,y) is the real image pair in the discriminator, and D(x,G(x)) is the generated image pair; The L1 distance loss is: L1 (G) = E x,y [yG(x)1]; The VGG perceptual loss is: where VGGl(·) represents the feature map extracted from the lth layer of the VGG-19 network, Z is the set of layers selected for computing the perceptual loss, and W l and H l They represent the width and height of the feature map of the lth layer respectively.

6. A label-free single-cell two-dimensional light scattering imaging modality amplification system, characterized in that: include: Light source excitation unit: equipped with laser light source and beam shaping components to provide excitation beam for single cells; Data imaging unit: including detection chip and image acquisition device, used to capture light scattering signals and convert them into digital images; Preprocessing unit: used to crop and filter the acquired images, extract the region of interest of single-cell light scattering and improve the image quality; Modality amplification unit: including an amplification model, used to amplify the two-dimensional light scattering image into a multi-modality image; Result evaluation unit: Classification or diagnostic analysis of multimodal images based on deep learning models.

7. The label-free single-cell two-dimensional light scattering imaging modality amplification system according to claim 6, characterized in that: The modality augmentation unit adopts an augmentation model based on convolutional neural networks to generate multimodal images with more information.

8. The label-free single-cell two-dimensional light scattering modality amplification system according to claim 6, characterized in that: In the post-processing unit, the pre-trained deep learning model includes the loss function L overall : Among them, L cGAN (G, D) is the conditional adversarial loss, L L1 (G) is the L1 distance loss, is the VGG perceptual loss, and α, β, and γ are weight coefficients.

9. The label-free single-cell two-dimensional light scattering modality amplification system according to claim 6, characterized in that: The result evaluation unit combines deep learning and traditional classification methods for cell type identification, health status assessment and disease diagnosis.

10. The label-free single-cell two-dimensional light scattering modality amplification system according to claim 6, characterized in that: The sample is excited by the light source excitation unit; the light scattering image is captured by the data imaging unit; the image is cropped and filtered by the preprocessing unit; the multimodal image is generated by the modal amplification unit; and the result evaluation unit is used for analysis to output the diagnosis result.

Images