A deep learning-based senescent cell microscopic image recognition method

Abstract

This invention discloses a deep learning-based method for recognizing senescent cells using microscopic images. The method includes: segmenting a multi-cell image to be identified into single-cell images, wherein the multi-cell image to be identified is either a bright-field image or a nuclear channel image; inputting the single-cell image to be identified into a senescent cell recognition network model to obtain the senescence recognition result, wherein the senescent cell recognition network model extracts features through multiple parallel classification network models. This invention adopts an end-to-end approach and constructs a senescent cell recognition network model based on microscopic images using deep learning, achieving the output of a senescent cell probability map and the proportion of senescent cells in the corresponding image from an input bright-field image or a nuclear channel image.

Description

Technical Field

[0001] This invention relates to a deep learning-based method for recognizing microscopic images of senescent cells, belonging to the field of image recognition technology. Background Technology

[0002] Cellular senescence involves functional decline at multiple levels, including the body, tissues, organs, cells, and molecules, and is closely related to malignant tumors and age-related diseases. Accurate identification of senescent cells helps to understand the relationship between senescent cells and diseases and plays an important role in studying the pathogenesis of age-related diseases. There is no unified standard for detecting and identifying senescent cells, and it is usually necessary to combine multiple biomarkers to identify senescent cells. Among them, during the aging process, the structure and morphology of cells undergo significant changes, exhibiting unique morphology. The topic of identifying senescent cells based on the morphological characteristics of cells has attracted people's attention. [Kusumoto D, Seki T, Sawada H, et al. Anti-senescent drug screening by deep learning-based morphology senescence scoring. Nature Communications, 2021, 12(1): 1-10].

[0003] Currently, only a few machine learning algorithms are used to identify senescent cells. Cell biology methods are commonly used, and two common biological characteristics for detecting cellular senescence are: first, growth arrest, where cells stop dividing and enter irreversible cell cycle arrest; and second, increased activity of senescence-associated β-galactosidase (SAβ-gal) [Hernandez-Segura A, Nehme J and Demaria M. Hallmarks of cellular senescence. Trends in cell biology, 2018, 28(6):436-453]. These biological detection methods typically require measuring changes in gene expression and protein markers, making the detection process complex and resource-intensive. In the detection of senescent cells, combining microscopic imaging and deep learning technologies for automatic identification can save significant time and is more convenient and efficient.

[0004] With technological advancements, deep learning has driven the analysis and solution of many complex problems and has been widely applied in fields such as image recognition and object detection, achieving significant progress in biomedical image processing [Skrede OJ, DeRaedt S, Kleppe A, et al. Deep learning for prediction of colorectal cancer outcome: a discovery and validation study. The Lancet, 2020, 395(10221):350-360]. In the field of cell biology, cell morphology images obtained through microscopy contain a wealth of biological information. Neural networks are highly sensitive to morphological changes and can capture minute changes imperceptible to humans. With the establishment of convolutional neural networks (CNNs) and the proposal of more models, the accuracy of image classification tasks has been greatly improved. However, compared to image recognition tasks in other fields, biomedical datasets are smaller and more difficult to obtain and label. Therefore, data processing in the early stages of model building is more important. Manual feature extraction relies on doctors' experience and is subject to human error, requiring further improvement. Summary of the Invention

[0005] In view of this, the present invention provides a method, system, computer device and storage medium for senescent cell microscopic image recognition based on deep learning. It adopts an end-to-end approach and constructs a senescent cell recognition network model based on microscopic images through deep learning. It realizes that the input is a bright field image or a nuclear channel image, and the output is a probability map of senescent cells and the proportion of senescent cells in the corresponding image.

[0006] The first objective of this invention is to provide a method for recognizing senescent cell microscopic images based on deep learning.

[0007] The second objective of this invention is to provide a deep learning-based microscopic image recognition system for senescent cells.

[0008] A third objective of this invention is to provide a computer device.

[0009] A fourth objective of this invention is to provide a storage medium.

[0010] The first objective of this invention can be achieved by adopting the following technical solution:

[0011] A method for identifying senescent cells using microscopic images, the method comprising:

[0012] The multi-cell image to be identified is segmented into single-cell images to be identified, wherein the multi-cell image to be identified is either a bright-field image or a nuclear channel image.

[0013] The single-cell image to be identified is input into the senescent cell recognition network model to obtain the senescence recognition result. The senescent cell recognition network model extracts features through multiple classification network models in parallel.

[0014] Furthermore, the classification network model is of three types: a first classification network model, a second classification network model, and a third classification network model.

[0015] The first classification network model is built based on ResNet34;

[0016] The second classification network model is built based on Resnext50;

[0017] The third classification network model is built based on the Swing Transformer.

[0018] Furthermore, the first classification network model includes a first convolutional layer, a first max pooling layer, a residual structure, a second max pooling layer, and a fully connected layer connected in sequence.

[0019] The kernel size of the first convolutional layer is 7×7, and the stride is 2;

[0020] The kernel size of the first max pooling layer and the second max pooling layer is 3×3; the first max pooling layer and the second max pooling layer are used for downsampling;

[0021] The residual structure consists of multiple first residual blocks.

[0022] Furthermore, the second classification network model includes a second residual block;

[0023] In the second residual block, group convolution is used for optimization;

[0024] The group convolutions consist of at least thirty-two groups, each group of which includes a second convolutional layer with 4 channels and a kernel size of 3×3.

[0025] Furthermore, the third classification network model includes a Win-T encoding structure, a LinearEmbedding layer, a module group, and a softmax layer connected in sequence.

[0026] The module group includes at least four Swing Transformer Block modules connected in sequence.

[0027] Furthermore, the segmentation of the multi-cell image to be identified into single-cell images to be identified specifically includes:

[0028] Obtain the binarized template image to be identified;

[0029] Based on the binarized template image to be identified, the multi-cell image to be identified is segmented to obtain the single-cell image to be identified;

[0030] The process of obtaining the binary template image to be identified specifically includes:

[0031] Background correction and binarization are performed on the multi-nucleus channel image to be identified to obtain the binarized image;

[0032] Noise is removed from the binarized image using an opening operation.

[0033] In the noise-removed image, constraints are set to remove connected components that do not belong to a single cell, resulting in a binarized template image to be identified.

[0034] Furthermore, the process of acquiring the senescent cell recognition network model specifically includes:

[0035] Multiple classification network models are trained multiple times using a dataset, which includes multiple single-cell bright-field images or multiple single-cell nuclear channel images;

[0036] The voting algorithm is used to fuse multiple trained classification network models, thereby obtaining multiple fused network models;

[0037] If the dataset consists of multiple single-cell bright-field images, the recognition accuracy of the fusion network model is calculated using the average fusion method.

[0038] If the dataset consists of multiple single-cell nucleus channel images, a weighted fusion method is used to calculate the recognition accuracy of the fusion network model;

[0039] The best-performing fusion network model was used as the senescent cell recognition network model.

[0040] The second objective of this invention can be achieved by adopting the following technical solution:

[0041] A microscopic image recognition system for senescent cells, the system comprising:

[0042] A segmentation unit is used to segment a multi-cell image to be identified into a single-cell image to be identified, wherein the multi-cell image to be identified is either a bright-field image or a nuclear channel image.

[0043] The recognition unit is used to input the image of the single cell to be recognized into the senescent cell recognition network model to obtain the senescence recognition result. The senescent cell recognition network model extracts features through multiple classification network models in parallel.

[0044] The third objective of this invention can be achieved by adopting the following technical solution:

[0045] A computer device includes a processor and a memory for storing a processor-executable program, wherein when the processor executes the program stored in the memory, it implements the above-described method for identifying senescent cell microscopic images.

[0046] The fourth objective of this invention can be achieved by adopting the following technical solution:

[0047] A storage medium storing a program that, when executed by a processor, implements the above-described method for recognizing senescent cell microscopic images.

[0048] The present invention has the following advantages over the prior art:

[0049] 1. This invention constructs a senescent cell recognition network model based on microscopic images through deep learning, realizing the output of senescent cell probability map and senescent cell percentage of the corresponding image from the input bright field image or nuclear channel image;

[0050] 2. This invention adopts an end-to-end approach. After the network model is built and trained, the user only needs to input a multi-cell bright-field image or a multi-cell nuclear channel image obtained by microscopic imaging at the input end. The user can then obtain the identification result of a single senescent cell and the proportion of senescent cells at the output end, which intuitively provides the senescence identification result image. It has the advantages of convenient identification and short time, and can also be used to assist in judgment.

[0051] 3. This invention can continuously update the network model using new input data, thereby improving the accuracy of the network model for long-term use. It can also be deployed directly on a local computer with a simple and easy-to-operate interface.

[0052] 4. This invention employs an average fusion method and a weighted fusion method, which improves the prediction accuracy of the senescent cell recognition network model. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0054] Figure 1 This is a flowchart illustrating the deep learning-based microscopic image recognition method for senescent cells according to Embodiment 1 of the present invention.

[0055] Figure 2This is a simplified flowchart of the deep learning-based microscopic image recognition method for senescent cells according to Embodiment 1 of the present invention.

[0056] Figure 3 This is a flowchart of the end-to-end output aging probability heatmap of Embodiment 1 of the present invention.

[0057] Figure 4 This is a schematic diagram of the data acquisition process in Embodiment 1 of the present invention.

[0058] Figure 5 This is a schematic diagram of image binarization processing in Embodiment 1 of the present invention.

[0059] Figure 6(a) is a schematic diagram of the bright field image ROC curve and its area under the curve (AUC) value in Embodiment 1 of the present invention.

[0060] Figure 6(b) is a schematic diagram of the ROC curve of the kernel channel image and its area under the curve (AUC) value in Embodiment 1 of the present invention.

[0061] Figure 6(c) is a schematic diagram of the bright field image confusion matrix of Embodiment 1 of the present invention.

[0062] Figure 6(d) is a schematic diagram of the kernel channel image confusion matrix of Embodiment 1 of the present invention.

[0063] Figure 7(a) is a schematic diagram of the kernel channel image output result of Embodiment 1 of the present invention.

[0064] Figure 7(b) is a schematic diagram of the bright field image output result of Embodiment 1 of the present invention.

[0065] Figure 8 This is a structural block diagram of the deep learning-based senescent cell microscopic image recognition system according to Embodiment 2 of the present invention.

[0066] Figure 9 This is a structural block diagram of the computer device according to Embodiment 3 of the present invention. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0068] Example 1:

[0069] like Figure 1 and Figure 2 and Figure 3As shown in the figure, this embodiment provides a method for recognizing senescent cell microscopic images based on deep learning. The method includes the following steps:

[0070] S1, cell culture, and induction of cell senescence.

[0071] Human alveolar adenocarcinoma basal epithelial cells (A549 cells) were cultured, and doxorubicin was added to induce senescence in the logarithmic growth phase of these cells.

[0072] Specifically, the human lung cancer cell line (A549) was purchased from the Tumor Cell Bank of the Chinese Academy of Medical Sciences and cultured in DMEM medium containing 10% fetal bovine serum, 100 μg / mL streptomycin, and 100 U / mL penicillin G at 37°C in an incubator containing 5% carbon dioxide. All experiments were performed during the logarithmic growth phase of these cells. The cells were digested with trypsin, transferred to cell culture dishes, and cultured for 24 hours. When the cells reached 50% confluence, doxorubicin was added to induce senescence. Senescent cells were identified by increased senescence-associated β-galactosidase (SA-β-Gal) staining and morphological characteristics, typically appearing as flattened, enlarged cell bodies.

[0073] In this embodiment, the specific steps for inducing senescence in this type of cell are as follows:

[0074] S11. When the cells in the culture dish reached 50% confluence, the cells were treated with (0-100 μM) doxorubicin for 24 h. The dose-dependent cytotoxicity induced by doxorubicin was determined by CCK8. 1 μM doxorubicin reduced cell viability to about 55% and was selected for subsequent experiments.

[0075] S12. Replace the old culture medium with fresh culture medium every 48 hours. On the fifth day, obtain senescent cells and use them for subsequent microscopic imaging, i.e., step S2.

[0076] In subsequent step S2, microscopic imaging showed that doxorubicin increased the percentage of SA-β-Gal positive cells from 5.8% (control group) to 72.9% (DOX-treated group).

[0077] S2, Microscopic imaging (imaging bright-field images and Hoechst-stained cells).

[0078] like Figure 4As shown, after obtaining senescent cells from step S1, the culture medium in the culture dish is replaced with serum-free DMEM, and Hoechst (20 μg / mL) staining is added. The culture dish is then placed back into the incubator and incubated for 30 minutes. After 30 minutes, the Hoechst staining solution is aspirated, and the cells in the culture dish are washed 1-2 times with serum-free DMEM or PBS. DMEM is then added again before microscopic imaging.

[0079] In this embodiment, bright-field images and nuclear channel images were captured using an Olympus microscope with an objective magnification of 40. Imaging conditions (including light intensity and exposure time) were controlled, and the images were stored in TIF file format with a resolution of 2048×2048. This allowed for the acquisition of the original datasets for the aging group and the control group.

[0080] It should be noted that bright-field images taken using an Olympus microscope are multi-cell bright-field images, and nuclear channel images taken using an Olympus microscope are multi-cell nuclear channel images. All nuclear channel images mentioned in this embodiment are fluorescence images.

[0081] S3. Obtain the binarized template image.

[0082] like Figure 5 As shown, in order to segment and obtain single-cell images, namely single-cell bright-field images and single-cell nuclear channel images, the multi-cell nuclear channel images are binarized to obtain binarized template images. The specific steps are as follows:

[0083] S31. Perform background correction and binarization on the multi-nucleus channel image to obtain the binarized image.

[0084] S311. For cases of uneven lighting or dirt on the lens, background correction is performed on the multi-nucleus channel image obtained in step S2 to remove background noise. Specifically, high-frequency components in the image are filtered out by Gaussian blur to obtain a background image, and the background image is subtracted from the original image to obtain a denoised image.

[0085] S312. Use the OTSU algorithm to obtain the binarization threshold. According to the grayscale characteristics of the image, determine the optimal threshold to maximize the inter-class variance. Perform binarization on the denoised image and display it to obtain the binarized image.

[0086] S32. Use the opening operation to remove noise from the binarized image.

[0087] The opening operation is used to remove noise from the binarized image. This involves first eroding and then dilating the image, and then filling small holes through morphological operations to obtain the noise-removed image.

[0088] S33. In the image after noise removal, set constraints to remove connected components that do not belong to a single cell, and obtain a binarized template image.

[0089] When acquiring image data, images of poor quality, such as overlapping cells, may be encountered. After removing image noise, each connected component is analyzed and processed to calculate the area (area), the ratio of the shortest to the longest diameter (dist_ratio), the smoothness, and the ratio of the area of ​​the largest inscribed circle to the area of ​​the smallest circumscribed circle (area_ratio) for each cell (connected component). By setting constraints, connected components that are clearly not single cells are removed. Each connected component in the image is traversed, and the range of the area (area) and the ratio of the shortest to the longest diameter (dist_ratio) is determined. If area < 7000, area > 29000, and dist_ratio < 0.6, the connected component is deleted, and the search continues for the next connected component until the traversal is complete. The removed image, i.e., the binarized template image, is then output.

[0090] In this embodiment, area calculates the total number of pixels in the connected region;

[0091] The formula for calculating dist_ratio is: That is, the ratio of the length of the minor axis of an ellipse having the same standard second central moment as the connected region to the length of the major axis of an ellipse having the same standard second central moment as the connected region;

[0092] The formula for calculating smoothness is: (Define distance(0) = distance(dislen), distance(dislen+1) = distance(1)), where distance stores the distance from all points on the edge of the i-th cell to the center point, and dislen is the number of all points on the edge of the connected component;

[0093] The formula for calculating area_ratio is: Where dismin = min(distance), dismax = max(distance).

[0094] S4, Single-cell image segmentation.

[0095] Based on the binarized template image, the multi-cell bright-field image and the multi-cell nuclear channel image are segmented to obtain the single-cell bright-field image and the single-cell nuclear channel image, thus obtaining the dataset. The specific steps are as follows:

[0096] S41. Use the binarized template image as a template for cell segmentation, and perform connected component analysis on the binarized template image (obtain the centroid coordinates of the connected components and mark the position of each cell).

[0097] S42, the binarized template image and its corresponding multi-cell bright-field image and multi-cell nuclear channel image are used as input, and the output is the segmented single-cell image, namely the single-cell bright-field image and the single-cell nuclear channel image, which are saved in the Brightfield and Nucleus folders respectively, and are used as the dataset for the senescent cell recognition network model.

[0098] It should be noted that most of the cells in the microscopic imaging results are senescent cells, which are labeled as the senescent group. Correspondingly, the segmented single-cell images are also labeled as senescent cells. Cells that were not treated with doxorubicin are labeled as control group cells.

[0099] S5. Construct a senescent cell recognition network model, wherein the senescent cell recognition network model extracts features through multiple parallel classification network models.

[0100] The general idea for constructing a network model for recognizing senescent cells is as follows: Based on the characteristics of ResNet34, Resnext50, and SwingTransformer neural network models, corresponding classification network models are constructed. Preprocessed image features are then obtained through these classification network models. The obtained image features and corresponding labels are then fed into the corresponding classification network models for training, resulting in trained classification network models. Finally, all trained classification network models are fused, and the optimal fused network model is selected as the senescent cell recognition network model.

[0101] In practical applications, the image input to the senescent cell recognition network model (input image) is preprocessed. Specifically, the preprocessing involves: first, randomly cropping, rotating, and horizontally flipping the input image, and then performing data augmentation. Next, the data-augmented input image is normalized, and its size is set to 224×224.

[0102] The specific steps for obtaining the senescent cell recognition network model are as follows:

[0103] S51. Train multiple classification network models multiple times using a dataset, wherein the dataset includes multiple single-cell bright-field images or multiple single-cell nuclear channel images.

[0104] In this embodiment, there are three classification network models: a first classification network model, a second classification network model, and a third classification network model. The first classification network model is built based on ResNet34; the second classification network model is built based on ResNext50; and the third classification network model is built based on Swing Transformer.

[0105] Furthermore, the first classification network model includes a first convolutional layer, a first max pooling layer, a residual structure, a second max pooling layer, and a fully connected layer connected in sequence; the kernel size of the first convolutional layer is 7×7 and the stride is 2; the kernel size of the first max pooling layer and the second max pooling layer is 3×3; the first max pooling layer and the second max pooling layer are used for downsampling; the residual structure is composed of multiple first residual blocks.

[0106] Furthermore, the second classification network model includes a second residual block; in the second residual block, group convolution is used for optimization; there are at least thirty-two groups of group convolutions, each group of group convolutions includes a second convolutional layer, the number of channels of the second convolutional layer is 4, and the kernel size of the second convolutional layer is 3×3.

[0107] Furthermore, the third classification network model includes a sequentially connected Swin-T encoding structure, a LinearEmbedding layer, a module group, and a softmax layer; the module group includes four sequentially connected Swin Transformer Block modules.

[0108] In the third classification network model, the Swin-T encoding structure extracts the features of the preprocessed input image and downsamples the height and width of the preprocessed input image by a factor of four. Then, the output channel dimension obtained by mapping through the Linear Embedding layer is equal to 96, and then passes through four Swin Transformer Block modules. Finally, the classification result is output through the softmax layer.

[0109] In this embodiment, 80% of the dataset is used as the training set, and the remaining 20% ​​is used as the test set. The training set is then divided into eight equal parts. Each time, one part is taken as the validation set, and the remaining seven parts are used as the training set. This process is repeated eight times to train the model, continuously adjusting and optimizing the parameters. Finally, the model's generalization ability is validated on the test set. TensorBoard is used to visualize the model training process, including the prediction accuracy (train_accuracy) and loss (train_loss) curves of each classification network model on the training set, and the validation accuracy (valid_accuracy), loss (valid_loss), and learning rate (learning_rate) curves on the validation set. The hyperparameters of the corresponding classification network model are fine-tuned to continuously optimize the network model and improve its prediction accuracy. The hyperparameters include the number of neurons, batch size, loss function, optimizer, and activation function.

[0110] S52. The voting algorithm is used to fuse multiple training classification network models to obtain multiple fused network models.

[0111] S53. As shown in Figure 6(a), if the dataset consists of multiple single-cell bright-field images, the average fusion method is used to calculate the recognition accuracy of the fusion network model; as shown in Figure 6(b), if the dataset consists of multiple single-cell nuclear channel images, the weighted fusion method is used to calculate the recognition accuracy of the fusion network model; the fusion network model with the best performance is used as the senescent cell recognition network model.

[0112] In step S53, the weights for weighted fusion are calculated based on the loss functions of the three classification network models, using the following formula:

[0113]

[0114]

[0115]

[0116]

[0117]

[0118]

[0119] In addition, this embodiment also evaluates the fusion network model by calculating the confusion matrix, F1 value, ROC curve and AUC value, so as to select the optimal fusion network model as the senescent cell recognition network model.

[0120] As shown in Figures 6(c) and 6(d), the metrics obtained from the confusion matrix include Accuracy, Precision, Recall, and Specificity.

[0121]

[0122]

[0123]

[0124]

[0125]

[0126] Where TP represents the number of correctly predicted positive samples, TN represents the number of correctly predicted negative samples, FN represents the number of incorrectly predicted negative samples, and FP represents the number of incorrectly predicted positive samples.

[0127] S6. Segment the multi-cell image to be identified into single-cell images to be identified, wherein the multi-cell image to be identified is either a bright-field image or a nuclear channel image.

[0128] In step S6, the segmentation is performed with reference to step S4.

[0129] S7. Input the single-cell image to be identified into the senescent cell recognition network model to obtain the senescence recognition result.

[0130] The bright field image of the single cell to be identified is used as the input of the corresponding senescent cell identification network model, and the output is the classification probability of the cell being identified as a senescent cell, i.e. the senescence identification result.

[0131] The image of the single cell nucleus channel to be identified is used as the input of the corresponding senescent cell recognition network model, and the output is the classification probability of the cell being identified as a senescent cell, i.e. the senescence recognition result.

[0132] Based on the classification probability, the cell state is evaluated, and a senescent cell probability map is output, as shown in Figure 7(a) or Figure 7(b).

[0133] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware, and the corresponding program can be stored in a computer-readable storage medium.

[0134] It should be noted that although the method operations of the above embodiments are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. On the contrary, the order of execution of the described steps may be changed. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0135] Example 2:

[0136] like Figure 8 As shown, this embodiment provides a deep learning-based microscopic image recognition system for senescent cells. The system includes a segmentation unit 801 and a recognition unit 802, and the specific functions of each unit are as follows:

[0137] The segmentation unit 801 is used to segment the multi-cell image to be identified into a single-cell image to be identified, wherein the multi-cell image to be identified is one of a bright field image type and a nuclear channel image type;

[0138] The recognition unit 802 is used to input the single-cell image to be recognized into the senescent cell recognition network model to obtain the senescence recognition result, wherein the senescent cell recognition network model extracts features through multiple classification network models in parallel.

[0139] Example 3:

[0140] like Figure 9 As shown, this embodiment provides a computer device, which includes a processor 902, a memory, an input device 903, a display device 904, and a network interface 905 connected via a system bus 901. The processor 902 provides computing and control capabilities. The memory includes a non-volatile storage medium 906 and internal memory 907. The non-volatile storage medium 906 stores an operating system, computer programs, and a database. The internal memory 907 provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium 906. When the computer program is executed by the processor 902, it implements the senescent cell microscopic image recognition method of Embodiment 1 described above, as follows:

[0141] The multi-cell image to be identified is segmented into single-cell images to be identified, wherein the multi-cell image to be identified is either a bright-field image or a nuclear channel image.

[0142] The single-cell image to be identified is input into the senescent cell recognition network model to obtain the senescence recognition result. The senescent cell recognition network model extracts features through multiple classification network models in parallel.

[0143] Example 4:

[0144] This embodiment provides a storage medium, which is a computer-readable storage medium, storing a computer program. When the computer program is executed by a processor, it implements the senescent cell microscopic image recognition method of Embodiment 1 above, as follows:

[0145] The multi-cell image to be identified is segmented into single-cell images to be identified, wherein the multi-cell image to be identified is either a bright-field image or a nuclear channel image.

[0146] The single-cell image to be identified is input into the senescent cell recognition network model to obtain the senescence recognition result. The senescent cell recognition network model extracts features through multiple classification network models in parallel.

[0147] It should be noted that the computer-readable storage medium in this embodiment can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0148] In this embodiment, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this embodiment, the computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying a computer-readable program. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable storage medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The computer program contained on the computer-readable storage medium can be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination thereof.

[0149] The computer-readable storage medium described above can be used to write computer programs for executing this embodiment in one or more programming languages ​​or combinations thereof. These programming languages ​​include object-oriented programming languages—such as Java, Python, and C++—and conventional procedural programming languages—such as C or similar programming languages. The program can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0150] In summary, this invention adopts an end-to-end approach and constructs a senescent cell recognition network model based on microscopic images through deep learning. It can input a bright-field image or a nuclear channel image and output the probability map of senescent cells and the proportion of senescent cells in the corresponding image.

[0151] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope disclosed in the present invention, based on the technical solution and inventive concept of the present invention, shall fall within the scope of protection of the present invention.

Claims

1. A method for recognizing senescent cells using microscopic images based on deep learning, characterized in that, The method includes: The multi-cell image to be identified is segmented into single-cell images to be identified, wherein the multi-cell image to be identified is either a bright-field image or a nuclear channel image. The single-cell image to be identified is input into the senescent cell recognition network model to obtain the senescence recognition result. The senescent cell recognition network model extracts features through multiple parallel classification network models. These classification network models are of three types: a first classification network model, a second classification network model, and a third classification network model. The first classification network model is based on... Resnet 34 is constructed; the second classification network model is based on Resnext 50 is constructed; the third classification network model is based on Swin Transformer To build; The process of obtaining the senescent cell recognition network model specifically includes: Multiple classification network models are trained multiple times using a dataset, which includes multiple single-cell bright-field images or multiple single-cell nuclear channel images; use voting The algorithm fuses multiple trained classification network models to obtain multiple fused network models; If the dataset consists of multiple single-cell bright-field images, the recognition accuracy of the fusion network model is calculated using the average fusion method. If the dataset consists of multiple single-cell nucleus channel images, a weighted fusion method is used to calculate the recognition accuracy of the fusion network model; The best-performing fusion network model was used as the senescent cell recognition network model.

2. The method for recognizing senescent cell microscopic images according to claim 1, characterized in that, The first classification network model includes a first convolutional layer, a first max pooling layer, a residual structure, a second max pooling layer, and a fully connected layer connected in sequence. The kernel size of the first convolutional layer is 7×7, and the stride is 2; The kernel size of the first max pooling layer and the second max pooling layer is 3×3; the first max pooling layer and the second max pooling layer are used for downsampling; The residual structure consists of multiple first residual blocks.

3. The method for recognizing senescent cell microscopic images according to claim 1, characterized in that, The second classification network model includes a second residual block; In the second residual block, group convolution is used for optimization; The group convolutions consist of at least thirty-two groups, each group of which includes a second convolutional layer with 4 channels and a kernel size of 3×3.

4. The method for recognizing senescent cell microscopic images according to claim 1, characterized in that, The third classification network model includes sequentially connected... Swin-T Encoding structure, Linear Embedding Layers, module groups softmax layer; The module group includes at least four modules connected in sequence. Swin Transformer Block Module.

5. The method for recognizing senescent cell microscopic images according to claim 1, characterized in that, The step of segmenting the multi-cell image to be identified into single-cell images to be identified specifically includes: Obtain the binarized template image to be identified; Based on the binarized template image to be identified, the multi-cell image to be identified is segmented to obtain the single-cell image to be identified; The process of obtaining the binary template image to be identified specifically includes: Background correction and binarization are performed on the multi-nucleus channel image to be identified to obtain the binarized image; Noise is removed from the binarized image using an opening operation. In the noise-removed image, constraints are set to remove connected components that do not belong to a single cell, resulting in a binarized template image to be identified.

6. A deep learning-based microscopic image recognition system for senescent cells, characterized in that, The system includes: A segmentation unit is used to segment a multi-cell image to be identified into a single-cell image to be identified, wherein the multi-cell image to be identified is either a bright-field image or a nuclear channel image. The recognition unit is used to input the single-cell image to be recognized into the senescent cell recognition network model to obtain the senescence recognition result. The senescent cell recognition network model extracts features through multiple parallel classification network models. These classification network models are of three types: a first classification network model, a second classification network model, and a third classification network model. The first classification network model is based on... Resnet 34 is constructed; the second classification network model is based on Resnext 50 is constructed; the third classification network model is based on Swin Transformer To build; The process of obtaining the senescent cell recognition network model specifically includes: Multiple classification network models are trained multiple times using a dataset, which includes multiple single-cell bright-field images or multiple single-cell nuclear channel images; use voting The algorithm fuses multiple trained classification network models to obtain multiple fused network models; If the dataset consists of multiple single-cell bright-field images, the recognition accuracy of the fusion network model is calculated using the average fusion method. If the dataset consists of multiple single-cell nucleus channel images, a weighted fusion method is used to calculate the recognition accuracy of the fusion network model; The best-performing fusion network model was used as the senescent cell recognition network model.

7. A computer device comprising a processor and a memory for storing a processor-executable program, characterized in that, When the processor executes the program stored in the memory, it implements the senescent cell microscopic image recognition method according to any one of claims 1-5.

8. A storage medium storing a program, characterized in that, When the program is executed by the processor, it implements the senescent cell microscopic image recognition method according to any one of claims 1-5.

Images