System and method for identifying senescent cells based on mitochondrial morphology

Abstract

The invention discloses a senescent cell recognition system and method based on mitochondrial morphology. The system is a classification network, and the input data of the system is a distribution image of mitochondria in subcellular level cells, the distribution image of mitochondria in cells, the shape and / or number of mitochondria distributed at different positions in a three-dimensional structure for displaying cells, and preferably the shape and number of mitochondria distributed at different positions in the three-dimensional structure for displaying cells. According to the senescence cell recognition system based on the mitochondrial morphology, the senescence cells jointly show mitochondrial morphological changes including mitochondrial distribution, mitochondrial shapes and mitochondrial number changes, the senescence cell recognition method is easy to conduct through computer vision, good robustness is achieved, and the senescence cell recognition system based on the mitochondrial morphology is suitable for being applied to senescence cell recognition. Compared with other senescence cell identification methods based on cell nucleus morphological characteristics, the senescence cell identification method based on the general characteristics of the senescence cells has good universality for different cell types.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics, and more specifically, relates to a senescent cell identification system and method based on mitochondrial morphology. Background Art

[0002] Accurately identifying senescent cells is crucial for elucidating the mechanisms of aging and developing anti-aging therapies such as senolytics. Image-based senescent cell identification has been a significant breakthrough at the intersection of biomedicine and artificial intelligence in recent years. Compared to traditional biochemical-based senescent cell identification methods, image-based methods offer the advantages of being non-invasive and high-throughput.

[0003] Existing image-based methods for identifying senescent cells include a senescent cell identification network based on nuclear morphology, proposed by Indra et al. (Nature Aging, 2, 742–755 (2022)). This method achieves over 94% accuracy in a basic aging model of fibroblasts using nuclear fluorescence images, but the accuracy drops to 70% when applied to other aging pathways in fibroblasts. Chinese patent document CN115457549A provides a cell-based microscopic image recognition method for identifying senescent cells. Using images of cell or nuclear channels, the method achieves over 98% accuracy in identifying senescent cells in A549 cells induced by doxorubicin. These methods all utilize readily available cell image information for senescent cell identification. However, due to the wide variety of cell types and complex physiological processes, nuclear morphology and overall cell morphology are significantly affected by factors such as cell type and the stage of the cell's physiological and biochemical process. Therefore, the transferability of these identification methods is greatly limited, making it difficult to generalize to other cell types and accurately identify individual senescent cells. Summary of the Invention

[0004] In response to the above-mentioned defects or improvement needs of the prior art, the present invention provides a senescent cell identification system and method based on mitochondrial morphology. Its purpose is to identify the degree of cell aging by extracting features from the input mitochondrial distribution image at the subcellular level of the cell, obtaining the mitochondrial shape, number and distribution characteristics, and having good generalization performance for different cell lines, thereby solving the technical problem that the prior art is limited to the identification of the degree of cell aging, has poor versatility, and requires separate development and training for each cell type.

[0005] To achieve the above objectives, according to one aspect of the present invention, a senescent cell identification system based on mitochondrial morphology is provided, which is a classification network whose input data is: a mitochondrial distribution image in cells at the subcellular level, wherein the mitochondrial distribution image in the cells displays the shape and / or number of mitochondria distributed at different positions in the three-dimensional structure of the cell, preferably displays the shape and number of mitochondria distributed at different positions in the three-dimensional structure of the cell.

[0006] Preferably, in the senescent cell identification system based on mitochondrial morphology, the mitochondrial distribution image in the cell is a slice sequence of preset positions in a three-dimensional image of the cell, which is specifically obtained according to the following method:

[0007] S1. Obtain a three-dimensional image of the cell to be identified at the subcellular level;

[0008] S2. For the three-dimensional cell image obtained in step S1, continuous transverse image slices are extracted at the upper, middle, and lower parts to obtain a slice sequence.

[0009] Preferably, in the senescent cell identification system based on mitochondrial morphology, mitochondrial fluorescence signals are collected in the three-dimensional image of the cells.

[0010] Preferably, in the senescent cell identification system based on mitochondrial morphology, the three-dimensional image is a two-dimensional image stack acquired by wide-field imaging, and the resolution thereof is 220 nm-280 nm.

[0011] Preferably, in the senescent cell identification system based on mitochondrial morphology, the mitochondrial distribution image in the cell is a slice sequence of preset positions in a three-dimensional image of the cell, which is specifically obtained according to the following method:

[0012] For a cell's mitochondrial fluorescence 2D image stack, if the total number of layers containing mitochondrial information is N, let the top layer be numbered 1 and the bottom layer be numbered N. Six images are extracted from the stack as the network dataset, with sequence numbers 1, 2, (N-1), N, and two consecutive images selected in the middle.

[0013] Preferably, in the senescent cell identification system based on mitochondrial morphology, the mitochondria in the cell mitochondrial image include image-enhanced mitochondrial fluorescence signals, and the mitochondrial fluorescence signals are image-enhanced using a rolling ball algorithm.

[0014] Preferably, the classification network of the senescent cell identification system based on mitochondrial morphology is based on an attention mechanism.

[0015] Preferably, in the senescent cell identification system based on mitochondrial morphology, the classification network is a ViT network.

[0016] According to another aspect of the present invention, a training method for the senescent cell recognition system based on mitochondrial morphology is provided, which uses the same steps as the input data for induced senescence and non-induced senescence to obtain mitochondrial distribution images in the cells at the subcellular level, and the sample labels are "young" and "senescent", respectively.

[0017] Preferably, the training method of the mitochondrial morphology-based senescent cell identification system is optimized using an adamw optimizer.

[0018] Preferably, the training method of the mitochondrial morphology-based senescent cell identification system uses a strategy of gradually reducing the learning rate.

[0019] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0020] The senescent cell identification system based on mitochondrial morphology provided by the present invention utilizes the mitochondrial morphological changes commonly exhibited by senescent cells, including changes in mitochondrial distribution, mitochondrial shape, and mitochondrial number. It is a method that can easily identify senescent cells through computer vision and has good robustness. It does not rely on the labeling effect of specific markers, but relies on changes in mitochondrial morphological characteristics to distinguish senescent cells. More importantly, compared with other senescent cell identification methods based on nuclear morphological characteristics, the present invention identifies senescent cells based on the common characteristics of senescent cells, and has good versatility for different cell types. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Comparison of mitochondria between young and old U2OS cells and young and old HeLa cells.

[0022] Figure 2 Schematic diagram of the structure of a senescent cell identification system based on mitochondrial morphology provided in an embodiment of the present invention.

[0023] Figure 3 These images demonstrate cell senescence using traditional biomarkers, as used in an example of the present invention. These images include a fluorescent image of the cell nucleus and a brightfield image stained with the SA-β-Gal marker. With the exception of the control group, all other senescence groups exhibited a deep blue SA-β-Gal marker, confirming cellular senescence.

[0024] Figure 4 Schematic diagram of a senescent cell recognition system training method based on mitochondrial morphology provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the following embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0026] The senescent cell recognition system based on mitochondrial morphology provided by the present invention is a classification network, preferably based on an attention mechanism, and specifically a ViT (Vision Transformer) network can be selected;

[0027] The input data is: a subcellular level mitochondrial distribution image in a cell, wherein the mitochondrial distribution image in the cell displays the shape and / or number of mitochondria distributed at different positions in the three-dimensional structure of the cell, preferably displays the shape and number of mitochondria distributed at different positions in the three-dimensional structure of the cell; mitochondria provide cells with the bioenergy required for cellular life activities and are widely present in various types of cells. It has been observed that the total amount of mitochondria in cells, cell distribution, mitochondrial morphology, and the relationship between mitochondria in different types of cells show similar changes in morphology, distribution, etc. as the cells age, such as Figure 1 As shown, the mitochondria of young cells exhibit a more complete morphology, with fragmented and dispersed distribution, while the mitochondria of senescent cells exhibit an enlarged morphology, with mitochondria interconnected in a network-like pattern. This pattern is similar in U2OS and HeLa cells. This similar pattern of changes in senescent cells provides a basis for universal senescent cell identification.

[0028] In a preferred embodiment, the mitochondrial distribution image in the cell is a slice sequence of a preset position of a three-dimensional image of the cell, which is obtained specifically according to the following method:

[0029] S1. Acquire a subcellular three-dimensional image of the cell to be identified; preferably, the mitochondrial fluorescence signal is captured in the 3D image of the cell; more conveniently, the 3D image is a stack of 2D images acquired using widefield imaging. In a preferred embodiment of the present invention, axial slices are extracted from the cell image, thereby reducing the requirement for axial resolution. Widefield imaging is used to acquire subcellular three-dimensional images, providing flexible imaging conditions and high speed, facilitating high-throughput senescent cell identification.

[0030] S2. For the three-dimensional cell image obtained in step S1, continuous transverse image slices are extracted at the upper, middle, and lower parts to obtain a slice sequence.

[0031] The diameter of mitochondria is generally between 0.5 and 10 microns, and the length is about 1 to 10 microns. The distribution and volume proportion of mitochondria in different types of cells vary greatly. The fluorescence signal in the image information other than mitochondria in the global 3D image of the cell is sparse, and it is impossible to effectively extract information such as mitochondrial morphology and distribution, and may even cause interference. Experiments have shown that the use of slice sequences for senescent cell identification has good accuracy, and can effectively reduce the amount of data that needs to be processed, thereby increasing the recognition speed, which is conducive to improving the recognition throughput.

[0032] Typically, the mitochondria in the cell mitochondrial image are labeled by fluorescent signals, and the resolution thereof is between 220nm and 280nm. In a preferred embodiment, the mitochondria in the cell mitochondrial image include image-enhanced mitochondrial fluorescence signals, and the mitochondrial fluorescence signals are image-enhanced using a rolling ball algorithm. The purpose of enhancing the mitochondrial fluorescence signal is to highlight the morphological characteristics of the mitochondria, which include length, width, area, curvature, shape complexity, and the like. Morphological characteristics not only directly reflect the activity of individual mitochondria, but also the distribution of mitochondria with different activities more accurately reflects the aging state of the cell. Therefore, accurate highlighting of morphological characteristics, accurate identification of mitochondrial composition in different states, and the relationship between different mitochondria will be automatically extracted in the classification network based on the attention mechanism. Based on these features, boundaries are established in the high-order parameter space, so that the senescent cell recognition system provided by the present invention can accurately identify senescent cells while having good generalization performance between different cell categories.

[0033] The training method of the senescent cell recognition system based on mitochondrial morphology provided by the present invention is:

[0034] It uses the same steps as the input data to obtain the mitochondrial distribution images in cells at the subcellular level for induced aging and non-induced aging, and the sample labels are "young" and "aging" respectively.

[0035] The adamw optimizer is used for optimization, and the preferred solution adopts a strategy of gradually reducing the learning rate to avoid overfitting and reduce the possibility of learning data noise, thereby improving the generalization ability of the model.

[0036] The following are examples:

[0037] This embodiment provides a senescent cell identification system based on mitochondrial morphology, such as Figure 2The figure shows the ViT (Vision Transformer) network, which segments the input image of mitochondrial distribution in cells into fixed-size image blocks (e.g., 32×32) and converts it into a sequence input using a linear embedding layer and a positional encoding module. The core of the network is a Transformer encoder composed of multiple stacked multi-head self-attention modules and feedforward neural network modules. This encoder efficiently extracts global features and captures long-range dependencies. Finally, the network extracts classification features and uses a multi-layer perceptron as the classification head for prediction.

[0038] The input data is: a subcellular level mitochondrial distribution image in a cell. The mitochondrial distribution image in the cell in this embodiment is a slice sequence of a preset position in a three-dimensional image of the cell, which is obtained specifically according to the following method, i.e., a "three-order two-layer" selection method for cell mitochondrial fluorescence images:

[0039] S1. Acquire a three-dimensional image of the cell to be identified at the subcellular level; wherein the three-dimensional image of the cell contains mitochondrial fluorescence signals;

[0040] In this embodiment, three-dimensional imaging is performed on cells stained with mitochondrial fluorescence to obtain three-dimensional images at the subcellular level of the cells to be identified. Specific fluorescence staining steps: preheat the culture medium to 37°C, prepare a culture medium with a final content of 100nM PK Mito DeepRed dye, aspirate the original culture medium of the cells, replace it with the diluted culture medium dye, incubate in the incubator for 15 minutes, and then wash the cells one to two times with preheated culture medium. Specific imaging steps: Use the Widefield mode of SuperVision's HiS-SIM microscope to perform mitochondrial fluorescence imaging of the cells, with an exposure time of 30ms, a step length of 0.5μm, a field of view of 1024×1024, and collect a 3D stack of mitochondrial fluorescence images of the cells. The mitochondrial fluorescence images collected by the wide-field fluorescence microscope are at the subcellular level, with a resolution of 100 times that of the wide-field microscope and a numerical aperture (NA) of 1.5.

[0041] S2. For the three-dimensional cell image obtained in step S1, continuous transverse image slices are extracted at the upper, middle, and lower parts to obtain a slice sequence.

[0042] The specific method in this embodiment is as follows:

[0043] For a stack of mitochondrial fluorescence 2D images of a cell, if the total number of layers containing mitochondrial information is N, let the top layer be numbered 1 and the bottom layer be numbered N. Six images are extracted from this stack as the network dataset, consisting of sequence numbers 1, 2, (N-1), N, and two consecutive images selected in the middle. Using only these six images can more completely describe the entire mitochondrial landscape of a cell without losing continuous mitochondrial information. This example images adherent cells, ignoring differences in cell axial position.

[0044] In the cell mitochondrial images described in this embodiment, mitochondria are labeled with fluorescent signals with a resolution of 220nm-280nm (widefield imaging); the mitochondrial fluorescence signals are enhanced using a rolling ball algorithm. The rolling ball algorithm was applied to six original images to produce mitochondrial fluorescence images with enhanced signal-to-background ratios. The rolling ball algorithm is based on the concept of a virtual sphere rolling across the image, smoothing background variations based on the radius of the sphere. By calculating the brightness of the portion of the sphere in contact with the image, a smoothed background image can be generated, which is then subtracted from the original image to produce an image with enhanced signal-to-background ratio.

[0045] An image segmentation network based on the Vision Transformer (ViT) architecture was constructed. Mitochondrial fluorescence images and manually delineated single-cell outline binary images were fed into the segmentation network as paired data. The network parameters were optimized until the in-library accuracy reached 98%. This trained single-cell segmentation network model was used to segment individual cell images. The single-cell segmentation mask, based on the layer with sequence number N, was then applied to five other images. Six segmented mitochondrial fluorescence images were obtained after masking.

[0046] The present invention provides a senescent cell recognition system based on mitochondrial morphology, and the training method is as follows: Figure 4 As shown in the figure, the cosine annealing strategy is used to regulate the learning rate during training, and the learning rate is dynamically adjusted to improve the convergence and stability of the model.

[0047] Obtain training data as follows:

[0048] 1. Reagents and Instruments

[0049] 1.1 Main reagents and materials

[0050] Etoposide (purchased from MedChemExpress), Doxorubicin (purchased from MedChemExpress), human osteosarcoma cell line (U2OS cell line), human cervical cancer cell line (HeLa cell line), cell senescence β-galactosidase staining kit (purchased from Beyotime Biotechnology Co., Ltd.), Hoechst 33342 (purchased from MedChemExpress)

[0051] 1.2 Main instruments

[0052] Biochemical incubator, cell culture incubator, biological safety cabinet, bright field microscope.

[0053] 2. Experimental Methods

[0054] 2.1 Induction of cell senescence

[0055] Cells in the logarithmic growth phase were trypsinized and transferred to a cell culture dish. After 24 hours of culture, drug-containing medium was added to induce senescence. The medium was replaced with the drug-containing medium every two days. Senescent cells were obtained after several days and used for subsequent microscopic imaging. Senescent samples were obtained using the same imaging and processing methods as the input images.

[0056] On the third day before obtaining fully senescent cells, cells were seeded from the same stock solution as a control group, and senescent samples were obtained following the same imaging and processing methods as the input images.

[0057] Different cell types were treated with different levels and durations of drug induction: 1) U2OS cells were induced with 2μM Etoposide, and Mccoy's 5A complete medium containing the drug was replaced every two days. Fully senescent U2OS cells were obtained after 10 days of drug addition. 2) U2OS cells were induced with 250nM Doxorubicin, and Mccoy's 5A complete medium containing the drug was replaced every two days. Fully senescent U2OS cells were obtained after 10 days of drug addition. 3) HeLa cells were induced with 3μM Etoposide, and DMEM complete medium containing the drug was replaced every two days. Fully senescent HeLa cells were obtained after 10 days of drug addition.

[0058] like Figure 3As shown, senescent cells were identified by staining with senescence-associated β-galactosidase (SA-β-Gal) (available from Beyotime Biotechnology Co., Ltd.) and cell morphology. Senescent cells typically appear dark blue after SA-β-Gal staining and have larger cell bodies. Cells labeled as "senescent" were grouped as the U2OS-Etoposide, U2OS-Doxorubicin, and HeLa-Etoposide groups. Cells not treated for senescence induction were labeled as "young" as the U2OS-Control and HeLa-Control groups.

[0059] The image acquisition and processing methods for all training data are the same as those for the input data.

[0060] The training data uses U2OS cell images, namely U2OS-Etoposide (aging) and HeLa-Control (young), and is trained according to the following method:

[0061] The training dataset was divided into 80% for training and 20% for validation. The network was trained for a total of 400 epochs, where an epoch refers to a complete forward and backward pass through all the data. Validation was performed every 10 epochs, and the model parameters with the highest accuracy on the validation set were retained. The loss function combined cross-entropy loss (CrossEntropyLoss) with supervised contrastive loss (Supervised Contrastive Loss). The former is used for classification tasks, while the latter enhances feature separability by aggregating features of samples of the same class and dispersing features of samples of different classes. During training, a cosine annealing strategy was used to regulate the learning rate. Every 200 epochs, the learning rate was gradually reduced from the initial value to 0.2 times and then increased again. This dynamic adjustment of the learning rate improved model convergence and stability. The resulting trained senescent cell recognition network model achieved 100% in-library accuracy. Using the same dataset and training method, a traditional CNN network achieved an accuracy of 75.56%. Accuracy is the percentage of true positives and true negatives predicted among all samples.

[0062] In order to verify the generalization performance of the model and its versatility for other cells, the trained ViT was used to identify samples from the U2OS-Doxorubicin group, Hela-Etoposide group, and HeLa-Control group, with accuracy rates of 96.36%, 97.73%, and 83.84%, respectively, showing good generalization and universal performance.

[0063] The recognition accuracy of the senescent cell recognition system using the "three-order two-layer" slice sequence:

[0064] The comparative example's senescent cell identification system differs only from the aforementioned embodiment in that the slice sequence is acquired using a continuous sampling method, which extracts ten consecutive image slices from the bottom of the cell image stack. A comparison with a senescent cell identification system using the aforementioned "three-step, two-layer" slice sequence as input and training data revealed that the three-step, two-layer method achieved 100% accuracy in identifying the U2OS-Etoposide group, while the continuous sampling method achieved 96.67% accuracy. This decrease in accuracy may be due to interference from irrelevant information introduced by the continuous sampling method.

[0065] The present invention provides a senescent cell identification system based on mitochondrial morphology, aiming to overcome the limitations of poor generalization of methods based on traditional biomarkers and nuclear morphology.

[0066] The method of the present invention comprises:

[0067] Step 1: Construct a senescence-induced model, covering multiple cell lines and multiple senescence-inducing methods, and perform mitochondrial fluorescence labeling respectively;

[0068] Step 2: Collect mitochondrial fluorescence images of young and senescent cells from the same cell line to train a deep learning network. By iteratively optimizing the loss function, a senescent cell recognition network is constructed.

[0069] Step 3: Identify senescent cells. After collecting mitochondrial fluorescence images of young and senescent cells under different aging models, input them into the senescent cell recognition network and obtain the recognition results.

[0070] Compared to traditional biomarker-based methods for identifying senescent cells, the present invention is more robust. The accuracy of traditional biomarker-based methods for identifying senescent cells is affected by various factors. For example, increased staining of the senescence-related enzyme β-galactosidase, a primary method for identifying senescent cells, increases in staining time and volume of staining fluid will result in increased staining in younger cells. Senescent cells often exhibit morphological changes, making them susceptible to analysis using computer vision methods. The present invention's method for identifying senescent cells based on mitochondrial morphology has the advantage of not relying on the labeling effect of specific markers, but instead relying on changes in mitochondrial morphological characteristics to distinguish senescent cells.

[0071] Compared with the method of identifying senescent cells using features such as nuclear morphology, the present invention has higher accuracy and generalization: the changes in nuclear morphology are relatively simple, mainly due to changes in the aspect ratio, perimeter, convexity, etc. of the nuclear outline, and the information dimension is low. Since the changes in the nuclear characteristics of senescent cells in different cell lines are not obvious, the recognition network constructed based on nuclear morphological features cannot be well generalized to senescent cells in other cell lines. The present invention identifies senescent cells based on the morphological features of mitochondria. Since mitochondria are directly related to cell aging, and mitochondria have different point-like and filamentous morphologies, as well as different features such as thickness, length, and curvature, its information dimension is wider, so the accuracy and generalization of identifying senescent cells based on mitochondrial morphological features are improved.

[0072] Compared to machine learning-based methods for identifying senescent cells, the present invention offers greater accuracy and generalizability. Traditional machine learning methods extract morphological features, converting image information into data such as perimeter and area, and then identify senescent cells based on models such as decision trees. These methods employ relatively simple logic and, therefore, offer lower accuracy. However, the present invention utilizes deep learning to construct a senescence recognition network, allowing for direct input of image information, extracting information at a higher level and depth. This allows for the identification of senescent cells from different cell lines and those induced by different mechanisms, with significantly improved accuracy.

[0073] This invention uses deep learning combined with mitochondrial morphology to accurately distinguish between the young and aging states of cells. It has high accuracy and strong generalization, is applicable to a variety of cell types and aging patterns, has universal applicability, and is easy to promote.

[0074] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A senescent cell identification system based on mitochondrial morphology, which is a classification network, characterized in that: The input data is: a subcellular level image of mitochondrial distribution in cells, wherein the image of mitochondrial distribution in cells displays the shape and / or quantity of mitochondria distributed in different positions in the three-dimensional structure of the cells, preferably displays the shape and quantity of mitochondria distributed in different positions in the three-dimensional structure of the cells.

2. The senescent cell identification system based on mitochondrial morphology according to claim 1, characterized in that: The mitochondrial distribution image in the cell is a slice sequence of a preset position of a three-dimensional image of the cell, and is specifically obtained according to the following method: S1. Obtain a three-dimensional image of the cell to be identified at the subcellular level; S2. For the three-dimensional cell image obtained in step S1, continuous transverse image slices are extracted at the upper, middle, and lower parts to obtain a slice sequence.

3. The senescent cell identification system based on mitochondrial morphology according to claim 2, wherein: Mitochondrial fluorescence signals are collected in the three-dimensional image of the cells.

4. The senescent cell identification system based on mitochondrial morphology according to claim 2, wherein: The three-dimensional image is a two-dimensional image stack obtained by wide-field imaging, and its resolution is 220nm-280nm.

5. The senescent cell identification system based on mitochondrial morphology according to claim 2, wherein: The mitochondrial distribution image in the cell is a slice sequence of a preset position of a three-dimensional image of the cell, and is specifically obtained according to the following method: For a stack of mitochondrial fluorescence two-dimensional images of a cell, if the total number of layers containing mitochondrial information is N, let the top layer be numbered 1 and the bottom layer be numbered N; 6 images are extracted from it as the network data set, which is composed of serial numbers 1, 2, (N-1), N, and two consecutive images selected in the middle.

6. The senescent cell identification system based on mitochondrial morphology according to claim 1, wherein: The mitochondria in the cell mitochondrial image include image-enhanced mitochondrial fluorescence signals, and the mitochondrial fluorescence signals are image-enhanced using a rolling ball algorithm.

7. The senescent cell identification system based on mitochondrial morphology according to claim 1, wherein: The classification network is based on an attention mechanism, and preferably the classification network is a ViT network.

8. The training method for a senescent cell recognition system based on mitochondrial morphology according to any one of claims 1 to 7, wherein: It uses the same steps as the input data to obtain the mitochondrial distribution images in cells at the subcellular level for induced senescence and non-induced senescence, and the sample labels are "young" and "senescent", respectively.

9. The training method for the senescent cell identification system based on mitochondrial morphology according to claim 8, characterized in that: The adamw optimizer is used for optimization.

10. The training method for the senescent cell identification system based on mitochondrial morphology according to claim 9, characterized in that: Use a strategy of gradually reducing the learning rate.

Images