Method and system for monitoring survival rate of magnetic labeled stem cells of articular cavity based on improved comparative learning

Abstract

The invention discloses a method and system for monitoring the survival rate of articular cavity magnetically labeled stem cells based on improved comparative learning, and relates to the technical field of biomedical engineering, medical image analysis and artificial intelligence crossing. The method comprises the following steps: acquiring multi-sequence magnetic resonance imaging data of a stem cell transplantation region loaded with magnetic nanoparticles at an articular cavity part; preprocessing the data, inputting the preprocessed data into a physical perception double-domain contrast learning network, and extracting features by the network through a spatial domain encoder and a frequency domain encoder which are parallel to each other; a physical consistency constraint based on a Bloch equation is innovatively introduced, a joint loss function is constructed, and the network is driven to distinguish microscopic magnetic susceptibility difference caused by intracellular iron aggregation and extracellular iron dispersion in learning; and finally, performing regression prediction by using the optimized features, and outputting the survival rate of the stem cells. According to the method, the dependence on a large amount of labeled data is reduced, and nondestructive and high-precision longitudinal monitoring on the survival state of the transplanted stem cells is realized.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of biomedical engineering, medical image analysis, and artificial intelligence, and more specifically to a method and system for monitoring the survival rate of intra-articular magnetically labeled stem cells based on improved contrastive learning. Background Technology

[0002] Osteoarthritis (OA) is a common degenerative joint disease that severely impacts patients' quality of life. Intra-articular injection of mesenchymal stem cells (MSCs) has become a promising treatment for OA due to its cartilage regeneration and immunomodulatory capabilities. However, the survival rate of stem cells in the joint microenvironment (such as hypoxia, inflammation, and mechanical shear stress) is often low, and their survival status directly determines the treatment outcome. Therefore, how to non-invasively, in real-time, and accurately monitor the survival of transplanted stem cells is a key technical bottleneck in clinical translation.

[0003] Currently, commonly used monitoring methods mainly include: 1. Invasive biopsy: Tissue is removed through arthroscopy or puncture for pathological analysis. This method is accurate but invasive, cannot provide continuous longitudinal monitoring, and is subject to sampling errors.

[0004] 2. Optical imaging (fluorescence / bioluminescence): High sensitivity, but limited penetration depth, making it difficult to apply to deep human joints (such as knee and hip joints), and the light signal decays severely over time.

[0005] 3. Magnetic Resonance Imaging (MRI) Tracing: Stem cells were labeled using superparamagnetic iron oxide nanoparticles (SPIOs). SPIOs served as a negative contrast agent at T2 / T2. Low signal (dark area) is generated on the weighted image. This is a highly promising non-destructive monitoring method.

[0006] However, existing MRI-based SPIO-labeled cell monitoring technologies face a significant challenge in distinguishing between live and dead cells: when labeled stem cells die, their cell membranes rupture, releasing intracellular SPIOs into the extracellular matrix or causing them to be engulfed by macrophages in the joint cavity. On conventional MRI images, whether it's a live labeled stem cell, a post-death SPIO, or a macrophage that has engulfed SPIOs, all will appear as low-signal areas. Traditional image processing methods or conventional supervised learning models (such as U-Net) primarily rely on pixel grayscale values ​​for segmentation, making it difficult to morphologically distinguish the subtle signal differences caused by this alteration in the microscopic magnetic environment, leading to overestimation or misjudgment of stem cell viability.

[0007] Furthermore, existing deep learning methods typically require a large amount of labeled data (i.e., they need to know the true live / dead ratio of each voxel), but this is almost impossible to obtain in live experiments, which limits the application of supervised learning methods.

[0008] Based on this, the present invention proposes a method and system for monitoring the survival rate of intra-articular magnetically labeled stem cells based on improved contrastive learning to solve the above problems. Summary of the Invention

[0009] To overcome the aforementioned deficiencies of the prior art, this invention provides a method and system for monitoring the survival rate of intra-articular magnetically labeled stem cells based on improved contrastive learning, in order to solve the problems existing in the background art.

[0010] This invention provides the following technical solution: a method for monitoring the survival rate of intra-articular magnetically labeled stem cells based on improved contrastive learning, comprising the following steps: S1. Acquire multi-sequence magnetic resonance imaging data of the subject's joint cavity, including T2-weighted imaging and T2-weighted imaging of the stem cell transplantation region loaded with magnetic nanomaterials. Weighted imaging and magnetic susceptibility-weighted imaging data; S2. The multi-sequence magnetic resonance imaging data is preprocessed, including motion artifact correction and magnetic field inhomogeneity correction, to obtain standardized three-dimensional voxel data; S3. Construct a physical perception dual-domain contrast learning network, the network including a parallel spatial domain encoder, a frequency domain encoder, and a projection head; S4. Input the standardized three-dimensional voxel data into the spatial domain encoder and the frequency domain encoder respectively, and extract the spatial domain feature vector and the frequency domain feature vector; S5. Based on the physical difference in transverse relaxation rate caused by the aggregated state of magnetic nanomaterials in living cells and the diffused state after death cells, a physical consistency constraint term is constructed, and combined with the multi-view contrast loss function, the physical perception dual-domain contrast learning network is trained to obtain an optimized feature representation with physical discriminative power. S6. Input the optimized feature representation into the survival rate regression prediction component and output the current survival rate value of the intra-articular stem cells.

[0011] As a further aspect of the present invention: In step S5, the total loss function of the physical perception dual-domain contrastive learning network... Defined as:

[0012] in, To estimate the loss for multi-view noise contrast, For physical consistency regularization loss, For frequency domain consistency loss, These are adjustable hyperparameter weighting coefficients.

[0013] As a further aspect of the present invention: the physical consistency regularization loss The calculation formula is:

[0014] in, This represents the number of training samples; The latent feature vector output by the encoder; This is a mapping function from features to physical parameters; and The first The actual observed apparent transverse relaxation rate and transverse relaxation rate of each sample voxel; It is the theoretical relaxation difference function based on the static dephase physical model.

[0015] As a further aspect of the present invention: the working process of the frequency domain encoder includes: S41. Perform a three-dimensional fast Fourier transform on the input standardized three-dimensional voxel data to convert it from the spatial domain to the frequency domain; S42. Decompose the frequency domain data into amplitude spectrum and phase spectrum, and apply a high-pass filter to the phase spectrum to enhance the high-frequency signal caused by the local magnetic susceptibility abrupt change caused by the aggregation of magnetic nanomaterials; S43. Adaptive weight allocation is performed on the filtered spectral features using a spectral attention mechanism; S44. Perform an inverse Fourier transform on the weighted spectral features to reconstruct them into spatial domain features, and fuse them with the output of the spatial domain encoder.

[0016] As a further aspect of the present invention: the survival rate regression prediction component adopts a self-attention aggregation module based on the Transformer architecture, and its prediction process is expressed as follows:

[0017] in, The fused multimodal feature vectors It is a multilayer perceptron. This is the predicted survival percentage for the output.

[0018] As a further aspect of the present invention: the magnetic nanomaterial is a superparamagnetic iron oxide nanoparticle with a surface modified by polylysine or polydopamine, and its particle size ranges from 10 nanometers to 200 nanometers; in the comparative learning training, the positive sample pairs are selected from at least one of the following: multiple sequence image pairs of the same region of interest at different echo times, and image pairs of the same region of interest generated by random data augmentation; the negative sample pairs are selected from at least one of the following: voxel pairs of different regions with a spatial distance exceeding a preset threshold, and reference sample pairs with significant differences in survival rate based on prior knowledge.

[0019] A joint cavity magnetic labeling stem cell survival monitoring system based on improved contrastive learning, characterized in that, for implementing the monitoring method as described in any one of claims 1 to 6, the system comprises: The data acquisition module is used to acquire multi-sequence imaging data of the joint cavity from the magnetic resonance imaging equipment; The preprocessing module is used to standardize, denoise, and correct the imaging data; The physical perception contrastive learning module, which incorporates the physical perception dual-domain contrastive learning network, is used to extract and optimize feature representations from preprocessed data. The survival rate analysis module is used to receive the optimized feature representation, obtain the stem cell survival rate through regression prediction, and generate a visualized survival rate distribution report.

[0020] An electronic device, characterized in that it comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the method as described in any one of the above descriptions.

[0021] A computer-readable storage medium, characterized in that the storage medium stores computer program instructions, which, when executed by a computer, cause the computer to perform the method as described in any one of the preceding descriptions.

[0022] The technical effects and advantages of this invention are as follows: This invention introduces a physical model of magnetic susceptibility based on the Bloch equation as a constraint into a contrastive learning framework, driving the network to learn the essential characteristics that distinguish between the aggregated state (living cells) and the diffuse state (dead cells / macrophages) of magnetic nanoparticles in cells. This fundamentally overcomes the problem of overestimation or misjudgment of survival rate caused by signal overlap in traditional MRI tracing methods. The core feature learning in this invention adopts a self-supervised contrastive learning paradigm, which does not require a large amount of hard-to-obtain voxel-level survival status annotation data. Only a small number of in vitro experimental calibration samples are needed to complete the training of the final regression model, which significantly reduces the data acquisition threshold and cost for clinical applications. This invention relies entirely on non-invasive magnetic resonance imaging data, which allows for multiple longitudinal scans of the same subject without interfering with the treatment process. This enables dynamic and quantitative assessment of stem cell survival over time, providing continuous and objective data support for efficacy evaluation and treatment plan optimization. The feature representations learned by the model in this invention are associated with explicit magnetic resonance physical parameters (such as differences in transverse relaxation rates), making the prediction results physically interpretable rather than simply "black box" outputs, thus increasing clinical trust. The dual-domain (spatial and frequency domain) feature extraction mechanism in this invention can capture signal characteristics more comprehensively and has better robustness to image noise and background tissue interference. This methodology is not limited to joint cavities and can be extended to magnetic labeling cells or nanomedicine monitoring scenarios in other parts of the brain, heart, etc. Attached Figure Description

[0023] The invention will now be further described with reference to the accompanying drawings.

[0024] Figure 1 A schematic diagram of the overall process of the non-destructive monitoring method for stem cell survival provided in the embodiments of this application; Figure 2 A detailed architecture diagram of the Physical Perception Dual-Domain Contrast Learning Network (PIDD-CL) provided in the embodiments of this application; Figure 3 This is a schematic diagram of the physical model of MRI signals corresponding to different distribution states of magnetic nanomaterials in living / dead cells in the embodiments of this application. Detailed Implementation

[0025] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. These embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0026] Example 1 This embodiment provides a complete implementation process for a monitoring method, such as... Figure 1 As shown, the specific steps include: Step S100: Multimodal MRI Data Acquisition and Preprocessing Image the joint cavity (e.g., knee joint) of mesenchymal stem cells labeled with superparamagnetic iron oxide nanoparticles using a clinical or research-grade MRI scanner (e.g., 3.0T or 7.0T). The following core sequences must be acquired: Multi-echo spin echo sequence: used to generate T2 mapping plots. Typical parameters are: TR=3000ms, TE=10,20,30,…,100ms.

[0027] Multi-echo gradient echo sequence: used to generate T2 Mapping plot. Typical parameters are: TR=500ms, TE=5,10,15,…,40ms, flip angle=20°.

[0028] High-resolution magnetic susceptibility weighted imaging sequences: used to provide phase and amplitude information sensitive to magnetic materials.

[0029] Data preprocessing is performed on computing devices (such as workstations), and the process is as follows: Format conversion and import: Convert raw DICOM data to NIfTI or NumPy array format.

[0030] Motion correction: All sequence images were rigidly registered using FSL’s MCFLIRT tool or ANTs registration tool to eliminate the effects of slight subject movements.

[0031] Magnetic field inhomogeneity correction: The N4 bias field correction algorithm (such as the SimpleITK library) is applied to eliminate low-frequency intensity inhomogeneities caused by coil sensitivity and other factors.

[0032] Image registration: T2 The map and SWI images are rigidly registered to the T2map space to ensure a one-to-one correspondence between voxels.

[0033] Standardization: The image intensity of each sequence is z-score standardized, or normalized to the [0,1] interval.

[0034] Region of Interest Extraction: A 3D ROI mask of the intra-articular injection area is drawn on T2-weighted images by a radiologist or using an automated segmentation algorithm (such as U-Net).

[0035] The final standardized set of multimodal 3D image patches is obtained:

[0036] Where B is the batch size, H, W, and D are the image patch sizes (e.g., 64x64x32), and C=3 is the number of modalities.

[0037] Step S200: Construct a physical perception dual-domain contrastive learning network Network architecture such as Figure 2 As shown, this is implemented using the PyTorch or TensorFlow framework. The specific construction is as follows: Spatial Domain Encoder : A 3D ResNet-18 is used as the backbone network, and its last fully connected layer is removed. The input is... ; The encoder outputs a spatial feature vector. .

[0038] Frequency domain encoder : FFT transform: for the input (and (Same) Perform a 3D Fast Fourier Transform:

[0039] Spectral decomposition: Calculating the amplitude spectrum and phase spectrum .

[0040] High-frequency enhancement: for phase spectrum A three-dimensional ideal high-pass filter is applied to filter out low-frequency phases associated with background tissue while retaining the localized, dramatic phase changes caused by SPIOs.

[0041] Spectrum Attention Module: This module processes the filtered spectrum... and Concatenate along the channel dimension to input a lightweight convolutional block. This block is first subjected to global average pooling. Then, it passes through two 1x1x1 3D convolutional layers (with ReLU activation in between) and a sigmoid function. Generate channel attention weights .

[0042] Weighted and inverse transform: applying attention weights to the original spectrum Multiply each channel and then perform an inverse Fourier transform (IFFT3D) to return to the spatial domain, resulting in a spatial feature map.

[0043] Feature extraction: The above feature maps are passed through a lightweight network consisting of 3 layers of 3D convolutions to output a frequency domain feature vector. ,in .

[0044] Projector head: Concatenate spatial and frequency domain features: .

[0045] Through a two-layer multilayer perceptron (MLP) projection head This yields low-dimensional feature vectors for contrastive learning. .

[0046] The MLP structure is as follows: .

[0047] Step S300: Model Training – Physical Perception Contrast Learning Pre-training This stage employs unsupervised / self-supervised methods to pre-train the encoder and projector.

[0048] Construct positive and negative sample pairs: Positive samples: For a voxel location within the same ROI, its value on T2map, T2 map, feature vectors in the three modes of SWI These are "positive sample pairs". At the same time, features extracted after random data augmentation (such as ±15° rotation, Gaussian noise addition, random small region masking) on ​​any modal image also constitute positive sample pairs with the features of the original image.

[0049] Negative samples: Voxel feature vectors from different spatial locations (Euclidean distance greater than a preset threshold, such as 10 pixels) within the same batch are automatically considered negative samples. Furthermore, known in vitro calibration data can be introduced: feature vectors from 100% live cell samples and 100% dead cell samples can be forced as hard negative sample pairs.

[0050] Define the loss function: The total loss function is:

[0051] in, .

[0052] Comparative loss Using the InfoNCE loss function, temperature parameters .

[0053] Physical consistency loss This is the core of the invention. First, the value of each voxel is calculated from the preprocessed data. Define a small physical subnetwork. (A three-layer MLP), the input is the feature vector. The output is the predicted relaxation ratio. .

[0054] The true ratio is .

[0055] Physical loss is defined as the mean squared error between the predicted and actual values, with an additional marginal loss term added to force a wider feature distance between known dead and live samples: in The edge value is m=0.5.

[0056] Regularization loss Apply L2 regularization to all network weights.

[0057] Training process: Using the Adam optimizer, with an initial learning rate of 1e-4 and a batch size of B=32.

[0058] Train for 200 epochs on a dataset containing a large amount of unlabeled intra-articular MRI data.

[0059] After training, retain the encoder. , and projector head The first layer discards the physical subnetwork. The final layer of the projection head yields a general feature extractor.

[0060] Step S400: Fine-tuning of survival rate regression prediction Prepare a fine-tuned dataset: Prepare an in vitro agarose gel model. Mix live / dead SPIO-labeled stem cells in known precise proportions (e.g., 100% live, 75% / 25%, 50% / 50%, 25% / 75%, 0% live) and perform MRI scans to obtain a small sample dataset (e.g., a total of 50 samples) labeled with the “gold standard” survival rate.

[0061] Constructing a regression network: Freeze the parameters of the pre-trained feature extractor. Follow this with a regression head, employing a single-layer Transformer encoder structure: First, the features... Survival rate predictions are obtained by mapping a linear layer, followed by layer normalization (LayerNorm) and a multi-head self-attention module with four attention heads, residual connections, and then passing the results through a two-layer MLP (256-dimensional hidden layer, 1-dimensional output layer, using sigmoid activation). .

[0062] Fine-tuning training: Train only the regression head. Use mean squared error loss. ,in This represents the true survival rate of in vitro samples. The SGD optimizer was used with a learning rate of 1e-3, trained for 50 epochs on a small dataset to prevent overfitting.

[0063] Step S500: Reasoning and Application For new intra-articular MRI data from patients or experimental animals, after the same preprocessing in step S100, the data is input into the trained complete pipeline (feature extractor + regression head). The network directly outputs the predicted average survival rate of stem cells within the ROI. Simultaneously, predictions can be made for each small image patch within the ROI, and the results can be mapped back to three-dimensional space to generate a "survival rate heatmap," visually displaying the distribution of surviving cells.

[0064] Example 2: Verification Experiment and Results To verify the effectiveness of the present invention, the following experiments were conducted: Animal model: A rat knee osteoarthritis model was established by intra-articular injection of SPIO-labeled bone marrow mesenchymal stem cells (n=10). A control group was also injected with an equal amount of inactivated dead cells (n=5).

[0065] Longitudinal imaging: 7.0T MRI scans were performed on days 1, 3, 7, 14, and 28 post-injection to acquire multimodal data.

[0066] Comparison method: Method A (Traditional): Based on T2 A semi-quantitative method for value thresholding.

[0067] Method B (Supervised Learning): A 3DCNN regression model is directly trained using the same in vitro data as in this invention.

[0068] Method C (the present invention): The method described in the present invention.

[0069] Gold standard: After each MRI scan, some animals were randomly sacrificed, and synovial tissue from the joints was collected for Prussian blue staining and cell viability fluorescence staining. The actual cell viability was calculated as the benchmark.

[0070] result: Quantitative analysis: The survival rate predicted by this invention (Method C) showed a Pearson correlation coefficient of 0.94 with the histological gold standard, and a root mean square error (RMSE) of 6.7%. The correlation coefficients of Methods A and B were 0.61 and 0.78, respectively, with RMSEs of 21.3% and 15.2%, respectively.

[0071] Specificity verification: On day 14 post-injection, in the dead cell control group, methods A and B produced significantly low signals due to macrophage phagocytosis of SPIOs, with misjudged survival rates of 45% and 38% on average, respectively; while the method of this invention accurately identified these signals as originating from extracellular iron, predicting a survival rate of 8% on average, which was highly consistent with the histological results (5%).

[0072] Feature visualization: t-SNE visualization shows that the features extracted by the pre-trained network of this invention result in clearly separated clusters of live cells and dead cells / macrophages in the embedding space, while the features of traditional CNNs are seriously overlapping.

[0073] The method of this invention is not limited to intra-articular stem cell monitoring, but is also applicable to: intracranial monitoring after stem cell transplantation for stroke; intracardiac survival monitoring after stem cell therapy for myocardial infarction; and drug release monitoring of tumor-targeting magnetic nanomedicines (drug release is often accompanied by the disintegration of nanoparticle structures, with similar physical processes).

[0074] System Implementation Those skilled in the art can develop and implement the described non-destructive stem cell survival monitoring system using Python, combined with the PyTorch deep learning framework and the SimpleITK / ANTs medical image processing library, based on the methods described above. The system modules include: a data I / O interface, a preprocessing pipeline, network model definition, training and inference scripts, and a results visualization module. This system can be deployed on a server or workstation equipped with a GPU for research or clinical auxiliary analysis.

[0075] In summary, this application, through an innovative physical perception dual-domain contrast learning algorithm, has for the first time achieved the precise quantification of the survival rate of magnetically labeled stem cells in vivo by utilizing the differences in the microscopic physical mechanisms behind MRI signals without requiring a large amount of labeled data. This solves a key pain point in regenerative medicine imaging assessment and has significant clinical application value.

[0076] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for monitoring the survival rate of intra-articular magnetically labeled stem cells based on improved contrastive learning, characterized in that, Includes the following steps: S1. Acquire multi-sequence magnetic resonance imaging data of the subject's joint cavity, including T2-weighted imaging and T2-weighted imaging of the stem cell transplantation region loaded with magnetic nanomaterials. Weighted imaging and magnetic susceptibility-weighted imaging data; S2. The multi-sequence magnetic resonance imaging data is preprocessed, including motion artifact correction and magnetic field inhomogeneity correction, to obtain standardized three-dimensional voxel data; S3. Construct a physical perception dual-domain contrast learning network, the network including a parallel spatial domain encoder, a frequency domain encoder, and a projection head; S4. Input the standardized three-dimensional voxel data into the spatial domain encoder and the frequency domain encoder respectively, and extract the spatial domain feature vector and the frequency domain feature vector; S5. Based on the physical difference in transverse relaxation rate caused by the aggregated state of magnetic nanomaterials in living cells and the diffused state after death cells, a physical consistency constraint term is constructed, and combined with the multi-view contrast loss function, the physical perception dual-domain contrast learning network is trained to obtain an optimized feature representation with physical discriminative power. S6. Input the optimized feature representation into the survival rate regression prediction component and output the current survival rate value of the intra-articular stem cells.

2. The method according to claim 1, characterized in that: In step S5, the total loss function of the physical perception dual-domain contrastive learning network is... Defined as: in, To estimate the loss for multi-view noise contrast, For physical consistency regularization loss, For frequency domain consistency loss, These are adjustable hyperparameter weighting coefficients.

3. The method according to claim 2, characterized in that: The physical consistency regularization loss The calculation formula is: in, This represents the number of training samples; The latent feature vector output by the encoder; This is a mapping function from features to physical parameters; and The first The actual observed apparent transverse relaxation rate and transverse relaxation rate of each sample voxel; It is the theoretical relaxation difference function based on the static dephase physical model.

4. The method according to claim 1, characterized in that: The workflow of the frequency domain encoder includes: S41. Perform a three-dimensional fast Fourier transform on the input standardized three-dimensional voxel data to convert it from the spatial domain to the frequency domain; S42. Decompose the frequency domain data into amplitude spectrum and phase spectrum, and apply a high-pass filter to the phase spectrum to enhance the high-frequency signal caused by the local magnetic susceptibility abrupt change caused by the aggregation of magnetic nanomaterials; S43. Adaptive weight allocation is performed on the filtered spectral features using a spectral attention mechanism; S44. Perform an inverse Fourier transform on the weighted spectral features to reconstruct them into spatial domain features, and fuse them with the output of the spatial domain encoder.

5. The method according to claim 1, characterized in that: The survival rate regression prediction component employs a self-attention aggregation module based on the Transformer architecture, and its prediction process is as follows: in, The fused multimodal feature vectors It is a multilayer perceptron. This is the predicted survival percentage for the output.

6. The method according to claim 1, characterized in that: The magnetic nanomaterial is a superparamagnetic iron oxide nanoparticle with a surface modified with polylysine or polydopamine, and its particle size ranges from 10 nanometers to 200 nanometers. In the contrastive learning training, the positive sample pairs are selected from at least one of the following: multiple sequence image pairs of the same region of interest at different echo times, and image pairs of the same region of interest generated by random data augmentation. The negative sample pairs are selected from at least one of the following: voxel pairs of different regions with a spatial distance exceeding a preset threshold, and reference sample pairs with significant differences in survival rate based on prior knowledge.

7. A joint cavity magnetic labeling stem cell survival monitoring system based on improved contrastive learning, characterized in that, For implementing the monitoring method as described in any one of claims 1 to 6, the system comprises: The data acquisition module is used to acquire multi-sequence imaging data of the joint cavity from the magnetic resonance imaging equipment; The preprocessing module is used to standardize, denoise, and correct the imaging data; The physical perception contrastive learning module has a built-in physical perception dual-domain contrastive learning network as described in claim 1, which is used to extract and optimize feature representations from preprocessed data; The survival rate analysis module is used to receive the optimized feature representation, obtain the stem cell survival rate through regression prediction, and generate a visualized survival rate distribution report.

8. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, which, when executed by the processor, implements the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The storage medium stores computer program instructions, which, when executed by a computer, cause the computer to perform the method as described in any one of claims 1 to 6.

Images