Multi-scale image feature extraction and prediction system for adipose-derived stem cell differentiation process

Abstract

The invention belongs to the technical field of biomechanics, and particularly relates to a multi-scale image feature extraction and prediction system for the adipose-derived stem cell differentiation process, which comprises the following steps: constructing a dynamic viscoelastic three-dimensional matrix for simulating a perivascular mechanical microenvironment, and constructing a multi-scale image feature extraction and prediction system for the adipose-derived stem cell differentiation process on the basis of super-resolution time sequence image data; by utilizing a double-channel deep neural network model of adipose-derived stem cells, taking a multi-scale morphological-mechanical characteristic set as input, and combining with chromatin conformation dynamic data, constructing a cross-scale prediction model taking matrix relaxation response as priori constraint; and outputting stage probability of differentiation from the adipose-derived stem cells to the skin fibroblasts, function maturity evaluation and an optimal mechanical intervention strategy. According to the invention, based on the unique perivascular source mechanical memory of the adipose-derived stem cells and the specific response of the adipose-derived stem cells to dynamic relaxation, accurate directional differentiation of the adipose-derived stem cells to the mechanical adaptive skin fibroblasts can be realized.

Description

Technical Field

[0001] This invention belongs to the field of biomechanics technology, specifically relating to a multi-scale image feature extraction and prediction system for the differentiation process of adipose-derived stem cells. Background Technology

[0002] Adipose-derived stem cells (ADSCs) are a type of mesenchymal stem cells derived from the perivascular microenvironment of adipose tissue. Due to their convenient acquisition, strong expansion capacity, low immunogenicity, and multi-lineage differentiation potential, they have become important seed cells in regenerative medicine. Recent studies have shown that ADSCs not only possess classic differentiation capabilities such as adipogenesis, osteoogenesis, and chondrogenesis, but can also efficiently transdifferentiate into skin fibroblasts, demonstrating great potential in chronic wound repair, scar regulation, and skin mechanical function reconstruction. Cell fate determination is not only regulated by biochemical factors, but also profoundly influenced by the mechanical signals of the microenvironment. Traditional research has focused on the effects of static matrix stiffness (such as elastic modulus) on stem cell differentiation. However, increasing evidence suggests that dynamic mechanical properties (such as stress relaxation and viscoelasticity) – especially parameters that simulate the time-dependent deformation behavior of tissues in vivo – play a decisive role in the differentiation guidance of ADSCs. For example, skin tissue exhibits significant stress relaxation behavior under physiological conditions (relaxation time constant of about 5–30 seconds), and ADSCs respond much more strongly to such dynamic signals than to static stiffness. Meanwhile, the development of super-resolution microscopy techniques (such as STED and SIM) has made it possible to observe cytoskeleton reconstruction, nuclear membrane deformation, and chromatin spatial conformation at the nanoscale; deep learning models (such as Swin-Unet and graph neural networks) have also been attempted for cell morphology analysis. However, existing technologies have not yet systematically integrated the unique biological identity of ADSCs, which originate from the biomechanical memory formed by the perivascular microenvironment and are highly sensitive to dynamic relaxation, into image feature definition, model architecture design, and prediction target construction.

[0003] Problems with existing technology: Although existing studies have attempted to combine mechanical stimulation with image analysis to predict stem cell differentiation, they generally suffer from five key drawbacks: First, they ignore the unique mechanosensory mechanisms of adipose-derived stem cells (ADSCs) arising from the low-stiffness, high-viscoelastic perivascular microenvironment, and still use regulatory strategies applicable to general mesenchymal stem cells, resulting in low differentiation efficiency and unstable phenotypes; second, the mechanical stimulation modes are singular, mostly limited to static matrix or single load types, making it difficult to reproduce the multimodal coupling of stretching-shearing-compression and time-dependent stress relaxation behavior experienced by the skin during movement or repair; third, existing deep learning models lack biological priors, and the extracted morphological features are disconnected from the ADSC-specific mechanical transduction pathways, making them neither interpretable nor able to guide intervention optimization; fourth, the predicted results are disconnected from mechanical regulation, lacking the ability to transform the differentiation state into executable mechanical dose instructions with uncertain boundaries, and lacking a closed-loop feedback mechanism to achieve dynamic correction of the differentiation trajectory; fifth, they fail to construct a cross-scale causal chain from mechanical stimulation to tissue mechanical adaptation, resulting in weak model generalization ability and difficulty in transferring to clinical scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-scale image feature extraction and prediction system for the differentiation process of adipose-derived stem cells. Based on the unique perivascular biomechanical memory of adipose-derived stem cells and their specific response to dynamic relaxation, this system can construct an intelligent system that integrates microenvironment simulation, multi-scale feature extraction, causal interpretable prediction, and closed-loop optimization, thereby achieving precise and directional differentiation of adipose-derived stem cells into biomechanically adapted skin fibroblasts.

[0005] The specific technical solution adopted by this invention is as follows: A multi-scale image feature extraction and prediction method for the differentiation process of adipose-derived stem cells includes the following steps: A dynamic viscoelastic three-dimensional matrix simulating the perivascular biomechanical microenvironment was constructed to specifically activate the endogenous biomechanical sensing pathways of adipose-derived stem cells. Multimodal mechanical loads were applied to adipose-derived stem cells on a dynamic viscoelastic three-dimensional matrix, and super-resolution temporal image data of the cells during differentiation were acquired simultaneously. Based on super-resolution temporal image data, a dual-pathway deep neural network model of adipose-derived stem cells was used to extract a set of multi-scale morphological-mechanical features that reflect their perivascular origin characteristics. By using a multi-scale morphological-mechanical feature set as input and combining it with dynamic chromatin conformation data, a cross-scale prediction model with matrix relaxation response as a prior constraint is constructed. The model outputs the stage probability of adipose-derived stem cells differentiating into skin fibroblasts, functional maturity assessment, and optimal mechanical intervention strategy.

[0006] According to another aspect of the present invention, the dual-path deep neural network model The first path encoder integrates a prior knowledge module of the perivascular microenvironment to enhance the sensitivity of feature extraction of the nuclear-cytoplasmic space compression ratio and the local protrusion morphology of the nuclear membrane; The second path decoder embeds a physical information constraint layer to ensure that the inverted cellular traction field strictly satisfies the constitutive response relationship of the dynamic viscoelastic matrix.

[0007] According to another aspect of the present invention, the loading frequency and amplitude of the multimodal mechanical load are synergistically modulated with the stress relaxation time constant of the dynamic viscoelastic three-dimensional matrix to match the time integration window of the mechanical signal for adipose-derived stem cells under physiological conditions, thereby maximizing their transcriptional activation efficiency toward the fibroblast phenotype.

[0008] According to another aspect of the present invention, the cross-scale prediction model generates a regulatory path with causal interpretability through an interpretable symbolic rule extractor, the path being characterized as follows: Matrix relaxation rate → YAP / TAZ nucleoplasmic translocation kinetics → COL1A1 promoter region chromatin openness → collagen secretion efficiency The quantitative mapping relationship is presented, along with confidence intervals based on Bayesian inference.

[0009] According to another aspect of the present invention, the multi-scale image feature extraction and prediction method for the differentiation process of adipose-derived stem cells further includes a quantitative characterization step of the mechanical activation kinetic fingerprint of adipose-derived stem cells: Based on the nuclear deformation response rate and myosin II activity dynamics in response to matrix relaxation stimulation within the initial 24 hours, a personalized mechanical response fingerprint was constructed for pre-screening of highly differentiated potential cell subpopulations.

[0010] According to another aspect of the present invention, the multi-scale image feature extraction and prediction method for the differentiation process of adipose-derived stem cells further includes a closed-loop regulation step: The optimal mechanical intervention strategy is transformed into real-time executable mechanical parameter instructions, which are fed back to the mechanical loading system. The prediction model weights are dynamically updated based on subsequently acquired image data to achieve online correction and stable maintenance of the differentiation trajectory.

[0011] According to another aspect of the present invention, a multi-scale image feature extraction and prediction system for the differentiation process of adipose-derived stem cells is also provided, comprising: The vascular perimechanical microenvironment simulation unit is configured to provide a three-dimensional culture substrate with controllable stress relaxation behavior and integrates a multi-axial, programmable mechanical loading device. The super-resolution in situ imaging unit is configured to perform sub-100-nanometer dynamic imaging of adipose-derived stem cells during mechanical stimulation. The ADSCs-specific feature extraction and prediction engine, with its built-in deep learning architecture based on the perivascular origin and dynamic relaxation sensitivity of adipose-derived stem cells as biological priors, is used to extract multi-scale morphomechanical features from imaging data and predict differentiation trajectories. The causal path visualization and human-machine collaborative decision-making unit is configured to transform prediction results into actionable mechanical intervention instructions with uncertainty boundaries, and support correction feedback from coaches or experimenters to dynamically optimize model decision weights.

[0012] According to another aspect of the present invention, the ADSCs-specific feature extraction and prediction engine is communicatively connected to a clinical skin tissue mechanics database to verify the elastic modulus, viscoelastic phase angle of differentiated fibroblasts and their mechanical adaptability to the host tissue, and thereby calibrate the optimal mechanical intervention strategy.

[0013] According to another aspect of the present invention, the vascular perimechanical microenvironment simulation unit adopts a photoresponsive or enzyme-responsive smart hydrogel, whose stress relaxation behavior can be programmed in real time by external light signals or biochemical signals to dynamically simulate the vascular perimechanical microenvironment under different physiological or pathological conditions.

[0014] According to another aspect of the present invention, the super-resolution in-situ imaging unit is coaxially integrated with the mechanical loading device, supporting STED or SIM super-resolution imaging while applying tensile, shear, or compressive loads, ensuring that the mechanical-morphological data are aligned in time and space.

[0015] According to another aspect of the present invention, an electronic device is also provided, the electronic device including a memory and a processor; the memory is used to store a program; the processor executes the program to implement the method described in any one of the foregoing.

[0016] According to another aspect of the present invention, a computer-readable storage medium is also provided, the storage medium storing a computer program that, when executed by a processor, implements the method described in any one of the preceding embodiments.

[0017] According to another aspect of the present invention, a computer program product is also provided, including a computer program that, when executed by a processor, implements the method described in any one of the preceding embodiments.

[0018] The technical effects achieved by this invention are as follows: This invention constructs a dynamic viscoelastic three-dimensional matrix simulating the perivascular biomechanical microenvironment and synergistically modulates multimodal mechanical loads (tension, shear, compression) and matrix stress relaxation time constants to match the physiological time integration window of adipose-derived stem cells (ADSCs) to mechanical signals. Simultaneously, using a dual-path deep neural network model customized for ADSCs, it extracts multi-scale morphomechanical features such as nuclear membrane local curvature evolution, actin orientation entropy, and traction gradient heterogeneity from super-resolution temporal images. By fusing dynamic chromatin conformation data, it constructs a cross-scale causal prediction model with matrix relaxation response as a priori. This achieves high-precision, interpretable prediction of the differentiation process of ADSCs into skin fibroblasts and can reverse-generate optimal mechanical intervention strategies with Bayesian confidence intervals, significantly improving differentiation efficiency and functional maturity. This solves the problems of regulatory blindness and phenotypic instability caused by neglecting the perivascular origin specificity of ADSCs.

[0019] This invention constructs a closed-loop, adaptive intelligent control system by integrating a vascular perimechanical microenvironment simulation unit, a super-resolution in situ imaging unit, an ADSCs-specific prediction engine, and a human-machine collaborative decision-making unit. It can not only pre-screen highly differentiated potential cell subpopulations based on individualized mechanical response fingerprints, but also acquire image data in real time during differentiation, dynamically update model weights, and correct mechanical parameters online to achieve stable guidance of the differentiation trajectory. Simultaneously, by linking with a clinical skin mechanics database, it verifies whether the elastic modulus and viscoelastic phase angle of the final product cells are compatible with the host tissue mechanics, and optimizes intervention strategies accordingly. This ensures that regenerated cells not only have the correct molecular phenotype but also possess biological functions compatible with host skin mechanics, providing a feasible, verifiable, and closed-loop optimized precision cell manufacturing paradigm for sports injury repair and individualized regenerative medicine. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0021] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0022] It should be noted that the terms "first," "second," etc., used in this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] According to embodiments of the present invention, a method embodiment for multi-scale image feature extraction and prediction of the differentiation process of adipose-derived stem cells is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0024] like Figure 1 As shown, the multi-scale image feature extraction and prediction method for the differentiation process of adipose-derived stem cells includes the following steps: S1. Construct a dynamic viscoelastic three-dimensional matrix that simulates the perivascular biomechanical microenvironment to specifically activate the endogenous biomechanical sensing pathways of adipose-derived stem cells. S2. Apply multimodal mechanical loads to adipose-derived stem cells on a dynamic viscoelastic three-dimensional matrix and simultaneously acquire super-resolution temporal image data of them during the differentiation process. S3. Based on super-resolution temporal image data, a dual-pathway deep neural network model of adipose-derived stem cells is used to extract a set of multi-scale morphological-mechanical features that reflect their perivascular origin characteristics. S4. Using a multi-scale morphological-mechanical feature set as input and combining it with chromatin conformation dynamic data, a cross-scale prediction model with matrix relaxation response as a prior constraint is constructed. The model outputs the stage probability of adipose-derived stem cells differentiating into skin fibroblasts, functional maturity assessment, and optimal mechanical intervention strategy.

[0025] According to step S1 above, in the human body, adipose-derived stem cells are located in the perivascular region of adipose tissue (i.e., the perivascular niche). This microenvironment has the mechanical characteristics of low elastic modulus, high viscoelasticity, and continuous dynamic remodeling, especially exhibiting significant stress relaxation behavior (i.e., under constant deformation, the internal stress of the material decays over time). This dynamic mechanical environment is fundamentally different from the static high-stiffness microenvironment derived from other mesenchymal stem cells such as bone marrow or umbilical cord.

[0026] Therefore, the mechanical response of a three-dimensional hydrogel matrix with controllable stress relaxation characteristics (such as a network structure based on dynamic covalent bonds or reversible non-covalent crosslinking) is determined not only by the initial stiffness but also by the relaxation time constant. When ADSCs are seeded in such a matrix, the cells sense the dynamic relaxation process of the matrix through mechanoreceptors such as integrins, triggering cytoskeleton reorganization and nuclear membrane deformation, thereby activating mechanotransduction pathways such as YAP / TAZ and PI3K / Akt.

[0027] Since ADSCs are pre-adapted to such dynamic signals during development, their response intensity is significantly higher than that of mesenchymal stem cells from other sources, thereby achieving selective activation of the endogenous mechanical sensing pathways of ADSCs, avoiding the blindness of general stiffness regulation strategies, and making differentiation induction more cell type specific.

[0028] In one specific embodiment, the dynamic viscoelastic three-dimensional matrix can be prepared using a photocrosslinkable hyaluronic acid-polyethylene glycol (HA-PEG) hybrid hydrogel system. For example, methacryloyl hyaluronic acid (MeHA) and multi-arm thiol-terminated polyethylene glycol (PEG-SH) are mixed in a certain molar ratio, a photoinitiator is added, and an initial crosslinking network is formed under ultraviolet light irradiation. By adjusting the degree of polymer substitution and molecular weight, the initial elastic modulus and stress relaxation time constant of the matrix can be independently controlled.

[0029] Furthermore, reversible dynamic bonds (such as imine bonds, borate ester bonds, or disulfide bonds) can be introduced into the network to give the material time-dependent stress relaxation behavior, that is, the internal stress decays nonlinearly with time under constant strain, and its evolution is determined by the viscoelastic constitutive relation of the material.

[0030] It should be noted that the specificity of this application refers to the fact that adipose-derived stem cells (ADSCs) exhibit a unique, preferential, and significantly stronger response to dynamic viscoelastic mechanical stimulation (especially stress relaxation behavior) compared to other types of mesenchymal stem cells (such as bone marrow MSCs, umbilical cord MSCs, etc.). This characteristic stems from their developmental experience of naturally being located in the perivascular microenvironment and is the basis for achieving precise regulation.

[0031] In step S2, skin tissue is subjected to multi-directional mechanical forces such as tension (e.g., limb movement), shear (e.g., clothing friction), and compression (e.g., sitting or lying pressure) under physiological conditions. Moreover, these loads often appear in a dynamic and coupled form. In order to realistically simulate this complex mechanical scenario, a multi-axial programmable mechanical loading device is integrated on the basis of a dynamic matrix to realize the independent or coordinated application of tensile, shear, and compressive loads.

[0032] Simultaneously, in-situ imaging of cells is performed using super-resolution microscopy. Super-resolution imaging breaks through the optical diffraction limit and can clearly distinguish the arrangement of cytoskeleton fibers, the nanoscale distribution of adhesion spots, and the subtle deformation of the nuclear membrane. These structures are the key mediators for the transmission of mechanical signals from the extracellular space to the cell nucleus. By continuously acquiring images at fixed time intervals, the dynamic evolution trajectory of cell morphology and mechanical state under mechanical stimulation can be obtained, providing a high-quality, high spatiotemporal resolution data foundation for feature extraction.

[0033] In step S3, image analysis methods typically extract macroscopic morphological parameters such as area and perimeter, making it difficult to capture the dynamics of subcellular structures directly related to mechanical transduction. However, the dual-path deep neural network model embeds the biological prior knowledge of ADSCs into the network architecture. The first approach focuses on semantic understanding of cell morphology. Through hierarchical feature extraction, it identifies morphological markers closely related to the perivascular origin, such as asymmetric compression of the nucleoplasmic space, abnormal local curvature of the nuclear membrane (reflecting uneven stress on the nuclear lamina), and dynamic deflection of the cell polar axis. These features are difficult to manifest in a static stiffness environment, but are significantly enhanced in a dynamically relaxed matrix due to the continuous adjustment of tension by the cells.

[0034] The second approach focuses on the physical inversion of mechanical states. By utilizing the spatial distribution of cell boundaries and adhesion spots, combined with the known viscoelastic constitutive relations of the matrix, and through a neural network layer constrained by physical information, the traction force field exerted by the cell on the matrix is ​​inverted. The above process not only outputs the magnitude of the force, but also reveals the spatial gradient and directional heterogeneity of the force, reflecting the organizational state of the tension network inside the cell.

[0035] The two paths mentioned above interact through a cross-scale feature fusion mechanism. The final output feature set can not only reflect the morphology of the cell, but also how the cell applies force and how the force affects the nuclear structure, thus forming a multi-dimensional characterization of the mechanical response state of ADSCs.

[0036] In step S4, a multi-scale morphological-mechanical feature set is used as input, combined with chromatin conformation dynamic data, to construct a cross-scale prediction model with matrix relaxation response as a prior constraint, and output the stage probability of adipose-derived stem cells differentiating into skin fibroblasts, functional maturity assessment, and optimal mechanical intervention strategy.

[0037] The optimal strategy refers to a multi-objective optimization solution that comprehensively considers differentiation efficiency, functional maturity, and host mechanical adaptability at the Pareto front.

[0038] Cell differentiation is essentially a reprogramming of gene expression programs, and mechanical signals can regulate gene accessibility by changing the spatial conformation of chromatin (such as stretching, compression or phase separation of chromatin regions). The aforementioned image-derived morphomechanical features are fused with chromatin conformation data across modalities.

[0039] Furthermore, the predictive model does not learn the relationship between input and output without constraints, but rather uses the matrix relaxation response as a prior constraint. That is, during training, the model is guided to prioritize feature combinations that are significantly related to the relaxation time constant. For example, the model learns that chromatin opening in the COL1A1 promoter region is only effectively triggered when the traction gradient and nuclear membrane curvature change occur synergistically within a specific time window. This constraint enables the model to have causal interpretability, rather than just outputting statistical correlation.

[0040] Therefore, the model can not only predict the probability of the current differentiation stage, but also assess functional maturity (such as collagen synthesis capacity and matrix remodeling activity), and recommend a set of mechanical parameters (such as relaxation rate and loading frequency) that maximize the target phenotype. Furthermore, this recommendation is not an isolated parameter, but a synergistic strategy that considers the coupling effect between parameters. For example, "moderate relaxation rate combined with low-frequency stretching" may produce a stronger differentiation-inducing effect than single-parameter optimization.

[0041] In summary, by transforming the unique perivascular origin of ADSCs into computable and embeddable prior knowledge, we have achieved full-chain technological synergy from microenvironment design to intelligent prediction, thus solving the technical defects of general stem cell differentiation regulation methods in ADSCs applications, such as low efficiency, unclear mechanisms, and unreliable regulation.

[0042] As an optional embodiment, the dual-path deep neural network model The first path encoder integrates a prior knowledge module of the perivascular microenvironment to enhance the sensitivity of feature extraction of the nuclear-cytoplasmic space compression ratio and the local protrusion morphology of the nuclear membrane; The second path decoder embeds a physical information constraint layer to ensure that the inverted cellular traction field strictly satisfies the constitutive response relationship of the dynamic viscoelastic matrix.

[0043] Furthermore, ADSCs are naturally located in the loose connective tissue around blood vessels in vivo. This microenvironment is characterized by low stiffness, high compliance, and continuous dynamic remodeling. In this environment, ADSCs are in a state of mild nuclear compression for a long time. Their nuclei often exhibit asymmetric deformation, and the nuclear membrane shows small blebs or depressions due to uneven tension of the cytoskeleton. These morphological features are not pathological manifestations, but rather adaptive phenotypes to the dynamic mechanical microenvironment, and can serve as sensitive indicators for identifying the mechanical activation state of ADSCs.

[0044] The encoder of the first path uses a prior knowledge module of the vascular microenvironment, which can be implemented in the following way: In the shallow convolution kernel initialization stage of the encoder, a filter with high response to high curvature edges (corresponding to nuclear membrane protrusions) and non-uniform density distribution (corresponding to nuclear-cytoplasmic compression asymmetry) is preset; Alternatively, in the feature map channel attention mechanism, channels related to the kernel region can be given higher weights to enhance the ability to model the spatial relationship between the nucleus and cytoplasm. Alternatively, through multi-task learning, in addition to the main task (differentiation stage classification), auxiliary tasks (such as nuclear membrane curvature regression and nuclear-to-cytoplasmic ratio estimation) can be added to guide the network to learn morphological semantics related to the perivascular origin.

[0045] The aforementioned effect is that the model no longer relies solely on the overall cell outline for judgment, but can actively identify and quantify the dynamics of subnuclear structures that are directly related to the mechanical memory of ADSCs. For example, when cells are cultured on a dynamically relaxed substrate, the frequency of local protrusions in the nuclear membrane increases, and the model can infer that the cell is undergoing effective mechanical activation, thereby improving the predictive sensitivity of early differentiation events.

[0046] Furthermore, cellular traction force is a mechanical quantity of the interaction between cells and the matrix. Its inversion usually relies on methods such as traction force microscopy (TFM), which requires the assumption that the matrix is ​​a linear elastic material. However, the matrix used is a dynamic viscoelastic material, whose stress is not only proportional to strain but also related to strain history, and obeys constitutive relations such as the generalized Maxwell model or the standard linear solid model.

[0047] A Physics-Informed Constraint Layer is embedded in the decoder of the second path. This layer is not an independent network module, but is integrated into the loss function or network output layer in the form of soft or hard constraints. Soft constraint method: Add a constitutive residual loss term to the training loss function, that is, calculate whether the inverted traction field and the matrix displacement field satisfy the known viscoelastic constitutive equation. If they do not satisfy, a penalty is applied. Hard constraint method: After the network outputs the traction force, the force field is forcibly corrected by a differentiable constitutive operator (such as an analytical / numerical solver based on finite element or Green's function) to ensure that it is physically consistent with the input displacement field.

[0048] It can ensure that the inversion results are not only reasonable in the image pixel space, but also valid within the framework of continuous medium mechanics. The effect is that the obtained traction force field can truly reflect the mechanical behavior of ADSCs in the dynamic matrix, such as capturing the dynamic process of cells applying high traction force in the early stage of matrix relaxation and then adjusting the force output as the matrix softens.

[0049] Furthermore, the two can be collaboratively embedded into the same dual-path architecture, and information complementarity can be achieved through cross-path feature interaction, for example: The localized protrusions of the nuclear membrane identified by the first pathway can serve as a marker of increased intracellular tension. The high-gradient traction field retrieved through the second path provides evidence of external mechanical input. After fusing the features of the two pathways, the model can more reliably determine whether the cell is in a state of effective mechanical activation, rather than a false positive signal caused by noise or non-specific deformation.

[0050] Based on the above, the core of being able to be directly customized for ADSCs is that if the same model is used for bone marrow MSCs, the latter lacks sensitivity to dynamic relaxation, and its nuclear morphology and traction response pattern are different, so the model performance will be significantly reduced, thus verifying the specific adaptation ability of this embodiment to ADSCs.

[0051] As an alternative embodiment, the loading frequency and amplitude of the multimodal mechanical load are synergistically modulated with the stress relaxation time constant of the dynamic viscoelastic three-dimensional matrix to match the time integration window of the mechanical signal for adipose-derived stem cells under physiological conditions, thereby maximizing their transcriptional activation efficiency toward the fibroblast phenotype.

[0052] The biological principle behind the time integration window for mechanical signals is as follows: The process of accumulating and integrating mechanical signals over time through the cytoskeleton-nuclear lamina-chromatin network is defined as mechanical time integration, specifically: When the matrix undergoes deformation or relaxation, the integrin-mediated focal adhesion complex is the first to sense the external force, triggering a phosphorylation cascade of signaling molecules such as FAK and RhoA. These signals are transmitted to the cell nucleus via actomyosin contraction, causing nuclear membrane deformation and changes in nuclear lamina tension; Changes in nuclear mechanics further regulate the spatial conformation of chromatin (such as stretching specific gene promoter regions), ultimately affecting the nucleoplasmic distribution of transcription factors and their DNA binding ability.

[0053] When the duration or rate of dynamic change of mechanical stimulation falls within a certain time scale (i.e., the time integration window), sufficient signal intensity can be effectively accumulated to cross the transcriptional activation threshold. For ADSCs, due to their long-term adaptation to the slowly remodeled mechanical environment in the perivascular microenvironment, their time integration window usually corresponds to the order of several seconds to tens of seconds—which happens to match the stress relaxation time constant of skin tissue (τ≈5–30s).

[0054] In this embodiment, the stress relaxation time constant (τ) of the dynamic viscoelastic three-dimensional matrix constitutes a programmable mechanical input system: The matrix relaxation time constant τ determines how quickly the cell's microenvironment softens, directly affecting the dynamic adjustment rhythm of cell traction force. The loading frequency f determines the repetition rate of external periodic stimuli (such as stretching); The loading amplitude A determines the intensity of each stimulus.

[0055] The time-scale coupling affects the integral effect of cells. For example: If the loading frequency is too high (e.g., f is greater than 1 Hz), the interval between single stimulations is shorter than the time required for cytoskeleton remodeling, resulting in the inability of the signal to accumulate effectively. If the matrix relaxes too quickly (τ is too small), the cells will face matrix softening before they can establish stable adhesion, making it difficult to form sustained tension. Conversely, if the relaxation is too slow (τ is too large) or the loading frequency is too low, the signal strength will be insufficient to trigger the downstream transcription process.

[0056] Therefore, this embodiment uses coordinated modulation of f, A, and τ to match the dynamic characteristics of external mechanical input (such as stress change rate and strain energy release rate) with the time constant of ADSCs' endogenous signal transduction, thereby ensuring that the mechanical signal is effectively integrated within the correct time window and maximizing the YAP / TAZ nuclear translocation efficiency and the promoter opening probability of fibroblast marker genes such as COL1A1.

[0057] Aligning the intrinsic mechanosensory dynamics of ADSCs with the extrinsic mechanoinput system in the time dimension enables precise temporal programming of cell fate, facilitating the efficient and stable acquisition of mechanofitted fibroblasts.

[0058] As an optional embodiment, the cross-scale prediction model generates a regulatory path with causal interpretability through an interpretable symbolic rule extractor, the path being characterized as follows: The quantitative mapping relationship between matrix relaxation rate → YAP / TAZ nucleoplasmic translocation kinetics → chromatin openness of COL1A1 promoter region → collagen secretion efficiency is presented, along with confidence intervals based on Bayesian inference.

[0059] By using an interpretable symbol rule extractor, explicit regulatory rules that conform to biological logic are extracted from data-driven models, thereby achieving a closed loop in the prediction, interpretation, and intervention process.

[0060] The extractor can be implemented based on any or a combination of the following techniques: Symbolic regression: Searching for concise mathematical expressions in a high-dimensional feature space through genetic programming or sparse regression, for example, discovering that the YAP nucleocytoplasmic ratio is ∝log(relaxation rate) / (1+τ); Causal graph learning: Using conditional independence tests (such as PC algorithms) or structural equation modeling, it infers directed dependencies between variables from time-series multi-omics data; Attention mechanism visualization and rule induction: Analyze the contribution weight of each feature to the output in the deep model, and combine domain knowledge to summarize if-then rules such as when the relaxation rate is greater than the threshold and the rate of change of nuclear membrane curvature is greater than the threshold, the chromatin opening probability increases significantly.

[0061] Regardless of the technology used, the output is a human-readable symbolic rule, rather than neural network weights or feature vectors.

[0062] Furthermore, the matrix relaxation rate, as an external mechanical input, determines the dynamic rate of change of cellular traction force; YAP / TAZ nucleocytoplasmic translocation kinetics: YAP / TAZ is a mechanically sensitive transcriptional coactivator, and its nuclear retention time and intensity are regulated by cell tension. Among them, YAP / TAZ is the main but not the only effector pathway. COL1A1 promoter region chromatin openness: After YAP / TAZ enters the nucleus, it binds to transcription factors such as TEAD, recruits chromatin remodeling complexes (such as SWI / SNF), and opens the promoter region of fibroblast genes such as COL1A1. Collagen secretion efficiency: After chromatin is opened, RNA polymerase can bind and initiate transcription, which is eventually translated into collagen and secreted into the extracellular matrix.

[0063] The path representation includes four aspects: mechanics, cellularity, epigeneity, and function, forming a complete causal chain. The model learns the quantitative mapping between each link (e.g., for every unit increase in relaxation rate, the YAP nucleocytoplasmic ratio increases by α, which in turn increases the COL1A1 opening probability by β), thus achieving a mechanistic explanation from input to output.

[0064] Furthermore, due to individual differences, measurement noise, and model uncertainty in biological systems, any prediction should be accompanied by uncertainty quantification, by using a Bayesian inference framework (such as Bayesian neural networks, Gaussian processes, or Monte Carlo) to calculate confidence intervals for each rule or mapping relationship.

[0065] Confidence intervals can reflect the degree of certainty between the current data and the model regarding causal relationships, thus avoiding the over-interpretation of weak evidence; In addition, it provides decision-makers with risk boundaries. For example, when the confidence interval is too wide, the system can suggest supplementing experiments or reducing the intensity of intervention to improve clinical safety.

[0066] Furthermore, by combining symbolic causal rule extraction with cross-scale biological pathways and embedding it into the ADSCs differentiation prediction model, causal pathways that conform to known biological logic can be automatically discovered and quantified in a data-driven manner.

[0067] Furthermore, the pathway is customized for the response mechanism of ADSCs to dynamic relaxation. If applied to other cell types (such as bone marrow MSCs), the extracted rules will be significantly different due to the different YAP response patterns, thus reflecting the cell type specificity in this embodiment.

[0068] As an optional embodiment, the multi-scale image feature extraction and prediction method for the differentiation process of adipose-derived stem cells also includes a quantitative characterization step of the mechanical activation kinetic fingerprint of adipose-derived stem cells: Based on the nuclear deformation response rate and myosin II activity dynamics in response to matrix relaxation stimulation within the initial 24 hours, a personalized mechanical response fingerprint was constructed for pre-screening of highly differentiated potential cell subpopulations.

[0069] Adipose-derived stem cells (ADSCs) naturally locate in the perivascular microenvironment of adipose tissue in vivo. This region is characterized by low stiffness, high viscoelasticity, and continuous dynamic remodeling. ADSCs long-term exposure to this microenvironment develop a preferential response to dynamic relaxation-type mechanical signals at both the epigenetic and cytoskeleton levels. This ability can be considered a mechanical activation kinetic fingerprint. This imprint is not permanent but is highly preserved in the early stages of in vitro culture (e.g., the first 24 hours), manifesting as follows: Rapid sensing of matrix stress relaxation; Dynamic adjustment of the actomyosin contractile system; The nuclear membrane and nuclear lamina are sensitive to changes in tension.

[0070] Furthermore, even among ADSC populations derived from the same donor, there is significant functional heterogeneity—some cells retain strong mechanical memory and respond rapidly to dynamic stimuli; while others forget their original mechanical identity due to in vitro expansion or microenvironmental disturbances and respond sluggishly.

[0071] Furthermore, two key dynamic indicators were monitored within the first 24 hours after cell seeding into the dynamic viscoelastic matrix: Nuclear deformation response rate: refers to the speed at which the cell nucleus undergoes shape changes (such as shortening of the long axis and increase of curvature) during matrix relaxation. This indicator reflects the mechanical coupling efficiency between the nuclear lamina and the cytoskeleton and is a key intermediate phenotype for the transmission of mechanical signals from the cell membrane to the nuclear chromatin. Dynamics of myosin II activity: Myosin II is a core motor protein for active cell contraction. Temporal changes in its phosphorylation level or fluorescent labeling intensity can characterize the cell's active tension adjustment strategy in response to matrix softening.

[0072] The basis for the above indicators is: They are located at key upstream nodes in the mechanotransduction pathway, earlier than changes in gene expression, and can be used for early prediction; Their dynamic changes have time-resolved characteristics (such as response delay, peak time, and decay rate), which reflect the intrinsic state of cells better than static snapshots. Together, they form a closed loop of perception and response: nuclear deformation reflects passive transmission, while myosin activity reflects active regulation. Their combined use can comprehensively characterize the cell's mechanical adaptability.

[0073] Based on the time-series trajectories of the two dynamic indicators mentioned above, several quantitative parameters are extracted through feature engineering, such as: The time required for nuclear deformation to reach 50% of its maximum amplitude; The time difference between the peak activity of myosin II and the onset of matrix relaxation; Phase hysteresis angle between nuclear deformation and myosin activity; The cross-correlation coefficient or dynamic coupling entropy of the two change curves.

[0074] These parameters are combined into a low-dimensional vector, which constitutes the individualized mechanical response fingerprint of the cell (or cell population). The fingerprint is a quantitative characterization of the degree to which ADSCs retain their original perivascular mechanical identity and can serve as an early biomarker of their differentiation potential.

[0075] Before formal differentiation is induced, the cell population is sorted or grouped based on this fingerprint: Cells whose fingerprint features fall within a high-response range (such as rapid nucleus transformation and strong myosin adjustment synchronicity) are identified as a highly differentiated potential subpopulation. Conversely, it is considered a low-response or mechanical amnesia subgroup.

[0076] The effect of the pre-screening mechanism is: Improve overall differentiation efficiency: Avoid including low-response cells in the induction system to reduce ineffective culture and resource waste; Enhance batch consistency: Reduce functional variability of the final product by screening for a more homogeneous starting cell population; Support for personalized treatment: In clinical applications, the regenerative potential of ADSCs from patients can be predicted based on their fingerprint characteristics, and differentiated culture strategies can be developed.

[0077] This embodiment uses the dynamic coupling relationship between nuclear deformation and myosin activity as the core criterion, rather than a single indicator. This multi-parameter time-series synergistic analysis strategy can more accurately reflect the intrinsic mechanical state of ADSCs.

[0078] As an optional embodiment, the multi-scale image feature extraction and prediction method for the differentiation process of adipose-derived stem cells also includes a closed-loop regulation step: The optimal mechanical intervention strategy is transformed into real-time executable mechanical parameter instructions, which are fed back to the mechanical loading system. The prediction model weights are dynamically updated based on subsequently acquired image data to achieve online correction and stable maintenance of the differentiation trajectory.

[0079] The closed-loop system in this embodiment consists of four functional modules: Sensing module: continuously acquires images of the morphology and mechanical state of ADSCs under the current mechanical environment through a super-resolution imaging unit; Prediction module: Based on the current image data, it uses a cross-scale prediction model to output an assessment of the current differentiation status and the optimal mechanical intervention strategy for the next stage; Execution module: Transforms the strategy into specific mechanical parameter instructions (such as setting the stretching frequency to 1.5Hz and the matrix relaxation time constant τ to 12s) and sends them to the multimodal mechanical loading system; Feedback and Learning Module: After implementing new parameters, it collects subsequent image data, evaluates whether the actual response meets expectations, and dynamically adjusts the internal parameters of the prediction model accordingly.

[0080] The above modules are used to mimic the homeostatic regulatory mechanisms in organisms (such as negative feedback regulation) to enable in vitro differentiation systems to resist environmental disturbances (such as cell batch differences and fluctuations in culture conditions) and adapt to individualization.

[0081] The ADSC population may exhibit heterogeneous responses, or slight shifts in external environmental factors (such as temperature and pH) may cause the actual differentiation trajectory to deviate from the model prediction. Therefore, the following correction procedure is performed in each regulation cycle (e.g., every 24 hours): S101. Input the newly acquired image into the prediction model to obtain an updated evaluation of the current state; S102. Calculate the deviation between the predicted state and the target state (e.g., the fiber maturity is 10% lower than expected). S103. Based on the direction and magnitude of the deviation, invoke a reinforcement learning strategy or a gradient-based optimizer to fine-tune the mechanical parameters of the next cycle (such as slightly increasing the relaxation rate or extending the loading time). S104. If the deviation persists, the model weight update mechanism will be triggered.

[0082] To prevent models from becoming obsolete due to long-term use, incremental learning or online fine-tuning strategies are employed. Retain the characteristics and outcomes of historical success stories as a memory bank; Calculate the prediction error for the newly collected data; If the error exceeds the preset threshold, the sample will be added to the training set for lightweight fine-tuning of the model. The update process can be completed in real time on edge computing devices without interrupting the cultivation process.

[0083] This mechanism enables the model to continuously evolve with the cell state, adapting to subtle differences in ADSCs from different donor sources and generations, and achieving personalized regulation.

[0084] The effects of closed-loop regulation include: Improved differentiation stability: Even if the initial cell state fluctuates, the system can approximate the target phenotype through multiple rounds of correction; Reduced human intervention: Eliminates the need for researchers to frequently observe and manually adjust parameters, enabling unattended intelligent cultivation; Supports complex trajectory tracking: It can not only maintain homeostatic differentiation, but also guide cells to evolve along a preset non-homeostatic path (such as first inducing myofibroblast phenotype, and then regressing to the resting fibroblast state). Accumulate high-quality data: The policy-response pairs generated during the closed-loop process can feed back into model training, forming a positive reinforcement loop.

[0085] This embodiment treats the prediction model itself as an updatable control component, rather than a fixed rule base, enabling it to continuously learn and adapt.

[0086] As an optional embodiment, a multi-scale image feature extraction and prediction system for the differentiation process of adipose-derived stem cells includes: The vascular perimechanical microenvironment simulation unit is configured to provide a three-dimensional culture substrate with controllable stress relaxation behavior and integrates a multi-axial, programmable mechanical loading device. The super-resolution in situ imaging unit is configured to perform sub-100-nanometer dynamic imaging of adipose-derived stem cells during mechanical stimulation. The ADSCs-specific feature extraction and prediction engine, with its built-in deep learning architecture based on the perivascular origin and dynamic relaxation sensitivity of adipose-derived stem cells as biological priors, is used to extract multi-scale morphomechanical features from imaging data and predict differentiation trajectories. The causal path visualization and human-machine collaborative decision-making unit is configured to transform prediction results into actionable mechanical intervention instructions with uncertainty boundaries, and support correction feedback from coaches or experimenters to dynamically optimize model decision weights.

[0087] The vascular perimechanical microenvironment simulation unit is the system's physical input layer, responsible for constructing a highly biomimetic mechanical microenvironment, including: Dynamic viscoelastic matrix generation module: It adopts a programmable hydrogel system (such as a network based on dynamic covalent bonds, host-guest interactions or enzyme response crosslinking) to achieve independent control of the stress relaxation time constant (τ) by adjusting the crosslinking density and bond exchange kinetics, covering the physiological range of skin tissue (approximately 5–30 seconds). Multi-axial mechanical loading device: integrates a micro linear actuator, a rotary shearing platform and a piezoelectric actuator, which can apply tensile, shear and compressive loads independently or in combination in three-dimensional space. The loading waveform (sine wave, square wave, random), frequency and amplitude can be programmed by software. Environmental coupling control: placed in a culture chamber with constant temperature, constant humidity, and 5% CO2 to ensure that mechanical intervention is compatible with the physiological state of cells.

[0088] Furthermore, the super-resolution in situ imaging unit is the system's sensing layer, responsible for acquiring high-fidelity dynamic images of cells without interfering with mechanical stimuli, including: Coaxial integration: The imaging optical path and the mechanical loading device are strictly aligned in space to ensure that the cell is always located in the center of the imaging field of view during the force process, avoiding data mismatch caused by displacement; Sub-100 nanometer resolution: Using stimulated emission depletion microscopy (STED) or structured light illumination microscopy (SIM), the optical diffraction limit is broken, and subcellular structures such as F-actin fiber arrangement, adhesion spot nanoclusters, and nuclear membrane curvature are clearly distinguished. Timing-synchronized acquisition: Through hardware trigger signals, the mechanical load waveform and image frame acquisition are synchronized at the microsecond level, ensuring that each frame of the image corresponds to the precise mechanical state; Multi-channel fluorescent labeling compatibility: Supports simultaneous labeling of the cytoskeleton, nucleus, and mechanosensitive proteins, providing multi-dimensional input for multi-scale feature extraction.

[0089] Furthermore, the ADSC-specific feature extraction and prediction engine serves as the intelligent center of the system, embedding the biological identity of ADSCs as a priori into the deep learning architecture: Customized network architecture: A dual-path neural network is adopted. The first path enhances the sensitivity to perivascular morphology such as nuclear-cytoplasmic compression asymmetry and local protrusions of the nuclear membrane. The second path embeds the matrix constitutive equation as a physical constraint to ensure that the traction force inversion conforms to the laws of viscoelastic mechanics. Cross-scale fusion mechanism: Image-derived morphological-mechanical features are fused with externally input chromatin accessibility data in a graph neural network to construct a mapping from mechanics to appearance to function; Online learning capability: Incremental fine-tuning based on newly acquired data to prevent the model from failing due to differences in cell batches; Edge deployment optimization: The model is lightweighted (e.g., knowledge distillation, quantization compression) and can run in real time on embedded GPUs or FPGAs, meeting the latency requirements of closed-loop control.

[0090] Furthermore, the causal path visualization and human-machine collaborative decision-making unit is the system's interaction and decision-making output layer, used to bridge the gap between AI prediction and human expert judgment.

[0091] As an optional embodiment, the ADSCs-specific feature extraction and prediction engine communicates with a clinical skin tissue mechanics database to verify the elastic modulus, viscoelastic phase angle of differentiated fibroblasts and their mechanical adaptability to the host tissue, and calibrates the optimal mechanical intervention strategy accordingly.

[0092] This clinical skin biomechanics database can be sourced from collaborating medical institutions (such as dermatology departments or burn centers in tertiary hospitals), and contains a large number of skin biomechanical parameters from healthy individuals and patients with different pathological conditions (such as diabetes, keloid scarring, and senile skin atrophy), including: Elastic modulus: reflects the stiffness of an organism; Viscoelastic phase angle (Loss angle, δ): Characterizes the energy dissipation capacity, i.e. the softness, stiffness and damping properties of a tissue under dynamic loading; Stress relaxation time constant (τ): describes the rate at which stress decays in a tissue under constant strain.

[0093] After ADSCs complete differentiation, the system measures the mechanical properties of the resulting fibroblasts and their secreted extracellular matrix (ECM) using atomic force microscopy (AFM) or microcolumn array traction microscopy (μTFM) and uploads the results to the prediction engine. The engine compares the measured values ​​with the skin mechanical distribution of the corresponding patient groups (such as age, gender, and medical history matching) in the database and calculates the mechanical fit error (such as modulus deviation percentage and phase angle difference).

[0094] If the error exceeds the preset threshold (e.g., greater than 10%), the system will automatically backtrack the mechanical intervention parameters during the differentiation process and calibrate the optimal mechanical intervention strategy for the next round of experiments (e.g., adjusting the matrix τ or loading frequency) through backpropagation or Bayesian optimization, so that the newly generated cells are closer to the target mechanical phenotype.

[0095] By introducing clinical mechanical data as feedback signals into the in vitro differentiation regulation loop, bidirectional calibration is achieved. If the mechanical properties of the regenerated skin do not match those of the host (e.g., too rigid or too soft), functional impairment (e.g., limited mobility, repeated damage) may still occur even if the cell phenotype is correct. Therefore, this connection is not only a data docking, but also a paradigm shift in in vitro manufacturing driven by clinical functional needs.

[0096] As an optional embodiment, the vascular perimechanical microenvironment simulation unit uses a light-responsive or enzyme-responsive smart hydrogel, whose stress relaxation behavior can be programmed in real time by external light signals or biochemical signals to dynamically simulate the vascular perimechanical microenvironment under different physiological or pathological conditions.

[0097] Specific implementation methods include: Photoresponsive hydrogels: Photosensitive groups (such as o-nitrobenzyl and spiropyran) are introduced into the polymer network, which undergo photolysis or isomerization under specific wavelength light irradiation, resulting in a transient change in crosslinking density, thereby regulating the relaxation rate; for example, 365 nm ultraviolet light irradiation can accelerate network dissociation, shortening τ from 20 s to 5 s, simulating the mechanical environment of rapid matrix degradation under inflammatory conditions. Enzyme-responsive hydrogel: Using matrix metalloproteinase (MMP) cleavable peptides as cross-linking points, when exogenous MMPs or cell-secreted MMPs are added, the network is locally degraded, achieving spatially and temporally controllable mechanical softening, simulating the dynamic remodeling process of ECM during wound repair.

[0098] The advantages of the aforementioned smart hydrogel are that mechanical relaxation behavior is no longer a fixed property, but a parameter that can be dynamically programmed by external signals. The system can automatically trigger a light source or inject an enzyme solution according to the intervention strategy output by the prediction model, and adjust the microenvironment state in real time without changing the culture medium.

[0099] Furthermore, the dynamic programming capability synergizes with the mechanical memory properties of ADSCs: cells not only respond to the current mechanical state, but also sense its rate of change, thereby activating different transcriptional programs.

[0100] As an optional embodiment, the super-resolution in-situ imaging unit is coaxially integrated with the mechanical loading device, supporting STED or SIM super-resolution imaging while applying tensile, shear, or compressive loads, ensuring that the mechanical-morphological data are spatiotemporally aligned.

[0101] Based on the above, the microscope stage is directly embedded through a mechanical loading device (such as a miniature tensile clamp or a shear flow cavity), and its axis of motion coincides with the optical axis of the objective lens. The loading components are made of flexible materials with high light transmittance (such as PDMS and quartz) to ensure that the high numerical aperture (NA greater than 1.4) required by STED / SIM is not obstructed; By using a hardware synchronization trigger, the clock signal of the mechanical waveform generator is locked with the camera exposure signal to achieve microsecond-level synchronization; During the imaging process, the cells are always under stress, and the correspondence between dynamic processes such as F-actin recombination and nuclear membrane deformation and the load phase (such as the stretching peak value and relaxation valley value) can be continuously observed.

[0102] Furthermore, each super-resolution image is accompanied by precise mechanical context (such as images acquired at the peak strain moment of 2 Hz stretching, matrix τ=12s), providing high-quality, unambiguous training data for subsequent AI models; if spatiotemporal misalignment occurs, the model may incorrectly attribute the morphology of the relaxation phase to the loading phase, leading to distorted causal inference.

[0103] It should be noted that the above modules can be implemented by software or hardware. For the latter, they can be implemented in the following ways, but are not limited to: all the above modules are located in the same processor; or, the above modules are located in different processors in any combination.

[0104] According to another aspect of the present invention, an electronic device is also provided, the electronic device including a memory and a processor; the memory is used to store a program; the processor executes the program to implement the method of any of the foregoing.

[0105] According to another aspect of the present invention, a computer-readable storage medium is also provided, the storage medium storing a computer program that, when executed by a processor, implements the method of any of the foregoing.

[0106] According to another aspect of the present invention, a computer program product is also provided, including a computer program that, when executed by a processor, implements the method described in any of the foregoing.

[0107] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.

Claims

1. A method for multi-scale image feature extraction and prediction of adipose-derived stem cell differentiation process, characterized in that, The method comprises the following steps: constructing a dynamic viscoelastic three-dimensional matrix simulating the perivascular mechanical microenvironment of blood vessels for specifically activating the endogenous mechanical sensing pathway of adipose-derived stem cells; applying multi-modal mechanical load to the adipose-derived stem cells on the dynamic viscoelastic three-dimensional matrix, and synchronously acquiring super-resolution time sequence image data of the adipose-derived stem cells in the differentiation process; based on the super-resolution time sequence image data, using a two-path deep neural network model of the adipose-derived stem cells, extracting a multi-scale morpho-mechanical feature set reflecting the perivascular origin characteristics of the adipose-derived stem cells; taking the multi-scale morpho-mechanical feature set as input, combining chromatin conformation dynamic data, constructing a cross-scale prediction model with the matrix relaxation response as a prior constraint, and outputting the stage probability of the adipose-derived stem cells to skin fibroblasts, the functional maturity evaluation and the optimal mechanical intervention strategy. 2.The method of claim 1, wherein: The two-path deep neural network model of the adipose-derived stem cells The first path encoder integrates the perivascular microenvironment prior knowledge module to enhance the sensitivity of feature extraction to the nuclear-cytoplasmic space compression ratio and the local convex shape of the nuclear membrane; The second path decoder embeds a physical information constraint layer to ensure that the cell traction field obtained by inversion strictly satisfies the constitutive response relationship of the dynamic viscoelastic matrix. 3.The method of claim 1, wherein: The loading frequency, amplitude and dynamic viscoelastic three-dimensional matrix stress relaxation time constant of the multi-modal mechanical load are coordinated to match the time integration window of the adipose-derived stem cells to the mechanical signal in the physiological state, thereby maximizing the transcriptional activation efficiency of the adipose-derived stem cells to the fibroblast phenotype. 4.The method of claim 1, wherein: The cross-scale prediction model generates a regulatory path with causal interpretability through an interpretable symbolic rule extractor, which is characterized by: a quantitative mapping relationship between the matrix relaxation rate, YAP / TAZ nuclear-cytoplasmic translocation dynamics, COL1A1 promoter region chromatin openness and collagen secretion efficiency with a confidence interval based on Bayesian inference. 5.The method of claim 1, wherein, It also includes a step of quantitatively characterizing the mechanical fingerprint of the mechanical activation of the adipose-derived stem cells: Based on the nuclear deformation response rate and myosin II activity dynamics of the adipose-derived stem cells to the matrix relaxation stimulus within the initial 24 hours, an individualized mechanical response fingerprint is constructed for pre-screening of a high-differentiation potential cell subpopulation. 6.The method of claim 1, wherein, It also includes a closed-loop regulation step: The optimal mechanical intervention strategy is converted into real-time executable mechanical parameter instructions, which are fed back to the mechanical loading system, and the prediction model weight is dynamically updated based on the subsequently collected image data to realize online correction and stable maintenance of the differentiation trajectory.

7. A multiscale image feature extraction and prediction system for adipose-derived stem cell differentiation process, operating a multiscale image feature extraction and prediction method for adipose-derived stem cell differentiation process according to any one of claims 1 to 6, characterized in that, It includes: a perivascular mechanical microenvironment simulation unit configured to provide a three-dimensional culture matrix with controllable stress relaxation behavior and integrate a multi-axial, programmable mechanical loading device; a super-resolution in-situ imaging unit configured to perform sub-hundred nanometer dynamic imaging of the adipose-derived stem cells during mechanical stimulation; an ADSC-specific feature extraction and prediction engine with a deep learning architecture that takes the perivascular origin characteristics and dynamic relaxation sensitivity of the adipose-derived stem cells as biological priors to extract multi-scale morpho-mechanical features from the imaging data and predict the differentiation trajectory; The cause-effect path visualization and human-machine collaborative decision unit is configured to convert the prediction results into actionable mechanical intervention instructions with uncertainty boundaries, and support the correction feedback of the coach or the experimental personnel to dynamically optimize the model decision weights.

8. The multi-scale image feature extraction and prediction system of fat-derived stem cell differentiation process according to claim 7, characterized in that: The ADSCs-specific feature extraction and prediction engine is in communication connection with a clinical skin tissue mechanics database, for verifying the elastic modulus, viscoelastic phase angle of the differentiated fibroblasts and the mechanical adaptability of the host tissue, and calibrating the optimal mechanical intervention strategy accordingly.

9. The multi-scale image feature extraction and prediction system of fat-derived stem cell differentiation processes according to claim 7, characterized in that: The perivascular mechanical microenvironment simulation unit adopts a light-responsive or enzyme-responsive intelligent hydrogel, whose stress relaxation behavior can be programmed in real time through external light signals or biochemical signals, to dynamically simulate the perivascular mechanical microenvironment under different physiological or pathological conditions.

10. The multi-scale image feature extraction and prediction system of fat-derived stem cell differentiation processes of claim 7, wherein: The super-resolution in-situ imaging unit is coaxially integrated with the mechanical loading device, supporting the STED or SIM super-resolution imaging while applying tensile, shear or compressive load, ensuring the spatiotemporal alignment of the mechanical-morphology data.

Images