Animal tissue slice path planning method based on machine learning

Abstract

This invention relates to a machine learning-based method for animal tissue slicing path planning, belonging to the field of intelligent medical image processing and pathological analysis technology. The method acquires the three-dimensional contour of the tissue and identifies its internal structure using a high-resolution machine vision system; constructs a nonlinear constitutive model by combining the tissue's physical properties; integrates the two to establish a digital twin model that can simulate the interaction between the cutting tool and the tissue; uses a physical information neural network to predict stress and deformation under different paths, generating an initial slicing path with deformation pre-compensation capabilities; and dynamically corrects the path based on actual feedback data during the slicing process to maintain thickness consistency and tissue integrity. Through the above technical solutions, this invention improves slicing quality and automation levels, and enhances the system's adaptability to complex biological samples.

Description

Technical Field

[0001] This invention belongs to the field of intelligent medical image processing and pathological analysis technology, specifically relating to a machine learning-based method for animal tissue slice path planning. Background Technology

[0002] With the continuous advancement of biomedical engineering and pathological diagnostic technologies, tissue sectioning technology plays a crucial supporting role in disease diagnosis, drug development, and life science research. Traditional tissue sectioning processes rely on high-precision automated sectioning equipment, aiming to obtain high-quality biological tissue samples through precise mechanical movements. In modern pathological analysis workflows, the stability and accuracy of section quality directly affect the effectiveness of subsequent staining observation and digital scanning. Therefore, improving the intelligence level of automated sectioning systems has become a key research focus in the field of precision medicine.

[0003] Machine learning-based path planning is a key step in achieving intelligent slicing. It primarily uses machine vision systems to acquire morphological features of tissue samples, then calculates the optimal slicing trajectory and feed parameters. This technology aims to reduce the frequency of manual intervention and improve slicing efficiency by automating the recognition of complex biological sample contours. In practical applications, ideal path planning must consider not only the geometric continuity of the slices but also the complex physical interactions between the mechanical actuator and the soft tissue sample.

[0004] Existing technologies typically employ static image analysis for path planning, neglecting the instantaneous tissue strain caused by blade compression during slicing, leading to deviations between the preset path and the actual cutting state. Lacking deep integration with nonlinear constitutive models such as viscoelasticity in biological tissues, traditional algorithms cannot predict the dynamic deformation trend of tissue under stress, resulting in physical damage such as uneven slice thickness or tissue fiber tearing. Planning logic relying solely on geometric contours struggles to achieve real-time synchronization between the digital twin space and the physical entity's state, failing to dynamically pre-compensate for positioning deviations during slicing, causing visual recognition results to lose their timeliness at the moment of physical contact. Therefore, a machine learning-based path planning method for animal tissue slicing is needed. Summary of the Invention

[0005] The purpose of this invention is to provide a machine learning-based method for planning the path of animal tissue slices, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a machine learning-based method for animal tissue slicing path planning, comprising the following specific steps:

[0007] Step 1: Acquire multi-angle images of animal tissue samples using a high-resolution machine vision system to obtain their three-dimensional contour information, and identify tissue boundaries and internal structural features based on a deep learning model; Step 2: Construct a nonlinear constitutive model incorporating viscoelastic response characteristics based on tissue type and physical properties to describe the dynamic deformation behavior of tissue under external mechanical action; Step 3: Fuse the three-dimensional contour information with the nonlinear constitutive model to establish a digital twin model of the tissue slicing process in virtual space. This model can simulate the instantaneous deformation state of the tissue during tool contact in real time; Step 4: Based on the digital twin model, use a physical information neural network to predict the stress distribution and corresponding deformation of the tissue under different slicing paths, and generate an initial slicing path with deformation pre-compensation capability accordingly; Step 5: During the slicing process, continuously collect actual cutting feedback data and transmit it back to the digital twin model for online updates, dynamically correcting subsequent slicing paths to maintain slice thickness consistency and tissue integrity.

[0008] Preferably, the machine vision system used in step 1 is equipped with multiple industrial-grade optical imaging units, which can perform non-destructive scanning of the surface and near-surface structure of the tissue without contacting it. The acquired image data is processed by semantic segmentation through a convolutional neural network to accurately extract the outer contour of the tissue and key anatomical landmarks.

[0009] Preferably, the nonlinear constitutive model constructed in step 2 sets different sets of mechanical parameters according to the type of tissue. The parameter set covers basic properties such as elastic modulus, Poisson's ratio, relaxation time constant and creep coefficient, and its numerical range is determined by experimental calibration to ensure that the model can accurately reflect the time-dependent deformation law of real biological tissue under stress conditions.

[0010] Preferably, the digital twin model established in step 3 not only includes geometric morphology information, but also integrates material properties, boundary constraints and tool kinematic parameters to form an interactive and evolving multiphysics coupled simulation environment, so that the virtual slicing process can synchronously map the tissue response state in the physical world.

[0011] Preferably, the physical information neural network used in step 4 incorporates a large number of synthetic data samples generated by finite element analysis during the training phase. These samples cover a variety of typical tissue types and slicing conditions, enabling the network to generalize and predict tissue deformation trends under unknown scenarios, thereby supporting the generation of optimized paths that balance safety and efficiency.

[0012] Preferably, the actual cutting feedback data collected in step 5 includes, but is not limited to, tool load signal, displacement deviation value and changes in surface reflectivity. These data are used to evaluate the current path execution effect and trigger the state correction mechanism inside the digital twin model, thereby achieving closed-loop path adaptive adjustment.

[0013] Preferably, the initial slicing path generation process comprehensively considers factors such as tissue hardness distribution, fiber orientation, and local curvature changes, ensuring minimal cutting resistance while avoiding the risk of tearing due to local stress concentration. The path trajectory is output in the form of a continuous and smooth spatial curve, which is compatible with the control command format of a six-degree-of-freedom precision feed platform.

[0014] Preferably, the digital twin model is connected to the physical slicing device via a low-latency communication link to ensure that the synchronization error between the virtual simulation results and the physical operation is controlled within a predetermined distance, thereby ensuring the effective implementation of the pre-compensation strategy.

[0015] Preferably, the loss function design of the physical information neural network incorporates conservation law constraints from biomechanical theory, enabling it to not only fit observed data during the learning process but also maintain consistent adherence to physical laws, thereby improving the robustness of the model under extreme or sparse data conditions.

[0016] Preferably, the method is applicable to soft tissue samples from various mammalian sources, including but not limited to liver, kidney, brain tissue and tumor tissue. When switching between different samples, only the corresponding constitutive model parameter set and visual recognition template need to be loaded, without retraining the entire system.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] 1. By integrating machine vision, biomechanical modeling and digital twin technologies, an intelligent slice planning system with the ability to predict tissue deformation and dynamically correct paths was constructed, overcoming the technical deficiency of traditional static image-driven path planning that cannot cope with instantaneous tissue deformation.

[0019] 2. This method improves the uniformity of slice thickness and the integrity of tissue structure, reduces sample damage caused by mechanical compression, and improves the reliability of pathological diagnosis and the reproducibility of scientific research data.

[0020] 3. By leveraging physical information neural networks and closed-loop feedback mechanisms, a path control paradigm shift from "passive following" to "active pre-adjustment" has been achieved, enhancing the adaptability and intelligence level of the automated slicing system to complex biological samples. Attached Figure Description

[0021] Figure 1This is a schematic diagram of the overall technical solution architecture according to the present invention;

[0022] Figure 2 This is a schematic diagram of the core principle framework for deformation prediction and path planning based on a digital twin model and a physical information neural network according to the present invention.

[0023] Figure 3 This is a logical flowchart illustrating the process of fusing three-dimensional contour information acquired from multi-angle images with a nonlinear constitutive model to establish an organizational digital twin model according to the present invention.

[0024] Figure 4 A flowchart illustrating the process of using a physical information neural network to predict tissue stress distribution and generate an initial slice path with deformation pre-compensation capability according to the present invention.

[0025] Figure 5 This is a schematic diagram of the multi-level interaction and data flow of the path adaptive dynamic correction based on the online update mechanism of actual cutting feedback data and digital twin model according to the present invention. Detailed Implementation

[0026] Example 1: Please attach Figure 1 To be continued Figure 5 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0027] In this embodiment, a machine learning-based animal tissue slicing path planning method is provided. This method achieves precise pre-compensation for dynamic deformation during animal tissue slicing through multi-dimensional data acquisition, high-precision biomechanical modeling, virtual reality-synchronized digital twin construction, and deep learning prediction of physical constraints. The method includes the following detailed steps:

[0028] In the method described above, step 1 involves acquiring multi-angle images of animal tissue samples using a high-resolution machine vision system to obtain their three-dimensional contour information, and identifying tissue boundaries and internal structural features based on a deep learning model. In practice, the machine vision system is equipped with multiple industrial-grade optical imaging units arranged in an array around the tissue to be sliced, forming a comprehensive visual monitoring field. Each optical imaging unit possesses high pixel density and a wide dynamic range, enabling non-destructive scanning of the surface and near-surface structures without contact with the tissue. To ensure image acquisition accuracy, the system pre-calibrates intrinsic and extrinsic parameters, using a calibration board to determine the relative spatial relationships between the cameras and distortion correction parameters. During acquisition, a synchronous triggering mechanism ensures that all cameras capture images at the same millisecond-level timestamp, thereby eliminating motion blur caused by minor vibrations or tissue relaxation.

[0029] Furthermore, the acquired multi-source image data is transmitted to a high-performance computing module, where semantic segmentation is performed using a convolutional neural network. This convolutional neural network employs a multi-layer feature extraction architecture, including convolutional layers, pooling layers, and skip connection structures, enabling it to classify each pixel in the image and accurately extract the tissue outline and key anatomical landmarks. The semantic segmentation process divides the complex biological tissue image into background areas, tissue surface areas, vascular distribution areas, and potentially lesion areas. By performing stereo matching and 3D reconstruction on the multi-angle segmentation results, the system generates high-precision 3D point cloud data of the tissue. This point cloud data describes the initial geometric shape of the tissue, providing a spatial reference for subsequent modeling.

[0030] In the method described above, step 2 involves constructing a nonlinear constitutive model incorporating viscoelastic response characteristics based on tissue type and physical properties. This model describes the dynamic deformation behavior of the tissue under external mechanical action. The nonlinear constitutive model sets different sets of mechanical parameters according to the tissue type. Considering that biological tissues are not ideal elastic bodies, they exhibit significant time dependence, hysteresis effects, and nonlinear elastic characteristics under stress. The constitutive model in this embodiment covers basic properties such as elastic modulus, Poisson's ratio, relaxation time constant, and creep coefficient. Specifically, the elastic modulus describes the tissue's resistance to deformation during the initial stress stage; Poisson's ratio reflects the lateral expansion effect of the tissue under longitudinal compression; for soft tissues, this value is typically close to 0.5, indicating near-incompressibility; the relaxation time constant describes the rate at which stress decays over time under constant strain; and the creep coefficient describes the trend of strain growth over time under constant stress.

[0031] To ensure the model accurately reflects the time-dependent deformation of real biological tissues under stress, the system's numerical range was determined through experimental calibration. In the preliminary experimental phase, step load and sinusoidal wave tests were performed on similar tissue samples using a micro-displacement loading device, recording the stress and strain curves over time. Curve fitting algorithms were then used to transform the experimental data into key physical parameters in the constitutive model. For example, the generalized Maxwell model was used to describe the viscoelastic characteristics of the tissue, representing it as a series-parallel combination of multiple springs and dampers, each combination corresponding to a specific relaxation mode. In this way, the constitutive model can simulate the stress redistribution process within the tissue and the residual deformation caused by viscous flow under different cutting speeds.

[0032] Furthermore, to clarify the specific influence and interrelationship of elastic modulus, Poisson's ratio, relaxation time constant, and creep coefficient on tissue deformation behavior in the nonlinear constitutive model, this embodiment establishes the following quantitative relationships:

[0033] The organization at all times strain response tensor It is composed of the superposition of elastic and viscous components, and can be expressed as:

[0034]

[0035] in, Let be the instantaneous elastic strain tensor, dimensionless, and governed by Hooke's law. Seek; For stress tensor (unit: Pascal, Pa). The fourth-order elastic stiffness tensor (unit: Pa) is derived from the elastic modulus. Compared with Poisson structure; The viscous strain tensor evolves over time; it is dimensionless and contributed by creep and stress relaxation processes.

[0036] For the isotropic biological soft tissue involved in this application, the specific correspondence between the fourth-order elastic stiffness tensor C and the elastic modulus E and Poisson's ratio ν is constructed using Lamé constants: First, Lamé's first constant λ and Lamé's second constant μ are defined, and they satisfy the following transformation relationship with the elastic modulus E and Poisson's ratio ν:

[0037]

[0038] Based on the Lamé constant mentioned above, the components of the fourth-order elastic stiffness tensor C can be expressed as:

[0039]

[0040] in, Let Kronecker function be used when i=j. =1, otherwise 0. Through this constructive relationship, those skilled in the art can directly calculate all components of the fourth-order elastic stiffness tensor C based on the calibrated elastic modulus E and Poisson's ratio ν, thereby clarifying the mapping relationship between elastic strain and stress. Simultaneously, the correspondence between the strain response tensor and the constitutive model can be clearly defined as follows:

[0041] The total strain response tensor ε is the elastic strain. With viscous strain The linear superposition of elastic strains, where elastic strains The stress tensor σ satisfies Hooke's law. (: indicates tensor double dot product operation), viscous strain The creep and stress relaxation equations are then iteratively calculated based on the current stress history and time step, thus providing a complete description of the instantaneous and delayed deformation response of the tissue during the stress process.

[0042] elastic modulus Poisson's ratio characterizes the ability of a microstructure to resist deformation during the initial stress stage, measured in Pascals (Pa). It is obtained by performing uniaxial compression experiments using a micro-displacement loading device and calculating the slope of the initial linear segment of the stress-strain curve. It reflects the transverse expansion effect of tissue under longitudinal compression, is dimensionless, and is determined by simultaneous measurement of longitudinal and transverse strain experiments for soft tissues that are close to incompressible.

[0043] Stress relaxation behavior is determined by the normalized relaxation modulus describe:

[0044]

[0045] In the formula, For a moment Stress (Pa) Initial stress (Pa); Let be the weight coefficient of the i-th relaxation mode, dimensionless, satisfying ; Let be the i-th relaxation time constant, in seconds (s), which describes the rate at which stress decays with time under constant strain; The smaller the value, the faster the stress relaxation. and All data were obtained through exponential fitting of experimental data under step load.

[0046] Creep behavior is determined by creep compliance. Characterization:

[0047]

[0048] In the formula, Let be the strain at time t; The applied constant stress (Pa); Instantaneous elastic flexibility Let j be the j-th delayed elastic compliance; Let j be the j-th delay time constant (s), which reflects how quickly creep reaches equilibrium. The steady-state viscosity coefficient (Pa·s) characterizes the strain growth trend over time under constant stress.

[0049] The term describes the viscous flow strain that is irreversible. , , , All data were obtained through nonlinear regression of constant stress creep experimental data.

[0050] elastic modulus relative to the initial value of creep compliance They are reciprocals of each other, that is... Poisson's ratio The anisotropic distribution of elastic stiffness under multiaxial stress states affects the following: When the bulk modulus approaches infinity, the microstructure exhibits incompressible properties; relaxation time constant With delay time constant Both factors jointly determine a material's dissipation capacity, and there is a positive correlation between them, i.e., high... Organizations often have higher The aforementioned quantitative relationships enable the constitutive model to accurately reproduce the instantaneous elastic response, short-term viscoelastic creep, and long-term relaxation behavior of tissues under external forces based on experimentally calibrated parameters, providing a complete mechanical description basis for subsequent digital twin models.

[0051] In the method, step 3 involves fusing the three-dimensional contour information with a nonlinear constitutive model to establish a digital twin model of the tissue slicing process in virtual space. This digital twin model not only includes geometric morphology information but also integrates material properties, boundary constraints, and tool kinematic parameters. The digital twin model is constructed by converting the three-dimensional contour obtained in step 1 into a physically meaningful finite element mesh using voxelization technology. Each mesh element is assigned the nonlinear constitutive parameters determined in step 2. The model defines the boundary constraints of the tissue, such as the adhesion and fixation state at the tissue base or the lateral support force field.

[0052] The digital twin model constructs an interactive and evolving multiphysics-coupled simulation environment, enabling the virtual slicing process to synchronously map the tissue response state in the physical world. Within this environment, the geometry of the virtual cutting tool is modeled according to actual physical parameters, including the cutting edge angle, coating friction coefficient, and back stiffness. The tool's kinematic parameters, such as feed rate, oscillation frequency, and entry angle, are input to the model in real time. When the virtual cutting tool contacts the tissue mesh, the system calculates the pressure distribution and friction on the contact surface based on contact mechanics algorithms. Due to the integration of a nonlinear constitutive model, the digital twin system can calculate the instantaneous displacement and strain fields of the tissue under tool compression in real time.

[0053] Furthermore, to clarify the influence of material properties, boundary constraints, and tool kinematic parameters on the simulation results in the digital twin model, as well as the correlation between parameters, this embodiment defines the following governing equations and parameter coupling relationships within the finite element framework:

[0054] Displacement field of an organization in virtual space (Unit: meter, m) Controlled by equilibrium equations:

[0055]

[0056] In the formula, Let be the stress tensor (Pa). Volumetric force density (N / m³) 3 In this model, gravity is the primary factor, and the values ​​are... , Tissue density (kg / m³) 3 The values ​​were determined by the water displacement method, ranging from 950 to 1100 kg / m³. 3 ; The acceleration vector due to gravity (m / s²) 2 For the specific control of the displacement field, this digital twin model adopts a control method combining finite element iterative solution and real-time feedback correction: In the offline simulation stage, for each simulation time step, the contact boundary conditions are first updated according to the preset motion trajectory of the tool, and the current contact force and friction force are applied as loads to the corresponding nodes of the finite element mesh; then, the current stress-strain relationship calculated by the nonlinear constitutive model is substituted into the above equilibrium equation, and the increment of the nodal displacement is solved by the Newton-Raphson iterative method, iterating until the equation residual converges to 1e -6 The system uses a preset threshold for N to obtain the displacement field distribution at the current moment, thus enabling pre-calculation of tissue deformation. During the online execution phase, when the actual cutting feedback data is transmitted back to the model, the system first compares the deviation between the actual cutting force and the model's predicted cutting force. If the deviation exceeds a 5% threshold, the system automatically adjusts the constitutive model parameters for the corresponding region. Then, it re-executes the aforementioned finite element iterative solution process, updates the predicted displacement field results, and corrects the subsequent slicing path based on the updated displacement field. This achieves dynamic closed-loop control of the displacement field, ensuring that the virtual simulation results are synchronized with the deformation state of the physical entity in real time.

[0057] Stress tensor Based on the current strain history, the material property field (elastic modulus) is calculated using the aforementioned nonlinear constitutive model. Poisson's ratio Weighting coefficients of relaxation mode , Relaxation time constant, creep parameters , , , Each finite element mesh element is assigned a spatial distribution to achieve heterogeneity modeling.

[0058] Boundary constraints are achieved through displacement boundary conditions. Force boundary conditions Application. Corresponding to the adhesion and fixation state at the tissue base. At the bottom node; the lateral support force field is equivalent to a linear spring constraint, providing restoring force. ,in The support stiffness coefficient (N / m) is derived from preliminary experiments, and its value range is [missing value]. to ; To support the initial reference position (m).

[0059] The quantitative correlation between tool kinematic parameters and contact response is as follows: the tool at feed rate (m / s) motion, The range of values ​​is given by the path planning module. to When the blade comes into contact with the tissue surface, the normal pressure on the contact surface... With tangential friction Expressed as:

[0060]

[0061] In the formula, Normal contact stiffness It is determined by the material's elastic modulus and the mesh size; The normal penetration depth (m) is taken as a positive value; It is a non-linear exponent, and for soft tissue it is usually taken as... to ;for The Coulomb coefficient of friction between the tool coating and the tissue is dimensionless and determined by sliding experiments. (Range) to ; The relative tangential velocity vector (m / s) between the tool and the tissue at the contact point; This is a small velocity regularization constant (m / s) to avoid the denominator being zero.

[0062] There is a coupling between tool kinematic parameters and material properties: feed rate The large blade angle leads to a significant strain rate effect, which enhances the microstructure stiffness through strain rate-related terms in the constitutive model (such as the sticking pot damping force being proportional to the strain rate); the blade angle affects the degree of local stress concentration, with a larger blade angle corresponding to higher stress concentration. A smaller cutting edge angle helps reduce cutting resistance. The aforementioned quantitative relationships enable the digital twin model to calculate structural deformation, stress distribution, and cutting force based on physically real parameter inputs, providing a high-fidelity simulation environment for subsequent path pre-compensation.

[0063] This simulation is dynamic, meaning that as the tool penetrates deeper, the stress concentration areas within the tissue shift, causing the tissue contour to undergo real-time, non-linear distortion.

[0064] In the method, step 4 involves using a physical information neural network (PEN) based on the digital twin model to predict the stress distribution and corresponding deformation of tissues under different slicing paths, and thereby generating an initial slicing path with deformation pre-compensation capability. During the training phase, the PSN incorporates a large number of synthetic data samples generated from finite element analysis, covering various typical tissue types and slicing condition combinations. Unlike traditional purely data-driven networks, the PSN in this embodiment embeds conservation law constraints from biomechanical theory into its loss function design. The loss function includes the mean square error between the predicted deformation value and the labeled value, as well as residual terms from the momentum conservation equation, mass conservation equation, and energy balance equation. This allows the network to not only fit observed data during the learning process but also maintain consistent adherence to physical laws.

[0065] When generating the initial slicing path, the physical information neural network rapidly scans the candidate path space. It predicts the displacement vector field of the tissue due to compression when the tool moves along a predetermined trajectory. If the prediction shows that the tissue experiences a 30-micrometer downward indentation in front of the cutting edge, the generated deformation pre-compensation path will automatically be corrected downward by 30 micrometers in the corresponding spatial coordinates, thus ensuring that the cutting depth when the cutting edge actually contacts the tissue is completely consistent with the preset slicing thickness. The generation process of the initial slicing path comprehensively considers factors such as tissue hardness distribution, fiber orientation, and local curvature changes. For tissues with obvious fibrous structures (such as muscle tissue), the neural network identifies the fiber axis and optimizes the path to reduce tearing stress caused by cross-fiber cutting. Finally, the path trajectory is output as a continuous and smooth spatial curve, which has undergone high-order Bezier smoothing and can be directly adapted to the control command format of a six-degree-of-freedom precision feed platform, ensuring that the acceleration and jerk of the robotic arm during movement are within safe thresholds.

[0066] Furthermore, this paper elucidates the specific influence of tissue stiffness distribution, fiber orientation, and local curvature changes on the generation of the initial slice path, as well as the correlation between these factors. This embodiment defines the following path optimization model:

[0067] Let the path of the slice to be planned be a spatial curve. ,in This is the arc length parameter (unit: meters, m). Let be the total path length (m). Path generation is achieved by solving a multi-objective optimization problem:

[0068]

[0069] In the formula, These are weighting coefficients, dimensionless, set manually or preset by the system based on the slicing mode (e.g., precision slicing or high-speed slicing), with values ​​ranging from [value range missing]. And satisfy .

[0070] Cost function for deformation compensation error (unit: ), defined as:

[0071]

[0072] in, The ideal position curve (m) where the desired slice plane intersects the tissue contour is determined by three-dimensional contour information; For in position The displacement compensation vector (m) caused by tissue deformation is generated by a physical information neural network based on the local tissue hardness field. (Pa) and fiber orientation vector The hardness field was predicted. Fiber orientation obtained by mapping ultrasound elastography data A unit vector (dimensionless) The main direction of muscle fibers or collagen bundles is extracted and characterized by polarization imaging combined with texture analysis algorithms.

[0073] The relationship between hardness and fiber orientation on the deformation vector is as follows: when the tool feed direction is perpendicular to the fiber orientation... included angle satisfy (i.e., when cutting vertically) The amplitude increases, causing significant extrusion deformation due to cross-fiber cutting; when (i.e., cutting along the fiber) The amplitude decreases. Specifically, the displacement amplitude... It can be modeled as:

[0074]

[0075] In the formula, The reference displacement (m) is obtained from the calibration experiment; The average hardness of the tissue is taken as the reference hardness value (Pa). The fiber anisotropy factor is dimensionless, ranging from 0.2 to 2.0, and is obtained through uniaxial compression comparison experiments. This formula indicates the tissue stiffness. The higher the fiber, the smaller the deformation displacement; the more perpendicular the fiber direction is to the cutting direction, the greater the deformation displacement.

[0076] The energy cost of cutting (unit: J·m) is based on the predicted stress field. calculate:

[0077]

[0078] in, Energy conversion factor ( The value is determined by the tool geometry and the fracture toughness of the material, and is determined through calibration cutting experiments. For position The predicted equivalent stress (Pa) at the location is output by the physical information neural network. The effective cutting area of ​​the tool ( ), which is a constant.

[0079] Cost of path curvature smoothing (unit: ), to suppress drastic path bending:

[0080]

[0081] Path curve The curvature is calculated as follows: ,unit The surface curvature of local tissues affects path feasibility: In areas with high curvature, if the path follows surface undulations, then... Enlargement, leading to The optimization process tends to deviate slightly from the surface in areas of abrupt curvature change to reduce path curvature, while ensuring cutting accuracy through deformation compensation terms.

[0082] The relationship between the factors is reflected in the fact that fiber orientation and hardness distribution jointly determine the deformation field, which in turn affects... The local curvature of the path is limited by the surface geometry of the tissue, and there is a trade-off between this and the deformation compensation requirements. This can be addressed by... and Adjustment. The quantitative model described above comprehensively considers mechanical, geometric, and energy consumption factors in the path generation process, outputting a three-dimensional tool trajectory that combines deformation pre-compensation capability and motion smoothness characteristics.

[0083] In the method described above, step 5 involves continuously collecting actual cutting feedback data during the slicing process and transmitting it back to the digital twin model for online updates, dynamically correcting subsequent slicing paths. The collected actual cutting feedback data includes, but is not limited to, tool load signals, displacement deviation values, and changes in tissue surface reflectivity. The tool load signal is acquired through a multi-dimensional force sensor mounted on the tool holder, reflecting the deviation between the actual cutting resistance and the predicted value; the displacement deviation value is confirmed using a laser interferometer or high-frequency visual tracking; and changes in tissue surface reflectivity are captured by a high-speed industrial camera to indirectly assess the degree of dehydration of the tissue surface or changes in its microstructure under pressure.

[0084] This real-time data is used to evaluate the current path execution effect and trigger the state correction mechanism within the digital twin model. If the actual measured cutting force is greater than the predicted value of the digital twin model, the system automatically determines that there are local calcification points or insufficient estimation of hardness parameters within the tissue, and then adjusts the elastic modulus parameter in the constitutive model in real time through a gain adjustment algorithm. This online update of parameters prompts the digital twin model to recalculate the deformation trend under the remaining path and make immediate corrections to the output of the physical information neural network. The digital twin model and the physical slicing device are connected via a low-latency communication link using the real-time industrial Ethernet protocol, ensuring that the synchronization error between the virtual simulation results and the physical operation is controlled within a predetermined distance, for example, within 10 micrometers. Through this closed-loop path adaptive adjustment, the system can respond in real time to unpredictable deformations caused by tissue heterogeneity during slicing, thereby maintaining extremely high consistency of slice thickness and tissue integrity.

[0085] In step 1, the vision system employs polarized illumination technology to address the complex reflective properties of animal tissue surfaces. By adding a polarizer in front of the light source and an analyzer in front of the lens, specular reflections from the tissue surface are effectively eliminated, thus clearly revealing the deep anatomical textures of the tissue. The acquired image data is preprocessed and transformed into feature maps with grayscale gradient distributions. The encoder in the deep learning model compresses the image into high-dimensional feature vectors, while the decoder is responsible for reconstructing the boundary curves of each anatomical layer of the tissue. This deep learning-based recognition method automatically ignores interference from bloodstains, moisture, or other impurities on the tissue surface, ensuring that the extracted contour information has extremely high robustness.

[0086] When constructing the constitutive model in step 2, the system also incorporates ultrasound elastography data as an aid, taking into account the tissue differences among individuals. The local stiffness distribution of the tissue is determined by measuring the propagation speed of ultrasound waves within it. This distribution field is then mapped onto the nonlinear constitutive model, transforming it from a homogeneous entity into a complex mechanical field with spatial variability. This fine-grained physical description provides a more solid data foundation for subsequent accurate predictions.

[0087] In the digital twin model maintenance process of step 3, the system establishes a time-synchronized virtual mirror. Each time the physical slicer completes a feed operation, the mesh in the virtual model undergoes corresponding cell deletion or attribute updates, simulating material removal. The model calculates the heat transfer process in the tissue; if localized temperature increases may lead to tissue protein denaturation or changes in deformation properties, these heat-induced parameter drifts are also reflected in the digital twin model, ensuring that the physical states of the virtual and physical spaces remain highly consistent.

[0088] During the path generation process in step 4, the physical information neural network employs a multi-objective optimization strategy. The optimization objectives include not only maintaining uniform thickness deformation compensation but also minimizing power consumption and maximizing tool life during the slicing process. By adjusting the weights of each term in the loss function, the system can switch between "high-speed slicing" and "high-precision slicing" modes according to experimental requirements. The generated path instruction set includes coordinated control of each joint of the six-degree-of-freedom platform. Through inverse kinematics, the spatial curve is transformed into an encoder pulse sequence for the motor, and a feedforward control algorithm is used to pre-compensate for the inertial effects of the mechanical structure.

[0089] During the dynamic correction in step 5, the response frequency of the feedback loop is set above 1000 Hz. This means that the system performs thousands of checks and corrections during each micrometer of cutting. When a deviation of the cutting trajectory from the predetermined plane is detected, the correction command is immediately applied to the piezoelectric ceramic actuator to provide micrometer-level rapid compensation for the tool position. This high-frequency closed-loop mechanism fundamentally solves the "drift" phenomenon caused by stress release within biological tissue.

[0090] This invention's method is applicable to soft tissue samples from various mammalian sources, including but not limited to liver, kidney, brain tissue, and tumor tissue. Due to the vastly different mechanical properties of various tissues, when switching between different samples, the operator only needs to load the corresponding constitutive model parameter set and visual recognition template from a pre-set database; the system can automatically complete the configuration adjustment without requiring time-consuming retraining for each new sample. This combination of versatility and flexibility enables this intelligent slice planning system to be widely applied in pathological analysis, pharmacokinetic studies, and high-end biomanufacturing.

[0091] Example 2: This example focuses on how the method of the present invention achieves high-quality slicing through refined local path optimization when processing highly heterogeneous tumor tissue samples. Tumor tissue typically contains sclerotic stroma regions and softened necrotic cores, and its mechanical response is extremely complex.

[0092] In step 1, in addition to acquiring the outer contour, the vision system also uses multispectral imaging technology to identify the blood vessel density and necrotic foci distribution on the tumor surface. These multispectral images are input into a deep learning model, which uses a specific spectral feature extraction algorithm to identify tissue regions with different hardness levels and generate hardness feature masks.

[0093] In step 2, the nonlinear constitutive model is further refined into regionalized parameter sets. For the identified hardened regions, the elastic modulus is set to a higher value, and the strain hardening coefficient is increased; for the softened regions, the viscous flow parameter is increased. This locally differentiated constitutive model construction can more realistically simulate the dramatic mechanical response generated when the tool passes through tumor tissue.

[0094] In step 3, the digital twin model introduces local stress concentration monitoring points. When the simulated tool approaches the boundary of the hardened region, the model issues a warning of potential stress abrupt changes. This warning information is transmitted in real time to the path planning module in step 4.

[0095] In step 4, the physical information neural network, based on the early warning from the digital twin model, incorporates "decelerated entry" and "small-angle bevel" strategies into the generated initial path. Before entering the hardened region, the system reduces the feed rate and adjusts the tool entry angle to decrease the normal force and increase the shear component. This path optimization effectively prevents tissue tearing caused by sudden changes in hardness.

[0096] In step 5, the feedback system focuses on monitoring the vibration spectrum of the cutting tool. The heterogeneity of tumor tissue often induces high-frequency vibrations. The feedback mechanism analyzes the amplitude and frequency of the vibration signal to determine in real time whether the cutting tool has deflected and drives the active vibration damping system to compensate. The online update mechanism automatically adjusts the pre-compensation amount of the next slice at the same position based on the actual feedback of each slice, thereby continuously accumulating experience during continuous slicing and improving the slicing success rate.

[0097] Example 3: This example describes how the method of the present invention utilizes digital twins and physical information neural networks to handle extreme deformation problems when slicing ultra-soft tissues (such as brain tissue). Brain tissue has extremely low stiffness and extremely high viscosity, and conventional slicing is prone to fluid-like drag deformation.

[0098] In step 1, the visual system uses a low-light mode combined with fluorescent labeling technology to precisely capture the subtle boundary features of brain tissue. Three-dimensional contour information is accurate to the micrometer level to capture minute collapses caused by gravity.

[0099] In step 2, the constitutive model emphasizes the viscous component, employing a higher-order relaxation modulus function to describe the long-delay response of brain tissue under stress. The creep coefficient is given a high initial weight to simulate pre-deformation that may occur in the tissue before tool contact.

[0100] In step 3, the digital twin model integrates a fluid-structure interaction algorithm. Since brain tissue behaves almost like a fluid in the slicing fluid, the model simulates the buoyancy support of the slicing fluid on the tissue and the fluid resistance generated by the tool movement. This fully coupled virtual environment can predict the three-dimensional displacement of the tissue in the liquid.

[0101] In step 4, the pre-compensation path generated by the physical information neural network exhibits a complex nonlinear curve. To address the "yielding" effect of brain tissue, the path planning pre-plans the tool to make small, synchronous movements along the expected deformation direction of the tissue before contacting the tissue, thereby reducing relative displacement and achieving smooth entry. The neural network calculates the optimal motion compensation vector by adhering to the laws of fluid dynamics conservation.

[0102] In step 5, the feedback system monitors the liquid film thickness on the tissue surface by utilizing changes in the reflectivity of the tissue surface, preventing tissue shrinkage due to dehydration. The online update mechanism fine-tunes the ambient liquid level or spray flow rate in real time, combined with dynamic path correction, to ensure that continuous sections with constant thickness can be obtained even under extreme soft conditions.

[0103] In the method described in this invention, in order to achieve accurate path planning, the conversion logic between the output value of the physical information neural network and the actual mechanical action is as follows:

[0104] Define the ideal coordinate sequence of the target slicing plane in three-dimensional space. When the physical information neural network predicts that at position coordinate point P, the tissue will generate a displacement vector D due to mechanical compression, then the deformation pre-compensation coordinate point P' calculated by the system is equal to the ideal coordinate point P minus the displacement vector D. Through this reverse superposition logic, when the tool moves to position P', due to the deformation D of the tissue, the actual contact point between the tool and the tissue exactly returns to the ideal coordinate point P.

[0105] The textual logic for adjusting the cutting speed is as follows: The system sets a base cutting speed V. The physical information neural network predicts the equivalent stress S at the current position in real time. When the equivalent stress S exceeds the set safety stress threshold T, the system adjusts the cutting speed to the product of V and a correction coefficient K. The correction coefficient K is a value less than 1, and its magnitude decreases as the stress S exceeds the threshold T. This textual logic ensures that the structural integrity is protected by reducing the speed in areas of stress concentration.

[0106] The multiphysics coupling calculation logic in the digital twin model is described as follows: Within each simulation step, the system first calculates the contact force distribution based on the tool displacement increment, and then uses this force distribution as input to the constitutive model to solve for the displacement field of each node in the microstructure. The result of the displacement field then changes the geometric boundary of the microstructure, thereby affecting the contact force calculation in the next step. This iterative process, described through a textual temporal logic, ensures a rigorous simulation of physical realism within the virtual environment.

[0107] In the control of a six-DOF precision feed platform, the coordinated movements of the joint motors follow these principles: The instantaneous normal direction of the cutting edge relative to the tissue is calculated based on the spatial tangent vector of the pre-compensated path. This normal direction requirement is then decomposed onto each rotary and displacement joint through inverse kinematics. The system ensures that at any given time, the tool's feed direction aligns with the local force equilibrium direction of the tissue, minimizing transverse cutting forces. If the movement of a joint approaches its physical limit, the path planning module automatically triggers replanning of the bypass trajectory, avoiding singular points by adjusting the tool's redundant degrees of freedom without altering the slicing plane.

[0108] To address the communication synchronization between the digital twin model and the physical slicing device, the system employs heartbeat detection and time truncation compensation logic. The main control unit periodically sends synchronization packets containing high-precision timestamps, and the receiving end predicts future tissue deformation trends based on the packet arrival time delay. If the communication delay exceeds a preset millisecond threshold, the system automatically enters a safe shutdown mode or uses an inertial navigation strategy based on historical data to continue executing a short path, ensuring operational continuity and safety.

[0109] In the training data generation stage of the physical information neural network, the finite element analysis process employs nonlinear large deformation theory. Each training sample contains response data of the tissue under different prestress states. These samples are input into the network after normalization. The neurons within the network map the input features through a nonlinear activation function, and their weight updates depend not only on the labeling error but also on the physical residuals. When the strain field predicted by the network does not satisfy the continuity equation, the physical loss term generates a large penalty value, forcing the network weights to converge in a direction that satisfies physical laws.

[0110] In practical applications of this method, operators monitor the entire slicing process through a graphical interface. The interface displays real-time images from the physical camera, a virtual rendering of the digital twin model, and a heatmap showing the discrepancy between the two. The heatmap visually demonstrates the degree of agreement between the model's predictions and the actual feedback. When a significant red area of ​​high deviation appears in the heatmap, the system automatically prompts the operator to check the constitutive parameter settings or tool wear status. This human-computer interaction mechanism provides an additional layer of protection for the system's robust operation.

[0111] The method adapts to different animal species through "parameter transfer learning." When the system switches from processing mouse tissue to processing rat or pig tissue, the physical information neural network utilizes existing knowledge of physical laws and is fine-tuned using a small number of specific species samples. This approach significantly shortens the system's deployment time, enabling it to quickly adapt to changing research and clinical needs.

[0112] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A machine learning-based method for animal tissue slicing path planning, characterized in that, The specific steps include the following: Step 1: Acquire multi-angle images of animal tissue samples using a high-resolution machine vision system to obtain the three-dimensional contour information of the animal tissue samples, and identify the tissue boundaries and internal structural features of the animal tissue samples based on a deep learning model; Step 2: Based on the tissue type and physical properties, construct a nonlinear constitutive model that includes viscoelastic response characteristics to describe the dynamic deformation behavior of the animal tissue sample under external mechanical action; Step 3: Integrate the three-dimensional contour information with the nonlinear constitutive model to establish a digital twin model of the tissue slicing process in virtual space. The digital twin model is used to simulate the instantaneous deformation state of the animal tissue sample during the contact of the cutting tool in real time. Step 4: Based on the digital twin model, use a physical information neural network to predict the stress distribution and corresponding deformation of the tissue under different slicing paths, and generate an initial slicing path with deformation pre-compensation capability accordingly. Step 5: During the slicing process, continuously collect actual cutting feedback data and send the actual cutting feedback data back to the digital twin model for online updates, dynamically correcting subsequent slicing paths to maintain slice thickness consistency and tissue integrity.

2. The method for animal tissue slicing path planning based on machine learning according to claim 1, characterized in that: The machine vision system used in step one is equipped with multiple industrial-grade optical imaging units, which are arranged in an array around the tissue to be sliced, forming a comprehensive visual monitoring field. During image acquisition, specular reflections on the tissue surface are eliminated using polarized illumination technology. A polarizer is installed in front of the light source, and an analyzer is installed in front of the lens, thereby revealing the deep anatomical texture of the tissue. The acquired image data is processed by semantic segmentation through a convolutional neural network. The convolutional neural network adopts a feature extraction architecture including convolutional layers, pooling layers, and skip connection structures to divide the image into background areas, tissue surface areas, blood vessel distribution areas, and lesion areas, accurately extracting the outer contour of the tissue and key anatomical landmarks. By performing stereo matching and 3D reconstruction on the multi-angle segmentation results, high-precision 3D point cloud data of the tissue is generated as a spatial reference for subsequent modeling.

3. The method for animal tissue slicing path planning based on machine learning according to claim 1, characterized in that: The nonlinear constitutive model constructed in step two sets different mechanical parameter sets according to the type of tissue. The mechanical parameter sets cover elastic modulus, Poisson's ratio, relaxation time constant and creep coefficient. The elastic modulus is used to describe the tissue's resistance to deformation during the initial stress stage. The Poisson's ratio reflects the lateral expansion effect of the tissue under longitudinal compression and its value approaches a predetermined incompressible state value. The relaxation time constant describes the rate at which stress decays over time under constant strain. The creep coefficient describes the trend of strain growth over time under constant stress. The numerical range of the mechanical parameter set was determined by experimental calibration. Step load test and sinusoidal wave test were performed on tissue samples of the same type using a micro-displacement loading device. The experimental data were then converted into physical parameters in the nonlinear constitutive model. The nonlinear constitutive model uses the generalized Maxwell model to describe the viscoelastic characteristics of the microstructure, equating the microstructure to a series and parallel combination of multiple springs and dampers, in order to simulate the redistribution of stress within the microstructure under different cutting speeds and the residual deformation caused by viscous flow.

4. The method for animal tissue slicing path planning based on machine learning according to claim 3, characterized in that: When constructing the nonlinear constitutive model, ultrasonic elastography data is also introduced as an aid. The local hardness distribution of the tissue is determined by the propagation speed of ultrasound in the tissue, and the local hardness distribution field is mapped into the nonlinear constitutive model, so that the nonlinear constitutive model has spatial variability characteristics. For heterogeneous tumor tissue samples, the nonlinear constitutive model is refined into a regional parameter set. For hardened regions, strain hardening coefficients are added, and for softened regions, viscous flow parameters are added, so as to realistically simulate the dramatic mechanical response generated when the tool passes through the tissue.

5. The method for animal tissue slicing path planning based on machine learning according to claim 1, characterized in that: The digital twin model established in step three transforms the three-dimensional contour information into a physically meaningful finite element mesh using voxelization technology, and assigns mechanical parameters from the nonlinear constitutive model to each mesh element. The digital twin model integrates material properties, boundary constraints, and tool kinematic parameters to form an interactively evolving multiphysics coupled simulation environment. The boundary constraints include the adhesion and fixation state at the bottom of the tissue and the lateral support force field. In the simulation environment, the virtual tool is modeled according to actual physical parameters, including the cutting edge angle, coating friction coefficient, and back stiffness. When the virtual tool contacts the tissue mesh, the pressure distribution and friction on the contact surface are calculated based on the contact mechanics algorithm, and the instantaneous displacement and strain fields of the tissue under tool compression are calculated in real time.

6. The method for animal tissue slicing path planning based on machine learning according to claim 5, characterized in that: The digital twin model establishes a time-synchronized virtual mirror. Whenever the physical slicing device completes a cutting action, the mesh in the virtual model performs corresponding cell deletion or attribute update to simulate the material removal process. The digital twin model calculates the heat transfer process of cutting in the tissue and reflects the parameter drift of the tissue protein denaturation or deformation characteristics caused by heat. For ultra-soft tissue samples, the digital twin model integrates a fluid-structure interaction algorithm to simulate the buoyancy support of the slicing fluid on the tissue and the fluid resistance generated by the movement of the cutting tool, so as to predict the three-dimensional displacement of the tissue in the liquid.

7. The method for animal tissue slicing path planning based on machine learning according to claim 1, characterized in that: The physical information neural network used in step four incorporates a large number of synthetic data samples generated by finite element analysis during the training phase. These synthetic data samples cover a variety of typical tissue types and slicing conditions. The loss function of the physical information neural network embeds conservation law constraints from biomechanical theory. These conservation law constraints include residuals from the momentum conservation equation, the mass conservation equation, and the energy balance equation, ensuring that the physical information neural network follows physical laws while fitting the observed data. When the strain field predicted by the physical information neural network does not satisfy the continuity equation, the network weights are forcibly adjusted through the penalty value generated by the physical loss term.

8. The method for animal tissue slicing path planning based on machine learning according to claim 7, characterized in that: When generating the initial slice path, the physical information neural network quickly scans the candidate path space and predicts the displacement vector field; The system defines the ideal coordinate sequence of the target slice plane in three-dimensional space, and calculates the deformation pre-compensation coordinate points according to the deformation pre-compensation logic. The deformation pre-compensation coordinate points are equal to the difference between the ideal coordinate points in the ideal coordinate sequence and the displacement vectors in the displacement vector field. The initial slice path generation process takes into account tissue hardness distribution, fiber orientation and local curvature changes, and the path trajectory is output as a continuous smooth space curve after high-order Bezier smoothing. A multi-objective optimization strategy is adopted in the path planning. The optimization objectives include maintaining uniform thickness deformation compensation, minimizing power consumption during the slicing process, and maximizing tool life. The system also sets a basic cutting speed and adjusts the cutting speed in real time based on the equivalent stress predicted by the physical information neural network. When the equivalent stress is greater than the preset safety stress threshold, the cutting speed is adjusted to the product of the basic cutting speed and the correction coefficient. The correction coefficient is a value less than 1 and its magnitude decreases as the stress exceeds the threshold.

9. The method for animal tissue slicing path planning based on machine learning according to claim 1, characterized in that: The actual cutting feedback data collected in step five includes tool load signal, displacement deviation value, and changes in surface reflectivity. The tool load signal is acquired by a multi-dimensional force sensor, the displacement deviation value is confirmed by a laser interferometer or high-frequency visual tracking, and the change in the reflectivity of the tissue surface is used to assess the degree of dehydration of the tissue surface or the change in micromorphology after pressure. The actual cutting feedback data is used to trigger the state correction mechanism inside the digital twin model. When there is a deviation between the actual measured cutting force and the predicted value, the parameters in the nonlinear constitutive model are adjusted in real time through the gain adjustment algorithm, and the physical information neural network is prompted to make immediate corrections to the path. The digital twin model and the physical slicing device are connected via a low-latency communication link using the real-time industrial Ethernet protocol. Through heartbeat detection and time truncation compensation logic, the synchronization error between the virtual simulation results and the physical operation is controlled within a predetermined distance threshold. The response frequency of the feedback loop is set above a predetermined frequency threshold. Multiple checks and corrections are performed during each micrometer of cutting. When the cutting trajectory is detected to deviate from the predetermined plane, a correction command is applied to the driver to perform micrometer-level rapid compensation of the tool position.

10. The method for animal tissue slicing path planning based on machine learning according to claim 1, characterized in that: The method achieves adaptation to different animal species through parameter transfer learning. When switching the species of tissue samples, the physical information neural network uses existing physical law knowledge and combines a small number of specific species samples for fine-tuning training, without having to retrain the entire system. Throughout the slicing process, the real-time images from the physical camera, the virtual rendering of the digital twin model, and the deviation heatmap reflecting the degree of agreement between the prediction and the feedback are displayed in real time through a graphical interface. In the control of the six-degree-of-freedom precision feed platform, the coordinated action of each joint motor follows the inverse kinematics logic, transforming the spatial curve into a pulse sequence of the motor, and using a feedforward control algorithm to compensate for the inertial effect of the mechanical structure, ensuring that the feed direction of the tool is consistent with the local force balance direction of the tissue.

Images