Electrical impedance imaging method, medium and equipment for dynamic muscle imaging

Abstract

The invention discloses an electrical impedance imaging method for muscle dynamic imaging, a medium and equipment, and belongs to the technical field of image processing. The method comprises the following steps: for a target, acquiring a muscle dynamic boundary voltage data sequence in a walking process; and inputting the muscle dynamic boundary voltage data sequence into a trained neural network model, reconstructing a corresponding muscle internal conductivity dynamic image sequence, and further obtaining an electrical impedance imaging result. The neural network model extracts spatial features of the muscle dynamic boundary voltage data sequence by using a convolutional layer, and predicts a time evolution mode of conductivity distribution caused by muscle activity by using a recurrent neural network layer. According to the method, the complex time sequence dynamic state of muscle activity can be effectively captured, and the electrical impedance imaging speed and the time resolution are improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and more particularly to an electrical impedance imaging method, medium and device for dynamic muscle imaging. Background Art

[0002] Electrical impedance tomography (EIT) is a novel imaging technique that infers the electrical conductivity, dielectric constant, and impedance of a body part from surface electrode measurements and uses them to generate tomographic images of that part. EIT offers advantages such as non-invasiveness, reproducibility, low cost, and functional imaging, leading to its widespread application in biomedicine, agricultural geology, and industrial testing. For example, in medicine, EIT can be used to monitor the functional status of organs such as the lungs and heart, as well as to detect changes in muscle tissue morphology and properties. The high temporal resolution of EIT enables continuous monitoring imaging to capture dynamic changes in muscle activity. This is invaluable for studying muscle responses and states during different exercise regimes. By monitoring changes in muscle impedance in real time, detailed analysis of the timing and pattern of muscle activity can be performed, revealing the dynamic characteristics of muscle contraction and relaxation.

[0003] EIT reconstructs an image of the body's electrical conductivity, or impedance distribution, by applying a small alternating current (typically at frequencies ranging from a few kHz to several hundred kHz) to the body's surface and then measuring the voltage difference between multiple locations. The basic principles of EIT include: 1) Current injection: Current is injected into the body through one or more selected pairs of electrodes. This current is a safe, low-amplitude alternating current. 2) Voltage measurement: As the current passes through different tissues, different voltage drops are generated due to the different electrical conductivities of these tissues (e.g., muscle, fat, blood, and air). Additional electrodes are used to measure these voltage changes. 3) Data acquisition: During a complete scan, current is injected through different electrode combinations, and the corresponding voltage responses are recorded. This process generates a large number of voltage-current data points, known as boundary measurement data. 4) Image reconstruction: A mathematical model is used to convert the boundary measurement data into an estimate of the internal conductivity distribution. This step involves complex computations because it requires solving a so-called "inverse problem."

[0004] The inverse problem involves inferring the conductivity distribution within an object (the system's characteristics) given the surface voltage measurement (output) and the applied current pattern (input). This is a significantly more difficult problem, primarily because multiple different conductivity distributions can produce the same surface voltage measurement, even small measurement noise can cause significant fluctuations in the solution (i.e., unstable solutions), and the nonlinear relationship between conductivity and voltage increases the complexity of the solution. To address this inverse problem, researchers have developed various reconstruction algorithms, such as iterative reconstruction algorithms. These algorithms are typically based on a physical model (such as a finite element model (FEM) or a boundary element model (BEM)) that describes the current conduction within the object being measured (the forward problem model). Then, through iterative optimization, a guess about the conductivity distribution within the model is continuously adjusted to minimize the difference between the boundary voltage calculated based on this guess (the forward problem solution) and the actual measured boundary voltage. Commonly used iterative algorithms include linearization methods and nonlinear iterative optimization methods. These iterative algorithms are theoretically mature and, under certain conditions, can produce reconstructed images with high spatial resolution. However, they have significant limitations when applied to dynamic muscle imaging, primarily due to the following reasons: 1) High computational complexity and time consumption. Each iteration requires solving one or more large linear equations (derived from the direct problem or the calculation of the Jacobian matrix), which is computationally intensive. For dynamic imaging requiring high temporal resolution (such as capturing millisecond-scale muscle contractions), these algorithms often cannot keep up with the reconstruction speed (single-frame reconstruction may take seconds or even minutes), making real-time or near-real-time monitoring impossible.

[0005] 2) Sensitive to initial guesses and prior information. The convergence and final results of the iterative algorithm may depend on the guessed value of the initial conductivity distribution and the selected prior model or regularization parameter. Improper selection may lead to slow convergence or trapping in local optimality.

[0006] 3) High model accuracy requirements. Algorithm performance depends on the accuracy of the problem model. For a complex, heterogeneous, and dynamically changing object like the human body, establishing accurate geometric and physical models is a challenge in itself.

[0007] To overcome the shortcomings of traditional iterative algorithms, particularly in dynamic imaging and real-time requirements, researchers have explored new approaches. For example, patent application CN111938641A proposes a bioelectrical impedance imaging method using an optimized adaptive extended Kalman filter (EKF). This approach aims to address the issues of low noise immunity, potential loss of accuracy, and filter divergence that may exist in traditional EKF (Extended Kalman Filter) applications in EIT, while also balancing reconstruction accuracy and computational complexity. However, although designed for dynamic systems, the EKF still requires matrix operations (such as the Jacobian matrix, covariance updates, and Kalman gain calculations), which limits its maximum achievable temporal resolution and may present a bottleneck for capturing rapid muscle activity (e.g., milliseconds). Furthermore, the EKF relies on simplified state transition models (such as random walks) and local linearization assumptions, making it difficult to accurately capture the complex, nonlinear, and patterned dynamic changes in muscles during activities such as the gait cycle. For another example, patent application CN119055216A proposes a method for estimating upper arm muscle force based on a dual-branch self-attention network. This solution more accurately estimates upper arm muscle force by fusing EIT image information with raw voltage frequency domain information and utilizing a self-attention mechanism. However, this solution relies on pre-reconstructed EIT images using traditional algorithms (such as Gauss-Newton), which becomes a speed bottleneck for the overall process and does not provide fully end-to-end dynamic imaging. Furthermore, while the attention mechanism can process sequences, it is less effective in simulating the temporal evolution and naturalness of state memory for muscle activity with clear sequential and phased characteristics, such as the gait cycle.

[0008] After analysis, the existing technology mainly has the following defects: 1) The conflict between imaging speed and temporal resolution. During exercise (such as walking, running, or rehabilitation training), muscles undergo rapid contractions and relaxations, causing significant changes in their internal conductivity distribution on timescales ranging from milliseconds to seconds. To accurately capture these transient changes, EIT imaging algorithms must have sufficiently high temporal resolution—that is, the ability to reconstruct a single image frame in a short period of time. However, while traditional EIT reconstruction algorithms, particularly numerical methods based on iterative optimization (such as the Gauss-Newton method and the conjugate gradient method), can achieve relatively high image quality, their computational process typically involves repeatedly solving a complex forward problem (calculating the boundary voltage from a given conductivity distribution) and an inverse problem (inverting the internal conductivity distribution from the measured boundary voltage). This iterative process is computationally intensive, resulting in long single-frame reconstruction times, typically on the order of seconds or even minutes. Such long computational delays fall far short of the requirements for real-time or near-real-time monitoring of rapid muscle dynamics. Consequently, the reconstructed image sequences fail to accurately reflect the true temporal course of muscle activity, losing critical dynamic information.

[0009] 2) Training data lacks physiological relevance and dynamic realism. Most current neural network imaging algorithms for EIT suffer from a common limitation during training: they typically rely on simulation data generated from simplified finite element models of regular geometric shapes (such as circles, rectangles, or triangles). These general-purpose models often fail to accurately simulate the complex anatomical structure of human tissue, especially in areas like the thigh that contain multiple tissues (muscle groups, fat, bone) and have irregular morphologies. More importantly, these models are typically static and fail to reflect the complex dynamic changes that occur in muscles during physiological activity (such as the different phases of the walking gait cycle). These changes, such as the increase in cross-sectional area and morphological changes (bulging) caused by muscle contraction, or the thinning and lengthening caused by stretching, can directly affect the distribution of electrical conductivity and current pathways, and are a significant source of EIT signal variation. Using training data generated by this overly simplified model that lacks dynamic and physiological relevance to train a neural network will result in an imaging algorithm that is fast. However, when applied to real, dynamic human muscle imaging tasks, its accuracy in reconstructed images, its ability to resolve specific muscle structures (such as quadriceps and hamstrings), and its ability to reflect the true dynamic behavior of muscles (such as contraction / relaxation) are greatly reduced, making it difficult to meet the needs of refined functional assessment.

[0010] 3) Insufficient capture of the temporal dependencies of muscle dynamics. Traditional iterative algorithms typically reconstruct each frame independently or use simple regularization constraints on adjacent frames. This makes it difficult to effectively utilize the strong temporal continuity and state-dependence inherent in muscle activity (such as the gait cycle). Summary of the Invention

[0011] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide an electrical impedance imaging method, medium and equipment for dynamic muscle imaging, which can not only meet the high temporal resolution requirements required for dynamic muscle imaging, but also accurately reconstruct the complex spatiotemporal conductivity changes of muscles during real movement using an EIT imaging algorithm.

[0012] According to a first aspect of the present invention, a method for electrical impedance tomography (EIT) for dynamic muscle imaging is provided. The method comprises the following steps: Aiming at the target, the dynamic boundary voltage data sequence of muscles during walking is collected; Inputting the muscle dynamic boundary voltage data sequence into a trained neural network model to reconstruct a corresponding muscle internal conductivity dynamic image sequence, thereby obtaining electrical impedance imaging results; The neural network model uses a convolutional layer to extract the spatial features of the muscle dynamic boundary voltage data sequence, and uses a recurrent neural network layer to predict the time evolution pattern of the conductivity distribution caused by muscle activity.

[0013] In one embodiment, the training set for training the neural network model is obtained according to the following steps: The gait cycle is divided into the resting state, the early stance phase, the end stance phase, and the swing phase according to the dynamic changes of the thigh muscles during walking. The resting state is characterized by the relaxation balance of the quadriceps and hamstrings. The early stance phase refers to the heel strike to the middle stage. The end stance phase refers to the toe lift-off. The swing phase refers to the foot lift-off to the ground again. Using the resting state as a benchmark model, multiple groups of models are obtained by constructing a finite element model, wherein the finite element model is composed of multiple finite element meshes, the skeleton is used as the geometric center, and the anterior quadriceps femoris and the posterior hamstring muscles dynamically evolve in a crescent-shaped form; each group of the multiple groups of models is divided into four, representing the resting state, the early stance phase, the late stance phase, and the swing phase respectively; During the dynamic phases of the early stance phase, the late stance phase, and the swing phase, the area sizes of the quadriceps and hamstrings were parameterized to simulate the continuous process of muscle morphology evolving with gait, and the electrical conductivity values of each tissue fluctuated randomly within the set normal physiological range. The surface voltage distribution of the multiple groups of models under the electrode excitation mode is calculated by finite element simulation, thereby obtaining a training set containing multiple groups of data, each group of data corresponding to the correspondence between the conductivity and voltage paired data of the four gait cycles of the resting state, the early stance phase, the late stance phase and the swing phase.

[0014] In one embodiment, the neural network model includes multiple convolutional layers, recurrent neural network layers and fully connected layers, each of the convolutional layers is connected to a nonlinear activation function and a pooling layer, the convolutional layer generates convolution features by performing convolution operations on input data and convolution kernels; the pooling layer generates a dimensionality reduction matrix by extracting set values from the convolution features; the recurrent neural network layer captures the dynamic correlation between the four gait cycles of the resting state, the early stance phase, the late stance phase and the swing phase through a memory unit mechanism.

[0015] In one embodiment, the recurrent neural network layer is a long short-term memory network layer.

[0016] According to a second aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, the steps of the electrical impedance imaging method for dynamic muscle imaging are implemented.

[0017] According to a third aspect of the present invention, a computer device is provided, comprising a memory and a processor, wherein a computer program capable of running on the processor is stored on the memory, wherein the processor implements the steps of the above-mentioned electrical impedance imaging method for dynamic muscle imaging when executing the computer program.

[0018] Compared with existing technologies, the advantages of the present invention lie in that, in addressing existing problems such as slow imaging speed, insufficient capture of temporal dependencies, and lack of physiological relevance between models and training data, the present invention provides a new electrical impedance muscle dynamic imaging method that can achieve high-speed, near-real-time muscle dynamic imaging. This method overcomes the computational speed bottlenecks of traditional iterative algorithms and filtering-based methods, leverages the advantages of deep learning end-to-end mapping, significantly shortens reconstruction time, and achieves the high temporal resolution required for monitoring rapid muscle dynamic changes. It effectively captures the complex temporal dynamics of muscle activity by introducing neural network structures specifically designed for processing sequential data (such as recurrent neural networks (RNNs), particularly long-short-term memory (LSTM) networks), enabling them to learn and utilize the continuous evolution of muscle conductivity distribution in the temporal dimension and long-range dependencies (for example, the association and transition patterns of muscle states at different stages of the walking gait cycle). Improving the physiological relevance and accuracy of imaging results: By constructing a dynamic simulation model based on physiological principles that can simulate the morphological and conductivity changes of target muscle groups (such as the quadriceps and hamstrings) during specific dynamic activities (such as the walking gait cycle), and using this model to generate a large amount of training data with high physiological relevance and dynamic realism, the neural network is trained. This enables the trained algorithm to more accurately reconstruct the spatiotemporal distribution of muscles during real-world movement, improving its ability to distinguish different tissue structures and reflect muscle functional states (contraction and relaxation).

[0019] Further features and advantages of the present invention will become apparent from the following detailed description of exemplary embodiments of the present invention with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and, together with the description, serve to explain the principles of the invention.

[0021] Figure 1 is a flow chart of an electrical impedance tomography method for dynamic muscle imaging according to one embodiment of the present invention; Figure 2 This is a diagram of the RNN-LSTM (Recurrent Neural Network-Long Short-Term Memory) neural network model architecture according to one embodiment of the present invention; Figure 3 is a schematic diagram of a process for constructing a training set according to one embodiment of the present invention; Figure 42 is a schematic diagram of the process of training an RNN-LSTM neural network model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that unless otherwise specifically stated, the relative arrangement of components and steps, numerical expressions and numerical values set forth in these embodiments do not limit the scope of the present invention.

[0023] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the invention, its application, or uses.

[0024] Technologies, methods, and equipment known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered part of the specification.

[0025] In all examples shown and discussed herein, any specific values should be interpreted as merely exemplary and not limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0026] It should be noted that like reference numerals and letters refer to like items in the following figures, and therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0027] To address the low temporal resolution and insufficient physiological relevance of existing electrical impedance tomography (EIT) techniques for dynamic muscle imaging, this paper proposes a deep learning-based EIT method. This method employs a recurrent neural network (RNN)-long short-term memory (LSTM) neural network model, comprising convolutional layers and LSTM (long short-term memory) layers. The convolutional layers (and pooling layers) process input boundary voltage data (e.g., data at each time step or within a short time window) to extract spatial features of the voltage distribution, effectively filtering information and capturing patterns related to the spatial structure of the electrical conductivity distribution across tissues such as muscle, fat, and bone at the current moment. The LSTM layers process the temporally ordered sequence of features extracted by the convolutional layer. Leveraging the LSTM's unique gating mechanisms (e.g., forget gate, input gate, output gate) and memory cells, this model effectively learns and memorizes the dynamic evolution and long-term dependencies of muscle conductivity distribution over time. For example, this model can capture the continuous changes in muscle contraction and relaxation during different phases of the gait cycle (e.g., stance and swing phases), while overcoming potential gradient issues associated with standard RNNs.

[0028] Specifically, see Figure 1 As shown, the provided electrical impedance imaging method for muscle dynamic imaging includes the following steps: Step S110: Design an RNN-LSTM neural network model based on the temporal characteristics of the continuous changes in muscle state over time during walking.

[0029] Considering that when applying EIT to dynamic muscle imaging, the continuous surface voltage signal has the temporal characteristics of reflecting the continuous changes in muscle state over time, it is necessary to rely on a network model that can process serialized continuous data. In this paper, a neural network model RNN-LSTM based on recurrent neural network (RNN) and long short-term memory (LSTM) units is designed. This model has advantages in time series modeling and is particularly suitable for processing dynamic test data generated during human walking.

[0030] RNNs are artificial neural networks that maintain a "memory" of past inputs through recurrent connections. Their core advantage lies in processing sequential data. Unlike feedforward neural networks, RNNs process inputs sequentially, one time step at a time. This allows the output at each moment to depend not only on the current input but also on the historical hidden state. This property makes them particularly well-suited for tasks with temporal dependencies, such as measuring the continuous voltage signal generated by muscle activity during walking.

[0031] Sherstinsky derived the mathematical foundation of RNN from the perspective of differential equations and modeled it as a nonlinear first-order inhomogeneous ordinary differential equation: (1) in, is the state vector, is the input vector, is a nonlinear vector function, is the bias vector, and t is a continuous time variable. After discretization by the backward Euler method, the time domain expression of RNN is: (2) (3) in, , , are the weight matrices for state loop, output feedback, and input mapping, respectively, is the state bias, G(·) is the activation function, represents the output / activation vector of the previous time step, is the state vector of the previous time step, represents the state vector of the current time step, represents the output / activation vector of the previous time step, Represents the current input vector of the time step, and n represents the time step. This recursive structure enables RNN to capture temporal patterns and is suitable for modeling the continuous evolution of muscle EIT data.

[0032] However, standard RNNs suffer from the vanishing / exploding gradient problem during training. When backpropagating through time, the gradients of long sequences can decay exponentially or diverge, making it difficult to learn long-term dependencies. This is particularly critical for dynamic muscle imaging, as impedance signals can exhibit temporal correlations across multiple gait cycles, necessitating a more stable model architecture.

[0033] To overcome the limitations of standard RNNs, researchers have proposed the LSTM unit, which models long-term dependencies by introducing memory cells and three types of gating mechanisms (forget gate, input gate, and output gate). Sherstinsky further theoretically demonstrated that the LSTM's constant error carousel (CEC) can effectively alleviate the vanishing gradient problem. The core equation of the LSTM is as follows: The forget gate controls the retention ratio of historical states: (4) Where n is the time step. Is the forget gate vector (controls the forget ratio of the previous cell state, 0 = forget everything, 1 = retain everything). is the weight matrix input to the forget gate, is the current input vector. is the weight matrix from the previous cell state to the forget gate. is the weight matrix from the previous hidden state to the forget gate. is the previous hidden state vector. is the forget gate bias vector.

[0034] The input gate regulates the update of the state by new information:

[0035] in, is the input gate vector (controls the candidate value update ratio). : Weight matrix input to the input gate. is the weight matrix from the previous cell state to the input gate. is the weight matrix from the previous hidden state to the input gate. is the input gate bias vector.

[0036] State updating refers to fusing the historical state with the new input: (7) in, is the current cell state vector (core memory unit). It is to forget old information (regulated by the forget gate). Indicates the addition of new information (adjusted by the input gate).

[0037] The output gate controls the observable output: (8) (9) in, Represents the output gate vector (controls the cell state output ratio). Represents the weight matrix from input to output gate. Represents the weight matrix from the current cell state to the output gate. Represents the weight matrix from the previous hidden state to the output gate. Represents the output gate bias vector. is the Sigmoid activation function, and ⊙ is element-wise multiplication. The LSTM gating mechanism enables it to simultaneously capture short-term fluctuations and long-term trends in muscle impedance data, providing key support for dynamic imaging.

[0038] See also Figure 2 As shown in the figure, the designed RNN-LSTM neural network model consists of multiple convolutional layers, recurrent neural network layers (for example, LSTM layers), and fully connected layers. Each convolutional layer is followed by a nonlinear activation function and a pooling layer. The number of convolutional layers and convolution kernel size can be set according to actual needs.

[0039] In addition to the LSTM layer, convolutional and pooling layers are essentially specially designed data preprocessing layers. Their core function is to filter the input data and extract useful information. The convolutional layer generates new features by convolving the original input data with a convolution kernel. Since this technology was originally designed for feature extraction from image datasets, the input data must be represented in the form of a structured matrix. The convolution kernel (or filter) can be considered a miniature window relative to the input matrix, with its coefficients organized in matrix form. This window slides over the input matrix, performing a convolution operation on each subregion it passes through. The result of all operations forms a convolution matrix, the eigenvalues of which are determined by the filter coefficients and their dimensionality. By applying different convolution kernels to the input data, multiple convolutional features with higher representation power can be generated. These features are superior to the original input features, effectively improving model performance. The convolutional layer is followed by a nonlinear activation function and a pooling layer. The pooling layer acts as a downsampling method, extracting specific values from the convolutional features to generate a reduced-dimensional matrix. Its operation is similar to that of a convolutional layer. A small sliding window is used to take the values of each convolutional feature block as input. A single new value is generated as output through a defined operation (such as max pooling or average pooling, which calculates the maximum or average value of each feature block, respectively). Ultimately, the pooling layer produces a new matrix that can be viewed as a compressed version of the convolutional layer's output features. The advantage of pooling is that it enhances stability: the pooled output value is insensitive to small changes in the input, which effectively improves the model's generalization ability.

[0040] Step S120: Acquire a training set by constructing a finite element simulation model, and train an RNN-LSTM neural network model based on a set loss function to learn the mapping relationship between dynamic voltage and dynamic conductivity.

[0041] To address the lack of physiological relevance in existing training data, this paper innovatively constructs a finite element simulation model that simulates muscle changes in the human lower limb thigh during specific dynamic activities. Based on physiological knowledge, this model divides the walking gait cycle into key phases (e.g., resting state, early stance phase, end stance phase, and swing phase). The finite element simulation model explicitly simulates the physiological behavior of major muscle groups (e.g., quadriceps and hamstrings) during these phases (e.g., concentric contraction, eccentric contraction, and relaxation), resulting in morphological changes (e.g., increases and decreases in cross-sectional area, thickness, width, and shape), as well as potential changes in electrical conductivity. Model parameters (e.g., muscle area geometry and electrical conductivity) are parametrically and adaptively adjusted across gait phases to reflect the dynamic contraction-stretch cycle of the muscles. The model also incorporates other relevant tissues (e.g., bone and fat layer), taking into account their relative positions and electrical properties. Furthermore, the data generated by the model incorporates physiological variation (e.g., individual differences in muscle size and random fluctuations in tissue electrical conductivity within the normal range). This dynamic, physiologically relevant simulation model can be used to generate large amounts of training data in batches using EIDORS simulation software or other software. The generated simulation data sequences containing spatiotemporal dynamic information (voltage sequences as input and conductivity distribution sequences as target outputs) are used to train the RNN-LSTM neural network model.

[0042] In one embodiment, based on the advantages of the RNN-LSTM model in capturing temporal patterns and processing serialized continuous data, the dynamic changes of the thigh muscles during human walking are innovatively simplified into four states, and this model is used to generate training data, thereby obtaining a neural network imaging algorithm dedicated to EIT dynamic imaging of lower limb muscle tissue.

[0043] Specifically, the walking gait cycle consists of two main phases: the stance phase (foot contact) and the swing phase (foot lift-off). The thigh muscles perform their motor functions through alternating periods of concentric contraction (muscle fiber shortening to generate force), eccentric contraction (muscle fiber lengthening to brake), and relaxation. The quadriceps muscles (represented by the rectus femoris and vastus lateralis) are primarily responsible for knee extension during the early stance phase to support body weight, while the hamstring muscles (such as the biceps femoris) dominate knee flexion control and forward swing control of the lower limb from the end of the swing phase to the beginning of the stance phase. This functional antagonism and spatiotemporal coordination results in the two major muscle groups forming an anteroposteriorly opposed mechanical structure in cross-section. In a two-dimensional cross-section perpendicular to the femur, this mechanical behavior manifests as real-time changes in muscle thickness, width, and overall morphology. The cross-sectional area (CSA) of a muscle reflects the deformation characteristics of a specific muscle during contraction or extension. Although the adductor muscles participate in hip joint stability, their mechanical contribution is manifested as static support in the two-dimensional cross-sectional model, and the dynamic deformation amplitude is small. Therefore, they can be classified as a secondary muscle group and simplified.

[0044] During the gait cycle, the cross-sectional area and morphological changes of the thigh muscles show significant temporal characteristics. Based on a two-dimensional cross-sectional model (centered on the femur, with the quadriceps crescent region on the anterior side and the symmetrical structure of the hamstrings on the posterior side), the mechanical behavior of the muscles can be divided into the following four stages: (1) Resting state The resting state serves as a baseline, demonstrating balanced relaxation of the quadriceps and hamstrings. At this point, the two muscle groups form a symmetrical crescent-shaped structure around the femur, with the CSA at baseline (no significant bulging or concavity), and a flattened elliptical cross-sectional profile. This state provides a geometric reference for subsequent morphological changes during contraction and extension.

[0045] (2) Early stance phase (from heel strike to mid-stance) The anterior quadriceps muscle contracts concentrically to drive knee extension to support body weight. During this process, the muscle fibers shorten along their longitudinal axis. Due to the nearly incompressible nature of muscle tissue (the principle of volume conservation), their cross-section bulges outward, increasing both thickness and width. Ultrasound studies have shown that the rectus femoris CSA increases by 5–15% during isometric contraction. Dynamic contraction during walking may induce similar morphological changes: the cross-section transforms from a flat crescent shape at rest to a thickened, convex shape, with the medial femoral region remaining fixed and the lateral belly expanding significantly. Simultaneously, the posterior hamstrings control knee extension velocity through eccentric contraction, passively lengthening the muscle fibers and resulting in a decrease in CSA of approximately 3–8%. Studies have shown that muscle fiber lengthening under tension causes the muscle contour to extend laterally, resulting in a thinner, more tapered crescent structure.

[0046] (3) End of stance phase (toes off the ground) As the center of gravity shifts forward, the quadriceps gradually relax, and the CSA returns to baseline due to passive stretching, returning to a flattened crescent shape. However, the bulging is less pronounced than during the early stance phase. At this point, the hamstrings switch to a concentric contraction mode, with muscle fibers shortening to drive knee flexion and provide propulsion for the forward swing of the lower limb. During contraction, the CSA increases by approximately 8-12%, and the cross-sectional shape thickens to a rounded, compact crescent shape, with the muscle belly bulging outward from the posterior femur. This phenomenon is consistent with the predictions of Blemker et al.'s computational model of muscle deformation: during concentric contraction, muscle fiber aggregation leads to localized stress concentration, causing the cross-sectional profile to become more circular.

[0047] (4) Swing phase (from foot leaving the ground to foot touching the ground again) During the early to mid-swing phase, both muscle groups are in a state of low activation, with CSA remaining at baseline and exhibiting a symmetrical crescent-shaped cross-section. Toward the end of the swing phase, the hamstrings contract eccentrically to decelerate the forward swing of the lower limb, resulting in a brief decrease in CSA of 2-5%. The morphology then bulges briefly before quickly returning to a flattened shape in preparation for foot contact. The quadriceps, on the other hand, undergo a slight eccentric activity before contact, with a 2-5% decrease in CSA due to pre-activation elongation, resulting in a slightly stretched, flattened morphology.

[0048] This rapid contraction-stretching cycle causes the CSA to continuously oscillate with the gait rhythm, dynamically shifting between thickening and widening and thinning and narrowing. Based on this gait cycle division, the dynamic EIT of the human thigh during the walking cycle is modeled, for example, using a finite element model constructed using EIDORS, a specialized EIT simulation software. The model consists of a 2,500-element mesh, with the skeleton serving as the geometric center. The quadriceps femoris and hamstrings on the anterior side dynamically evolve in a crescent-shaped pattern, and the morphological parameters of muscle, fat, and bone adaptively adjust with each gait phase.

[0049] Specifically, an automated modeling process was developed for simulation model construction. Using the resting state as the baseline model, a customized script was written to batch generate 4,000 models in EIDORS. Each model group was divided into four groups, representing the resting state, early stance phase, end stance phase, and swing phase. During the dynamic phases (early stance phase, end stance phase, and swing phase), the quadriceps and hamstring areas were parameterized to simulate the continuous evolution of muscle morphology with gait. The muscle cross-sectional areas at rest were physiologically variable by ±50% across the models. Furthermore, the electrical conductivity of each tissue was randomly varied within the normal physiological range (muscle: 0.2-0.5 S / m, fat: 0.02-0.06 S / m, bone: 0.008-0.02 S / m) to enhance data diversity. Finite element simulations were used to calculate the surface voltage distribution of each model under a 16-electrode excitation mode, resulting in a training set consisting of 4,000 data sets, each corresponding to conductivity-voltage paired data from four phases. The construction process of the training set can be found in Figure 3 shown.

[0050] After obtaining the training set, the RNN-LSTM model is trained until the set loss standard is met, such as minimizing the loss function as the training optimization goal.

[0051] Specifically, the RNN-LSTM model training process fully combines the temporal characteristics of gait cycle data with deep learning technology. The model uses LSTM (Long Short-Term Memory) as its core architecture, and uses its unique "memory unit" mechanism to capture the dynamic relationship between the four phases of gait (early stance phase, late stance phase, swing phase, and resting state). Figure 4As shown in the figure, during training, data is fed into the model in batches, with each batch containing complete gait sequences for 32 model groups. Each sequence includes paired voltage input and conductivity output data for four models. The input voltage features are derived from the surface voltage distribution in a 16-electrode adjacent excitation and adjacent measurement mode. This involves exciting a single pair of electrodes and measuring the remaining electrodes in pairs, for a total of 208 voltage data points. The output target is the conductivity distribution matrix of muscle, fat, and bone at the corresponding stage.

[0052] The RNN-LSTM model is trained using a standard deep learning training process, such as the Adam optimizer and the mean squared error (MSE) loss function. Backpropagation (including backpropagation along the LSTM time steps) is used to adjust network weights, enabling the model to learn the complex nonlinear mapping from dynamic voltage input to dynamic conductivity output. Specifically, the initial learning rate is set to 0.001, and the parameter update step size is adaptively adjusted to balance convergence speed and stability. The mean squared error (MSE) loss function directly measures the difference between the model's predicted conductivity distribution and the true simulated value. The entire training process involves 125 iterations. In each iteration, the model sequentially performs forward propagation, loss calculation, and backpropagation. First, the voltage data is processed by a convolutional layer to extract spatial features. Feature dimensionality reduction is then achieved through max pooling. The data is then organized into sequences of four time steps and fed into the LSTM layer to generate conductivity predictions. The mean squared error loss function is then calculated based on the predicted and true conductivity values. Finally, the error signal is transmitted backward along the model topology through the backpropagation mechanism (including gradient backpropagation along the time steps of the LSTM layer) to update the network model weight parameters and obtain optimized model parameters.

[0053] Step S130: For the actually collected muscle dynamic boundary voltage data sequence, the trained RNN-LSTM neural network model is used to obtain the imaging result.

[0054] After the RNN-LSTM neural network model is trained, it can be directly used to process real-time or offline muscle dynamic boundary voltage data sequences collected from EIT hardware systems (such as wearable devices), and quickly reconstruct the corresponding dynamic image sequences of muscle internal conductivity, achieving high-speed, high-accuracy and highly physiologically relevant muscle dynamic visualization.

[0055] It should be noted that the neural network model training process described in the present invention can be performed offline on a server or in the cloud. By embedding the trained model into an electronic device, real-time muscle electrical impedance imaging can be achieved. The electronic device can be a terminal device or a server, including any terminal device such as a mobile phone, tablet computer, personal digital assistant (PDA), point of sale (POS), in-vehicle computer, or smart wearable device (smartwatch, virtual reality glasses, virtual reality helmet, etc.). Servers include, but are not limited to, application servers or web servers, and can be standalone servers, cluster servers, or cloud servers.

[0056] To further validate the effectiveness of this invention, measurements of the human thigh were taken under static conditions using commercial EIT equipment and processed using this invention. The results demonstrate that this invention can successfully distinguish major thigh tissues (e.g., bone, fat, quadriceps, and hamstrings), demonstrating basic tissue resolution capabilities and laying the foundation for subsequent applications in complete dynamic imaging. By combining a specially designed RNN-LSTM network architecture with high-quality training data generated from a physiologically based dynamic muscle activity simulation model, this invention achieves fast, accurate, and physiologically relevant electrical impedance dynamic muscle imaging.

[0057] In summary, compared with the prior art, the present invention has the following advantages: 1) This paper designs a specific hybrid neural network architecture (RNN-LSTM), employing convolutional neural network layers to extract the spatial features of the EIT boundary voltage at each time point. Long short-term memory (LSTM) units within a recurrent neural network (RNN) are specifically selected to process the time series of these spatial features. LSTM effectively processes sequential data, capturing temporal dependencies (such as the evolution of muscle states within the gait cycle) and overcoming the vanishing / exploding problem of standard RNNs. This hybrid RNN-LSTM architecture effectively integrates the spatial and temporal information of voltage data, enabling fast end-to-end mapping from dynamic voltage measurement sequences to internal conductivity image sequences, significantly improving the speed of dynamic imaging.

[0058] 2) Unlike existing static or simplified models, this invention constructs a physiologically based dynamic simulation model, designing a finite element simulation model that simulates the physiological changes of specific muscle groups (e.g., the quadriceps and hamstrings) during specific dynamic activities (e.g., the walking gait cycle). This simulation model divides the target dynamic activity (walking gait) into multiple key physiological phases (e.g., resting state, early / late stance phase, and swing phase). Based on the known physiological behavior of the muscles in each phase (e.g., concentric / eccentric contraction, relaxation, etc.), the geometry (e.g., CSA, thickness, shape) and / or conductivity of the muscle regions in the model are parametrically and adaptively adjusted, thereby highly realistically simulating the dynamic changes of the muscles.

[0059] 3) The present invention utilizes dynamic, physiologically relevant simulation models to generate a large-scale training dataset containing spatiotemporal dynamic information and physiological variability, and uses this dataset to train the designed RNN-LSTM neural network model, which can learn the real dynamic laws of muscles and improve imaging accuracy and physiological relevance.

[0060] 4) This invention utilizes end-to-end deep learning mapping. After training, image reconstruction primarily involves the forward propagation of the neural network, which is inherently faster than the EKF method in terms of computational speed. Although designed for dynamic systems, the EKF still requires matrix operations (Jacobian, covariance updates, and Kalman gain calculations), which limits its maximum achievable temporal resolution. However, this invention more easily enables near-real-time or even real-time dynamic muscle imaging, better meeting the needs of capturing rapid muscle activity. This speed advantage stems primarily from the end-to-end deep learning architecture itself and its design specifically tailored for dynamic imaging. Furthermore, the architecture of this invention is independent of precision selection and can also apply mixed-precision techniques based on deployment requirements. This invention is completely end-to-end, going directly from voltage data to dynamic images, avoiding any time-consuming traditional reconstruction preprocessing steps, thus offering a fundamental advantage in imaging speed.

[0061] 5) The LSTM network used in this invention, through its unique gating mechanism and recurrent structure, can directly learn complex long-term and short-term temporal dependencies from data, without the need for pre-simplified physical models or linearization. This allows for more effective modeling and prediction of the true evolution of muscle states over time. The introduction of LSTM provides a structural advantage in processing muscle dynamic EIT data with strong temporal continuity. Through its gradually updated hidden states and memory cells, it can better capture the correlations and dynamic trends between frames, more naturally simulating the temporal evolution process and state memory. This is more effective for understanding muscle activity with clear sequence and stage characteristics, such as the gait cycle.

[0062] 6) This invention innovatively constructs a dynamic simulation model that reflects the physical changes (CSA, morphology, and possibly electrical conductivity) caused by muscle contraction (concentric / eccentric) and relaxation during walking. By dividing the gait cycle into physiological phases and adjusting the model based on muscle function parameters, training data is generated that closely approximates the actual physiological process. Using this unprecedented, high-fidelity dynamic physiological simulation data for training, the neural network learns the true spatiotemporal electrical behavior of muscles during real-world movement. This directly translates into significantly improved reconstructed image accuracy, enabling clearer discrimination between different muscle groups (such as the quadriceps and hamstrings) and their states (bulging, thinning, etc.) during different phases of movement, resulting in dynamic imaging results with greater physiological significance and clinical value. This is difficult to achieve with existing methods that rely on general, static, or simplified model training.

[0063] The present invention may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present invention.

[0064] A computer-readable storage medium can be a tangible device that can hold and store instructions used by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure within a groove on which instructions are stored, and any suitable combination thereof. As used herein, a computer-readable storage medium is not to be construed as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse passing through a fiber optic cable), or an electrical signal transmitted through an electrical wire.

[0065] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions to be stored in the computer-readable storage medium in each computing / processing device.

[0066] The computer program instructions for performing the operations of the present invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages such as Smalltalk, C++, Python, and conventional procedural programming languages such as "C" or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, the state information of the computer-readable program instructions is used to personalize an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), so that the electronic circuit can execute the computer-readable program instructions, thereby implementing various aspects of the present invention.

[0067] Various aspects of the present invention are described herein with reference to flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present invention. It should be understood that each block of the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, can be implemented by computer-readable program instructions.

[0068] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine, so that when these instructions are executed by the processor of the computer or other programmable data processing device, a device is generated that implements the functions / actions specified in one or more blocks in the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium, where these instructions cause the computer, programmable data processing device, and / or other device to operate in a specific manner. Thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing various aspects of the functions / actions specified in one or more blocks in the flowchart and / or block diagram.

[0069] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operational steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to implement the functions / actions specified in one or more blocks in the flowchart and / or block diagram.

[0070] The flowcharts and block diagrams in the accompanying drawings show the possible implementation architecture, functions and operations of the systems, methods and computer program products according to multiple embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, program segment or part of an instruction, and the module, program segment or part of the instruction contains one or more executable instructions for implementing the specified logical function. In some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart, can be implemented by a dedicated hardware-based system that performs the specified function or action, or can be implemented by a combination of dedicated hardware and computer instructions. It is well known to those skilled in the art that implementation by hardware, implementation by software, and implementation by a combination of software and hardware are all equivalent.

[0071] While various embodiments of the present invention have been described above, the foregoing description is intended to be illustrative, non-exhaustive, and not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is selected to best explain the principles of the embodiments, their practical applications, or technological improvements in the marketplace, or to enable others skilled in the art to understand the embodiments disclosed herein. The scope of the present invention is defined by the appended claims.

Claims

1. An electrical impedance tomography method for dynamic muscle imaging, characterized in that: The following steps are involved: Aiming at the target, the dynamic boundary voltage data sequence of muscles during walking is collected; Inputting the muscle dynamic boundary voltage data sequence into a trained neural network model to reconstruct a corresponding muscle internal conductivity dynamic image sequence, thereby obtaining electrical impedance imaging results; The neural network model uses a convolutional layer to extract the spatial features of the muscle dynamic boundary voltage data sequence, and uses a recurrent neural network layer to predict the time evolution pattern of the conductivity distribution caused by muscle activity.

2. The method according to claim 1, characterized in that The training set for training the neural network model is obtained according to the following steps: The gait cycle is divided into the resting state, the early stance phase, the end stance phase, and the swing phase according to the dynamic changes of the thigh muscles during walking. The resting state is characterized by the relaxation balance of the quadriceps and hamstrings. The early stance phase refers to the heel strike to the middle stage. The end stance phase refers to the toe lift-off. The swing phase refers to the foot lift-off to the ground again. Taking the resting state as a benchmark model, multiple groups of models are obtained by constructing a finite element model, wherein the finite element model is composed of multiple finite element meshes, the skeleton is used as the geometric center, and the quadriceps femoris on the front and the hamstring muscle on the back are dynamically evolved in a crescent shape; Each of the multiple groups of models is divided into four groups, representing the resting state, the early stance phase, the end of the stance phase and the swing phase respectively; During the dynamic phases of the early stance phase, the late stance phase, and the swing phase, the area sizes of the quadriceps and hamstrings were parameterized to simulate the continuous process of muscle morphology evolving with gait, and the electrical conductivity values of each tissue fluctuated randomly within the set normal physiological range. The surface voltage distribution of the multiple groups of models under the electrode excitation mode is calculated by finite element simulation, thereby obtaining a training set containing multiple groups of data, each group of data corresponding to the correspondence between the conductivity and voltage paired data of the four gait cycles of the resting state, the early stance phase, the late stance phase and the swing phase.

3. The method according to claim 2, characterized in that The neural network model includes multiple convolutional layers, recurrent neural network layers and fully connected layers. Each convolutional layer is connected to a nonlinear activation function and a pooling layer. The convolutional layer generates convolution features by performing convolution operations on input data and convolution kernels; the pooling layer generates a dimensionality reduction matrix by extracting set values from the convolution features; the recurrent neural network layer captures the dynamic correlation between the four gait cycles of the resting state, the early stance phase, the late stance phase and the swing phase through a memory unit mechanism.

4. The method according to claim 1, wherein The recurrent neural network layer is a long short-term memory network layer.

5. The method according to claim 2, characterized in that During the training of the neural network model, data is input in batches, each batch contains the complete gait sequences of the multiple models, and each sequence contains the voltage characteristics and conductivity output pairing data of the four gait cycles of the resting state, the early stance phase, the late stance phase and the swing phase. The voltage characteristics come from the surface voltage distribution of multiple electrodes under adjacent excitation and adjacent measurement modes. The output target of the neural network model is the muscle, fat and bone conductivity distribution matrix corresponding to the gait cycle.

6. The method according to claim 1, characterized in that The loss function for training the neural network model adopts mean square error loss, which is used to measure the difference between the predicted conductivity distribution and the true value.

7. The method according to claim 3, characterized in that The pooling layer uses a maximum pooling operation to perform feature dimensionality reduction.

8. The method according to claim 2, characterized in that The finite element model is constructed using the simulation software EIDORS.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer device comprising a memory and a processor, wherein a computer program capable of being run on the processor is stored in the memory, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

Images