A deep learning driven embryonic cell three-dimensional configuration generation processing method

Abstract

A deep learning driven embryo cell three-dimensional configuration generation processing method, comprising: acquiring an original embryo cell two-dimensional microscopic image sequence, and preprocessing the original embryo cell two-dimensional microscopic image sequence to obtain a standard input frame, a three-dimensional implicit field representation space is composed of a plurality of discrete spatial coordinate points and corresponding density values; the preprocessed standard input frame is input into a pre-trained space-time graph convolutional neural network module, and a multi-scale spatial feature map and a time sequence dynamic feature vector are extracted through the space-time graph convolutional neural network module.The present application collects two-dimensional projection images at multiple time points through a confocal microscope or a light sheet fluorescence microscope, combines a non-local mean filtering algorithm with a histogram equalization process, forms a standardized sequence with uniform resolution, pixel depth value and time stamp index, eliminates optical shot noise and enhances the visibility of low-contrast region structure, and provides a high-quality data basis for subsequent feature extraction.

Description

Technical Field

[0001] This invention belongs to the field of embryonic cell processing technology, specifically a deep learning-driven method for generating and processing three-dimensional embryonic cell configurations. Background Technology

[0002] Embryonic cell 3D conformation reconstruction and dynamic evolution analysis are research techniques in reproductive medicine, developmental biology, and cell engineering. Precise, continuous, and high-fidelity 3D spatial structure modeling of embryonic cells is a crucial foundation for elucidating embryonic development mechanisms, assessing embryonic status, and optimizing assisted reproductive technologies. Currently, mainstream techniques in the industry primarily rely on microscopic imaging equipment to acquire 2D image sequences of embryonic cells, combined with traditional image reconstruction algorithms to reconstruct the 3D structure of cells, applied in scenarios such as embryo screening and cell development research. However, with the increasing demand for high-precision cell analysis, the industry's requirements for embryonic cell imaging data quality, 3D conformation reconstruction accuracy, and spatiotemporal dynamic feature capture capabilities are continuously rising. The limitations of traditional embryonic cell 3D conformation generation and processing technologies are gradually becoming apparent, exhibiting the following technical problems: First, existing technologies mostly employ conventional filtering and grayscale adjustment methods to process embryonic microscopic images, which cannot effectively filter out optical shot noise generated during microscopic imaging. Furthermore, they struggle to specifically enhance the structural information of low-contrast regions in embryonic cells, resulting in image sequences with residual noise, blurred edges, and uneven brightness levels. Moreover, traditional processing methods cannot achieve unified calibration of image resolution and pixel depth across multiple time points, nor can they regularize the temporal index. The imaging dataset has a low degree of standardization and contains a large amount of invalid interference information, directly leading to poor reliability of the foundational data for subsequent feature extraction and reducing the overall accuracy of 3D model reconstruction.

[0003] Secondly, traditional reconstruction models can only extract static spatial features of embryonic cells, and cannot effectively explore the deep-dimensional spatial relationships of embryonic cells during continuous development, making it difficult to capture the dynamic evolution of cell morphology over time. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, the present invention provides a deep learning-driven method for generating and processing three-dimensional embryonic cell configurations, so as to at least partially solve the above-mentioned technical problems.

[0005] The technical solution adopted in this invention is as follows: This invention proposes a deep learning-driven method for generating and processing three-dimensional embryonic cell configurations, comprising: Step 1: Obtain a two-dimensional microscopic image sequence of primitive embryonic cells, and preprocess the two-dimensional microscopic image sequence of primitive embryonic cells to obtain a standard input frame; Step 2: Construct a three-dimensional implicit field representation space based on voxel meshes. The three-dimensional implicit field representation space consists of several discrete spatial coordinate points and their corresponding density values. Step 3: Input the preprocessed standard input frame into the pre-trained spatiotemporal graph convolutional neural network module, and extract multi-scale spatial feature maps and temporal dynamic feature vectors through the spatiotemporal graph convolutional neural network module; Step 4: Use a multi-head self-attention mechanism to fuse the spatial feature map and the temporal dynamic feature vector to generate a joint feature tensor containing global context information; Step 5: Input the joint feature tensor into the decoder network. The decoder network adopts a structure of alternating upsampling layers and convolutional layers to restore the geometric details of the three-dimensional space layer by layer. Step 6: Iteratively update the density values ​​and surface normal vector parameters of each spatial coordinate point in the three-dimensional implicit field representation space through the neural radiation field rendering loss function and optimization algorithm until convergence is obtained to obtain the initial three-dimensional configuration model. Step 7: Apply topology optimization constraints to smooth the surface and fill the voids in the initial three-dimensional configuration model, and output the final three-dimensional configuration model of the embryonic cell.

[0006] In one embodiment of the present invention, the step of obtaining the two-dimensional microscopic image sequence of primitive embryonic cells specifically includes: Two-dimensional projection images of embryonic cells at different developmental stages were acquired using confocal microscopy or light-sheet fluorescence microscopy at different time slices, with each time slice corresponding to a specific physiological time point and the corresponding focal length position information recorded. The acquired raw data were denoised using a nonlocal mean filtering algorithm to eliminate optical shot noise while preserving high-frequency details at cell edges. Histogram equalization was then performed to enhance the visibility of internal cell structures in low-contrast areas, resulting in a standardized two-dimensional microscopic image sequence with uniform resolution, uniform pixel depth values, and uniform timestamp indexes.

[0007] In one embodiment of the present invention, the step of constructing a three-dimensional implicit field representation space based on a voxel mesh specifically includes: A three-dimensional Cartesian coordinate system is defined, with its origin located at the centroid of the embryonic cell. The x-axis, y-axis, and z-axis correspond to the lateral, longitudinal, and vertical defocusing directions of the microscope stage, respectively. A cubic region surrounding the entire embryonic cell and its surrounding medium is delineated within the three-dimensional Cartesian coordinate system, and this cubic region is uniformly divided into N×N×N basic voxel units, where N is a positive integer and is set according to the required 3D reconstruction accuracy. Each basic voxel unit is assigned a unique three-dimensional spatial coordinate index (i,j,k) and an initialized scalar field value phi(i,j,k). The scalar field value is used to characterize whether the voxel is located inside or outside the cell. When phi(i,j,k) is greater than a preset threshold, it is determined to be an intracellular region; otherwise, it is a background region, thereby establishing a discretized three-dimensional implicit field data carrier.

[0008] In one embodiment of the present invention, the specific architecture of the spatiotemporal graph convolutional neural network module includes: The network consists of a residual connection block, a 3D spatial convolutional subnetwork, and a long short-term memory recurrent subnetwork, arranged in series. The residual connection block receives standard input frames from the preprocessing module and directly transmits high-resolution shallow features through short jump paths to reduce gradient vanishing. The 3D spatial convolutional subnetwork contains three 3D convolutional layers, each followed by a batch normalization layer and a ReLU activation function, used to infer spatial correlations in the depth dimension from 2D images. The long short-term memory recurrent subnetwork receives the feature sequence output from the previous layer, controls the information flow through forget gates, input gates, and output gates, captures the morphological evolution of embryonic cells over time, and outputs a time-dependent dynamic feature encoding sequence.

[0009] In one embodiment of the present invention, the specific implementation of the multi-head self-attention mechanism for fusing spatial feature maps and temporal dynamic feature vectors is as follows: The spatial feature map is flattened into a one-dimensional vector sequence as the query matrix Q, and the temporal dynamic features are vectorized into a key matrix K and a value matrix V. The attention weight matrix A is calculated using the formula: softmax(Q*K^T / sqrt(d_k)), where d_k is the feature dimension of the key matrix. The attention weight matrix A is multiplied by the value matrix V to obtain a weighted context vector. The context vector is added to the original temporal dynamic feature vector and further transformed using a multilayer perceptron to generate a joint feature tensor aligned with the dimension of the spatial feature map, thereby achieving deep fusion of global semantic information and local geometric features.

[0010] In one embodiment of the present invention, the structure of the decoder network with alternating stacked upsampling layers and convolutional layers is specifically as follows: First, the spatial dimension of the joint feature tensor is doubled and the number of channels is halved through transpose convolution. Then, two consecutive 3D convolutional layers are connected. The kernel size of each 3D convolutional layer is 3×3×3, the stride is 1, and the padding is 1 to ensure that the receptive field covers adjacent voxels. The above upsampling and convolution combination module is repeated three times. After each upsampling, the spatial resolution is doubled. Finally, the predicted volume data with the same spatial size as the 3D implicit field representation is obtained. Skip connections are introduced at the input of each upsampling module to directly stitch the spatial feature map of the corresponding layer into the current upsampled feature map to recover high-frequency edge information.

[0011] In one embodiment of the present invention, in the step of iteratively updating the neural radiation field rendering loss function and optimization algorithm, the optimization algorithm used is the Adam adaptive moment estimation algorithm, whose learning rate is dynamically adjusted using a cosine annealing strategy; the neural radiation field rendering loss function consists of three parts: L = L_rec + λ1L_geo + λ2L_reg; where L_rec is the reconstruction loss, calculated by comparing the differences between the virtual 3D view synthesized from different perspectives and the real multi-angle 2D image; L_geo is the geometric consistency loss, which measures the surface smoothness by calculating the continuity of the scalar field gradient between adjacent voxels; L_reg is the regularization loss, which limits the fluctuation range of the scalar field value to prevent overfitting; λ1 and λ2 are weight coefficients, and the optimal values ​​are determined by cross-validation.

[0012] In one embodiment of the present invention, the specific process of applying topology optimization constraints to perform surface smoothing and void filling on the initial three-dimensional configuration model includes: First, the value surface of the initial three-dimensional configuration model is extracted, that is, the set of voxels where all scalar field values ​​are at a preset zero threshold, forming a triangular mesh model. Laplace smoothing iteration is performed on the triangular mesh model. In each iteration, the current position of the vertex is moved a certain proportion of the distance to the average position of its neighboring vertices to eliminate the step-like jagged effect caused by discretization. Then, the connected regions in the triangular mesh model are detected, and isolated small triangular clusters enclosed inside the cell body are identified. These regions are regarded as voids. For the identified void regions, bridging surfaces are constructed based on the surrounding triangular mesh boundaries and filled.

[0013] In one embodiment of the invention, a curvature estimation step is included before performing the Laplace smoothing iteration: For each vertex in the triangular mesh model, the discrete differential operator at its location is calculated to obtain the principal curvature and the mean curvature. Based on the calculated curvature distribution, the intensity parameter of the smoothing iteration is adaptively adjusted. The smoothing coefficient is reduced in the high curvature region to preserve details, and the smoothing coefficient is increased in the low curvature region to quickly remove noise. After a finite number of iterations, the topological properties of the mesh are checked to ensure that no mesh self-intersection or flipping occurs. If an anomaly is detected, the process reverts to the state of the previous iteration step.

[0014] In one embodiment of the present invention, after outputting the final three-dimensional embryonic cell model, the method further includes model format conversion and visualization export steps: The processed triangular mesh model is converted into a stereolithography file format (STL) or a polygon file format (OBJ) for easy reading by 3D printing or professional medical image processing software. Simultaneously, the normal vector distribution of the model surface is calculated, and a pseudo-color mapping map is generated based on the normal vectors to highlight the fine structural regions within the cells. All key parameters, including but not limited to voxel mesh size, optimal number of iterations, loss function convergence curve, and confidence score, are packaged and stored in a metadata file, which is then associated with the final 3D configuration model file to form a complete digital archive of embryonic cells.

[0015] The beneficial effects of the technical solution of this invention are as follows: This invention acquires two-dimensional projection images at multiple time points using a confocal microscope or a light-sheet fluorescence microscope. It combines nonlocal mean filtering algorithm with histogram equalization processing to form a standardized sequence with uniform resolution, pixel depth value and timestamp index. This eliminates optical shot noise and enhances the visibility of low-contrast area structures, providing a high-quality data foundation for subsequent feature extraction.

[0016] This invention divides the embryonic cell and its surrounding medium into N×N×N discrete units in a three-dimensional Cartesian coordinate system defined by voxel meshes. Each unit is assigned a spatial coordinate index and an initial scalar field value, thus establishing an implicit field data carrier that can characterize the distribution inside and outside the cell, and realizing the mapping from continuous physical space to discrete computational space.

[0017] The spatiotemporal graph convolutional neural network module of this invention receives standardized input frames, retains shallow high-resolution features through residual connection blocks, infers spatial correlations in the depth dimension using a three-dimensional spatial convolutional sub-network, and then controls the information flow through the forget gate, input gate, and output gate in the long short-term memory recurrent sub-network to capture morphological evolution patterns and generate a dynamic feature encoding sequence with temporal dependence.

[0018] The multi-head self-attention mechanism of this invention flattens the spatial feature map into a query matrix Q, and vectorizes the temporal dynamic features into a key matrix K and a value matrix V. The attention weight matrix A is obtained by calculating softmax(Q*K^T / sqrt(d_k)). The weighted context vector is obtained by multiplying it with the value matrix V. Then, it is transformed by a multilayer perceptron to achieve deep fusion of global semantic information and local geometric features, generating a joint feature tensor containing complete context information. The decoder network uses transposed convolution to expand the spatial dimension and halve the number of channels. Then, it connects two consecutive 3×3×3 three-dimensional convolutional layers and repeats the upsampling and convolution combination module three times. Each time, the spatial resolution is increased by a factor of two, and finally, the predicted volume data with the same size as the implicit field is obtained. At the same time, the spatial feature map of the corresponding level is stitched into the current upsampled feature map through skip connections to restore high-frequency edge information.

[0019] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0020] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a schematic diagram of the framework of the deep learning-driven embryonic cell three-dimensional configuration generation and processing method proposed in the embodiments of the present invention. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0022] The following describes a deep learning-driven method for generating three-dimensional embryonic cell configurations according to an embodiment of the present invention, with reference to the accompanying drawings.

[0023] like Figure 1 As shown, this embodiment of the invention provides a deep learning-driven method for generating and processing three-dimensional embryonic cell configurations, including: Step 1: Obtain the two-dimensional microscopic image sequence of the original embryonic cells and preprocess the two-dimensional microscopic image sequence of the original embryonic cells to obtain the standard input frame; Step 2: Construct a three-dimensional implicit field representation space based on voxel meshes. The three-dimensional implicit field representation space consists of several discrete spatial coordinate points and their corresponding density values. Step 3: Input the preprocessed standard input frame into the pre-trained spatiotemporal graph convolutional neural network module, and extract multi-scale spatial feature maps and temporal dynamic feature vectors through the spatiotemporal graph convolutional neural network module; Step 4: Use a multi-head self-attention mechanism to fuse the spatial feature map and the temporal dynamic feature vector to generate a joint feature tensor containing global context information; Step 5: Input the joint feature tensor into the decoder network. The decoder network adopts a structure of alternating upsampling layers and convolutional layers to restore the geometric details of the three-dimensional space layer by layer. Step 6: Iteratively update the density values ​​and surface normal vector parameters of each spatial coordinate point in the three-dimensional implicit field representation space through the neural radiation field rendering loss function and optimization algorithm until convergence is obtained to obtain the initial three-dimensional configuration model. Step 7: Apply topology optimization constraints to smooth the surface and fill the voids in the initial 3D configuration model, and output the final 3D configuration model of the embryonic cell.

[0024] In specific applications, this invention acquires a two-dimensional microscopic image sequence of original embryonic cells and preprocesses it to obtain a standard input frame, eliminating optical speckle noise and background interference, ensuring that the signal-to-noise ratio of subsequent network input data meets the distribution requirements of model training, and providing a consistent input benchmark for feature extraction.

[0025] Next, a three-dimensional implicit field representation space based on voxel grids is constructed. This space serves as the carrier for all subsequent deep learning operations. By mapping continuous spatial coordinates to discrete voxel units and assigning initial density values, the system establishes a potential structural framework in memory. This allows the network to dynamically adjust the attributes of each spatial point according to the implicit semantics of the two-dimensional image, thereby adapting to the irregular and variable morphological changes of embryonic cells.

[0026] Subsequently, the preprocessed standard input frames are fed into a pre-trained spatiotemporal graph convolutional neural network module. This module, designed to address the morphological evolution of embryonic cells over time, utilizes concatenated residual connections, a 3D spatial convolutional sub-network, and a long short-term memory recurrent sub-network to achieve deep feature extraction from the data. The 3D spatial convolutional layer captures local texture and contour information in a single frame, while the long short-term memory network focuses on processing temporal dependencies, identifying structural changes caused by cell division and migration, ensuring that the extracted features encompass both the current geometric morphology and the historical evolutionary trajectory.

[0027] Furthermore, after extracting multi-scale spatial feature maps and temporal dynamic feature vectors, a multi-head self-attention mechanism is used to fuse them. In this stage, the system flattens the spatial features into a query matrix and transforms the temporal features into key and value matrices. By calculating attention weights, adaptive alignment between global contextual information and local high-frequency details is achieved. The multi-head mechanism allows the model to simultaneously focus on information associations in different subspaces, such as mapping specific temporal evolution patterns to local cell membrane boundaries, generating a joint feature tensor containing rich semantic information. Through weighted aggregation, the representational ability of key structural features is strengthened, providing a highly robust input signal for subsequent 3D decoding.

[0028] Furthermore, the generated joint feature tensor is then fed into the decoder network, which adopts an architecture that alternates between upsampling layers and convolutional layers. The sampling operation gradually expands the spatial dimension of the feature map to restore the original resolution. The convolutional layers are used to extract the geometric patterns in these high-resolution features. Skip connections are introduced into the decoder structure to directly stitch the shallow features of the encoder layers into the corresponding upsampling layers, ensuring that the final output three-dimensional geometric details are sharp enough to accurately reflect the microstructure of embryonic cells, such as the cell nucleus boundary and cytoplasmic texture.

[0029] Specifically, a neural radiation field rendering loss function and optimization algorithm are introduced to iteratively update the density values ​​and surface normals in the 3D implicit field. The optimization process combines an adaptive moment estimation algorithm with a cosine annealing strategy to dynamically adjust the learning rate, balancing convergence speed and stability. The loss function consists of reconstruction loss, geometric consistency loss, and regularization loss. The reconstruction loss constrains the consistency between the synthetic viewpoint and the real viewpoint, the geometric consistency loss ensures the smooth transition of neighborhood voxel gradients, and the regularization loss prevents numerical oscillations. By minimizing these composite losses, the network gradually corrects the parameters of the implicit field until it converges to obtain a preliminary 3D configuration model that conforms to the observation data.

[0030] Finally, a triangular mesh composed of zero-level sets is extracted, and Laplacian smoothing iteration is performed. The smoothing intensity is adaptively adjusted according to the vertex curvature to eliminate step noise while maintaining sharp features. Subsequently, mesh connectivity is checked, internal voids are filled, and broken regions are bridged to ensure the generation of a closed, watertight surface. The final output 3D model of embryonic cells not only has an accurate geometric shape but also maintains good topological integrity, which can be used for subsequent volumetric quantization analysis or 3D printing demonstration.

[0031] In one specific implementation, the steps for obtaining a two-dimensional microscopic image sequence of primitive embryonic cells include: acquiring two-dimensional projection images of embryonic cells at different developmental stages under different time slices using a confocal microscope or a light-sheet fluorescence microscope, wherein each time slice corresponds to a specific physiological time point, and the corresponding focal length position information is recorded; denoising the acquired raw data using a nonlocal mean filtering algorithm to eliminate optical shot noise while preserving high-frequency details at cell edges; and then performing histogram equalization to enhance the visibility of the internal cell structure in low-contrast areas, forming a standardized two-dimensional microscopic image sequence with uniform resolution, uniform pixel depth value, and uniform timestamp index.

[0032] The specific steps for constructing a three-dimensional implicit field representation space based on voxel meshes include: defining a three-dimensional Cartesian coordinate system with its origin located at the centroid of the embryonic cell, and the x-axis, y-axis, and z-axis corresponding to the lateral movement direction, longitudinal movement direction, and vertical defocus direction of the microscope stage, respectively; delineating a cubic region surrounding the entire embryonic cell and its surrounding medium within the three-dimensional Cartesian coordinate system, and uniformly dividing the cubic region into N×N×N basic voxel units, where N is a positive integer and is set according to the required three-dimensional reconstruction accuracy; assigning a unique three-dimensional spatial coordinate index (i,j,k) and an initialized scalar field value phi(i,j,k) to each basic voxel unit. The scalar field value is used to characterize whether the voxel is located inside or outside the cell. When phi(i,j,k) is greater than a preset threshold, it is determined to be an intracellular region; otherwise, it is a background region, thereby establishing a discretized three-dimensional implicit field data carrier.

[0033] In practical applications of this invention, during the initial stage of embryonic cell 3D reconstruction, the quality of the acquired raw data determines the reliability of the subsequent 3D model. Acquiring time-slice images of different developmental stages using confocal microscopy or light-sheet fluorescence microscopy is fundamental to establishing temporal-spatial correlation. Each time slice not only records the two-dimensional projection morphology at a specific physiological moment but also implicitly contains depth distribution clues in the vertical dimension by recording focal length information, providing a practical optical sampling basis for subsequent 3D reconstruction. Addressing the common optical shot noise in microscopic imaging, directly inputting raw data leads to feature distortion extracted by the neural network. A nonlocal mean filtering algorithm is introduced to perform denoising. This algorithm performs a weighted average by searching for all patch regions similar to the current pixel block in the image. Compared to traditional Gaussian blur, it can suppress random noise while perfectly preserving high-frequency details of the cell membrane edge structure, avoiding the smooth loss of key geometric information. Next, histogram equalization addresses the low contrast problem caused by uneven fluorescent labeling or light attenuation. By remapping the grayscale distribution of the entire image, the internal fine structures originally hidden in the background gain sufficient dynamic range, thereby improving visibility. The standardized two-dimensional microscopic image sequence formed after the above processing has a uniform resolution, pixel depth value and timestamp index. The consistency of data format eliminates the domain offset effect caused by fluctuations in equipment parameters, provides a homogeneous input stream for deep learning models, and ensures that the feature extraction module does not need to adjust the preprocessing logic for each acquisition, thereby improving the versatility and robustness of the system.

[0034] Furthermore, in the process of transforming two-dimensional images into three-dimensional space, constructing a three-dimensional implicit field representation space based on voxel meshes is a key step in achieving geometric continuity. First, a three-dimensional Cartesian coordinate system is defined, with the origin set at the centroid of the embryonic cell. The coordinate axes are aligned with the lateral, longitudinal, and vertical defocusing directions of the microscope stage to ensure the correspondence between the virtual space and the real physical space, avoiding structural misalignment due to coordinate deviations. Subsequently, a cubic region surrounding the embryonic cell and its surrounding medium is delineated within the coordinate system. This region encompasses all potential structural variations and is uniformly divided into N×N×N basic voxel units, where the value of N is determined based on the required 3D reconstruction accuracy. A higher N value signifies higher spatial resolution, enabling the capture of finer morphological details. Each basic voxel unit is assigned a unique three-dimensional spatial coordinate index (i,j,k) and initialized with a scalar field value phi(i,j,k). The scalar field value serves as a parameter characterizing voxel attributes, and its magnitude determines whether a point belongs to the cell interior or the external background. By setting a preset threshold, when phi(i,j,k) is greater than the threshold, it is determined to be an intracellular region, and otherwise it is a background region, thus discretizing the continuous biological tissue morphology into a computable mathematical field.

[0035] In one specific implementation, the spatiotemporal graph convolutional neural network module includes the following architecture: a residual connection block, a three-dimensional spatial convolutional subnetwork, and a long short-term memory recurrent subnetwork arranged in series. The residual connection block receives standard input frames from the preprocessing module and directly transmits high-resolution shallow features through short jump paths to reduce gradient vanishing. The three-dimensional spatial convolutional subnetwork contains three three-dimensional convolutional layers, each followed by a batch normalization layer and a ReLU activation function, used to infer spatial correlations in the depth dimension from two-dimensional images. The long short-term memory recurrent subnetwork receives the feature sequence output from the previous layer, controls the information flow through forget gates, input gates, and output gates, captures the morphological evolution of embryonic cells over time, and outputs a dynamic feature encoding sequence with temporal dependence.

[0036] The specific implementation of the multi-head self-attention mechanism for fusing spatial feature maps and temporal dynamic feature vectors is as follows: the spatial feature map is flattened into a one-dimensional vector sequence as the query matrix Q, and the temporal dynamic feature vectors are vectorized into a key matrix K and a value matrix V; the attention weight matrix A is calculated using the formula: softmax(Q*K^T / sqrt(d_k)), where d_k is the feature dimension of the key matrix; the attention weight matrix A is multiplied by the value matrix V to obtain a weighted context vector; the context vector is added to the original temporal dynamic feature vector, and further transformed through a multilayer perceptron to generate a joint feature tensor aligned with the dimension of the spatial feature map, thereby achieving deep fusion of global semantic information and local geometric features.

[0037] In specific applications, this invention constructs a spatiotemporal graph convolutional neural network module that includes residual connection blocks, a three-dimensional spatial convolutional subnetwork, and a long short-term memory recurrent subnetwork. It simultaneously captures local geometric details and global temporal dynamics. The standard input frame is first connected to the residual connection block. The module directly transmits high-resolution shallow features without nonlinear transformation to subsequent layers through short jump paths, ensuring that the subtle information of the bottom edge and texture can still be preserved after multiple convolution operations, providing basic support for high-level semantic understanding.

[0038] Subsequently, the data enters the 3D spatial convolutional sub-network. Although the input is a 2D image sequence, by introducing a virtual spatial channel in the third dimension or fusing displacement information from adjacent frames, the three 3D convolutional layers can infer spatial correlations in the depth direction from pixel-level data. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. Batch normalization accelerates model convergence and reduces internal covariate bias, while ReLU introduces nonlinear factors to enhance the model's ability to express complex nonlinear mappings, enabling the network to construct feature maps rich in spatial context information. These feature maps not only contain the cell outlines at the current time step but also implicitly contain potential structural continuity.

[0039] Meanwhile, the long short-term memory recurrent subnetwork receives the above-mentioned spatiotemporal feature sequences and uses its unique forget gate, input gate, and output gate structure to regulate the information flow. The forget gate determines which irrelevant information to discard from the historical state, the input gate updates the new content at the current moment, and the output gate controls the state that is finally passed to the next time step. This can effectively solve the problem of long-term memory loss in long sequence dependencies, accurately capture the evolutionary laws of division, migration, and morphological reconstruction of embryonic cells over time, and thus output a dynamic feature encoding sequence with strong temporal dependence.

[0040] Furthermore, to organically combine spatial structure with temporal dynamics, a multi-head self-attention mechanism is employed to fuse spatial feature maps and temporal dynamic feature vectors. The flattened spatial feature map is used as the query matrix Q, representing the local location information that needs to be enhanced; the temporal dynamic feature vectors are transformed into a key matrix K and a value matrix V, representing the global temporal evolution pattern. The attention weight matrix A is obtained by calculating softmax(Q*K^T / sqrt(d_k)), quantifying the contribution of different time steps to a specific spatial location.

[0041] In one specific implementation, the structure of the decoder network with alternating stacked upsampling and convolutional layers is as follows: First, the spatial dimension of the joint feature tensor is doubled and the number of channels is halved through transpose convolution. Then, two consecutive three-dimensional convolutional layers are connected. The kernel size of each three-dimensional convolutional layer is 3×3×3, the stride is 1, and the padding is 1 to ensure that the receptive field covers adjacent voxels. Subsequently, the above upsampling and convolution combination module is repeated three times. After each upsampling, the spatial resolution is doubled, and finally, the predicted volume data with the same spatial size as the three-dimensional implicit field representation is obtained. Skip connections are introduced at the input of each upsampling module to directly stitch the spatial feature map of the corresponding layer into the currently upsampled feature map to recover high-frequency edge information.

[0042] In practical applications, this invention utilizes a transposed convolution operation to expand the spatial dimension of the joint feature tensor, doubling the physical size of the features while halving the number of channels. This achieves spatial sparsity expansion of information and channel dimension compression integration. Subsequently, the data enters a processing unit consisting of two consecutive 3D convolutional layers. Each 3D convolutional layer uses a small 3×3×3 convolutional kernel, configured with a stride of 1 and padding of 1. This ensures that each output voxel is associated only with its immediate neighboring input voxels. The local receptive field ensures that the network maintains computational efficiency when processing large-scale 3D data, while avoiding boundary blurring caused by excessively large nonlocal dependencies.

[0043] Furthermore, each iteration doubles the spatial resolution. After four rounds of processing, the spatial resolution of the predicted volume data finally reaches a size that is completely consistent with the three-dimensional implicit field representation space, thus ensuring a one-to-one correspondence between the output grid and the underlying data carrier. The recovery of high-frequency information depends on the skip connections introduced at the input of each upsampling module. Since the downsampling process often leads to the attenuation or loss of high-frequency details of texture and edge, the skip connections stitch the original spatial feature map of the corresponding level in the encoder to the current upsampled feature map. The shallow features retain the fine contours and sharp boundaries of the cell membrane and cell nucleus structure, making up for the deficiency of deep semantic features, which have strong global semantics but lack local geometric precision.

[0044] In one specific implementation, the optimization algorithm used in the step of iteratively updating the neural radiation field rendering loss function and optimization algorithm is the Adam adaptive moment estimation algorithm, whose learning rate is dynamically adjusted using a cosine annealing strategy. The neural radiation field rendering loss function consists of three parts: L = L_rec + λ1L_geo + λ2L_reg; where L_rec is the reconstruction loss, calculated by comparing the differences between the virtual 3D view synthesized from different perspectives and the real multi-angle 2D image; L_geo is the geometric consistency loss, which measures the surface smoothness by calculating the continuity of the scalar field gradient between adjacent voxels; L_reg is the regularization loss, which limits the fluctuation range of the scalar field value to prevent overfitting; λ1 and λ2 are weight coefficients, and their optimal values ​​are determined through cross-validation.

[0045] In practical applications, the optimization process of this invention relies on the Adaptive Moment Estimation Algorithm (Adam) to perform parameter iteration. This algorithm combines first-order and second-order moment estimation of the gradient, enabling differentiated step size adjustments for different parameter update requirements, thus maintaining training stability when processing high-dimensional sparse voxel data. Simultaneously, a cosine annealing strategy is introduced to dynamically adjust the learning rate. Initially, a larger learning rate is used to quickly traverse the flat regions of the loss surface, and then the learning rate is gradually reduced as iterations progress, allowing the model to converge finely deep within the search space, avoiding getting trapped in local optima or generating numerical oscillations.

[0046] Furthermore, to comprehensively constrain the generation quality of the 3D implicit field, a composite objective function L = L_rec + λ1L_geo + λ2L_reg is constructed, which includes reconstruction loss, geometric consistency loss, and regularization loss. The reconstruction loss L_rec serves as the driving term, synthesizing 2D projected images from different virtual perspectives through the rendering pipeline. These images are then compared pixel-level with acquired real multi-angle 2D microscopic images to calculate the mean square error or perceptual loss, forcing the density distribution in the implicit field to accurately reproduce the observed cell morphology and texture details. However, relying solely on color information for matching can easily lead to surface jitter or non-manifold structures. Therefore, a geometric consistency loss L_geo is introduced as an auxiliary constraint. This loss quantifies the surface smoothness by calculating the continuity of the scalar field φ(i,j,k) gradient between adjacent voxels, encouraging smooth transitions in density changes between adjacent spaces, eliminating staircase artifacts caused by discrete sampling, and ensuring the geometric topology rationality and continuity of the generated cell surface.

[0047] In one specific implementation, the process of applying topology optimization constraints to perform surface smoothing and void filling on the initial three-dimensional configuration model includes: first, extracting the value surface of the initial three-dimensional configuration model, i.e., the set of voxels whose scalar field values ​​are at a preset zero threshold, to form a triangular mesh model; performing Laplace smoothing iteration on the triangular mesh model, in each iteration moving the current position of the vertex to the average position of its neighboring vertices by a certain proportion to eliminate the step-like jagged effect caused by discretization; then detecting the connected domains in the triangular mesh model, identifying isolated small triangular clusters enclosed inside the cell body, and treating these regions as voids; for the identified void regions, constructing bridging surfaces based on the surrounding triangular mesh boundaries and filling them.

[0048] In practical applications, this invention transforms implicit mathematical fields into explicit geometric structures, providing a foundation for subsequent topological operations. Due to the inherent characteristics of voxelization and the minute density fluctuations generated during neural radiation field rendering, the generated triangular mesh surface often exhibits high-frequency jagged edges and an unnatural roughness. Based on local geometric neighborhood relationships, in each iteration, the current vertex position is moved a specific proportion of the average position of its neighboring vertices. A smoothing strategy based on divergence operators can uniformly distribute vertex spacing, eliminating the step-like effect caused by discrete sampling, making the curvature changes of the cell surface more consistent, thereby improving the visual quality and geometric rationality of the model while maintaining overall morphological invariance.

[0049] Furthermore, while achieving surface smoothness, connected regions in the triangular mesh model are detected. All triangular faces are traversed using depth-first search or breadth-first search algorithms to identify isolated clusters of small triangles that are completely closed and not connected to the external host. These isolated regions correspond to cavities or noise artifacts within embryonic cells that have not been effectively reconstructed. For the identified voids, a boundary repair strategy is employed to fill them. Specifically, the edge boundary ring of the void region is located, and the consistency of the normal vectors of the surrounding triangular meshes is analyzed to determine the inner and outer orientations. Subsequently, based on the boundary ring, a bridging surface with the minimum area is constructed, or polygon patches are directly generated to merge the isolated faces with the host mesh. This ensures that the final output 3D configuration model is a closed, watertight, and topologically correct polyhedron, meeting the stringent requirements for geometric integrity in subsequent biomechanical simulations or quantitative volume analysis.

[0050] In one specific implementation, before performing the Laplace smoothing iteration, a curvature estimation step is included: for each vertex in the triangular mesh model, the discrete differential operator at its location is calculated to obtain the principal curvature and the mean curvature; based on the calculated curvature distribution, the intensity parameter of the smoothing iteration is adaptively adjusted, reducing the smoothing coefficient in high curvature regions to preserve details and increasing the smoothing coefficient in low curvature regions to quickly remove noise; after a finite number of iterations, the topological properties of the mesh are checked to ensure that no mesh self-intersection or flipping occurs, and if an anomaly is detected, the process reverts to the state of the previous iteration step.

[0051] After outputting the final 3D embryonic cell model, the process includes model format conversion and visualization export steps: converting the processed triangular mesh model into a stereolithography file format (STL) or a polygon file format (OBJ) for 3D printing or professional medical image processing software to read; simultaneously, calculating the normal vector distribution on the model surface and generating a pseudo-color mapping map based on the normal vectors to highlight the fine structural regions within the cells; and packaging all key parameters, including but not limited to voxel mesh size, optimal number of iterations, loss function convergence curve, and confidence score, into a metadata file, which is then associated with the final 3D model file to form a complete digital archive of the embryonic cell.

[0052] In practical applications, this invention reduces the smoothing coefficient at cell edges and in high-curvature regions of membrane folds to limit vertex displacement and preserve subtle morphological structures. Conversely, it increases the smoothing coefficient in low-curvature, flat regions of the cell body to accelerate noise attenuation and promote surface smoothing. After a preset number of iterations, a topological integrity check is performed to detect whether the mesh exhibits self-intersections, normal vector flips, or non-manifold edge anomalies. If any of these geometrically illegal conditions are detected, the system automatically reverts to the previous valid iteration state, ensuring the mathematical rigor and topological correctness of the final output model and preventing rendering errors or analysis deviations caused by numerical errors.

[0053] Furthermore, after generating and optimizing the 3D configuration model, a format conversion and visualization export process is executed to achieve standardized data application. The internally represented triangular mesh model is converted into the universal stereolithography file STL or polygon file OBJ format. Both formats are widely compatible with 3D printing equipment and mainstream medical image processing software such as 3DSlicer or MeshLab, supporting downstream biomechanical simulation, physical modeling, or teaching demonstration needs. Simultaneously, the normal vector distribution of each vertex on the model surface is calculated, and a pseudo-color mapping map is generated using the spatial orientation differences of the normal vectors. High curvature regions typically correspond to different rates of normal change. Color coding visually highlights the fine structures, fissures, or protrusions within cells, assisting researchers in quickly identifying key morphological features.

[0054] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0055] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A deep learning-driven method for generating and processing three-dimensional embryonic cell configurations, characterized in that, include: Step 1: Obtain a two-dimensional microscopic image sequence of primitive embryonic cells, and preprocess the two-dimensional microscopic image sequence of primitive embryonic cells to obtain a standard input frame; Step 2: Construct a three-dimensional implicit field representation space based on voxel meshes. The three-dimensional implicit field representation space consists of several discrete spatial coordinate points and their corresponding density values. Step 3: Input the preprocessed standard input frame into the pre-trained spatiotemporal graph convolutional neural network module, and extract multi-scale spatial feature maps and temporal dynamic feature vectors through the spatiotemporal graph convolutional neural network module; Step 4: Use a multi-head self-attention mechanism to fuse the spatial feature map and the temporal dynamic feature vector to generate a joint feature tensor containing global context information; Step 5: Input the joint feature tensor into the decoder network. The decoder network adopts a structure of alternating upsampling layers and convolutional layers to restore the geometric details of the three-dimensional space layer by layer. Step 6: Iteratively update the density values ​​and surface normal vector parameters of each spatial coordinate point in the three-dimensional implicit field representation space through the neural radiation field rendering loss function and optimization algorithm until convergence is obtained to obtain the initial three-dimensional configuration model. Step 7: Apply topology optimization constraints to smooth the surface and fill the voids in the initial three-dimensional configuration model, and output the final three-dimensional configuration model of the embryonic cell.

2. The deep learning-driven embryonic cell three-dimensional configuration generation and processing method according to claim 1, characterized in that, The steps for obtaining the two-dimensional microscopic image sequence of primitive embryonic cells specifically include: Two-dimensional projection images of embryonic cells at different developmental stages were acquired using confocal microscopy or light-sheet fluorescence microscopy at different time slices, with each time slice corresponding to a specific physiological time point and the corresponding focal length position information recorded. The acquired raw data were denoised using a nonlocal mean filtering algorithm to eliminate optical shot noise while preserving high-frequency details at cell edges. Histogram equalization was then performed to enhance the visibility of internal cell structures in low-contrast areas, resulting in a standardized two-dimensional microscopic image sequence with uniform resolution, uniform pixel depth values, and uniform timestamp indexes.

3. The deep learning-driven embryonic cell three-dimensional configuration generation and processing method according to claim 1, characterized in that, The steps for constructing the three-dimensional implicit field representation space based on voxel mesh specifically include: A three-dimensional Cartesian coordinate system is defined, with its origin located at the centroid of the embryonic cell. The x-axis, y-axis, and z-axis correspond to the lateral, longitudinal, and vertical defocusing directions of the microscope stage, respectively. A cubic region surrounding the entire embryonic cell and its surrounding medium is delineated within the three-dimensional Cartesian coordinate system, and this cubic region is uniformly divided into N×N×N basic voxel units, where N is a positive integer and is set according to the required 3D reconstruction accuracy. Each basic voxel unit is assigned a unique three-dimensional spatial coordinate index (i,j,k) and an initialized scalar field value phi(i,j,k). The scalar field value is used to characterize whether the voxel is located inside or outside the cell. When phi(i,j,k) is greater than a preset threshold, it is determined to be an intracellular region; otherwise, it is a background region, thereby establishing a discretized three-dimensional implicit field data carrier.

4. The deep learning-driven embryonic cell three-dimensional configuration generation and processing method according to claim 1, characterized in that, The specific architecture of the spatiotemporal graph convolutional neural network module includes: The network consists of a residual connection block, a 3D spatial convolutional subnetwork, and a long short-term memory recurrent subnetwork, arranged in series. The residual connection block receives standard input frames from the preprocessing module and directly transmits high-resolution shallow features through short jump paths to reduce gradient vanishing. The 3D spatial convolutional subnetwork contains three 3D convolutional layers, each followed by a batch normalization layer and a ReLU activation function, used to infer spatial correlations in the depth dimension from 2D images. The long short-term memory recurrent subnetwork receives the feature sequence output from the previous layer, controls the information flow through forget gates, input gates, and output gates, captures the morphological evolution of embryonic cells over time, and outputs a time-dependent dynamic feature encoding sequence.

5. The deep learning-driven embryonic cell three-dimensional configuration generation and processing method according to claim 4, characterized in that, The specific implementation of the multi-head self-attention mechanism for fusing spatial feature maps and temporal dynamic feature vectors is as follows: The spatial feature map is flattened into a one-dimensional vector sequence as the query matrix Q, and the temporal dynamic features are vectorized into a key matrix K and a value matrix V. The attention weight matrix A is calculated using the formula: softmax(Q*K^T / sqrt(d_k)), where d_k is the feature dimension of the key matrix. The attention weight matrix A is multiplied by the value matrix V to obtain a weighted context vector. The context vector is added to the original temporal dynamic feature vector and further transformed using a multilayer perceptron to generate a joint feature tensor aligned with the dimension of the spatial feature map, thereby achieving deep fusion of global semantic information and local geometric features.

6. The deep learning-driven embryonic cell three-dimensional configuration generation and processing method according to claim 1, characterized in that, The structure of the decoder network, which alternately stacks upsampling layers and convolutional layers, is specifically as follows: First, the spatial dimension of the joint feature tensor is doubled and the number of channels is halved through transpose convolution. Then, two consecutive 3D convolutional layers are connected. The kernel size of each 3D convolutional layer is 3×3×3, the stride is 1, and the padding is 1 to ensure that the receptive field covers adjacent voxels. The above upsampling and convolution combination module is repeated three times. After each upsampling, the spatial resolution is doubled. Finally, the predicted volume data with the same spatial size as the 3D implicit field representation is obtained. Skip connections are introduced at the input of each upsampling module to directly stitch the spatial feature map of the corresponding layer into the current upsampled feature map to recover high-frequency edge information.

7. The deep learning-driven embryonic cell three-dimensional configuration generation and processing method according to claim 1, characterized in that, In the step of iteratively updating the neural radiation field rendering loss function and optimization algorithm, the optimization algorithm used is the Adam adaptive moment estimation algorithm, whose learning rate is dynamically adjusted using a cosine annealing strategy. The neural radiation field rendering loss function consists of three parts: L = L_rec + λ1L_geo + λ2L_reg; where L_rec is the reconstruction loss, calculated by comparing the differences between the virtual 3D view synthesized from different perspectives and the real multi-angle 2D image; L_geo is the geometric consistency loss, which measures the surface smoothness by calculating the continuity of the scalar field gradient between adjacent voxels; L_reg is the regularization loss, which limits the fluctuation range of the scalar field value to prevent overfitting; λ1 and λ2 are weight coefficients, and the optimal values ​​are determined through cross-validation.

8. The deep learning-driven embryonic cell three-dimensional configuration generation and processing method according to claim 1, characterized in that, The specific process of applying topology optimization constraints to perform surface smoothing and void filling on the initial 3D configuration model includes: First, the value surface of the initial three-dimensional configuration model is extracted, that is, the set of voxels where all scalar field values ​​are at a preset zero threshold, forming a triangular mesh model. Laplace smoothing iteration is performed on the triangular mesh model. In each iteration, the current position of the vertex is moved a certain proportion of the distance to the average position of its neighboring vertices to eliminate the step-like jagged effect caused by discretization. Then, the connected regions in the triangular mesh model are detected, and isolated small triangular clusters enclosed inside the cell body are identified. These regions are regarded as voids. For the identified void regions, bridging surfaces are constructed based on the surrounding triangular mesh boundaries and filled.

9. The deep learning-driven embryonic cell three-dimensional configuration generation and processing method according to claim 1, characterized in that, Before performing the Laplace smoothing iteration, a curvature estimation step is also included: For each vertex in the triangular mesh model, the discrete differential operator at its location is calculated to obtain the principal curvature and the mean curvature. Based on the calculated curvature distribution, the intensity parameter of the smoothing iteration is adaptively adjusted. The smoothing coefficient is reduced in the high curvature region to preserve details, and the smoothing coefficient is increased in the low curvature region to quickly remove noise. After a finite number of iterations, the topological properties of the mesh are checked to ensure that no mesh self-intersection or flipping occurs. If an anomaly is detected, the process reverts to the state of the previous iteration step.

10. The deep learning-driven embryonic cell three-dimensional configuration generation and processing method according to claim 1, characterized in that, After outputting the final three-dimensional model of embryonic cells, the process also includes model format conversion and visualization export steps: The processed triangular mesh model is converted into a stereolithography file format (STL) or a polygon file format (OBJ) for easy reading by 3D printing or professional medical image processing software. Simultaneously, the normal vector distribution of the model surface is calculated, and a pseudo-color mapping map is generated based on the normal vectors to highlight the fine structural regions within the cells. All key parameters, including but not limited to voxel mesh size, optimal number of iterations, loss function convergence curve, and confidence score, are packaged and stored in a metadata file, which is then associated with the final 3D configuration model file to form a complete digital archive of embryonic cells.

Images