Unconditional automatic generation method for long-range projection neuron morphology
By generating biologically plausible neuron morphologies using a denoised diffusion probability model and a skeletonized connection algorithm, the problems of low computational efficiency and insufficient diversity in existing technologies are solved, achieving efficient, diverse, and biologically plausible neuron morphology generation.
Patent Information
- Application Number
- CN202511458030.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-13
- Publication Date
- 2026-01-09
AI Technical Summary
Existing methods for generating neuron morphology are difficult to effectively capture the overall spatial distribution patterns of neurons at the structural level, have low computational efficiency, and rely on real data conditions that limit the diversity and biological plausibility of generated morphologies.
An end-to-end point cloud generation module based on the denoising diffusion probability model (DDPM) is used to generate point clouds. Combined with an L1-medial skeleton extractor and growth-guided connector, biologically reasonable dendritic connection topology is constructed through iterative optimization and dynamic principal component analysis. Branch smoothing optimization is performed using an improved ResNet-18 network.
It achieves efficient, diverse, and biologically plausible unconditional neuron generation, reduces computational complexity, and improves generation efficiency and morphological accuracy, making it suitable for various long-range projection neuron datasets.
Smart Images
Figure CN121303207A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer-aided biological modeling technology, specifically relating to a neuron morphology generation method based on a global-to-local hierarchical framework, which is suitable for generating realistic biological neuron data in large-scale brain network models. Background Technology
[0002] Existing neuron morphology generation methods can be divided into two categories: The first category is rule-based generation methods, which generate a tree-like structure of neurons by sampling predefined growth rules or morphological statistics. Specifically, it directly synthesizes neurons based on statistics such as branch length and angle distribution, or uses algorithms such as fractal growth models to simulate the bifurcation and expansion process of neural neurites. This method has the advantage of high computational efficiency and is suitable for large-scale generation needs.
[0003] The second category is data-driven models, which learn the morphological distribution of real neurons through deep learning technology. Typical examples include the MorphVAE model (which decomposes neurons into 3D path sequences from cell body to terminal, and then splices them together after generating paths using a variational autoencoder) and the MorphGrower model (which generates paired branches layer by layer and constructs a tree structure by combining global conditional constraints). The neuron morphology generated by this type of method has an advantage in terms of biological realism and can more closely resemble the characteristics of real data distribution.
[0004] Current neuron morphology generation technologies face three main problems: First, at the overall structural level, existing methods rely on branch-level generation strategies (such as path-by-path or layer-by-layer generation patterns) to effectively capture the overall spatial distribution patterns of neurons. The fundamental reason for this is the cumulative effect of errors caused by the branch generation order and the lack of a global directionality maintenance mechanism in long-range projective neurons.
[0005] Secondly, in terms of computational efficiency, layer-by-layer generation methods, represented by MorphGrower, need to handle the exponentially increasing combination relationships between adjacent branches, resulting in training time increasing O(2^3) with neuron complexity. k The computational bottleneck stems from the combinatorial explosion problem caused by the increase in the number of branch layers k.
[0006] Finally, at the level of generation mechanism, existing models such as MorphGrower rely on real neurons as conditional inputs, which are essentially conditional distribution modeling rather than unconditional generation frameworks. This strong dependence on real data references severely restricts the diversity of generation patterns.
[0007] In summary, existing technical solutions suffer from the following core defects: First, rule-driven methods, due to oversimplification of biological complexity, lack morphological diversity and are difficult to adapt to the needs of multiple types of neurons, requiring manual parameter adjustment and thus failing to scale efficiently. Second, data-driven models ignore inter-branch conditional dependencies, leading to error accumulation and failure in complex structure modeling when generating long-range projective neurons. Moreover, layer-by-layer generation strategies experience a surge in computational complexity due to the exponential increase in the number of branches with the number of layers, significantly limiting the generation efficiency of high-complexity neurons. Furthermore, the conditional dependence of existing models on real samples severely restricts the degree of freedom in generation and generalization ability. Finally, existing methods generally suffer from discontinuous branch morphology, directional deviation, or rigidity, lacking structural smoothness and topological rationality consistent with biological growth mechanisms. Summary of the Invention
[0008] Based on the aforementioned deficiencies of the existing technology, this invention provides an unconditional automatic generation method for the morphology of long-range projection neurons, which is significantly superior to the existing technology in terms of core indicators such as generation efficiency, diversity, and biological rationality.
[0009] To achieve the above objectives, the present invention provides an unconditional automatic generation method for the morphology of long-range projection neurons, comprising the following steps: The overall structure prediction step uses a point cloud generation module based on the denoised diffusion probability model (DDPM) to generate point clouds from end to end. The skeletonization connection step first involves the L1-medial skeleton extractor extracting the central axis features from the point cloud using an iterative optimization algorithm for skeletonization, and then the growth-guided connector equipped with a direction-aware weight calculation unit and a dynamic principal component analysis (PCA) direction update module to realize the biologically reasonable dendritic connection topology construction. The branch smoothing optimization step uses an improved ResNet-18 network as the backbone, which integrates multi-scale one-dimensional convolutional layers and residual connection structures, and is equipped with a noise injection unit for training data augmentation.
[0010] In one embodiment, the point cloud generation module based on the Denoising Diffusion Probability Model (DDPM) adopts DiT-3D as the backbone network architecture, which includes multi-level three-dimensional convolutional layers and attention mechanisms. It gradually transforms Gaussian noise into a coarse-grained three-dimensional point cloud representation of neurons. The dimensionality transformation of the point cloud data is realized through cascaded voxelization-devoxelization processing units. The output end uses a farthest point sampler to standardize the generated point cloud to a fixed number of points.
[0011] In one embodiment, the overall structure prediction step includes: The training dataset is constructed by standardizing the real neuron point cloud through farthest point sampling (FPS). The forward diffusion process gradually degrades the real point cloud into Gaussian noise through a first preset number of iterations of noise injection; The reverse process uses the DiT-3D network architecture to extract hierarchical features from the voxelized point cloud, and learns the mapping relationship from noise to point cloud through Transformer blocks; During the inference phase, random noise is denoised through a second preset number of iterations to generate a coarse-grained neuron point cloud that conforms to the true distribution.
[0012] In one embodiment, the L1-medial skeleton extractor uses an iterative optimization algorithm to extract central axis features from the point cloud for skeletonization, including: The L1-median skeleton extraction algorithm is applied to compress the generated point cloud into a density-optimized skeleton point set through iterative optimization. In each iteration, the L1 distance constraint between the skeleton point and the original point is calculated, and a repulsive force term is introduced to prevent the skeleton point from clustering.
[0013] In one embodiment, the growth-guided connector enables biologically rational dendritic connection topology construction, including: The growth-guided connector implements a growth-guided graph connection algorithm. After determining the cell body position through local density calculation, it uses an improved minimum spanning tree algorithm for topology construction. The connection weight calculation integrates Euclidean distance and principal direction alignment, where the principal direction is dynamically updated through sliding window PCA to ensure that branch growth conforms to the overall projection pattern of neurons. A branch number constraint mechanism is designed to enforce the biological rule of at most bifurcation during each connection.
[0014] In one embodiment, the branch smoothing optimization step includes: The preprocessed third set of real branch datasets are used to generate training samples by adding multi-scale noise blocks. The noise block parameters include random location intervals [s_k, e_k), amplitude coefficient a, and three-dimensional direction vector u_k. The smoothing network employs a convolutional architecture that includes residual blocks, each containing dilated convolutions to capture long-range geometric features; During training, the Adam optimizer is used, and the nonlinear mapping from noisy branches to true branches is learned through the MSE loss function; During inference, the branches generated by the skeleton reconstruction are processed segment by segment, and an overlapping sliding window strategy is used to ensure the continuity of the branch connections.
[0015] In one embodiment, it further includes: The verification process involves evaluating the global morphological similarity of the output tree topology and optimized branch geometry by calculating the Chamfer distance (CD) and Earth Mover distance (EMD). Simultaneously, it analyzes 12 morphometric indicators, including width, depth, height, maximum degree, maximum branch order, path length, path distance, path angle, branch angle, logarithmic maximum curvature, tree asymmetry, and cell exit angle, to ensure biological rationality.
[0016] The unconditional automatic generation method for the morphology of long-range projection neurons described in this invention has the following significant advantages: 1. This invention proposes a hierarchical generation framework based on global to local, which reduces computational complexity by predicting the overall structure of neurons through a denoising diffusion probability model, enhances biological rationality by combining topology optimization algorithm, and optimizes the naturalness of branch morphology by using an auxiliary convolutional network.
[0017] 2. This invention combines a point cloud-based global modeling strategy, which does not rely on real samples as conditions. This efficient, unconditional generation process is completely autonomous, expanding the diversity of the generation space. It decouples computational resource consumption from neuron complexity, ensuring stable training and inference times even when dealing with highly hierarchical long-range projective neurons.
[0018] 3. This invention creatively employs a growth direction-guided skeletonized connection algorithm, dynamically adjusting branch extension directions through principal component analysis (PCA) to avoid the scattered connection problem of the traditional minimum spanning tree (MST) algorithm. The auxiliary smoothing network repairs mutations and distortions in generated branches through a pre-trained model, making it closer to the continuity and smoothness of real neurons.
[0019] 4. This invention has universality across neuron types. The model has been validated on various long-range projection neuron datasets (IT, CT, PT). The generated morphology is consistent with the distribution of real data in key morphological indicators such as maximum path distance, branch angle, and curvature, proving its wide applicability. Attached Figure Description
[0020] The accompanying drawings described herein are for illustrative purposes only and are not intended to limit the scope of the invention in any way. Furthermore, the shapes and proportions of the components in the drawings are merely illustrative to aid in understanding the invention and do not specifically limit the shapes and proportions of the components. Those skilled in the art, guided by the teachings of this invention, can select various possible shapes and proportions to implement the invention according to specific circumstances.
[0021] Figure 1 This is a schematic diagram of the overall architecture of a method for unconditionally and automatically generating the morphology of long-range projective neurons according to an embodiment of the present invention. Figure 2 This is a schematic diagram and example of the process effect of the present invention; Figure 3 This is a schematic diagram showing the local optimization comparison of each step in the present invention; Figure 4 This is a comparison of the neuron morphology generated by the present invention with that generated by existing methods. Detailed Implementation
[0022] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.
[0023] The first embodiment of the present invention provides an unconditional automatic generation method for the morphology of long-range projection neurons, comprising the following steps: The overall structure prediction step uses a point cloud generation module based on the denoised diffusion probability model (DDPM) to generate point clouds from end to end. The skeletonization connection step first involves the L1-medial skeleton extractor extracting the central axis features from the point cloud using an iterative optimization algorithm for skeletonization, and then the growth-guided connector equipped with a direction-aware weight calculation unit and a dynamic principal component analysis (PCA) direction update module to realize the biologically reasonable dendritic connection topology construction. The branch smoothing optimization step uses an improved ResNet-18 network as the backbone, which integrates multi-scale one-dimensional convolutional layers and residual connection structures, and is equipped with a noise injection unit for training data augmentation.
[0024] In one embodiment, the point cloud generation module based on the Denoising Diffusion Probability Model (DDPM) adopts DiT-3D as the backbone network architecture, which includes multi-level three-dimensional convolutional layers and attention mechanisms. It gradually transforms Gaussian noise into a coarse-grained three-dimensional point cloud representation of neurons. The dimensionality transformation of the point cloud data is realized through cascaded voxelization-devoxelization processing units. The output end uses a farthest point sampler to standardize the generated point cloud to a fixed number of points.
[0025] In one embodiment, the overall structure prediction step includes: The training dataset is constructed by standardizing the real neuron point cloud through farthest point sampling (FPS). The forward diffusion process gradually degrades the real point cloud into Gaussian noise through a first preset number of iterations of noise injection; The reverse process uses the DiT-3D network architecture to extract hierarchical features from the voxelized point cloud, and learns the mapping relationship from noise to point cloud through Transformer blocks; During the inference phase, random noise is denoised through a second preset number of steps to generate a coarse-grained neuron point cloud that conforms to the true distribution. In one specific embodiment, the first preset number is 40,000, and the second preset number is 1,000.
[0026] In one embodiment, the L1-medial skeleton extractor uses an iterative optimization algorithm to extract central axis features from the point cloud, including: The L1-median skeleton extraction algorithm is applied to compress the generated point cloud into a density-optimized skeleton point set through iterative optimization. In each iteration, the L1 distance constraint between the skeleton point and the original point is calculated, and a repulsive force term is introduced to prevent the skeleton point from clustering.
[0027] In one embodiment, the growth-guided connector enables biologically rational dendritic connection topology construction, including: The growth-guided connector implements a growth-guided graph connection algorithm. After determining the cell body position through local density calculation, it uses an improved minimum spanning tree algorithm for topology construction. The connection weight calculation integrates Euclidean distance and principal direction alignment, where the principal direction is dynamically updated through sliding window PCA to ensure that branch growth conforms to the overall projection pattern of neurons. A branch number constraint mechanism is designed to enforce the biological rule of at most bifurcation during each connection.
[0028] In one embodiment, the branch smoothing optimization step includes: The preprocessed third set of real branch datasets are used to generate training samples by adding multi-scale noise blocks. The noise block parameters include random location intervals [s_k, e_k), amplitude coefficient a, and three-dimensional direction vector u_k. The smoothing network employs a convolutional architecture that includes residual blocks, each containing dilated convolutions to capture long-range geometric features; During training, the Adam optimizer is used, and the nonlinear mapping from noisy branches to true branches is learned through the MSE loss function; During inference, the branches generated by the skeleton reconstruction are processed segment by segment, and an overlapping sliding window strategy is used to ensure the continuity of the branch connections. In one specific embodiment, the third preset number is 17067.
[0029] In one embodiment, it further includes: The verification process involves evaluating the global morphological similarity of the output tree topology and optimized branch geometry by calculating the Chamfer distance (CD) and Earth Mover distance (EMD). Simultaneously, it analyzes 12 morphometric indicators, including width, depth, height, maximum degree, maximum branch order, path length, path distance, path angle, branch angle, logarithmic maximum curvature, tree asymmetry, and cell exit angle, to ensure biological rationality.
[0030] This invention innovatively decouples global morphology generation from local optimization. It circumvents the sequential generation bottleneck of traditional methods through a point cloud diffusion model and overcomes the challenge of converting discrete point clouds into continuous dendrites using a skeletonized connection algorithm. Combined with a data-driven branch optimization network, it achieves a significant improvement in morphological naturalness. The entire system runs on a single A100 GPU, achieving end-to-end generation through a parallel pipeline design. The complete generation cycle for each neuron is controlled within 20 minutes, representing an efficiency improvement of over 55% compared to existing methods.
[0031] This invention innovatively achieves significant efficiency improvements and breakthroughs in generation quality in the field of neuron morphology generation. Its core advantage stems from a hierarchical structural design from global to local. Compared to existing generation methods based on branch sequence modeling, this framework directly predicts the overall point cloud structure of neurons through a denoised diffusion probability model, abandoning the computational paradigm of layer-by-layer decomposition generation in traditional methods. This global modeling strategy reduces training time by approximately 55% and maintains stable computational resource consumption even as neuron complexity increases, effectively overcoming the computational bottleneck problem caused by the exponential growth of the number of branches in existing methods.
[0032] In terms of generation quality, this method achieves multi-dimensional accuracy improvement through a three-stage collaborative optimization mechanism. First, the coarse-grained point cloud generated globally by the diffusion model fully preserves the spatial topological features of neurons. Second, the skeletonized connection algorithm guided by the growth direction reconstructs the discrete point cloud into a tree-like structure that conforms to biological laws, avoiding the orientation disorder problem caused by the traditional minimum spanning tree algorithm. Finally, the innovative branch smoothing network eliminates abrupt distortions of local fibers through a pre-trained mapping mechanism, making key morphological indicators (such as maximum path distance, median branch angle, logarithm of curvature, etc.) highly aligned with the distribution of real data. Experimental data show that this method improves the nearest neighbor classification accuracy by an average of 12% compared to the existing best model, and also shows significant advantages in core indicators such as coverage and distribution similarity.
[0033] Of particular note is that this technology achieves, for the first time, unconditional generation capabilities without the need for real-world sample references. Compared to existing models that rely on conditional inputs, this method overcomes the limitations of the generation space through global probability distribution modeling, providing more flexible synthetic data support for large-scale brain network construction. Simultaneously, the end-to-end generation process avoids the tedious process of manual parameter tuning, requiring only a single GPU to rapidly synthesize complex neurons, significantly reducing the computational resource threshold. This efficient and high-precision generation capability provides reliable technical support for neuroscience research in areas such as brain connectome modeling and disease mechanism analysis.
[0034] The following will combine Figures 1-4 A specific embodiment of the present invention will be described below.
[0035] First, the overall structure is modeled and generated using diffusion probability modeling. The original 3D morphological data of neurons is preprocessed, and the farthest-point sampling method is used to uniformly sample the point clouds of different neurons to a fixed number of points, constructing a standardized training dataset. The 3D Denoising Diffusion Probability Model (DDPM) is used to learn the overall point cloud distribution of neurons. Gaussian noise is progressively added to the initial point cloud during the diffusion process, and the DiT-3D network parameters are trained through a reverse denoising process. Specifically, in the forward diffusion stage, multiple noise injections are performed on the real point cloud according to a pre-defined noise coefficient table until it degenerates into a Gaussian distribution. In the reverse generation stage, the network progressively reconstructs a structured point cloud from random noise by predicting the noise components at each step. The coarse-grained point cloud output in this stage fully preserves the spatial distribution characteristics of neurons but lacks detailed structure.
[0036] The process then proceeds to the growth-oriented skeleton connection stage. First, the L1-median skeleton extraction algorithm is applied to topologically simplify the generated point cloud. By iteratively optimizing the positions of skeleton points, they are moved towards the geometric center of the point cloud while maintaining the minimum repulsion distance between them. This results in a sparse skeleton point set reflecting the neuron's main branch orientation, completing the skeletonization process. Next, a growth-oriented tree-like connection algorithm is employed, using the skeleton point with the highest local density as the cell body starting point. Hierarchical connections are achieved by dynamically calculating the growth direction weights. Specifically, in each connection step, principal component analysis of the currently connected point set determines the main branch growth direction. The Euclidean distance and direction cosine value of candidate connection points are combined as a weighting index, prioritizing connections to nodes that conform to axial growth patterns. This process continues until all skeleton points form a complete tree topology, ensuring that the branch orientation conforms to the spatial projection patterns of biological neurons.
[0037] In the branch smoothing optimization stage, a dedicated convolutional neural network is constructed to refine the morphology of the generated branches. During training data preparation, multi-scale noise perturbations are applied to the real neuron branches: positional offsets along the normal direction are superimposed in randomly selected segments to form low-quality branch samples containing geometric distortions. The network adopts a residual structure design, mapping the distorted branches back to the original morphology through end-to-end learning, focusing on repairing unnatural bending abrupt changes and local dilation. In the inference stage, the skeleton branches generated in the previous stage are cut into independent segments and input into the network. After processing through multiple layers of one-dimensional convolution, a smooth and continuous neural fiber morphology is output. This process effectively eliminates traces of algorithm generation and restores the natural biomechanical characteristics of axons and dendrites.
[0038] The entire generation process adopts a cascaded architecture, with the output of each stage serving as the input for the next stage. The number of point clouds generated by the global diffusion model is set to 2048, and the diffusion step size uses a cosine scheduling strategy. The directional weight coefficient γ in the skeleton connection stage is fixed at 1.2, and the main direction update frequency is set to update once every 10 nodes connected. The branch optimization network is based on the ResNet-18 architecture, containing 8 residual blocks and a total of 18 convolutional layers, and is trained using the mean squared error loss function. The parameters of each module are determined through separate training, and finally combined to form a complete morphology generation network. This implementation method achieves high-fidelity reconstruction of complex neuron morphologies while ensuring generation efficiency.
[0039] It should be understood that the above description is for illustrative purposes and not for limitation. Many embodiments and applications beyond the provided examples will be apparent to those skilled in the art upon reading the above description. Therefore, the scope of this teaching should not be determined by reference to the above description, but rather by reference to the foregoing claims and the full scope of their equivalents. For purposes of completeness, all articles and references, including patent applications and publications, are incorporated herein by reference. The omission of any aspect of the subject matter disclosed herein in the foregoing claims is not intended as a waiver of that subject matter, nor should it be construed as an indication that the applicant has not considered that subject matter as part of the disclosed application subject matter.
Claims
1. A method for unconditionally and automatically generating the morphology of long-range projection neurons, characterized in that, Includes the following steps: The overall structure prediction step uses a point cloud generation module based on the denoised diffusion probability model (DDPM) to generate point clouds from end to end. The skeletonization connection step first involves the L1-medial skeleton extractor extracting the central axis features from the point cloud using an iterative optimization algorithm for skeletonization, and then the growth-guided connector equipped with a direction-aware weight calculation unit and a dynamic principal component analysis (PCA) direction update module to realize the biologically reasonable dendritic connection topology construction. The branch smoothing optimization step uses an improved ResNet-18 network as the backbone, which integrates multi-scale one-dimensional convolutional layers and residual connection structures, and is equipped with a noise injection unit for training data augmentation.
2. The unconditional automatic generation method for the morphology of long-range projection neurons as described in claim 1, characterized in that, The point cloud generation module based on the Denoising Diffusion Probability Model (DDPM) adopts DiT-3D as the backbone network architecture, which includes multi-level three-dimensional convolutional layers and attention mechanisms. It gradually transforms Gaussian noise into a coarse-grained three-dimensional point cloud representation of neurons. The dimensionality transformation of the point cloud data is realized through cascaded voxelization-devoxelization processing units. The output end uses the farthest point sampler to standardize the generated point cloud to a fixed number of points.
3. The unconditional automatic generation method for the morphology of long-range projection neurons as described in claim 2, characterized in that, The overall structure prediction step includes: The training dataset is constructed by standardizing the real neuron point cloud through farthest point sampling (FPS). The forward diffusion process gradually degrades the real point cloud into Gaussian noise through a first preset number of iterations of noise injection; The reverse process uses the DiT-3D network architecture to extract hierarchical features from the voxelized point cloud, and learns the mapping relationship from noise to point cloud through Transformer blocks; During the inference phase, random noise is denoised through a second preset number of iterations to generate a coarse-grained neuron point cloud that conforms to the true distribution.
4. The unconditional automatic generation method for the morphology of long-range projection neurons as described in claim 1, characterized in that, The L1-medial skeleton extractor uses an iterative optimization algorithm to extract the central axis features from the point cloud for skeletonization, including: The L1-median skeleton extraction algorithm is applied to compress the generated point cloud into a density-optimized skeleton point set through iterative optimization. In each iteration, the L1 distance constraint between the skeleton point and the original point is calculated, and a repulsive force term is introduced to prevent the skeleton point from clustering.
5. The unconditional automatic generation method for the morphology of long-range projection neurons as described in claim 1, characterized in that, Growth-guided connectors enable biologically rational dendritic connection topology construction, including: The growth-guided connector implements a growth-guided graph connection algorithm. After determining the cell body position through local density calculation, it uses an improved minimum spanning tree algorithm for topology construction. The connection weight calculation integrates Euclidean distance and principal direction alignment, where the principal direction is dynamically updated through sliding window PCA to ensure that branch growth conforms to the overall projection pattern of neurons. A branch number constraint mechanism is designed to enforce the biological rule of at most bifurcation during each connection.
6. The unconditional automatic generation method for the morphology of long-range projection neurons as described in claim 5, characterized in that, The branch smoothing optimization step includes: The preprocessed third set of real branch datasets are used to generate training samples by adding multi-scale noise blocks. The noise block parameters include random location intervals [s_k, e_k), amplitude coefficient a, and three-dimensional direction vector u_k. The smoothing network employs a convolutional architecture that includes residual blocks, each containing dilated convolutions to capture long-range geometric features; During training, the Adam optimizer is used, and the nonlinear mapping from noisy branches to true branches is learned through the MSE loss function; During inference, the branches generated by the skeleton reconstruction are processed segment by segment, and an overlapping sliding window strategy is used to ensure the continuity of the branch connections.
7. The unconditional automatic generation method for the morphology of long-range projection neurons as described in claim 1, characterized in that, Also includes: The verification process involves evaluating the global morphological similarity of the output tree topology and optimized branch geometry by calculating the Chamfer distance (CD) and Earth Mover distance (EMD). Simultaneously, it analyzes 12 morphometric indicators, including width, depth, height, maximum degree, maximum branch order, path length, path distance, path angle, branch angle, logarithmic maximum curvature, tree asymmetry, and cell exit angle, to ensure biological rationality.