Neuron tracking method and device, computer equipment and storage medium

Abstract

The invention discloses a neuron tracking method and device, computer equipment and a storage medium, which are applied to the technical field of computers, and the method comprises the following steps: obtaining a three-dimensional fluorescent microscopic image; extracting image blocks based on the three-dimensional fluorescence microscopic image; inputting the image block into a pre-constructed neuron tracking model, predicting a curvature vector of neuron fibers in the image block, and predicting the position and direction of a next path point; carrying out legality detection on the predicted path points according to a preset standard, and if the predicted path points are illegal, stopping tracking; if yes, re-extracting an image block according to the predicted path point position and direction, and iteratively predicting a subsequent path until a neuron tracking target is completed; and automatic tracking of the neuron fibers is realized.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular, to a neuron tracking method, apparatus, computer device, and storage medium. Background Art

[0002] In the field of life sciences, the neuron reconstruction task in three-dimensional fluorescence microscopy images is an important part of mapping the whole-brain mesoscopic connection map. Its goal is to determine the position information and connection relationships of neurons in the images, and to analyze the high-level cognitive functions of the brain from the principles of neural circuits. In the neuron tracking task, a mainstream process is to perform segmentation first and then fragment connection. The segmentation can be automatically completed by a deep learning-based segmentation model, and the fragment connection is basically completed manually. If the recall rate of the segmentation result in the first stage is insufficient, a large number of annotation points need to be manually placed during fragment connection to complete the fragment connection operation.

[0003] In the field of autonomous driving, the algorithm needs to identify the road ahead based on real-time monitoring images and control the vehicle to move forward. In the early days, researchers used traditional image processing algorithms to design and extract road-related feature information from monitoring images, and combined with the motion information of the vehicle at the previous moment for comprehensive judgment to obtain the control scheme of the vehicle at the next moment. With the rise of deep learning algorithms, artificial neural networks are now used to fuse images and vehicle motion states to obtain deeper information for more accurate judgment. In the neuron tracking task, neuron fibers can be analogized to the lane lines of vehicle travel, so some mature automatic path-finding algorithms can be migrated and used. The difference is that most of the visual images processed by autonomous driving algorithms are two-dimensional images, while the images processed by neuron tracking algorithms are three-dimensional. Due to the different data modalities, they cannot be directly migrated and used.

[0004] To overcome these defects, the present application proposes a neuron tracking method, apparatus, computer device, and storage medium. Summary of the Invention

[0005] The purpose of the present application is to provide a neuron tracking method, apparatus, computer device, and storage medium, aiming to solve the above problems.

[0006] To achieve the above purpose, the present application provides the following technical solutions:

[0007] In the first aspect, the present application provides a neuron tracking method, and the steps include:

[0008] Obtain a three-dimensional fluorescence microscopy image;

[0009] Extract image patches based on the three-dimensional fluorescence microscopy image;

[0010] Input the image patch into a pre - constructed neuron tracing model to predict the curvature vector of neuron fibers in the image patch, as well as the position and direction of the next path point;

[0011] Conduct a legality check on the predicted path points according to preset criteria. If it is judged illegal, terminate the tracing; if it is judged legal, re - extract the image patch according to the position and direction of the predicted path points, and iteratively predict the subsequent paths until the neuron tracing goal is completed.

[0012] In a second aspect, the present application provides a neuron tracing device, and the specific content includes:

[0013] An acquisition module: acquire a three - dimensional fluorescence microscopic image;

[0014] An extraction module: extract image patches based on the three - dimensional fluorescence microscopic image;

[0015] A tracing module: input the image patch into a pre - constructed neuron tracing model to predict the curvature vector of neuron fibers in the image patch, as well as the position and direction of the next path point;

[0016] A verification module: conduct a legality check on the predicted path points according to preset criteria. If it is judged illegal, terminate the tracing; if it is judged legal, re - extract the image patch according to the position and direction of the predicted path points, and iteratively predict the subsequent paths until the neuron tracing goal is completed.

[0017] In a third aspect, the present application provides a computer device, which includes a processor and a memory coupled to the processor. Among them, the memory stores program instructions for implementing a neuron tracing method; the processor is used to execute the program instructions stored in the memory to implement a neuron tracing.

[0018] In a fourth aspect, the present application provides a storage medium storing program instructions that can be run by a processor, and the program instructions are used to execute a neuron tracing method.

[0019] The present application provides a neuron tracing method, device, computer device, and storage medium, which have the following beneficial effects:

[0020] (1) By predicting the input image patch through a neuron tracing model, the spatial features of neuron fibers are effectively extracted, avoiding the limitations of traditional two - dimensional algorithms on three - dimensional data; according to the mid - line proximity constraint condition, that is, using the characteristic that the brightness of the mid - line of neuron fibers in the fluorescence image is higher than that of the edge, the ability to correct the error of deviating from the mid - line is achieved, and the robustness to low - quality images is significantly improved.

[0021] (2) The method proposed in this application realizes the end-to-end automatic tracking of neuron fibers. Compared with the traditional manual annotation and semi-automatic pruning methods, the need for manual intervention is significantly reduced.

[0022] (3) During the training stage of the neuron tracking model, labeled data of various complex scenarios in the whole brain are used, enabling the model to learn diverse feature expressions. Therefore, the algorithm can adapt to various complex scenarios under optical microscopes without relying on specific premises, solving the problem of excessive scene limitations in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a schematic flowchart of a neuron tracking method according to Embodiment 1 of this application;

[0024] Figure 2 It is an image of the training set in the neuron tracking model according to Embodiment 1 of this application;

[0025] Figure 3 It is a schematic diagram of the principle of the parametric curve and the neuron tracking model according to Embodiment 1 of this application;

[0026] Figure 4 It is a schematic diagram for verifying the neuron tracking model according to Embodiment 1 of this application;

[0027] Figure 5 It is a comparison diagram of the experimental effects according to Embodiment 1 of this application;

[0028] Figure 6 It is an effect diagram of Embodiment 1 of this application on a more difficult complete tracking task;

[0029] Figure 7 It is a flowchart of a neuron tracking method according to Embodiment 1 of this application;

[0030] Figure 8 It is a schematic structural diagram of a neuron tracking device according to Embodiment 2 of this application;

[0031] Figure 9 It is a schematic structural diagram of a computer device according to Embodiment 3 of this application;

[0032] Figure 10 It is a schematic structural diagram of a storage medium according to Embodiment 4 of this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0033] It should be understood that the specific embodiments described herein are only used to explain this application and are not used to limit this application.

[0034] The following analyzes the solutions in the prior art in combination with related technologies.

[0035] In the existing implementation solutions for the light microscope system, one category is completed by using the method of first segmentation and then manual annotation. Another category starts from a seed point, classifies the pixel points around the fiber, and traverses and tracks them using the breadth-first strategy. There are also other methods that classify all pixel points within a certain image range, identify the overall neuron fibers, and then use pruning algorithms to remove the incorrect parts. However, these practices all have certain limitations. For example, they have poor anti-interference ability and are not suitable for images with poor quality; there are certain prerequisite restrictions, such as the need for cell bodies to exist within the image; they are not adaptable to various complex scenarios in the whole brain; and the computational cost is not suitable for large-scale neuron reconstruction.

[0036] In addition, on the electron microscope system, an algorithm based on deep learning has been proposed in combination with an automatic path-finding algorithm. This method calculates the curvature information of the neuron fiber where the current position is located and adaptively predicts and determines the direction and step size of the next movement. Although this method has good effects and application prospects, it still has the following limitations: There are significant differences in the content and scale of the images captured by electron microscopes and optical microscopes; the electron microscope captures the cell membrane of the neuron fiber. The "proxy point" controlled by this algorithm moves forward inside the fiber, and uses the cell membrane as the "lane edge" for the boundary condition of error correction. The optical microscope does not have this prior condition.

[0037] Therefore, in view of the above problems, the present application proposes a neuron tracking method, device, computer device, and storage medium, which realizes the idea of "center line approximation" and reduces the cost required to connect neuron segments.

[0038] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0039] Embodiment 1

[0040] Please refer to Figure 1 , which is a schematic flowchart of a neuron tracking method according to Embodiment 1 of the present application; the steps include:

[0041] S1: Obtain a three-dimensional fluorescence microscopic image.

[0042] S2: Extract image patches based on the three-dimensional fluorescence microscopic image.

[0043] In this embodiment, the initial point on the neuron fiber in the three-dimensional fluorescence microscopic image is used as the current path point.

[0044] Taking the current path point as the center and the path point direction as one of the coordinate axes, intercept the image block; the position and direction information of the current path point are included in the image block.

[0045] S3: Input the image block into a pre-constructed neuron tracking model to predict the curvature vector of neuron fibers in the image block, as well as the position and direction of the next path point.

[0046] In this embodiment, step S3 specifically includes steps S31 to S33, which will be described in detail below.

[0047] S31: Use the brightest path tracking algorithm based on the A* optimal path algorithm to label neuron data, use the labeled neuron data as the training set of the neuron tracking model, and obtain supervision signals from the training set.

[0048] Please refer to Figure 2 , which is the training set image in the neuron tracking model of Embodiment 1 of this application. Utilizing the feature that the center line is the brightest, a batch of neuron data is saturatedly labeled as the training set of the algorithm using the brightest path tracking algorithm based on the A* optimal path algorithm, for example: 20 macaque brain axon trunk data and 10 macaque brain axon terminal data. The neuron segments in the training set image are instance labels rather than semantic labels, and the mask of each segment can be accessed.

[0049] To describe the motion state of points on neurons, perform B-Spline interpolation of degree 4 on the neuron segments in the training set to obtain parameterized curve segments; the formula is expressed as:

[0050] γ(t)∈C 3 ,t∈[0,1];

[0051] In geometry, Frame is used to represent the motion state of a point on a curve. In this embodiment, the Frenet-Serret Frame is used to represent the motion state of a point on the parameterized curve, and the Double-Reflection method is used to convert the curve Frame into the Rotation Minimizing Frame to ensure that the rotation of points on the curve is minimized during the motion process. The formula is expressed as:

[0052]

[0053] Among them, γ(t) is the parameterized curve, t∈[0,1] is a point on the curve; γ′ is the first derivative of γ, and γ″ is the second derivative of γ; the unit vector is the tangent vector, representing the motion direction of the path point on the curve; is calculated from the first derivative γ′ of the curve; the unit vector Calculated from the first derivative t′, where t′ is the derivative of ; κ is the curvature value representing the amplitude of motion change; the vector is the curvature vector, obtained by multiplying the scalar curvature value κ by the unit vector and representing the direction and amplitude of the change in the motion direction; the unit vectors and are the basis vectors of the plane where the vector is located, orthogonal to and coplanar with ; k1 and k2 are scalar coefficients.

[0054] S32: In the training stage of the neuron tracking model, using the true value of the curvature of the labeled neuron fibers, optimize the parameters of the neuron tracking model by calculating the loss function between the predicted curvature value and the true value.

[0055] Calculate the gradient using the mean square error MSE as the loss function. The formula is:

[0056]

[0057] where k i is the true curvature value; is the predicted curvature value.

[0058] S33: In the inference stage of the neuron tracking model, input the image patch into the neuron tracking model, obtain the curvature vector of the current path point, and predict the position and direction of the next path point based on the curvature vector.

[0059] Please refer to Figure 3 , which is a schematic diagram of the principle of the parametric curve and the neuron tracking model in Embodiment 1 of this application. Figure 3 (a) shows the parametric curve, Figure 3 (b) shows the schematic diagram of the principle of the neuron tracking model. Take a point on the segment as the initial point, and take an image patch with its coordinate center and its and as the spatial basis vectors as the visual information input of the model. At this time, the representing the motion direction of the initial point is parallel to one of the standard coordinate axes of the image patch, so the motion direction information of the initial point is implicitly contained in the image patch.

[0060] Take the image patch as the input of the neuron tracking model and predict the curvature vector of the current path point; according to the current path point direction and the predicted curvature vector, perform a second-order Taylor expansion on the parametric curve of the fiber where the current path point is located. The formula is expressed as:

[0061]

[0062] Among them, is the parametric curve for second-order Taylor expansion; and describe the local approximation of the coordinate position of the curve at s and the local approximation of the moving direction of the curve at s.

[0063] Calculate the step size according to the adaptive step size strategy, substitute the step size into the second-order Taylor expansion formula of the parametric curve to obtain the position and direction vector of the next path point; the formula of the adaptive step size strategy is expressed as:

[0064]

[0065] Among them, the step size Δs is negatively correlated with the modulus of the predicted curvature vector ; f is the step size scaling factor; d is an empirical value constant set according to the training set; p is the physical size corresponding to the side length of the input image patch, and s min is the minimum movement step size.

[0066] After determining the step size Δs, calculate the position and direction vector of the next path point, re-obtain the image patch with the new position and direction vector, and perform a legality check on the predicted path point.

[0067] It should be noted that the neuron tracking model in this embodiment is a three-dimensional convolutional neural network, which is mainly used for feature extraction and vector prediction. The feature extraction uses a standard residual network module, with 32, 64, and 128 feature channels respectively for the three levels, the convolutional kernels are all 3*3*3, and the activation function is ReLU. The vector prediction is implemented using a fully connected layer. The three-dimensional feature map output by the convolutional layer is flattened into a sequence through an average pooling layer and sent to the fully connected layer to output the predicted curvature vector. Because the curvature vector is coplanar with the image patch coordinate basis vectors and and one of the dimension values is always zero, so the fully connected layer only needs to predict the values of the other two non-zero dimensions. Interpolate and parameterize the foreground segments produced in the segmentation stage to obtain the training data pairs composed of the image patch and the direction vector

[0068]

[0069] S4: Perform a legality check on the predicted path points according to the preset criteria. If it is judged illegal, terminate the tracking; if it is judged legal, re-extract the image patch according to the position and direction of the predicted path points, and iteratively predict the subsequent paths until the tracking target of the neuron is completed.

[0069] In this embodiment, the criteria for determining whether a path point is legal include, but are not limited to: whether the path point is within the image range, whether the path point conforms to other known neuron structures (such as branches, junctions), whether the direction of the path point is reasonable, etc. The legality of the predicted path points is detected based on a preset legality criterion. For example, check whether the path point exceeds the boundary of the image, or whether the direction of the path point is consistent with the direction of the known neuron structure. According to the position and direction of the predicted path points, image patches are re-extracted.

[0070] The legal predicted path points are added to the current path, and the path information is updated. At the same time, check whether the tracking target is reached, such as reaching a specific neuron structure, tracking to the end of the neuron, etc.; if the tracking target is not reached, continue to iteratively predict the path; when the tracking target is reached, stop the iterative process and output the tracked neuron path, which can include information such as the position, direction, and image features of the path points.

[0071] Please refer to Figure 4 , which is a schematic diagram for verifying the neuron tracking model in Embodiment 1 of this application. Starting from the end point of the obtained nerve segment as the initial point, the path is predicted outward, and all passed path points are recorded. At each step, check whether the brightness value of the path point where the current initial point is located is reasonable, and use backward tracking for verification. Starting from the current initial point, use the neuron tracking model to track backward, and judge whether the error between the backward tracking path and the forward tracking path is reasonable to ensure that the tracked path is always on the neuron foreground. At the same time, the position relationship between the current initial point and the end points of other segments in the neighborhood is judged in real time, and whether there is connectivity is judged according to aspects such as whether the position distance is close and whether the segment direction angles are the same.

[0072] The annotation mask in the training set has centerline consistency with the original image, and the supervision signal is obtained from the training set and used for supervised learning training. In addition, initial points deviating from the fiber are introduced, so that the outliers have larger curvature vectors pointing to the fiber centerline. With the training data with centerline consistency and appropriate outlier data, the model is constrained to learn the error correction characteristic of "tending to the midline".

[0073] Please refer to Figure 5 , which is a comparison diagram of the experimental effects in Embodiment 1 of this application. Figure 5 The data scenarios from left to right in are: dendrite, axon, axon terminal, and noise data. The first row in the figure is the foreground segments extracted by the MOST method, and a large number of broken segments need to be manually connected subsequently; the second row is the segments completed by the method proposed in this application. It can be seen that by using this method, the segments can be connected quickly and accurately. Compared with the traditional manual tracking and connecting of neuron segments, this application improves the efficiency.

[0074] In scenarios of different complexities, the recall rates of the extracted neuron signals are shown in the following table. Among them, the segmentation model refers to a developed deep learning-based segmentation model used for neuron segmentation.

[0075] Dendrite Axon Axon terminal Noise data MOST method 0.505 0.819 0.787 0.571 Segmentation model 0.849 0.858 0.715 0.904 Segmentation model + Tracking algorithm 0.882 0.961 0.950 0.938

[0076] Please refer to Figure 6 , which is the effect diagram of Embodiment 1 of this application in a more difficult complete tracking task. In the four scenarios of dendrites, axons, axon terminals, and noise data, the success rates of using the neuron tracking model to track the entire neuron fiber are statistically analyzed. Short, medium, and long respectively represent neurons of different lengths, and the bar chart represents the success rate of using the path tracking model to completely track the neuron fiber. It can be seen that on relatively short paths, the success rate of the present invention is higher; in complex scenarios, the number of failures due to the termination of tracking caused by strict error control is relatively large. Please refer to Figure 7 , which is the flowchart of a neuron tracking method according to Embodiment 1 of this application. Taking a point on the neuron in the three-dimensional fluorescence microscopic image as a path point, intercepting an image block with the segment direction and this path point as the center; inputting the image block into the neuron tracking model to predict the direction vector and curvature information corresponding to the fiber curve in the image block. Further combining the coordinate, direction, and curvature information can calculate the direction vector and curvature information of the next step. After obtaining the new coordinates and direction, judge the legality through the verification strategy. If it is judged to be legal, then judge whether the tracking target has been reached at this time. If not, update the coordinates and direction and repeat the process for tracking; if the new coordinates and direction are illegal or the tracking target has been reached, then terminate the tracking.

[0077] In summary, Embodiment 1 of this application significantly improves the accuracy, efficiency, and scene adaptability of neuron tracking through key technologies such as neuron tracking models, generalization training, and path correction, while reducing the manual and computational costs. In addition, for the first time, the "lane line pathfinding" idea in the field of autonomous driving is migrated to the neuron tracking task in optical microscopes; compared with the prior art, this application has fewer prerequisite restrictions, can effectively reduce the cost of connecting neuron segments, and shows high efficiency in the neuron reconstruction task.

[0078] Embodiment 2

[0079] Please refer to Figure 8 , which is the structural schematic diagram of a neuron tracking device according to Embodiment 2 of this application; the specific content includes:

[0080] Acquisition module: Acquire three-dimensional fluorescence microscopic images;

[0081] Extraction module: Extract image blocks based on the three-dimensional fluorescence microscopic images;

[0082] Tracking module: Input the image block into a pre-constructed neuron tracking model to predict the curvature vector of neuron fibers in the image block, as well as the position and direction of the next path point.

[0083] Verification module: Perform a legality check on the predicted path points according to a preset standard. If it is determined to be illegal, terminate the tracking; if it is determined to be legal, re-extract the image block according to the position and direction of the predicted path points, and iteratively predict the subsequent paths until the neuron tracking goal is completed.

[0084] Embodiment 3

[0085] Please refer to Figure 9 , which is a schematic structural diagram of a computer device according to Embodiment 3 of this application. The computer device 50 includes a processor 51 and a memory 52 coupled to the processor 51.

[0086] The memory 52 stores program instructions for implementing the above-mentioned neuron tracking method.

[0087] The processor 51 is used to execute the program instructions stored in the memory 52 to implement a neuron tracking.

[0088] Among them, the processor 51 can also be called a CPU (Central Processing Unit, central processing unit).

[0089] The processor 51 may be an integrated circuit chip with signal processing capabilities. The processor 51 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc.

[0090] Embodiment 4

[0091] Please refer to Figure 10, which is a schematic structural diagram of the storage medium according to Embodiment 4 of the present application. The storage medium of the embodiment of the present application stores a program file 61 that can implement all the above methods. Among them, the program file 61 can be stored in the above storage medium in the form of a software product, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the methods in various embodiments of the present invention. The foregoing storage medium includes: various media that can store program codes such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs, or devices such as computers, servers, mobile phones, and tablets.

[0092] It should be noted that in this article, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, device, article or method. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, device, article or method including that element.

[0093] The above are only the preferred embodiments of the present application, and do not limit the patent scope of the present application. Any equivalent structural or equivalent process transformation made by using the content of the specification and drawings of the present application, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present application.

[0094] Although the embodiments of the present application have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present application. The scope of the present application is defined by the appended claims and their equivalents.

[0095] Of course, the present invention can also have other various embodiments. Based on this embodiment, other embodiments obtained by those of ordinary skill in the art without any creative labor belong to the scope protected by the present invention.

Claims

1. A neuron tracing method, characterized in that, Including: Obtain a three-dimensional fluorescence microscopic image; Extract image patches based on the three-dimensional fluorescence microscopic image; Input the image patches into a pre-constructed neuron tracking model to predict the curvature vector of neuron fibers in the image patches, as well as the position and direction of the next path point; Perform a legality check on the predicted path points according to a preset standard. If it is judged illegal, terminate the tracking; if it is judged legal, re-extract image patches according to the position and direction of the predicted path points, and iteratively predict subsequent paths until the neuron tracking goal is completed.

2. The neuron tracing method according to claim 1, wherein, In the step of extracting image patches based on the three-dimensional fluorescence microscopic image, it specifically includes the following steps: Take the initial point on the neuron fiber in the three-dimensional fluorescence microscopic image as the current path point; Taking the current path point as the center and the path point direction as one of the coordinate axes, intercept the image patch; the image patch includes the position and direction information of the current path point.

3. The neuron tracking method according to claim 1, wherein In the step of inputting the image patches into a pre-constructed neuron tracking model to predict the curvature vector of neuron fibers in the image patches, as well as the position and direction of the next path point, it specifically includes the following steps: Use the brightest path tracking algorithm based on the A* optimal path algorithm to label neuron data, and use the labeled neuron data as the training set of the neuron tracking model to obtain a supervision signal from the training set; In the training stage of the neuron tracking model, use the true value of the labeled neuron fiber curvature, and optimize the parameters of the neuron tracking model by calculating the loss function between the predicted curvature value and the true value; In the inference stage of the neuron tracking model, input the image patch into the neuron tracking model, obtain the curvature vector of the current path point, and predict the position and direction of the next path point according to the curvature vector.

4. A neuron tracing method according to claim 3, characterized in that In the step of using the brightest path tracking algorithm based on the A* optimal path algorithm to label neuron data, and using the labeled neuron data as the training set of the neuron tracking model to obtain a supervision signal from the training set, it specifically includes the following steps: Interpolate the neuron data in the training set to obtain parameterized curve segments; the formula is expressed as: γ(t) ∈ C 3 , t ∈ [0, 1]; Among them, γ(t) is the parameterized curve, and t ∈ [0, 1] is a point on the curve; γ′ is the first derivative of γ, and γ″ is the second derivative of γ; the unit vector is the tangent vector, representing the motion direction of the path point on the curve, which is calculated from the first derivative γ′ of the curve; the unit vector is calculated from the first derivative t′, where t′ is the derivative; κ is the curvature value representing the amplitude of motion change; the vector is the curvature vector, which is obtained by multiplying the scalar curvature value κ by the unit vector and represents the change direction and amplitude of the motion direction; the unit vectors and are the basis vectors of the plane where the vector is located, orthogonal to and coplanar with ; k1 and k2 are scalar coefficients.

5. A neuron tracing method according to claim 4, wherein In the step of using the true value of the labeled neuron fiber curvature in the training stage of the neuron tracking model, and optimizing the parameters of the neuron tracking model by calculating the loss function between the predicted curvature value and the true value, it specifically includes the following steps: Use the mean square error MSE as the loss function to calculate the gradient, and the formula is: where k i is the true value of curvature; is the predicted curvature value.

6. A neuron tracking method according to claim 5, wherein In the step of inputting the image patch into the neuron tracking model in the inference stage of the neuron tracking model, obtaining the curvature vector of the current path point, and predicting the position and direction of the next path point according to the curvature vector, it specifically includes the following steps: Take the image patch as the input of the neuron tracking model and predict the curvature vector of the current path point; according to the direction of the current path point and the predicted curvature vector, perform a second-order Taylor expansion on the parameterized curve of the fiber where the current path point is located, and the formula is expressed as: Among them, is a parametric curve for second-order Taylor expansion; and describe the local approximation of the coordinate position of the curve at s and the local approximation of the moving direction of the curve at s; Calculate the step size according to the adaptive step size strategy, substitute the step size into the second-order Taylor expansion of the parametric curve to obtain the position and direction vector of the next path point; the formula of the adaptive step size strategy is expressed as: Among them, the step size Δs is negatively correlated with the modulus of the predicted curvature vector ; f is the step size scaling factor; d is an empirical value constant set according to the training set; p is the physical size corresponding to the side length of the input image patch, and s min is the minimum movement step size; Re-obtain the image patch according to the position and direction vector of the next path point, and perform a legality check on the predicted path point.

7. A neuron tracing method according to claim 1, characterized in that, The neuron tracking model is a three-dimensional convolutional neural network, including: an input layer, a plurality of convolutional layers, a pooling layer, and a fully connected layer; Flatten the three-dimensional feature map output by the convolutional layer into a sequence through the average pooling layer, and output the predicted direction vector and curvature information by the fully connected layer.

8. A neuron tracking device, characterized in that, Including: Acquisition module: acquire three-dimensional fluorescence microscopy images; Extraction module: extract image patches based on the three-dimensional fluorescence microscopy images; Tracking module: input the image patches into a pre-constructed neuron tracking model to predict the curvature vector of neuron fibers in the image patches, as well as the position and direction of the next path point; Verification module: perform a legality check on the predicted path point according to a preset standard. If it is judged illegal, terminate the tracking; if it is judged legal, re-extract the image patch according to the position and direction of the predicted path point, and iteratively predict the subsequent path until the neuron tracking target is completed.

9. A computer device, characterized in that, The computer device includes a processor and a memory coupled to the processor. Among them, the memory stores program instructions for implementing a neuron tracking method according to any one of claims 1-7; the processor is configured to execute the program instructions stored in the memory to implement a neuron tracking.

10. A storage medium, characterized in that, Store program instructions that can be run by a processor, and the program instructions are used to execute a neuron tracking method according to any one of claims 1-7.

Images