A method and device for nerve fiber reconstruction based on centerline feature guidance

Abstract

The application discloses a nerve fiber reconstruction method based on a center line feature guide, and comprises the following steps: S1, constructing a double-branch parallel convolutional neural network, and synchronously outputting a nerve fiber segmentation probability graph and a center line probability graph by using two mutually independent decoding devices; S2, extracting an initial neuron skeleton; S3, generating a topological guide signal; and S4, calculating the topological degree of each node of the initial neuron skeleton based on the topological guide signal, and taking the topological degree of the node as a reference to perform selective path tracking and backtracking pruning, and obtaining a nerve fiber reconstruction result with consistent topology. The application further discloses a nerve fiber reconstruction device based on the center line feature guide. The application realizes high-accuracy nerve fiber path tracking and structure reconstruction, and can be widely applied to the technical field of computer vision and medical image processing.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and medical image processing technology, and in particular to a method and apparatus for neural fiber reconstruction based on centerline features. Background Technology

[0002] In recent years, advancements in molecular marker technology and microscopic imaging have made it possible to acquire neuronal images at the whole-brain scale. Due to the widespread distribution and overall sparseness of neuronal fibers throughout the brain, extracting high-precision neuronal topology from massive amounts of images has become a core challenge in neuroinformatics. To improve processing efficiency, existing research has largely focused on removing non-fiber background and data redundancy, but significant bottlenecks remain in the accuracy of structural tracking in key local regions.

[0003] Existing neural fiber reconstruction methods suffer from fundamental flaws when dealing with densely overlapping local regions. When multiple fibers are closely attached to or cross each other, the segmentation responses of different fibers often exhibit a "spurious adhesion" state due to limitations in optical imaging resolution and classification errors in segmentation algorithms. Subsequent skeletonization algorithms, when processing such adhered regions, cannot distinguish between genuine physical connections and connectivity ambiguities generated by the algorithm, leading to numerous erroneous topological mergings in the reconstruction results. Currently, most mainstream error correction strategies rely on simple length threshold filtering. This single criterion cannot effectively distinguish between genuine short physiological branches and spurious branches, easily resulting in the accidental deletion of real structures.

[0004] A deeper problem lies in the lack of a verification mechanism in current technology capable of automatically identifying and locating topological conflict regions. While region segmentation features can fully cover fiber morphology, they tend to be over-connected at fiber overlaps; centerline features, while possessing strong spatial localization exclusivity, are prone to signal breakage in low signal-to-noise ratio regions. Currently, no research has effectively combined the response differences of these two types of heterogeneous features—that is, using centerline features as asymmetric topological prior verification evidence to adaptively guide path refinement within segmented regions. Due to the lack of effective path backtracking and topological consistency verification methods, the algorithm cannot accurately block false tracking paths at fiber overlaps while maintaining backbone connectivity.

[0005] Therefore, developing a neural fiber reconstruction method based on centerline features to achieve automatic identification and precise backtracking and pruning of topological conflict regions may become a key method to improve the accuracy of automatic reconstruction of large-scale neural images. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the above-mentioned background technology and provide a neural fiber reconstruction method and device based on centerline feature guidance. This method automatically identifies and backtracks topological conflicts in the pruning and segmentation results by establishing asymmetric topological prior constraints, thereby achieving highly accurate neural fiber path tracking and structural reconstruction.

[0007] This invention provides a neural fiber reconstruction method based on centerline features, comprising: constructing a dual-branch parallel convolutional neural network and synchronously outputting a neural fiber segmentation probability map. Probability plot with center line The segmentation probability map is then subjected to three-dimensional refinement to extract the initial neuron skeleton. Based on the difference in response intensity between the segmentation probability map and the centerline probability map, a topology guidance signal is generated. Calculate the topological degree of each node in the initial neuronal skeleton, where the topological degree is the number of neighboring skeleton points in the three-dimensional neighborhood of the node; identify nodes with a topological degree of 1 as leaf nodes, and determine whether the leaf nodes are located at the topological guidance signal. Within the marked area; for leaf nodes located within the marked area, a backtracking pruning operation is performed, iteratively deleting nodes along the main trunk direction until a preset stopping condition is met, thereby obtaining a topologically consistent neural fiber reconstruction result.

[0008] Furthermore, the dual-branch parallel convolutional neural network is trained using a multi-task joint loss function, which consists of a segmentation branch loss function. With centerline topology branch loss It is composed of a weighted linear combination, and its formula is as follows: ,in, To balance the hyperparameters and satisfy .

[0009] Furthermore, the segmentation branch loss With centerline topology branch loss Both employ a hybrid loss function design, including weighted binary cross-entropy loss and Dice loss.

[0010] Furthermore, the generation of topology guidance signals The process includes: calculating the metric: Generate a topology-guided mask according to the following formula. : ,in, The preset structure confidence threshold is used. This is the preset topology missing threshold.

[0011] Furthermore, the backtracking pruning operation includes: when leaf nodes Probability of satisfying the centerline Below the preset fine-tuning trigger threshold When this occurs, backtracking pruning is triggered; the deletion decision function for backtracking pruning is: .

[0012] Furthermore, in the backtracking pruning operation, the stopping condition for iteratively deleting nodes is: the topological degree of the current node is greater than or equal to 3, or the centerline probability of the current node is greater than or equal to the fine-tuning trigger threshold. .

[0013] Furthermore, the three-dimensional thinning process employs a three-dimensional parallel thinning algorithm, which iteratively identifies and deletes simple points that do not change the topological properties of the target until a central skeleton with a single-pixel width and equivalent to the original neural fiber target topology is extracted.

[0014] Furthermore, the shared encoder extracts hierarchical features through a multi-layer 3D convolutional structure to form a feature pyramid; the decoder branch gradually restores spatial resolution through transposed convolution and fuses the features of the corresponding layer of the shared encoder through skip connections.

[0015] Furthermore, during the backtracking pruning operation or path tracing process, the centerline probability map is used as a voxel-level connectivity constraint. When the centerline probability of a candidate node is lower than a preset connectivity threshold, the node is prohibited from being included in the tracing path or is determined to be a false branch.

[0016] The present invention also provides a nerve fiber reconstruction device based on centerline features, having a computer program that can execute a nerve fiber reconstruction method based on centerline features.

[0017] The present invention provides a method and apparatus for neural fiber reconstruction guided by centerline features, which has the following beneficial effects: First, by introducing centerline features as topological constraints, this invention overcomes the problem of "false adhesion" that traditional segmentation algorithms easily produce at fiber overlaps, and significantly improves the topological correctness of the tracing path.

[0018] Second, by using an asymmetric difference measurement mechanism to generate guidance signals, the strong constraint of centerline features on the structure can be transformed into adaptive error correction power, which can identify complex topological conflicts without human intervention.

[0019] Third, the dynamic backtracking-based pruning strategy achieves "directional blasting" of pseudo-branches, changing the limitation of traditional algorithms that can only handle isolated noise points. Through the dual closed-loop constraints of node degree and centerline response, it ensures that the true topological features of neurons are fully preserved while filtering out false paths.

[0020] Fourth, this invention focuses on optimizing the tracking logic, has good compatibility with the front-end segmentation model, and has strong practical value. Attached Figure Description

[0021] Figure 1 This is an overall flowchart of the neural fiber reconstruction method based on centerline features according to the present invention; Figure 2 This is an architecture diagram of the dual-branch parallel convolutional neural network in the neural fiber reconstruction method based on centerline features of this invention; Figure 3 This is a schematic diagram of the neural fiber reconstruction device based on centerline features according to the present invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments, but these embodiments should not be construed as limiting the present invention.

[0023] This invention addresses the technical shortcomings in 3D neural fiber reconstruction, such as spurious adhesions caused by dense fiber overlap, deviations of tracking paths from the actual structure, and a lack of adaptive error correction mechanisms. It provides a neural fiber reconstruction method guided by centerline features. This invention aims to achieve highly accurate neural fiber path tracking and structural reconstruction by establishing asymmetric topological prior constraints and automatically identifying and backtracking topological conflicts in the pruning and segmentation results.

[0024] Example 1 To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: First, a dual-branch parallel convolutional neural network is constructed. This network extracts hierarchical features of the three-dimensional neural image through a shared encoder, and uses two independent decoders to synchronously output segmentation probability maps of the nerve fibers. Probability plot with center line The dual-branch parallel convolutional neural network, which serves as a multi-task network, employs a weighted joint loss function during the training phase.

[0025] And satisfy the weighting coefficient The aim is to guide the network to prioritize learning the long-range connectivity features of neural fibers by increasing the loss weight of the central branch, thus providing a strong topological prior for subsequent tracking. Secondly, the segmentation probability map... Perform three-dimensional refinement processing to extract the initial neuronal skeleton. Based on this, a topology guidance signal is generated by quantitatively analyzing the difference between the segmentation response and the centerline response. Specifically, the differences between the two (segmentation probability map) are calculated. Probability plot with center line Pixel-by-pixel difference metric:

[0026] Preset structure confidence threshold Topology missing threshold .

[0027] When the spatial location simultaneously satisfies and When the location is determined to have a risk of false connectivity caused by segmentation and adhesion, it is marked as a topological conflict region, and a corresponding topological guidance mask is generated. This step transforms missing information about the centerline into negative evidence of segmentation errors through response comparison of heterogeneous features. Subsequently, this invention utilizes the topological guidance signal to perform selective path tracing and pruning of the initial neuronal skeleton based on node degree. The algorithm calculates the number of nodes in the skeleton using a 26-neighborhood operator. Topology ,definition Let each leaf node represent the end of the path. For all leaf nodes, a deletion decision function is introduced. :

[0028] in The threshold for triggering the pruning is set. If the decision result is true, the branch to which the node belongs is determined to be a pseudo-path caused by fiber adhesion, and this is used as the starting point for backtracking. Finally, dynamic path backtracking pruning is performed. For the selected leaf nodes to be removed, the algorithm executes the following recursive backtracking logic: continuous pruning iteration: after deleting the current node, trace its unique adjacent predecessor node along the skeleton towards the trunk; dynamic update of topology: for each node deleted, update the topology of its predecessor node in real time. If the predecessor node is transformed into a new leaf node after pruning (i.e., the topology becomes 1), it is included in the continuous pruning sequence; backtracking stopping constraint: continue backtracking pruning until a node that meets any stopping condition is encountered. Firstly, topological degree This means reaching the actual branch point or main path of the nerve fiber to prevent over-pruning from damaging normal structures; secondly, the probability of the centerline. That is, backtracking to the confidence region with strong topological evidence.

[0029] Through this dynamic backtracking mechanism, the present invention can completely block ambiguous connections while maintaining the true bifurcation structure of nerve fibers, and output topologically consistent reconstruction results.

[0030] Example 2 like Figure 1 As shown in the figure, this embodiment provides a neural fiber reconstruction method based on centerline features, the specific steps of which are as follows: S1, Dual-branch parallel prediction stage: like Figure 1 and Figure 2 As shown, this embodiment constructs a two-branch 3D convolutional neural network based on an encoder-decoder architecture. The shared encoder serves as the core of feature extraction, downsampling the input 3D neural image through consecutive 3D residual convolutional blocks to extract high-dimensional semantic features. To address the gradient vanishing problem caused by the extreme sparsity of neural fibers in 3D space, each convolutional layer in the encoder is followed by batch normalization and corrected linear units (ReLU).

[0031] In this architecture, the feature stream is decoupled during the deep decoding stage, flowing into two parallel decoder branches. The segmentation decoder branch recovers image resolution through upsampling and skip connections, aiming to generate a neural fiber segmentation probability map. This diagram provides complete information about fiber morphological coverage. The centerline decoder branch focuses on extracting the geometric center path of the fiber, outputting a centerline probability map. Because the centerline is anatomically unique and exclusive, it is more topologically distinct than segmentation masks when dealing with areas of fiber overlap.

[0032] In the model training phase, to enhance the network's ability to perceive topology, this embodiment designs an asymmetric weighted joint loss function:

[0033] in, and A constant can be chosen based on the data. By increasing the loss weight of the centerline branch, the network is forced to prioritize fitting the topological framework of the neural fibers during the optimization process. In the calculation... In this embodiment, Gaussian blurring is applied to the centerline truth label, and its distribution function is expressed as:

[0034] in For the true centerline coordinates, To control the standard deviation of the influence range, this approach enables the model to learn topological features with continuous probability gradients, providing a stable numerical basis for subsequent response difference quantification.

[0035] S2, Structure Extraction Stage: This stage performs the refinement and connectivity protection processes carried out during the structure extraction stage. For example... Figure 1 As shown, the aim is to segment the probability map. Convert to an initial skeleton that is one pixel wide. This embodiment employs a three-dimensional parallel thinning algorithm, which iteratively peels away the boundary layer of the target. The core of the thinning process lies in strictly identifying "simple points," that is, pixels whose deletion does not change the local topological connectivity. The determination of simple points is based on connectivity analysis within a 26-neighborhood, and must meet two conditions: first, deleting the point does not destroy the original connected components; second, deleting the point does not create new holes. In each iteration cycle, the algorithm identifies all boundary pixels that meet the definition of simple points and performs parallel deletion. The mathematical expression for the iterative process is as follows:

[0036] in, Indicates the first The skeleton set after the next iteration This represents the set of boundary pixels of the current target. To ensure the skeleton is precisely centered on the fiber, the algorithm follows a symmetrical stripping principle during deletion to avoid geometric shifts in the skeleton. The above iterative process is repeated until no more simple points can be deleted from the image. The final initial skeleton is then obtained. It is a connected graph with a width of one pixel. Although the skeleton retains the trunk and branch structure of nerve fibers, the segmentation and adhesion caused by the limitation of microscopic imaging resolution often introduces redundant closed loops or spurious branches in the fiber overlapping areas, which constitute the processing objects in the subsequent refinement stage.

[0037] S3, Topology Guiding Signal Generation Stage: This phase involves generating topology guidance signals and locating conflict areas. For example... Figure 1 As shown, the core processing approach lies in quantifying the "logical conflict" between segmentation responses and centerline responses. In areas where nerve fibers intersect or are densely clustered, segmentation branches often exhibit overconnectivity due to their morphological coverage characteristics; while centerline branches, adhering to path uniqueness constraints, show a significant decrease in response intensity at overlap points. This heterogeneous response difference provides physical evidence for identifying false branches. The system calculates pixel-by-pixel response difference metrics. To capture this conflict:

[0038] In regions with high metric values, the segmentation results indicate the presence of structure, which is not supported by centerline features. Based on this logic, this embodiment generates a topology-guided mask by setting a dual-threshold discrimination criterion. .set up The structural confidence threshold is (preferably 0.8). The topology missing threshold (preferably 0.2) is used, and the mask generation function is defined as follows:

[0039] Generated binary mask The algorithm transforms complex spatial response conflicts into intuitive logical signals, identifying all potential topological error regions. In practical applications, to enhance the connectivity of the mask and eliminate background noise, the algorithm further refines the mask... Perform morphological closing operations or small-area filtering based on connected component volume. This step successfully transforms high-dimensional topological prior information into negative evidence that can be used to guide skeleton refinement, enabling precise navigation of spurious adhesion regions.

[0040] S4, Topology Consistency Backtracking Refinement Phase: This phase involves execution path tracing and topology refinement based on node degree backtracking. Figure 1 In obtaining the topology guidance mask Subsequently, this embodiment focuses on the initial skeleton. Selective path refinement based on graph theory features is performed. First, the algorithm scans the skeleton. And calculate each node Topology That is, the total number of adjacent skeleton points of a node in its 26-neighborhood:

[0041] in This is an indicator function. Based on the topological degree, nodes are divided into leaf nodes (...). ), path nodes and branch nodes The refinement process starts with leaf nodes and identifies pseudo-paths that need to be undone by executing a deletion decision function. The decision function is expressed as follows:

[0042] in To refine the trigger threshold, when a leaf node is determined to be deleted, the algorithm initiates an iterative backtracking process. The backtracking process traces the unique adjacent nodes of the current node along the skeleton path towards the main trunk. And remove it. The algorithm is dynamically updated as each node is removed. The topological degree distribution within the neighborhood. If the updated... If a node becomes a new leaf node, its deletion decision is checked repeatedly. The backtracking process has a clear stopping boundary: when the algorithm backtracks to the topological degree... The bifurcation point (representing the point of reaching the actual fiber trunk), or the probability of entering the centerline. Tracking stops when the system recovers to a high-response confidence region. This dynamic backtracking mechanism effectively distinguishes between real physiological short branches and algorithm-generated pseudo-branches, completely eliminating ambiguous paths caused by fiber adhesions without disrupting real connectivity, and ultimately outputting a topologically consistent reconstruction result.

[0043] Example 3 See Figure 3 The present invention relates to a nerve fiber reconstruction device guided by centerline features, comprising the following components: Dual-branch parallel prediction module: Constructs a dual-branch parallel convolutional neural network, extracts hierarchical features from the 3D neural image through a shared encoder, and synchronously outputs segmentation probability maps of neural fibers using two independent decoders. Probability plot with center line ; Structure extraction module: For the segmentation probability map Perform three-dimensional refinement processing to extract the initial neuronal skeleton. ; Topology guidance signal generation module: based on the segmentation probability map With the centerline probability diagram The pixel-by-pixel response intensity difference generates a topology guidance signal. ; Backtracking Refinement Module: Based on the aforementioned topology guidance signal Calculate the initial neuronal skeleton Each node Topology , with nodes Topology To obtain topologically consistent neural fiber reconstruction results, selective path tracing and backtracking pruning are performed for reference.

[0044] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0045] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A method for neural fiber reconstruction based on centerline features, characterized in that: Includes the following steps: S1. Construct a dual-branch parallel convolutional neural network. Extract hierarchical features of the 3D neural image through a shared encoder and simultaneously output segmentation probability maps of the nerve fibers using two independent decoders. Probability plot with center line ; S2, the segmentation probability map Perform three-dimensional refinement processing to extract the initial neuronal skeleton. ; S3, Based on the segmentation probability map With the centerline probability diagram The pixel-by-pixel response intensity difference generates a topology guidance signal. ; S4. Based on the topology guidance signal Calculate the initial neuronal skeleton Each node Topology , with nodes Topology To obtain topologically consistent neural fiber reconstruction results, selective path tracing and backtracking pruning are performed for reference.

2. The neural fiber reconstruction method based on centerline features according to claim 1, characterized in that: The specific process of step S1 is as follows: S11. Construct a dual-branch parallel convolutional neural network based on an encoder-decoder architecture. The shared encoder downsamples the input three-dimensional neural image through continuous 3D residual convolutional blocks to extract high-dimensional semantic features. Each convolutional layer in the encoder is followed by a batch normalization and correction linear unit. S12. The feature stream of the encoder-decoder architecture is decoupled after deep decoding and enters two parallel decoder branches. The segmentation decoder branch restores the image resolution and generates a segmentation probability map through upsampling and skip connections. The centerline decoder branch outputs a centerline probability map by extracting the geometric center path of the nerve fibers. .

3. The neural fiber reconstruction method based on centerline features according to claim 2, characterized in that: Step S1 further includes: S13. Train a dual-branch parallel convolutional neural network using an asymmetric weighted joint loss function to enhance its ability to perceive topological structures. (Asymmetric weighted joint loss function) The formula is as follows: in, and Take a constant based on the data. For branch loss, For centerline topology branch loss, By increasing the loss weight of the centerline branch, the bi-branch parallel convolutional neural network is forced to prioritize fitting the topological skeleton of the neural fibers during the optimization process.

4. The neural fiber reconstruction method based on centerline features according to claim 3, characterized in that: In step S13, the centerline topology branch loss The calculation process is as follows: Gaussian blurring is applied to the centerline truth labels, and its distribution function is expressed as: in, For the true centerline coordinates, To control the standard deviation of the influence range.

5. The neural fiber reconstruction method based on centerline features according to claim 4, characterized in that: The specific process of step S2 is as follows: S21. Use a three-dimensional parallel thinning algorithm to determine pixels whose local topological connectivity is not changed after deletion. S22. Identify all boundary pixels that meet the definition within each iteration cycle. The mathematical expression for the iteration process is as follows: in, Indicates the first The skeleton set after the next iteration This is the set of boundary pixels of the current target; S23. Perform the symmetrical stripping principle and delete the boundary pixels in parallel to avoid geometric offset of the skeleton; S24. Repeat the above iterative process until there are no more removable boundary pixels in the image, finally obtaining the initial neuron skeleton as a single pixel wide. Connected graph.

6. The neural fiber reconstruction method based on centerline features according to claim 5, characterized in that: The specific process of step S3 is as follows: S31. Calculate the response intensity difference metric for each pixel. Detecting false branches along the centerline: S32. Generate a topology guidance signal by setting a dual-threshold discrimination criterion. .

7. The neural fiber reconstruction method based on centerline features according to claim 6, characterized in that: The specific process of step S32 is as follows: S321, Let The structural confidence threshold. Topology missing threshold, topology guiding signal The generating function is defined as: Generated topology guidance signal Mark all potential topology error areas; S322, Topology Guidance Signal By performing morphological closing operations or small-area filtering based on connected domain volume, high-dimensional topological prior information is transformed into negative evidence that can be used to guide skeleton refinement, enabling precise navigation of false adhesion regions.

8. The neural fiber reconstruction method based on centerline features according to claim 7, characterized in that: The specific process of step S4 is as follows: S41. Scan the initial neuronal skeleton. And calculate each node Topology That is, the total number of adjacent skeleton points of a node in its 26-neighborhood: in, For indicator functions; S42. Taking leaf nodes as the starting point, the deletion decision function is used to identify pseudo-paths that need to be undone in order to perform the refinement process. The deletion decision function is expressed as follows: in, To refine the trigger threshold; S43. When a leaf node is determined to be deleted, the iterative backtracking procedure is started to trace the unique adjacent node of the current node along the skeleton path towards the main trunk. And remove them; the algorithm is dynamically updated as each node is removed. Topological degree distribution within the neighborhood; S44, if the updated If it becomes a new leaf node, then it is repeatedly checked whether it meets the deletion decision until the backtracking process stops; S45. Without disrupting the true connectivity, eliminate ambiguous paths caused by fiber adhesion and output topologically consistent neural fiber reconstruction results.

9. The neural fiber reconstruction method based on centerline features according to claim 8, characterized in that: In step S43, the backtracking logic is as follows: Continuous pruning iteration: After deleting the current node, trace its unique adjacent predecessor node along the skeleton towards the trunk; Dynamic update of topology: For each node deleted, the topology of its predecessor node is updated in real time. If the predecessor node is transformed into a new leaf node after pruning (i.e., the topology becomes 1), it is included in the continuous pruning sequence. Backtracking Stop Constraint: Continuously backtrack and prune until a node that satisfies any stopping condition is encountered. Firstly, when backtracking to topological degree... The bifurcation connection point; secondly, the probability diagram of the center line. That is, backtracking to a confidence region with strong topological evidence, and then the tracking stops.

10. A neural fiber reconstruction device guided by central axis features, comprising a computer program, characterized in that: The computer program is capable of performing the neural fiber reconstruction method based on centerline features as described in any one of claims 1 to 9.

Images