Optimization of Traced Neuron Skeleton Using Lasso-Based Model

Shiwei Li, Tingwei Quan, Xu Cheng, Qing Huang, Hongtao Kang, Yijun Chen, Anan Li, Ling Fu, Qingming Luo, Hui Gong, Shaoqun ZengFrontiers in Neuroanatomy 2019 · DOI

TL;DR

This paper introduces a novel two-part method to refine automatically traced neuronal skeletons, addressing common inaccuracies in tortuous structures and at branch points. By leveraging L1-norm minimization for smooth tracing and a specialized model for branch point detection, the approach significantly improves the accuracy and consistency of neuronal reconstructions, reducing manual editing efforts.

What Problem Does This Paper Address?

Reconstructing the intricate 3D structure of neurons from microscopy images is crucial for understanding brain function. This process typically involves tracing neuronal "skeletons" – simplified representations of neurites. However, existing automated tracing methods often struggle with two key challenges:

  1. Tortuous Structures: Neurons have highly winding and curved segments (neurites). Standard tracing algorithms, which often favor smoothness, can fail to accurately capture these sharp turns, leading to distorted skeletons.
  2. Branch Points: Where neurites split, the complex interplay of signal intensity, diameter changes, and overlapping structures makes it difficult for algorithms to precisely pinpoint the true branching location.

These inaccuracies hinder quantitative analysis and require extensive manual correction, slowing down research. This paper aims to automate and improve this refinement process.

Methods

The authors propose a two-stage optimization approach:

  1. Optimized Neurite Skeleton Detection (L1-Norm Minimization): This method tackles tortuous structures. It's inspired by the Lasso (Least Absolute Shrinkage and Selection Operator) technique, which is known for promoting sparsity. The core idea is to model the neurite as a sequence of points. The optimization aims to:
  2. Keep skeleton points within regions of high image intensity (local maxima).
  3. Minimize a term related to the curvature of the skeleton. By using an L1-norm penalty on curvature-related terms, the model encourages sparsity in curvature, meaning most points are in smooth regions, but it can accurately identify and follow points in tortuous areas. This is a key departure from L2-norm methods that enforce overall smoothness and struggle with sharp turns.
  4. Optimized Branch Point Detection: This method focuses on accurately locating branch points. It recognizes that a branch point is where multiple neurite segments meet. The approach involves:
  5. First, using the L1-norm method to obtain an optimized skeleton for the neurites connected to a potential branch point.
  6. Then, modeling the local branch structure as three segments meeting at a point.
  7. A specialized optimization model (Equation 5) is used, incorporating information from the surrounding segments and an orthonormal coordinate system derived from the neurite's direction.
  8. To improve robustness, the final branch point position is determined by averaging the results from optimizing different segment combinations around the branch point.

These two models are integrated to process an initial skeleton and produce an optimized version.

Datasets

The authors validated their method on several datasets:

  • MOST (Micro-Optical Sectioning Tomography) dataset: Used for initial testing and demonstrating the method's effectiveness on real neuronal data.
  • BigNeuron project datasets: A large-scale, open-source initiative providing diverse neuronal images. This was used to show the method's applicability to various tracing algorithms and large-scale reconstructions.
  • Synthetic datasets: Generated by the authors to specifically test the model's performance on controlled tortuosity and noise levels, allowing for a clear comparison with other methods.

Key Insights and Ideas

  • L1-norm for Tortuosity: The paper highlights that L1-norm minimization is inherently better suited than L2-norm methods for capturing the sparse nature of sharp turns in neuronal structures. This is a crucial insight for improving tracing accuracy in complex morphologies.
  • Branch Point Averaging: The strategy of averaging optimized branch point positions derived from different segment combinations provides a more robust and accurate final branch point location, mitigating issues caused by local noise or structural ambiguities.
  • Unifying Traced Skeletons: The method demonstrates the ability to bring skeletons traced by different algorithms closer to a consistent, optimized representation, reducing inter-algorithm variability and improving downstream analysis.
  • Reducing Manual Effort: By automating the correction of common tracing errors, the method significantly reduces the need for laborious manual editing, accelerating the overall reconstruction pipeline.

Key Results

  • Tortuous Structures: The L1-norm method successfully identified tortuous neurite skeletons, outperforming an L2-minimization-based method, especially in noisy conditions and with varying degrees of tortuosity (Figures 3 & 4).
  • Branch Points: The optimized branch point detection method showed improved accuracy, with optimized points deviating less from real positions compared to initial traces, particularly in challenging cases with signal intensity variations or small angles (Figure 5).
  • Consistency Across Algorithms: When applied to skeletons generated by different tracing tools (NeuroGPS-Tree, NeuronStudio, APP2), the optimization method produced consistent results, reducing the average distance between optimized skeletons from different initial traces (Figure 6 & Table 3).
  • Large-Scale Data: The method was successfully applied to very large datasets (e.g., 4295 × 7401 × 1625 voxels), demonstrating its scalability and effectiveness in correcting complex structures in massive reconstructions (Figure 8).
  • Morphological Features: Measurements showed that the optimized reconstructions led to slight changes in total length and local branch angles, suggesting a more accurate representation of the neuron's morphology (Table 2).

Limitations and Future Work

  • Multifurcation Points: The current model is primarily designed for bifurcations (two-way splits). While multifurcation points (more than two splits) can be approximated by dividing them into bifurcations, this may reduce accuracy. Future work could address direct multifurcation optimization.
  • Input Skeleton Quality: The method relies on an initial skeleton. If the input has significant errors, such as spurious links, the optimization performance might be affected. Manual checking of input data is still sometimes necessary.
  • Dense Neuron Environments: In scenarios with densely packed neurons, the method's reliance on local signal intensity might be influenced by neighboring neurons. Future work aims to address this by incorporating neuron segmentation and shape identification to isolate the target neuron's morphology.

Summary updated

Abstract

Reconstruction of neuronal morphology from images involves mainly the extraction of neuronal skeleton points. It is an indispensable step in the quantitative analysis of neurons. Due to the complex morphology of neurons, many widely used tracing methods have difficulties in accurately acquiring skeleton points near branch points or in structures with tortuosity. Here, we propose two models to solve these problems. One is based on an L1-norm minimization model, which can better identify tortuous structure, namely, a local structure with large curvature skeleton points; the other detects an optimized branch point by considering the combination patterns of all neurites that link to this point. We combined these two models to achieve optimized skeleton detection for a neuron. We validate our models in various datasets including MOST and BigNeuron. In addition, we demonstrate that our method can optimize the traced skeletons from large-scale images. These characteristics of our approach indicate that it can reduce manual editing of traced skeletons and help to accelerate the accurate reconstruction of neuronal morphology.
BibTeX
@inproceedings{li2019optimization,
  author = {Shiwei Li and Tingwei Quan and Xu Cheng and Qing Huang and Hongtao Kang and Yijun Chen and Anan Li and Ling Fu and Qingming Luo and Hui Gong and Shaoqun Zeng},
  title = {{Optimization of Traced Neuron Skeleton Using Lasso-Based Model}},
  booktitle = {Frontiers in Neuroanatomy},
  year = {2019},
  doi = {10.3389/fnana.2019.00018},
}
Figure