Abstract
We describe our preliminary results regarding the tracking of dendrites spreading from a neuron in confocal microscope images. When using a small number of image layers, we obtain good results by combining an EM-based local estimate of the probability that an image pixel belongs to a neuron filament with the global tree properties of the complete set of dendrites. The optimal tree is obtained with a modified minimum-spanning tree procedure. This approach extends naturally to complete data volumes and provides robust reconstruction even in noisy environments.
Introduction
Understanding brain function requires precise mapping of neuron morphology—the intricate, tree-like structures called dendrites. Traditionally, researchers have relied on manual or semi-automated tracing through stacks of confocal microscope images. However, automation is difficult due to discontinuities in filaments, non-uniform lighting artifacts, and the high complexity of dense, overlapping branches.
Methodology
The tracking problem is framed as Bayesian inference, where the hidden state is the true dendritic tree location and the visible state is pixel intensity. The approach follows a two-stage pipeline.
Local Probability Estimation
Instead of simple thresholding, the algorithm uses a localized version of the Expectation-Maximization (EM) algorithm. It assumes that in any small sub-window, pixel intensities follow a mixture of two Gaussian distributions: one for dark filaments and one for the lighter background. By re-estimating these parameters locally using integral images, the model adapts to lighting variations across the slide, producing a probability map of neuron filaments.
Global Structural Optimization
To convert these probabilities into a coherent structure, a modified Minimum Spanning Tree (MST) algorithm is employed:
- Anchor Points: Local maxima in the probability map are identified as potential nodes.
- Scoring: A log-likelihood score is defined for every potential edge. "Cheap" edges pass through high-probability pixels, while "expensive" ones pass through the background.
- Tree Building: A greedy approach connects anchors into a tree maximizing total likelihood.
- Pruning: Subtrees with negative total integrated scores are removed to eliminate noise.
Results and Discussion
The method produces delineations that compare favorably to manual tracing. A key insight is the enforcement of "tree-ness"; by requiring a tree structure, the algorithm can bridge gaps where local evidence is weak but global connectivity is strong.
The pruning and scoring mechanism effectively handles substantial gaps, retaining edges that connect high-scoring subtrees even if the edge itself has a poor local score. By maintaining a probabilistic framework throughout, the method retains more information from raw data than traditional de-noising techniques.
Conclusion and Future Work
While currently tested primarily in 2D, the framework is designed for 3D volumes. Future work will address "false intersections" in 2D projections and improve computational scaling using Kd-trees. Additionally, replacing simple intensity models with machine learning classifiers (e.g., SVMs) to recognize cylindrical filament patterns could further enhance robustness.
Summary updated