Most neurological analysis algorithms consist of three steps: image pre-processing, neural segmentation, and parameter extraction. Image pre-processing produces an optimized image for further analysis. Optimization usually consists of image enhancement and noise reduction, for which there are many well-developed methods. Parameter extraction depends on the methods used for neural segmentation. Therefore, neural segmentation is the key step that differentiates neurological analyzing algorithms.
There are two major approaches to segmenting neurons from pre-processed images: global processing (GP) and local exploration (LE). GP means applying image processing methods to the entire image as a whole, while LE refers to iteratively processing image section until the whole image has been analyzed. Key steps in GP often include edge detection and skeletonization. GP is subject to regional image noise that remains after image pre-processing. As a result, algorithms applying GP require clear raw images. Furthermore, since GP does not differentiate critical points or regions of interest, it is more difficult to obtain the parameter of interest, especially for describing regional morphology such as branching points and branching degree. There are numerous algorithms using a GP-approach. Most are too complicated and time-consuming for novice users to make simple analysis.
LE begins at specified starting points, and for each starting point, the neuron across the point is traced within a region. A point on the traced portion is picked as the next starting point and this continues until the whole image has been processed. Therefore, the two key processes in LE are starting point detection and regional neuron tracing. Most LE algorithms require unser input for the starting point (e.g., Imaris) or both the start and end points (e.g., ImageJ). These manual processes for starting point detection are time-consuming and subjective to inter-operator differences. The algorithm implemented in Neurient can detect starting points automatically by choosing the local maxima along a set of grid lines put on the image. However, a series of sophisticated methods has to be applied to avoid noise, threshold unqualified seed points, prevent multiple tracings due to multiple starting points are picked for one neurite. Chaohong Wu (2010) implemented a method to detect seed lines considering the symmetric properties of a line revealed in frequency domain. However, the parameters correlated to frequency domain cannot be intuitively obtained by an untrained user. Therefore the robustness of this algorithm for different kinds of neuro-imaging is limited.
In terms of local neuron tracing, most algorithms involve the processes of direction detection and starting point determination for the next steps. Methods include local gradients, cross-correlation kernel (Neurient), edge-based tracing (Yong Zhang et al. 2007), and energy minimization (NeuronJ, NeuriteIQ).
Based on this information, our team decided to design a program specifically for our sponsor’s corneal nerve morphological study. Three of the main goals of our design were to be automatic, fast, and free of human input. We chose a LG-approach since an alternative program using GP-approach has been designed by another team invited by our sponsor.