Automatic segmentation of Drosophila neurons for content based retrieval using a minimum spanning tree approach

A critical part in our knowledge of the brain is the understanding of the structure and morphology of neurons. In recent years, we have witnessed a tremendous effort in developing libraries of neuronal structures that can be used for multiple purposes. These critical studies include modeling the brain connectivity and understanding how cellular structure regulates brain function. Presently, methods for tracing neuronal structures from microscopy images of neurons are not only tedious, but prone to user bias; hence they are not suitable for practical purposes. State of the art automatic neuron tracing algorithms fail to give satisfactory results for neuron images with low contrast, discontinuous filaments, and complex branching. In this paper, we develop Tree2Tree, an automatic neuron segmentation and morphology generation algorithm. Our algorithm uses a global tree linking strategy to connect smaller medial trees of visible components of the neuron in order to give an optimal explanation of the brightness patterns of the neuron branches. We intend to utilize these graph-based segmentation results in the retrieval of similar neurons from an extensive database of segmented neurons. We have tested our algorithm on confocal microscopy images of Drosophila neurons and our segmentation results are congruent with (within ±5.3 pixel RMSE) the manually obtained ground truth.