Shape Descriptors Comparison for Cell Tracking

Abstract

New microscope technologies are enabling the acquisition of large volumes of live cell image data. Accurate temporal object tracking is required to facilitate the analysis of this data. One principle component of cell tracking is correspondence, matching cells between consecutive frames. This component can be enhanced by incorporating shape metrics into the tracking model. The measure of shape similarity between two objects can be accomplished using Fourier descriptors, derived from a one dimensional shape signature. The type of one dimensional shape signature affects the quality of the resulting Fourier descriptors when dealing with noisy object boundaries. We present a comparison of several different methods of converting a two dimensional object boundary into a one dimensional shape signature suitable for computing Fourier descriptors. The Fourier descriptors are evaluated on two shape datasets with ground truth and used as a similarity feature for performing object tracking on a time-lapse series of NIH 3T3 cells. Experimental results show that for noisy object boundaries, Fourier descriptors constructed from the R-Theta Binning centroid distance shape signature presented here perform better than Fourier descriptors constructed from the other evaluated shape signatures.