Image segmentation accuracy depends on many acquisition and algorithmic factors that determine the confidence of detecting segments of interest. We compared segmentation results of nine different segmentation techniques applied to two different cell lines and five different sets of imaging conditions illustrated in Figure 1.
Due to differences in imaging conditions and segmentation algorithms, we observed significant variability in the results of segmentation. We quantified and compared the results with a similarity index metric that evaluates the degree of underestimating or overestimating a cell object. The results of our study show that commonly used threshold-based segmentation techniques are less accurate than k-means clustering with multiple clusters. Segmentation accuracy varies with imaging conditions that determine the sharpness of cell edges, and with geometric features of a cell. Based on this observation, we propose a method that quantifies cell edge character, to provide an estimate of how accurately an algorithm will perform. The results of this study will assist the development of criteria for evaluating inter-laboratory comparability.
The Problem: The initial project problem was to determine whether image analysis techniques existed (or could be developed by the ITL groups) which would assist in the automatic differentiation of the A10 and 3T3 cell lines. The larger problem at hand was to determine what factors (and interactions) in the image, the imaging, and the image processing algorithms are important, and what the optimal settings of such factors were that permitted maximal extraction of information and maximal differentiation of cell lines. In particular, it was important to determine and characterize the limits of existing popular image segmentation techniques, and then exceed those limits by development of improved algorithms and methodology.
Experiment Design: An experiment design was constructed to extract a random and representative set of images that would serve as valid surrogate for the larger set. The number of factors that complicated the comparison of the 2 cell lines was large: cell lines (2: A10 and 3T3), wells (3), exposure time (3: low, medium, high), optical filter (2: optimal and sub-optimal), image fields (50), image field reps (3), and day (3).
Data Analysis: A variety of data analysis techniques--both quantitative and graphical--were applied to the data. Such techniques showed dominant factors, interactions, and optimal settings. Imaging conditions played a major role in the performance of an algorithm some algorithms performed well under certain conditions but fared poorly under alternate conditions. Statistical methodologies were developed and applied which ferreted out the robustly best (and worst) algorithms.
This activity led to the following outcomes:
Paper: Alden A. Dima, John T. Elliott, James J. Filliben, Michael Halter, Adele Peskin, Javier Bernal, Benjamin L. Stottrup, Marcin Kociolek, Mary C. Brady, Hai C. Tang, and Anne L. Plant: Comparison of segmentation algorithms for flourescence microscopy images of cells. Cytometry Part A, 79A: 545-559, 2011. (download pdf)