Cell image segmentation (CIS) is critical for quantitative imaging in cytometric analyses. The data derived after segmentation can be used to infer cellular function. To evaluate CIS algorithms, first for dealing with comparisons of single cells treated as two-dimensional objects, a misclassification error rate (MER) is defined as a weighted sum of the false negative rate and the false positive rate. Then, all cells’ MERs are aggregated to constitute a new measure called the total error rate, which statistically takes account of the sizes of the cells in such a way that the weight on the result for an algorithm is higher if larger cells are not segmented correctly. This total error rate is used to measure the performance level of CIS algorithms. It was tested by applying ten CIS algorithms taken from the image processing toolkit ImageJ to 106 cells in our database. Furthermore, these cells with different sizes were manually segmented to be treated as the ground-truth cells. The test results were supported by the primitive pairwise comparison between two algorithms’ MERs on all cells.
Paper
Full text
A New Measure in Cell Image Segmentation Data Analysis
Semantic Scholar · Biology · 2012
Abstract
Cell image segmentation (CIS) is critical for quantitative imaging in cytometric analyses. The data derived after segmentation can be used to infer cellular function. To evaluate CIS algorithms, first for dealing with comparisons of single cells treated as two-dimensional objects, a misclassification error rate (MER) is defined as a weighted sum of the false negative rate and the false positive rate. Then, all cells’ MERs are aggregated to constitute a new measure called the total error rate, which statistically takes account of the sizes of the cells in such a way that the weight on the result for an algorithm is higher if larger cells are not segmented correctly. This total error rate is used to measure the performance level of CIS algorithms. It was tested by applying ten CIS algorithms taken from the image processing toolkit ImageJ to 106 cells in our database. Furthermore, these cells with different sizes were manually segmented to be treated as the ground-truth cells. The test results were supported by the primitive pairwise comparison between two algorithms’ MERs on all cells.