Yeast cell detection in color microscopic images using ROC‐optimized decoloring and segmentation

Abstract

Abstract For the detection and evaluation of yeast cell microscopic images, decoloring and segmentation is crucial. Usually, a standard operation of color conversion to grayscale is used and subsequently, a thresholding algorithm is applied to yield the segmented cells. However, both types of these operations are controlled merely by fixed empirical parameters which are prone to inaccuracy. The authors illustrate a common problem of such a nonoptimal decoloring. They then develop a novel pipeline of algorithms for yeast cell color phantom generation. In this study, the authors extend Grundland' s decoloring approach and couple it with an adaptive thresholding operation. For the generated phantom, the parameters of these operations are optimized using the ROC‐based F1‐measure. It is showed that this measure is a better choice than the Precision and Recall measure alone. Computer experiments with various combinations of the operation coupling and different phantom backgrounds are accomplished, and the optimum values of all the controllable parameters are found. For the color phantom representing the most realistic conditions, the final fully optimized 4D combination of the improved Grundland' s decoloring and locally adaptive thresholding ensures more efficient cell detection performance than parameterless or nonoptimized combinations.