Efficient nucleus segmentation of white blood cells mimicking the human perception of color

In this study, we present a nucleus segmentation proposal of white blood cells (WBCs) using chromatic features. It is human inspired on perception of color: a person locates the nucleus of the WBCs by the chromatic contrast between the nucleus and the other elements of the blood smear. To implement that, we segment the nucleus by selecting the pixels with high chromatic variance. First, an unsupervised neural network, which was trained offline to recognize different colors is applied to the images. Thereby, the hue of the pixels is normalized, and the chromatic variance is accurately computed. Processing the hue and using the unsupervised neural network the brightness and staining variations are robustly estimated. In previous related works, the color components are processed separately as uncorrelated intensity channels, and the mathematical operations are selected intuitively. Unlike that, we use color as a feature without separating the hue components, keeping their correlation, so the formal treat becomes systematic. Experiments use the RGB and HSV spaces with three public image databases: ALL-IDB2, CellaVision, and JTSC. A pixel-level segmentation evaluation is performed by comparing the segmented images with the ground truth. Our proposal competes with current methods since the values in accuracy, specificity, precision, sensitivity, dice coefficient, kappa index, and true positive rate all are similar to or improved upon the state of the art. The performance of our approach is classified as excellent regarding the kappa index value, and it detects at least 80% of the cells with an average dice coefficient larger than 0.9