Digital pathology and microscopic image analysis play an important role in cell morphology research. In particular, the effective segmentation of White Blood Cells (WBCs) remains a challenging problem due to the blurring boundaries of WBCs under rapid staining, as well as the adhesion between leukocytes and other cells. In this paper, we propose a novel WBC (including nuclei and cells) segmentation algorithm based on both sparsity and geometry constraints. Specifically, we first construct a sparse image representation via combining the HSL color space and the RGB color channels, followed by the use of a sparsity constraint to only preserve useful information from the nuclei features. In addition, we introduce a robust model fitting strategy (i.e., the geometry constraint) to detect cells. Our model fitting strategy is able to significantly improve the robustness of the proposed segmentation algorithm against outliers that could seriously contaminate WBCs. The experimental results show that the proposed algorithm presents clear advantages over the state-of-the-art WBC segmentation algorithms in terms of accuracy.

Introduction

White blood cells (WBCs) [1], [2] are important defense cells in human blood that consists of five kinds of cells, i.e., neutrophils, basophils, eosinophils, monocytes, and lymphocytes. The WBC segmentation is a challenging task for a variety of medical diagnosis applications. For example, the visual examination of WBCs in blood smears collected under a bright field microscope can be used to diagnose various diseases, such as septic bacterial inflammation, uremia, and various kinds of leukaemia. A number of WBC segmentation methods have been proposed in recent years. In general, existing methods can be divided into two distinct categories: supervised vs. unsupervised WBC segmentation methods. The supervised WBC segmentation methods [3]–[۶] formulate the WBC segmentation problem as a multi-class classification problem. They are typically solved in a two-step manner, i.e., first extracting image features and then classifying the extracted features. However, these supervised methods require a large number of annotated training samples, which are often manually labeled and difficult to obtain. Particularly, the manual pixelwise segmentation process is tedious and error-prone for the abundant fine structures in the blood cell imagery. In addition, the training and test images are assumed to be visually similar to minimize the domain shift between training and test images. In practice, this assumption could negatively impact the generalization abilities of these supervised algorithms.