Leukocyte Segmentation via End-to-End Learning of Deep Convolutional Neural Networks

Identification and analysis of leukocytes (white blood cells, WBC) in blood smear images play a vital role in the diagnosis of many diseases, including infections, leukemia, and acquired immune deficiency syndrome (AIDS). However, it remains difficult to accurately segment and identify leukocytes under variable imaging conditions, such as variable light conditions and staining degrees, the presence of dyeing impurities, and large variations in cell appearances, e.g., size, color, and shape of cells. In this paper, we propose an end-to-end leukocyte segmentation algorithm that uses pixel-level prior information for supervised training of a deep convolutional neural network. Specifically, a context-aware feature encoder is first introduced to extract multi-scale leukocyte features. Then, a feature refinement module based on the residual network is designed to extract more discriminative features. Finally, a finer segmentation mask of leukocytes is reconstructed by a feature decoded based on the feature maps. Quantitative and qualitative comparisons of real-world datasets show that the proposed method achieves state-of-the-art leukocyte segmentation performance in terms of both accuracy and robustness.