MSS-WISN: Multiscale Multistaining WBCs Instance Segmentation Network

Accurate segmentation and detection (instance segmentation) of white blood cells (WBCs) from whole slide images remains a challenging task, as the WBCs vary widely in shapes, sizes, and colors caused by different cell subtypes and various staining techniques. In this paper, we propose a novel framework for end-to-end segmentation and detection of WBCs that are on multiple scales and stained by different techniques. We name the framework the multi-scale and multi-staining WBC instance segmentation network (MSS-WISN). The MSS-WISN consists of two parts: 1) a feature extraction network for strengthening the feature expression and minimizing the impact of different staining techniques, and 2) a feature fusion network for highlighting salient features and thereby eliminating the effect of scale variations. To verify the effectiveness of the MSS-WISN, we build a new dataset containing 302 Magenta stained images (collected by Tianjin Medical University) and 242 Wright stained images (from a public dataset). Experiments show that the proposed framework outperforms other state-of-the-art methods in terms of WBC detection and WBC segmentation, achieving the highest F1-Score (0.901) and Dice (0.902).