Using Stain Decomposition for Nucleus Segmentation on Multisource H&E Slide Images

Hung-Chun Hsu, Hung-Wen Tsai, Maria Gabrani, Pau-Choo Chung, Antonio Foncubierta-Rodríguez, Yu-Ting Wu

Abstract: The deep learning technique is widely used for whole slide image (WSI) analysis due to its speed and relatively objective results. Nucleus instance segmentation, in which the nuclei in the WSI are individually identified, can provide many useful parameters to assist and accelerate the diagnosis process for pathologists. Hematoxylin and eosin stain (H&E stain) is commonly used on pathology tissue slides to observe the behavior and structure information of cells. However, the color-oriented features of the nuclei vary widely depending on the dye and staining protocol employed and also the scanning system used to obtain the digital WSIs. Accordingly, the proposed lightweight network module attached to the front of an arbitrary network can provide a more robust segmentation performance irrespective of staining protocols employed and scanning circumstances. Notably, the proposed module improves the performance for unseen color-oriented features and thus can benefit the realworld applications, where the WSIs may be collected from multiple hospitals and laboratories.