Direction-Guided Watershed for Adherent Cell Instance Segmentation

Accurate segmentation of adherent cells remains a critical challenge in computational pathology. While relation-based modeling excels at pixel-wise prediction, its reliance on watershed post-processing renders it susceptible to noise, potentially leading to over- or under-segmentation in complex adherent regions. To address this, we introduce directional guidance to augment the watershed algorithm via two complementary modules: a Directional Disparity Boundary Discernment (DDBD) module and an Oriented-Attentive Watershed Augmentation (OAWA) module. Together with high-gradient pixel screening, DDBD forms a dual-cue edge screening by assessing directional disparity, effectively discriminating true boundaries from noise. OAWA then accentuates these valid boundaries by weighting the neighborhoods along their dominant gradient directions, guiding the watershed to focus on salient ridges. Extensive experiments on the CytoDArk0 dataset demonstrate state-of-the-art performance across multiple magnifications, improving the accuracy and robustness of adherent cell instance segmentation.