From Pixels to Precision—A Dual-Stream Deep Network for Pathological Nuclei Segmentation

Segmenting cell nuclei in histopathological images is an extremely important process for computational pathology, affecting not only the accuracy of a disease diagnosis but also the analysis of biomarkers and the assessment of cells performed on a large scale. Although many deep learning models can take out global and local features, it is still difficult to find a good balance between semantic context and fine boundary precision, especially when nuclei are overlapping or have changed shapes. In this paper, we put forward a novel deep learning model named Dual-Stream HyperFusionNet (DS-HFN), which is capable of explicitly representing the global contextual and boundary-sensitive features for the robust nuclei segmentation task by first decoupling and then fusing them. The dual-stream encoder in DS-HFN can simultaneously acquire the semantic and edge-focused features, which can be later combined with the help of the attention-driven HyperFeature Embedding Module (HFEM). Additionally, the dual-decoder concept, together with the Gradient-Aligned Loss Function, facilitates structural precision by making the segmentation gradients that are predicted consistent with the ground-truth contours. On various benchmark datasets like TNBC and MoNuSeg, DS-HFN not only achieves better results than other 30 state-of-the-art models in all evaluation metrics but also is less computationally expensive. These findings indicate that DS-HFN provides a capability for accurate nuclei segmentation, which is essential for clinical diagnosis and biomarker analysis, across a wide range of tissues in digital pathology.