Answers are grounded in this paper and cite their sources.
Glomerular segmentation in whole-slide histopathology is essential for quantitative renal assessment, yet it is fundamentally challenged by local texture discriminability, spatial heterogeneity, and the need for multi-scale mapping of morphological priors. While convolutional neural networks address local textures via inductive biases, and Vision Transformers model global contexts through self-attention, neither paradigm simultaneously resolves these three challenges without incurring prohibitive computational overhead. Furthermore, clinical translation is frequently bottlenecked by human data scarcity. Here, we propose DualSeg, a dual-stage hybrid framework that unifies convolutional and attention capabilities through a synergistic local-to-global refinement strategy. The encoder employs Wave-Swin Blocks with dynamic propagation windows for hierarchical local feature extraction, followed by Vision Recurrent Weighted Key Value Blocks utilizing a novel Z-Shift operator for efficient, linear-complexity global context modeling. A lightweight decoder based on Non-negative Matrix Factorization subsequently fuses multi-scale features while suppressing background noise. On the murine KPIs and human HuBMAP datasets, DualSeg achieved mean DSCs of 92.98% and 89.79%, respectively. To evaluate crossspecies and cross-center generalizability under human data scarcity, we performed zero-shot inference on the KPMP dataset using models trained exclusively on murine data. DualSeg demonstrated superior generalization, validating its ability to leverage conserved mammalian morphological priors across staining domains. Code is available at https://github.com/unskye/DualSeg.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: