Abstract
Accurate segmentation of glomeruli in kidney histopathology images is vital for diagnosing renal diseases but remains challenging due to several limitations in current CNN-based methods. First, multiscale learning struggles to preserve the critical features of small glomeruli, as uniform processing dilutes their discriminative information and hinders effective separation from surrounding tissue. Second, limited interaction between consecutive layers leads to incomplete feature fusion, weakening the representation of complex glomeruli structures. Finally, traditional upsampling techniques introduce artifacts like checkerboards and blurred edges, further compromising segmentation accuracy. To address these challenges, we propose GlomNet, a simple yet effective glomeruli segmentation network designed for whole-slide images (WSIs) with diverse staining protocols, enabling accurate segmentation for Kidney disease diagnosis. GlomNet integrates three key components: Local-Global Cris-Cross Former (LGC(2)-Former) efficiently captures global contextual information and fine-grained local details by processing the input feature map through parallel global and local branches. It then aggregates these contextual cues along both horizontal and vertical directions, enhancing spatial awareness and improving segmentation accuracy of complex glomeruli structures. Hierarchical Multi-head Feature Aggregation (HMFA) promotes richer feature extraction and effective multi-scale fusion by enabling mutual guidance between consecutive layers, addressing the underutilization of complementary feature information. Feature-Refined Upsampling (FRU) resolves issues related to checkerboard artifacts and blurry edges in traditional upsampling methods, while improving resolution and maintaining feature integrity during the decoding phase for more accurate segmentation. Experimental results demonstrate that GlomNet achieves state-of-the-art performance across multiple benchmarks, including NEPTUNE, HuBMAP-1, and HuBMAP-2, showcasing its robustness across diverse histological conditions, with significant potential for real-world applications in the diagnosis of kidney disease and medical image analysis.