<p>Renal pathology, serving as the gold standard for diagnosing and treating kidney diseases, relies on the precise segmentation of glomerular structures in microscopic images. Under pathological conditions, glomeruli often exhibit thinning of the capsule wall and morphological abnormalities, leading to blurred glomerular boundaries and disrupted structural features in pathological images. Existing Mamba-based methods, which depend on directional scanning, still fall short in capturing multi-scale spatial relationships and extracting complex morphological features. When applied to two-dimensional images, the directional dependency issue remains inadequately addressed, thereby limiting the performance of fine segmentation. To address these challenges, this paper proposes the MambaDS U-Net segmentation method. By introducing a multi-scale feature extraction module within the Visual State Space Block, it sequentially extracts multi-scale features based on selective structured state-spaces. This enhances the model’s ability to utilize image spatial structural information during feature extraction. Furthermore, addressing the geometric distortion issues inherent in existing sampling methods during lesion segmentation, this paper incorporates positional information constraints on top of dynamic sampling. This ensures the preservation of spatial geometric relationships within lesion regions during sampling, preventing distortion or deformation of critical anatomical structures. Experimental results demonstrate that the proposed method achieves a Dice score of 88.87% for the 5/6Nx lesion category on the KPIs dataset. On the HubMap dataset, it improves the Precision by 1.8% compared with the second-best method. compared with existing state-of-the-art methods. These results indicate that MambaDS U-Net provides robust and competitive segmentation performance for pathological glomeruli exhibiting complex morphological variations.</p>