Leveraging advanced feature extraction for improved kidney biopsy segmentation

Medical image segmentation faces critical challenges in renal histopathology due to the intricate morphology of glomeruli characterized by small size, fragmented structures, and low contrast against complex tissue backgrounds. While the Segment Anything Model (SAM) excels in natural image segmentation, its direct application to medical imaging underperforms due to (1) insufficient preservation of fine-grained anatomical details, (2) computational inefficiency on gigapixel whole-slide images (WSIs), and (3) poor adaptation to domain-specific features like staining variability and sparse annotations. To address these limitations, we propose V-SAM, a novel framework enhancing SAM's architecture through three key innovations: (1) a V-shaped adapter that preserves spatial hierarchies via multi-scale skip connections, recovering capillary-level details lost in SAM's aggressive downsampling; (2) lightweight adapter layers that fine-tune SAM's frozen encoder with fewer trainable parameters, optimizing it for histopathology textures while avoiding catastrophic forgetting; and (3) a dynamic point-prompt mechanism enabling sub-pixel refinement of glomerular boundaries through gradient aware localization. Evaluated on the HuBMAP Hacking the Human Vasculature and Hacking the Kidney datasets, V-SAM achieves state-of-the-art performance, surpassing 89.31%, 97.65% accuracy, 86.17%, 95.54% F1-score respectively. V-SAM sets a new paradigm for adapting foundation models to clinical workflows, with direct applications in chronic kidney disease diagnosis and biomarker discovery. This work bridges the gap between SAM's generalizability and the precision demands of medical imaging, offering a scalable solution for resource constrained healthcare environments.