Kidney Segmentation of Histopathological Images with Edge-Aware U-Net to Support Medical Diagnosis and Treatment Planning

Accurate segmentation of renal anatomical structures is essential for informed clinical decision-making in nephropathology, supporting precise diagnosis, treatment planning, and longitudinal monitoring of kidney diseases. In this work, we propose an Edge-Aware U-Net architecture with Boundary-Sensitive Optimization, specifically designed to address the challenges of fine anatomical boundary delineation in histopathological images. Comprehensive benchmarking against state-of-the-art models including U-Net, Attention U-Net, and ResUNet demonstrates robust quantitative performance alongside strong potential for clinical deployment. The proposed model achieves superior boundary preservation, reflected by a high structural similarity index (SSIM: 0.9473), while maintaining computational efficiency with an average inference time of 52 ms per image. It further outperforms existing methods across key image quality metrics, including PSNR (17.269 dB), MAE (0.0266), and RMSE (0.0321). Clinical validation indicates statistically significant improvements in glomerular detection (p < 0.01) and effective tubulointerstitial differentiation (F1-score: 0.891), with consistent performance observed across multiple staining protocols.