Improving kidney segmentation in pathological images: a multiscale approach to resolve fragmentation and incomplete boundaries

Kidney biopsy plays a crucial role in the diagnosis and treatment of kidney disorders, as well as in the evaluation of the suitability of the donor for transplantation. Accurate and early detection of glomeruli and microvascular structures is essential to ensure precise diagnoses, develop effective treatment strategies, and facilitate reliable donor evaluations. However, current methods for segmenting pathological images often fail, resulting in incomplete and fragmented representations of kidney structures due to broken boundaries, poor integration of features, and limited scalability. These problems reduce the accuracy and reliability of existing approaches. To address these challenges, we introduce InFeNet, an innovative multiscale feature enhancement network specifically designed to improve the segmentation of functional tissue units in renal pathology. InFeNet combines advanced Residual Networks with multi-scale feature extraction and refinement, incorporating correlations among spatial features to optimize feature maps. By harmonizing spatial and high-level representations, InFeNet achieves unparalleled precision and reliability, ensuring accurate segmentation of glomeruli and microvascular structures. The results showed that InFeNet achieved equivalent or superior performance on two publicly available datasets: on the HuBMAP-Hacking the Human Vasculature dataset, it reached an accuracy of 97.2% and an F1-Score of 96.1%, whereas on the HuBMAP-Hacking the Kidney dataset, it achieved an accuracy of 98.8% and an F1-Score of 95.5% on PAS-stained images. We evaluated InFeNet on two datasets, where it outperformed state-of-the-art models, demonstrating superior accuracy and predictive performance. Comparison of encoders from two families further validated its effectiveness, highlighting its potential for practical implementation in digital renal pathology to improve diagnostic and therapeutic outcomes in large-scale medical imaging. Its superior performance and precision make it ideal for real-world applications in automated medical diagnosis, treatment planning, and more, offering reliable and efficient healthcare outcomes. © 2025 Elsevier B.V., All rights reserved.