Pediatric kidney disease poses a significant global health challenge, often progressing silently and requiring precise diagnostic intervention. Traditional renal biopsy analysis is time-consuming, labor-intensive, and subject to human variability. To address these challenges, KFBIO partnered with the Children’s Hospital of Zhejiang University School of Medicine to develop an innovative artificial intelligence (AI) system designed to assist pathologists in diagnosing pediatric kidney diseases with unprecedented accuracy and speed.

The Challenge: Diagnosing Pediatric Kidney Disease

Chronic Kidney Disease (CKD) affects millions of children worldwide, with symptoms ranging from hematuria and proteinuria to growth retardation. Accurate diagnosis typically requires renal biopsy and detailed examination by specialist nephropathologists—a process that is both tedious and prone to inconsistency. With a growing shortage of pathologists globally and an increasing number of pediatric CKD cases, there is an urgent need for intelligent, automated diagnostic tools.

Introducing APKD: The AI-Based Pediatric Kidney Diagnosis System

In response, KFBIO and Zhejiang Children’s Hospital co-developed the AI-based Pediatric Kidney Diagnosis (APKD) system—a deep learning framework capable of segmenting, classifying, and quantifying kidney structures from whole slide images (WSIs). The system was trained on a robust dataset of 2,935 pediatric patients and 93,932 manually annotated histological structures, making it one of the most comprehensive pediatric kidney AI models to date.

How APKD Works: A Multi-Module AI Framework

APKD integrates four specialized AI modules:

1. Kidney Structure Segmentation – Identifies arteries, tubules, glomeruli, and glomerular tufts.

2. Nuclei Segmentation – Detects and counts cell nuclei within tissue samples.

3. Glomerular Crescent Classification – Distinguishes between crescentic and non-crescentic glomeruli, key for diagnosing progressive nephritis.

4. Mesangial/Endothelium Region Segmentation – Pinpoints areas of hypercellularity in PAS-stained images.

The system uses a customized SCNet with ResNet50 backbone, which outperformed five other state-of-the-art models with a mean average precision (mAP) of 94% across all kidney structures—including an exceptional 99% accuracy for glomerulus detection.

Key Performance Highlights

• 99% glomerulus detection accuracy – Near-perfect alignment with pathologist annotations.

• Strong correlation with manual detection – Spearman correlation coefficient of 0.98, ICC of 0.98.

• 5.5x faster than manual analysis – APKD processes slides in an average of 7 seconds vs. 37 seconds by pathologists.

• Comprehensive feature extraction – Quantifies 58 pathological features, including glomerular size, cellularity, crescent formation, and tubular metrics.

Clinical Validation and Pediatric Insights

The model was validated on an independent cohort of 1,625 retrospective cases, showing consistent and reliable performance across various pediatric kidney diseases, including IgA nephropathy, HSPN, and mesangial proliferative glomerulonephritis. Additionally, APKD enabled novel quantitative insights, such as the correlation between glomerular size and patient age—a finding consistent with earlier adult studies but newly confirmed in children.

Why This Collaboration Matters

Most AI renal pathology models have been trained on adult data, leaving a gap in pediatric-specific diagnostics. This collaboration marks a significant step toward child-focused AI pathology, accounting for developmental variations in kidney structure and disease presentation.

KFBIO’s high-throughput KF-PRO-400 slide scanners provided the high-resolution digital slides necessary for training APKD, while Zhejiang Children’s Hospital contributed deep clinical expertise and one of the largest pediatric renal biopsy databases in China.

Future Directions and Ongoing Research

While APKD represents a major advance, the team continues to expand the model’s capabilities. Future work includes:

• Multicenter data collection across Chinese hospitals

• Integration of additional stains (Silver, Trichrome)

• Segmentation of podocytes and other specific cell types

• Incorporation of clinical data for holistic diagnostic support

Conclusion

The collaboration between KFBIO and Zhejiang Children’s Hospital has yielded a powerful, validated AI tool that enhances the accuracy, consistency, and efficiency of pediatric kidney disease diagnosis. By combining cutting-edge AI with deep clinical insight, APKD not only supports pathologists but also paves the way for more personalized and proactive pediatric nephrology care.

Access the published dataset and learn more:
https://github.com/ChunyueFeng/Kidney-DataSet

Feng, C., Ong, K., Young, D. M., Chen, B., Li, L., Huo, X., Lu, H., Gu, W., Liu, F., Tang, H., Zhao, M., Yang, M., Zhu, K., Huang, L., Wang, Q., Marini, G. P. L., Gui, K., Han, H., Sanders, S. J., Li, L., Yu, W., & Mao, J. (2024). Artificial intelligence-assisted quantification and assessment of whole slide images for pediatric kidney disease diagnosis. Bioinformatics, btad740. https://doi.org/10.1093/bioinformatics/btad740