Kidney Week

Abstract: TH-OR022

TEAMKidney: A Cross-Species Deep Learning Framework for Automated Quantification of Kidney Ultrastructure in Transmission Electron Microscopy Images

Session Information

• Artificial Intelligence in Kidney Care: From Pathology to Prediction
  October 22, 2026 | Location: Room 711, Convention Center
  Abstract Time: 05:00 PM - 05:10 PM

Category: Artificial Intelligence, Digital Health, and Data Science

• 300 Artificial Intelligence, Digital Health, and Data Science

Authors

• Zou, Anqi, Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States

Anqi Zou, PhD

• Tan, Winston, Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States

Winston Tan

• Ji, Jiayi, Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States

Jiayi Ji

• Rojas-Miguez, Florencia, Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States

Florencia Rojas-Miguez, MD, MPH

• Chen, Hui, Boston Medical Center, Boston, Massachusetts, United States

Hui Chen

• Henderson, Joel M., Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States

Joel M. Henderson, MD, PhD

• Fan, Xueping, Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States

Xueping Fan, PhD

• Lu, Weining, Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States

Weining Lu, MD, MS, FASN

• Zhang, Chao, Boston University Chobanian & Avedisian School of Medicine, Boston, Massachusetts, United States

Chao Zhang, MSc, PhD

Background

Transmission electron microscopy (TEM) is an essential tool for quantifying glomerular ultrastructural changes in kidney diseases and is widely used in both clinical diagnosis and biomedical research. However, TEM image analysis remains highly labor-intensive and often suffers from inter-operator variability due to the lack of dedicated computational approaches. We developed TEAMKidney, a deep learning framework for accurate and scalable quantification of kidney ultrastructure in TEM images across species, magnifications, and imaging platforms.

Methods

We assembled 12,991 TEM images from patients’ kidney biopsies representing diverse kidney diseases and from multiple animal models. To address the scarcity of accurately annotated training data and the complexity of ultrastructural segmentation, we developed a two-stage framework. First, we used self-supervised learning combined with HRNet-based semantic segmentation to reduce the manual annotation burden while generating high-quality training labels. Second, we developed a TEM-tailored panoptic segmentation model to improve segmentation of complex ultrastructural features, particularly podocyte foot processes. Model performance was evaluated across different scenarios and compared with expert pathology assessments.

Results

TEAMKidney outperformed existing methods for ultrastructural segmentation and quantification. The framework demonstrated robust adaptation across multiple species and is the first deep learning approach for kidney TEM analysis with cross-species generalizability. It also maintained strong performance across different magnifications and imaging platforms. Quantitative measurements generated by TEAMKidney showed close agreement with expert pathology assessments used in clinical evaluation protocols across multiple kidney diseases.

Conclusion

TEAMKidney substantially reduces reliance on manual tracing while preserving expert-level accuracy, providing a scalable and reproducible framework for kidney TEM image analysis. This approach may accelerate standardized ultrastructural quantification of kidney biopsies and support both clinical nephropathology and biomedical research applications.

Funding

• NIDDK Support