Kidney Week

Abstract: SA-PO0049

Tubulopia: Topology-Aware Segmentation of the Tubular Basement Membrane Enables Quantitative Analysis in Periodic Acid-Schiff-Stained Kidney Histology

Session Information

• Artificial Intelligence: Kidney Imaging, Pathology, and Precision Phenotyping
  October 24, 2026 | Location: Exhibit Hall A, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Artificial Intelligence, Digital Health, and Data Science

• 300 Artificial Intelligence, Digital Health, and Data Science

Authors

• Xu, Jingyi, University of Florida Department of Electrical & Computer Engineering, Gainesville, Florida, United States

Jingyi Xu, MS

• La Rosa, Patricio Salvatore, Bayer CropScience LLC, St. Louis, Missouri, United States

Patricio Salvatore La Rosa, PhD

• Abdelazim, Haitham M., University of Florida Department of Medicine, Gainesville, Florida, United States

Haitham M. Abdelazim

• Paul, Anindya S., University of Florida Department of Medicine, Gainesville, Florida, United States

Anindya S. Paul, PhD

• Hodgin, Jeffrey B., University of Michigan Department of Pathology, Ann Arbor, Michigan, United States

Jeffrey B. Hodgin, MD, PhD

• Barisoni, Laura, Duke University Department of Pathology, Durham, North Carolina, United States

Laura Barisoni, MD, FASN

• Tomaszewski, John E., University at Buffalo Department of Pathology and Anatomical Sciences, Buffalo, New York, United States

John E. Tomaszewski, MD

• Sarder, Pinaki, University of Florida Department of Medicine, Gainesville, Florida, United States

Pinaki Sarder, PhD, FASN

Group or Team Name

• KPMP

Background

Tubular basement membrane (TBM) remodeling is associated with tubular epithelial injury, interstitial fibrosis, and chronic kidney disease progression. TBM remodeling spans a heterogeneous morphologic continuum, and these changes are structurally complex and difficult to quantify reproducibly by manual evaluation. We developed and evaluated Tubulopia, a topology-aware framework for automated TBM segmentation, refinement, and quantitative analysis in periodic acid-Schiff-stained whole slide images (PAS-WSIs) of kidney biopsies.

Methods

Tubulopia integrates an HRNet (High-Resolution Network) backbone with OCR (Object-Contextual Representation) and topology-aware loss functions to promote continuous TBM contour segmentation, which is essential for reliable downstream morphologic measurement. Initial predictions are further refined through post-processing and adaptive parameter selection at the local tissue level. The model was trained on KPMP PAS-WSIs using manual annotations from 20 tubule-rich ROIs, with 16 used for training and 4 reserved for held-out testing. Selected ROIs captured diverse TBM morphologies, including thin, thickened, multilayered, and irregular TBMs. Performance was evaluated using Dice and clDice scores.

Results

On the held-out test ROIs, Tubulopia achieved a mean Dice score of 0.71 and a mean clDice score of 0.76, demonstrating preserved TBM continuity across morphologically heterogeneous TBMs despite boundary-level variability. The framework also generated quantitative outputs, including TBM thickness maps and thickness distributions within and across tubules.

Conclusion

Tubulopia provides a topology-aware framework for automated segmentation and clinically interpretable quantitative analysis by preserving TBM contour continuity and generating visual overlays, thickness maps, and remodeling measures. As a web-based platform, it supports collaborative review, interactive refinement, and scalable evaluation of TBM remodeling in multi-center studies.

Funding

• NIDDK Support