Kidney Week

Abstract: TH-OR021

Tubular Morphometric Signatures (TMS)-Based FSGS/Minimal Change Disease (MCD) Patient Retrieval

Session Information

• Artificial Intelligence in Kidney Care: From Pathology to Prediction
  October 22, 2026 | Location: Room 711, Convention Center
  Abstract Time: 04:50 PM - 05:00 PM

Category: Artificial Intelligence, Digital Health, and Data Science

• 300 Artificial Intelligence, Digital Health, and Data Science

Authors

• Fan, Fan, Department of Biomedical Engineering, Emory University and Georgia Institute of Technology, Atlanta, Georgia, United States

Fan Fan, MS

• Nair, Viji, University of Michigan Department of Internal Medicine, Ann Arbor, Michigan, United States

Viji Nair, PhD

• Smith, Cathy, University of Michigan Department of Internal Medicine, Ann Arbor, Michigan, United States

Cathy Smith, PhD

• Eichinger, Felix H., University of Michigan Department of Internal Medicine, Ann Arbor, Michigan, United States

Felix H. Eichinger

• McCown, Phillip J., University of Michigan Department of Internal Medicine, Ann Arbor, Michigan, United States

Phillip J. McCown, MS, PhD

• Demeke, Dawit S., University of Michigan Department of Pathology, Ann Arbor, Michigan, United States

Dawit S. Demeke, MD

• Wang, Bangchen, Department of Pathology, Division of AI & Computational Pathology, Duke University, Durham, North Carolina, United States

Bangchen Wang, MD, PhD

• Ozeki, Takaya, Nagoya University Graduate School of Medicine, Department of Nephrology, Nagoya, Japan

Takaya Ozeki, MD, PhD

• Jacobs, Jackson, Department of Biomedical Engineering, Emory University and Georgia Institute of Technology, Atlanta, Georgia, United States

Jackson Jacobs

• Liu, Qian, Children’s Hospital of Philadelphia Research Institute, Philadelphia, Pennsylvania, United States

Qian Liu, MPH

• Bitzer, Markus, University of Michigan Department of Internal Medicine, Ann Arbor, Michigan, United States

Markus Bitzer, MD

• Mariani, Laura H., University of Michigan Department of Internal Medicine, Ann Arbor, Michigan, United States

Laura H. Mariani, MD, MS, FASN

• Lafata, Kyle Jon, Duke University Department of Radiation Oncology, Durham, North Carolina, United States

Kyle Jon Lafata, PhD

• Holzman, Lawrence B., Department of Medicine, Division of Nephrology and Hypertension, University of Pennsylvania, Philadelphia, Pennsylvania, United States

Lawrence B. Holzman, MD

• Kretzler, Matthias, University of Michigan Department of Internal Medicine, Ann Arbor, Michigan, United States

Matthias Kretzler, MD

• Madabhushi, Anant, Department of Biomedical Engineering, Emory University and Georgia Institute of Technology, Atlanta, Georgia, United States

Anant Madabhushi, PhD

• Zee, Jarcy, Children’s Hospital of Philadelphia Research Institute, Philadelphia, Pennsylvania, United States

Jarcy Zee, PhD, FASN

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

Jeffrey B. Hodgin, MD, PhD

• Barisoni, Laura, Department of Pathology, Division of AI & Computational Pathology, Duke University, Durham, North Carolina, United States

Laura Barisoni, MD, FASN

• Janowczyk, Andrew, Department of Biomedical Engineering, Emory University and Georgia Institute of Technology, Atlanta, United States

Andrew Janowczyk, PhD

Background

The clinical, morphologic, and treatment heterogeneity among patients with FSGS/MCD poses a major barrier to understanding disease trajectories. Here, we evaluated whether tubular pathomic–derived features could support precision medicine by enabling nephrologists to identify patients with similar morphologic characteristics and explore their disease trajectories.

Methods

We applied image processing and deep learning algorithms to periodic acid–Schiff–stained whole-slide images from 246 NEPTUNE/CureGN FSGS/MCD patients with visual descriptor scoring and longitudinal clinical data. Tubular substructures—epithelium, nuclei, basement membrane, and lumen—were segmented, from which 99 pathomic features were extracted. Using these features, we computationally derived 14 TMS via non-negative matrix factorization. Tubule-level TMS was summarized at the patient level using mean.
Nearest-neighbor retrieval was performed using k-nearest neighbors (KNN) with Euclidean distance to identify patients with similar TMS profiles. For the MCD cohort (Fig. 1), each adult participant (query case) was compared to five adult participants with the most similar TMS-derived biopsy profiles. Comparisons were made across (a) biopsy visual descriptor scores [% global/segmental sclerosis, interstitial fibrosis and tubular atrophy (IFTA), inflammation, and arteriosclerosis], (b) clinical characteristics, and (c) treatment exposure.

Results

While expected similarities were observed between the query participant and nearest neighbors in tubular visual scoring, substantial variability remained in other tubulointerstitial, vascular, and glomerular features across cases. Divergent clinical trajectories were observed, often accompanied by differences in treatment exposure.

Conclusion

The pathomic signature–based patient retrieval framework enables structured comparison of therapeutic strategies and associated responses within phenotypically matched groups, supporting case-based interpretation, hypothesis generation, and exploratory evaluation of treatment decisions.

Acknowledgment

ChatGPT (OpenAI) was used for language editing and writing assistance. Authors reviewed and take full responsibility for the final content.

Funding

• NIDDK Support