Kidney Week

Abstract: SA-PO0013

Deep Learning to Estimate Interstitial Fibrosis from Trichrome Whole-Slide Images

Session Information

• Intelligent Imaging and Omics: Phenotyping and Risk Stratification
  November 08, 2025 | Location: Exhibit Hall, 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

• Athavale, Ambarish, University of California San Diego, La Jolla, California, United States

Ambarish Athavale, MD, MBBS

• Patel, Tushar, University of Illinois Chicago, Chicago, Illinois, United States

Tushar Patel, MD

• Kulkarni, Hemant, M&H Research LLC, San Antonio, Texas, United States

Hemant Kulkarni, MD, MBBS

Background

Visual assessment of Interstitial fibrosis (IF) leads to significant interobserver variability. Combining deep learning and morphometry, we developed a method (TRI_IF) to estimate IF from whole slide images (WSI) of kidney biopsies.

Methods

Cortical area in trichrome WSI from NEPTUNE study was manually annotated and masked. Since fibrotic area stains blue on trichrome, the TRI_IF algorithm 1)estimates pixel-wide blue-to-red intensity ratios (BRIR) 2) estimates slide-specific threshold for binarization of the image by sliding over a range of 0.6-1.5 to approximate the consensus nephropathologists ground truth IF. The deep learning (DL) model used for choosing the binarization threshold was validated on a separate test set. IF was estimated as a proportion of the white pixels from a binarized image using the DL-model-derived BRIR threshold. The IF estimates were then compared with nephropathologists using agreement analyses (R² for continuous IF scores and weighted Cohen’s kappa for categorized IF scores). Association of the categorized IF estimates was done with two clinical endpoints: a) composite of end-stage renal disease (ESRD) or 40% reduction of estimated glomerular filtration rate (eGFR) b) eGFR slope per year.

Results

We included a total of 315 kidney biopsies and cortical annotation led to 739 biopsied cores. The trained Xception regression model IF estimates showed excellent agreement (Table 1) with nephropathologists estimate (Pearsons correlation of 0.93 for continuous IF and Cohens Kappa of 0.91 for categorized IF). IF scores were strongly associated with time to ESRD/<40%eGFR and rates of eGFR decline (Figure 1).

Conclusion

The TRI_IF algorithm combining deep learning and morphometry for automated estimation of IF on Trichrome WSI showed good agreement with pathologist’s estimate of IF and association with clinical outcomes.

Agreement analysis

Funding

• NIDDK Support

Digital Object Identifier (DOI)

• doi: 10.1681/ASN.2025y3rrnqct