Deep learning driven quantification of interstitial fibrosis in kidney biopsies

Yi Zheng, Clarissa A. Cassol, Saemi Jung, Divya Veerapaneni, Vipul C. Chitalia, Kevin Ren, Shubha S. Bellur, Peter Boor, Laura Barisoni, Sushrut S. Waikar, Margrit Betke, Vijaya B. KolachalamamedRxiv 2021 · DOI

Introduction

Interstitial fibrosis and tubular atrophy (IFTA) are critical indicators of disease chronicity and prognosis in renal pathology. Currently, the "gold standard" for assessing kidney damage involves a pathologist's manual review of trichrome-stained biopsies. However, this process is inherently subjective. Pathologists must mentally integrate complex histological features across an entire tissue sample to provide a semi-quantitative grade. This leads to significant inter-pathologist variability, where experts may disagree on the severity of scarring, potentially impacting clinical decision-making. To address this, there is a pressing need for objective, reproducible, and automated tools that can provide standardized IFTA assessments.

Methodology: The glpathnet Framework

This research introduces glpathnet, a deep learning (DL) framework designed to mimic the diagnostic workflow of a nephropathologist. When reviewing a slide, a pathologist typically "pans" across the tissue at low magnification to understand the global context and then "zooms in" to high magnification to inspect specific microscopic details.

The glpathnet architecture replicates this through a dual-branch approach:

  1. Local Branch: This branch processes high-resolution image patches, capturing fine pathological structures.
  2. Global Branch: This branch analyzes a downsampled version of the entire Whole Slide Image (WSI), providing the necessary holistic context.

To ensure these two perspectives work in tandem, the model utilizes a Feature Pyramid Network (FPN) and a specialized interaction layer. This allows the local and global branches to share information, ensuring that microscopic findings are interpreted within the context of the overall kidney core. The final IFTA grade—categorized as Minimal (≤10%), Mild (11–25%), Moderate (26–50%), or Severe (>50%)—is determined by an ensemble branch that aggregates features from both arms.

Experimental Setup and Datasets

To validate the model's robustness and generalizability, the study utilized two distinct datasets:

  • OSUWMC (Internal): 67 WSIs from The Ohio State University Wexner Medical Center, used for training and cross-validation.
  • KPMP (External): 28 WSIs from the Kidney Precision Medicine Project, serving as an independent test set to evaluate real-world performance.

Establishing a reliable "ground truth" is challenging due to human subjectivity. To mitigate this, five independent nephropathologists reviewed the slides. The reference estimate for each slide was determined by majority voting. The study found a moderate agreement among the experts (Kappa = 0.622 ± 0.071), further highlighting the necessity for an automated, objective benchmark.

Results and Performance

The glpathnet model demonstrated strong predictive capabilities, achieving an accuracy of 71.8 ± 5.3% on the internal OSUWMC dataset and 65.0 ± 4.2% on the external KPMP dataset. Notably, the DL framework significantly outperformed traditional machine learning methods like WND-CHARM, which relies on generic image descriptors and only achieved accuracies between 21% and 35%.

Beyond accuracy, the model is highly efficient. Once trained, it can process a massive whole-slide image and predict an IFTA grade in approximately 30 seconds, a fraction of the time required for manual review by multiple experts.

Explainability and Clinical Implications

A critical component of clinical adoption is "explainability." The researchers used Class Activation Maps (CAMs) to visualize the regions the AI prioritized when making a prediction. These heatmaps revealed that the model focuses on the same pathological regions—areas of collagen deposition and tubular loss—that human pathologists use to determine IFTA grades. This alignment between AI logic and human expertise is vital for building clinical trust.

While the results are promising, the study notes certain limitations. The current model is optimized for trichrome-stained slides; future iterations will need to accommodate other common stains like H&E and PAS. Additionally, expanding the dataset will help the model capture a broader spectrum of rare kidney pathologies.

Conclusion

The glpathnet framework represents a significant step toward the digital quantification of renal chronicity. By combining microscopic detail with global context, the model provides a regional and contextual estimation of IFTA that mimics professional pathological assessment. As a clinical decision-support tool, this technology has the potential to reduce variability in diagnosis and enhance the precision of prognostic indicators in nephrology.

Summary updated

Abstract

Abstract Interstitial fibrosis and tubular atrophy (IFTA) on a renal biopsy are strong indicators of disease chronicity and prognosis. Techniques that are typically used for IFTA grading remain manual, leading to variability among pathologists. Accurate IFTA estimation using computational techniques can reduce this variability and provide quantitative assessment by capturing the pathologic features. Using trichrome-stained whole slide images (WSIs) processed from human renal biopsies, we developed a deep learning (DL) framework that captured finer pathological structures at high resolution and overall context at the WSI-level to predict IFTA grade. WSIs (n=67) were obtained from The Ohio State University Wexner Medical Center (OSUWMC). Five nephropathologists independently reviewed them and provided fibrosis scores that were converted to IFTA grades: <=10% (None or minimal), 11-25% (Mild), 26-50% (Moderate), and >50% (Severe). The model was developed by associating the WSIs with the IFTA grade determined by majority voting (reference estimate). Model performance was evaluated on WSIs (n=28) obtained from the Kidney Precision Medicine Project (KPMP). There was good agreement on the IFTA grading between the pathologists and the reference estimate (Kappa=0.622±0.071). The accuracy of the DL model was 71.8±5.3% on OSUWMC and 65.0±4.2% on KPMP datasets, respectively. Identification of salient image regions by combining microscopic and WSI-level pathological features yielded visual representations that were consistent with the pathologist-based IFTA grading. Our approach to analyzing microscopic- and WSI-level changes in renal biopsies attempts to mimic the pathologist and provides a regional and contextual estimation of IFTA. Such methods can assist clinicopathologic diagnosis. Translational statement Pathologists rely on interstitial fibrosis and tubular atrophy (IFTA) to indicate chronicity in kidney biopsies and provide a prognostic indicator of renal survival. Although guidelines for evaluation of IFTA exist, there is variability in IFTA estimation among pathologists. In this work, digitized kidney biopsies were independently reviewed by five nephropathologists and majority voting on their ratings was used to determine the IFTA grade. Using this information, a deep learning model was developed that captured microscopic and holistic features on the digitized biopsies and accurately predicted the IFTA grade. The study illustrates that deep learning can be utilized effectively to perform IFTA grading, thus enhancing conventional clinicopathologic diagnosis.
BibTeX
@misc{zheng2021deep,
  author = {Yi Zheng and Clarissa A. Cassol and Saemi Jung and Divya Veerapaneni and Vipul C. Chitalia and Kevin Ren and Shubha S. Bellur and Peter Boor and Laura Barisoni and Sushrut S. Waikar and Margrit Betke and Vijaya B. Kolachalama},
  title = {{Deep learning driven quantification of interstitial fibrosis in kidney biopsies}},
  year = {2021},
  doi = {10.1101/2021.01.03.21249179},
  howpublished = {medRxiv},
}
Figure