Abstract: SA-PO723
Prediction of Baseline Renal Function in Lupus Nephritis Using Deep Learning on Histology Images
Session Information
- Pathology and Lab Medicine: Clinical
November 09, 2019 | Location: Exhibit Hall, Walter E. Washington Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Pathology and Lab Medicine
- 1602 Pathology and Lab Medicine: Clinical
Authors
- Moindrot, Olivier, Owkin, Paris, France
- Moll, Solange, University Hospital of Geneva, Geneva, Switzerland
- Maussion, Charles, Owkin, Paris, France
- Tew, Wei, Genentech Inc, South San Francisco, California, United States
- Courtiol, Pierre, Owkin, Paris, France
- Wainrib, Gilles, Owkin, Paris, France
- Prunotto, Marco, F. Hoffmann-La Roche Ltd., Basel, Switzerland
Background
Lupus nephritis (LN) is a heterogeneous disease that might be an ideal case to use artificial intelligence such as deep learning (DL) to provide a more precise and objective assessment of biopsy. The aim of the study is to assess DL performance on kidney biopsy whole-slide images in predicting baseline renal function in LN stage II-IV patients.
Methods
Kidney biopsies from 98 LN patients were used to train an agnostic DL algorithm called Chowder using multitask learning. Additionally, a Mask Region-based Convolutional Neural Network instance segmentation model was used to segment renal structures, trained using expert annotations and validated before applying it to patient biopsies. Histomic features from the algorithm predictions were extracted to predict renal function. Cross-validation AUC was calculated.
Results
Chowder was able to predict baseline eGFR and creatinine (Table). The structure segmentation model has an average precision of 0.68 and 0.76 for proximal tubules and glomeruli, respectively (Figure). Histomic feature of inflammation area outperforms Chowder in proteinuria and tubulointerstitial fibrosis prediction (Table).
Conclusion
Preliminary results show that DL on renal biopsy can predict baseline renal function. Histomic features from structure segmentation prediction provided additional insights into histological manifestations. This methodology can be extended to disease outcome prediction.
Renal function prediction performances. AUC values shown.
| Renal Function | ISN/RPS Grade | Chowder | Histomics (inflammation area) |
| eGFR (<60 ml/min/1.73m2) | 0.70 | 0.91 | 0.77 |
| Serum Creatinine (>83 µmol/l) | 0.64 | 0.74 | 0.74 |
| Proteinuria (>0.35 g/l) | 0.82 | 0.62 | 0.75 |
| Tubulointerstitial fibrosis (>10%) | 0.69 | 0.55 | 0.75 |
An example of structure segmentation prediction on a test image in validation set. Classes and confidence values shown.
Funding
- Commercial Support –