Abstract: PO1762
Prediction of Histologic Class Using Deep Learning on Renal Biopsies from a Trial of Obinutuzumab for Proliferative Lupus Nephritis
Session Information
- Glomerular Diseases: Lupus and Membranous
October 22, 2020 | Location: On-Demand
Abstract Time: 10:00 AM - 12:00 PM
Category: Glomerular Diseases
- 1202 Glomerular Diseases: Immunology and Inflammation
Authors
- He, Bryan D., Stanford University, Palo Alto, California, United States
- Cascino, Matthew, Genentech Inc, South San Francisco, California, United States
- Schindler, Thomas, F Hoffmann-La Roche AG, Basel, Basel-Stadt, Switzerland
- Austin, Cary D., Genentech Inc, South San Francisco, California, United States
- Tew, Wei, Genentech Inc, South San Francisco, California, United States
- Garg, Jay P., Genentech Inc, South San Francisco, California, United States
- Zou, James, Stanford University, Palo Alto, California, United States
- Prunotto, Marco, F Hoffmann-La Roche AG, Basel, Basel-Stadt, Switzerland
Background
Glomerular lesions in lupus nephritis (LN) are classified according to the International Society of Nephrology and the Renal Pathology Society classification system and significant disagreement between pathologists can occur on histopathologic lesions. The aim of this study was to assess if deep learning on renal biopsy whole-slide images could be used to predict class III vs. IV status among patients enrolled in a randomized trial of obinutuzumab for the treatment of proliferative LN (NCT02550652).
Methods
Baseline biopsies from 84 of the 126 patients randomized were available for analysis. From each hematoxylin and eosin (H&E) slide, patches of 512x512 pixels were extracted resulting in an average of 500 patches per slide. An Inception v3 neural network (NN) with weights pretrained on the ImageNet dataset was used to make a prediction for each patch, which were then combined to make a prediction for the patient (Fig. A). From the initial weights, all slide’s layers were further fine-tuned using the cross-entropy loss between the model’s prediction and the patient’s true class. To evaluate the trained model, 25% of patients were held-out as a test set.
Results
The NN was able to classify the held-out patients with an area under the receiver operating characteristic of 0.82 (95% CI 0.60 - 1.00). Patches associated with class III vs. IV prediction could be extracted from each patient to provide interpretation (Fig. B).
Conclusion
These preliminary results showed that deep learning on renal biopsies can predict LN histologic class. The predictive patches provided additional interpretation. Such objective classification method has potential value to help minimize reading variability between pathologists.
(A) Computational pipeline for predicting the lupus nephritis histologic class of a patient from a whole-slide image. Examples of whole-slide image and patches were shown. (B) Examples of patches identified by the model as being most indicative of class III or IV.