Abstract: FR-PO020
Self-Supervised Learning Applied to Kidney Histomorphology in Whole Slide Images
Session Information
- AI, Digital Health, Data Science - II
November 03, 2023 | Location: Exhibit Hall, Pennsylvania Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Augmented Intelligence, Digital Health, and Data Science
- 300 Augmented Intelligence, Digital Health, and Data Science
Authors
- Pandit, Krutika, New York University, New York, New York, United States
- Coudray, Nicolas, New York University, New York, New York, United States
- Claudio Quiros, Adalberto, University of Glasgow, Glasgow, Glasgow, United Kingdom
- Surapaneni, Aditya L., New York University, New York, New York, United States
- Rosenberg, Avi Z., Johns Hopkins University, Baltimore, Maryland, United States
- Susztak, Katalin, University of Pennsylvania, Philadelphia, Pennsylvania, United States
- Grams, Morgan, New York University, New York, New York, United States
- Tsirigos, Aristotelis, New York University, New York, New York, United States
Background
Current approaches to characterize medical imaging data typically use some degree of supervision, relying heavily on expert annotation. Pathologist reads can have high inter-observer variability. A supervised model built on these reads will mimic the idiosyncrasies of the expert and may perpetuate inter-observer inconsistencies. Supervised approaches also require time-consuming and meticulous labeling on the input images. Given the limitations, supervised learning may not be optimal for unbiased and generalizable characterization of histologic patterns in the kidneys, which are complex.
Methods
Histomorphological Phenotype Learning leverages the latest advances in self-supervised learning (SSL), the Barlow-Twins method, to learn representations of image tiles that are later clustered by the Leiden method, a state-of-the-art community detection algorithm. This strategy was used to discover de novo clusters of morphologically similar tissue from 254 whole slide images (WSIs) of the kidney without any prior diagnostic information from pathologists. These clusters were then projected onto validation (N = 254), test (N = 255), and external validation (N = 113) sets, visually inspected across sets, and correlated with expert-provided histology scores. Additionally, elastic net models were trained to predict histology scores using proportion of patients’ tiles in each cluster.
Results
Visual inspection of representative tiles from clusters revealed comparable histologic patterns across sets. Clusters that were visually indicative of disease were positively correlated with fibrosis; clusters representing healthy tissue were negatively correlated with fibrosis (Figure 1). The best-performing elastic net model used 17 clusters to predict percent fibrosis (R2 in test set = 0.81).
Conclusion
In this study, we demonstrated the application of SSL to characterize WSIs of the kidney. Features extracted from this approach were associated with and demonstrated good prediction of expert-provided histology scores. Subsequent work will link clusters to gene expression and metabolic processes.
Funding
- NIDDK Support