Abstract: SA-PO910
Quantification of Globotriaosylceramide (GL3) in Peritubular Capillary (PTC) Endothelial Cells (ECs) in Kidney Biopsies from Patients with Fabry Disease (FD) Using Machine Learning
Session Information
- Pathology and Lab Medicine - 2
October 26, 2024 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Pathology and Lab Medicine
- 1800 Pathology and Lab Medicine
Authors
- Smerkous, David D., University of Washington, Seattle, Washington, United States
- Mauer, Michael, University of Minnesota Twin Cities, Minneapolis, Minnesota, United States
- Amedson, Alex, University of Washington, Seattle, Washington, United States
- Wang, Zian, University of Washington, Seattle, Washington, United States
- Kolakowski, Lexeigh Aspen, University of Washington, Seattle, Washington, United States
- Delombaerde, Vanessa, University of Washington, Seattle, Washington, United States
- Dastvan, Frank, University of Washington, Seattle, Washington, United States
Background
Intracellular GL3 accumulation is a hallmark of FD. Clearance of kidney PTC EC has been accepted by FDA as a valid clinical trial endpoint. Current light microscopy PTC EC GL3 inclusion quantitative methods are insensitive for partial clearance, time consuming, and subjective. We aimed to develop an automated electron microscopy (EM) approach for quantification of GL3 inclusions.
Methods
A pre-trained vision transformer (ViT) model was developed using masked auto-encoders (MAE) on ~190K EM images from human kidneys. ~200 EM images of PTC at 5000x from FD patients and controls biopsies were segmented for EC cytoplasm, GL3 inclusions and nuclei using a segmentation utility to serve as the ground truth. Segmentation models were trained on ground truth images using the ViT with minor modification and tested on validation segmentations.
Results
The train and test dices (shown respectively) for PTC EC segmentation were 0.82 ± 0.12 and 0.80 ± 0.07; for GL3 inclusions were 0.83 ± 0.30 and 0.84 ± 0.24; and for EC nuclei were 0.84 ± 0.26 and 0.87 ± 0.13. The intersection over union (IOU; 1=perfect; 0=none) of predicted and ground truth masks for GL3 inclusions, EC and nuclei ranged between 0.8 - 0.87 in test images. Fraction of EC occupied by GL3 [Vv(Inc/EC)] in the Fabry images studied was 0.097 ± 0.117 in manually segmented images and 0.107 ± 0.127 by model prediction. Vv(Inc/EC) in control (non-Fabry) biopsies was 0 by either manual segmentation or model prediction. The absoloue error of segmentation in test images was 0.03 ± 0.07.
Conclusion
Pre-training vision models using self-supervised methods is an effective way to develop performant and robust segmentation models, from large pools of unlabeled data. As exemplified by our GL3 quantification in PTC EC, these models can then be used with limited supervised training data. This approach provides a rapid, automated reproducible approach for GL3 quantification in PTC for clinical trial endpoints, research studies, or clinical purposes.
Raw, ground truth and model predictions for EC and GL3 inclusions
Funding
- Other NIH Support