Abstract: FR-PO029
Integration of CODEX and Brightfield Histology for Cell Type Segmentation and Classification Using Deep Learning
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
- Lucarelli, Nicholas, University of Florida Department of Biomedical Engineering, Gainesville, Florida, United States
- Winfree, Seth, University of Nebraska Medical Center Department of Medicine, Omaha, Nebraska, United States
- Sabo, Angela R., Indiana University Department of Medicine - Division of Nephrology and Hypertension, Indianapolis, Indiana, United States
- Border, Samuel, University of Florida Department of Biomedical Engineering, Gainesville, Florida, United States
- Barwinska, Daria, Indiana University Department of Medicine - Division of Nephrology and Hypertension, Indianapolis, Indiana, United States
- Laszik, Zoltan G., University of California San Francisco Department of Pathology and Laboratory Medicine, San Francisco, California, United States
- Eadon, Michael T., Indiana University Department of Medicine - Division of Nephrology and Hypertension, Indianapolis, Indiana, United States
- El-Achkar, Tarek M., Indiana University Department of Medicine - Division of Nephrology and Hypertension, Indianapolis, Indiana, United States
- Jain, Sanjay, Washington University in St Louis Department of Medicine Division of Nephrology, St. Louis, Missouri, United States
- Sarder, Pinaki, University of Florida Department of Medicine - Quantitative Health, Gainesville, Florida, United States
Background
Cell types in a biopsy provide information on disease processes or organ health. Multiplex imaging technologies like CODEX provide spatial context to protein expression and detect cell types in a tissue sample. New CODEX workflows allow for hematoxylin and eosin (H&E) staining on the same sections. Deep learning can automate the process of image analysis, saving time. We seek to segment and classify cell nuclei from renal tissue sections using deep learning with CODEX generated cell labels as a ground truth.
Methods
Images consisted of brightfield H&E whole slide images (WSIs) from two institutions, collected from human reference kidneys. Nuclei were segmented using deep learning, and CODEX markers were measured for each nucleus. Cells and their markers were clustered in an unsupervised manner and assigned labels according to upregulated markers and biological priors. Cell types included: proximal tubules, distal tubules connecting tubules and collecting ducts, thick ascending limb, podocytes, endothelium, vessels, and immune cells. Cell maps were used to train a Deeplab V3+ semantic segmentation model. Classification was assessed in hold-out slides from CODEX generated sections.
Results
Two segmentation models were trained on WSIs from each institution. For the model trained on 3 sections containing ~3.9M cells, we achieved a balanced accuracy of 0.68, and for the model trained on ~350k cells from 11 sections, we achieved 0.75.
Conclusion
We were able to automatically segment and classify nuclei from various cell types directly from H&E stained WSIs. In future work, we seek to extend these segmentations to typical WSIs in renal pathology, with no prior molecular interrogation.
Funding
- NIDDK Support