Abstract: FR-PO995
Unsupervised Machine-Learning Cytometry of High-Dimensional Image Data from Fluorescently Labeled Mesoscale Kidney Tissue Automates Quantitation and Uncovers Unique Cellular Populations
Session Information
- Pathology and Lab Medicine: Basic
November 08, 2019 | Location: Exhibit Hall, Walter E. Washington Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Pathology and Lab Medicine
- 1601 Pathology and Lab Medicine: Basic
Authors
- Winfree, Seth, Indiana University School of Medicine, Indianapolis, Indiana, United States
- Mcnutt, Andrew, Indiana University, Indianapolis, Indiana, United States
- Ferkowicz, Michael J., Indiana University, Indianapolis, Indiana, United States
- El-Achkar, Tarek M., Indiana University, Indianapolis, Indiana, United States
Background
The cytometric analysis of fluorescent mesoscale kidney imaging datasets presents unique challenges in segmentation, measurement and analysis. To address these challenges, we developed the tissue cytometry tool, Volumetric Tissue Exploration and Analysis (VTEA). VTEA streamlines nuclei segmentation, measurements and enables flow cytometry-like analysis in mesoscale 3D images of kidney tissue. However, as we add 1) additional fluorescent markers with novel imaging modalities and 2) imaging metrics, including texture and spatial characteristics, flow cytometry-like approaches become inadequate under the strain of these higher dimensional data. Here our goal was to implement and demonstrate the need and utility of unsupervised analysis of higher dimensional tissue cytometry data.
Methods
Imaging datasets were collected from fluorescently labeled mouse and human kidney tissue with confocal fluorescence microscopy and up to 8 independent fluorophores. Image processing and analysis was performed with the recent version of VTEA that incorporates Java based libraries for clustering of datasets (e.g. K-mean, Gaussian-mixtures) and dimensionality reduction tools (e.g. PCA and tSNEs).
Results
We extended our cytometry tool, VTEA, incorporating unsupervised machine learning approaches with publicly available Java libraries. Using clustering, we identify cellular population of cells not readily identified with manual gating. The mapping of high dimensional cytometry data to lower dimensions facilitates rapid visual identification of cell sub-population not readily apparent. Lastly, we demonstrate the advantage of looking at our imaging data with both dimensionality reduction and clustering-an approach that has been exploited by the omics fields.
Conclusion
Our work demonstrates the need and utility for high dimensional analysis of mesoscale kidney imaging data and suggests we gain insight with the application of unsupervised machine learning approaches to image analytics, especially in the framework of the interactive exploratory platform VTEA. Lastly, our work underlines the importance of reusable software and the power of an open software community found in the NIH supported ImageJ community.
Funding
- NIDDK Support