Category: Assay Development and Screening
Flow cytometry has long been used routinely in clinical diagnoses, especially in hematological disorders. The introduction of imaging flow cytometry (IFC) brings even greater diagnostic potential thanks to its high content and high-throughput capability, where hundreds of thousands of images of individual cells are captured in minutes.
We extended previous open-source protocols for high-content IFC data analysis [1,2] to detect and classify abnormal cells in collected blood samples of leukemic and allergic patients in 2 independent clinical trials. We utilized both classical machine learning and deep learning to examine high dimensional feature spaces extracted from IFC images, and found out that bright field (forward scatter) and dark field (side scatter) signals in fact contained valuable morphological information that is often overlooked by conventional image analysis methods.
We were able to use this label-free information to distinguish leukemic blasts from normal lymphocytes and granulocytes in leukemia patients; while in an independent project, the same IFC workflow was applied to classify 6 morphological phenotypes of red blood cells during 42-day storage. We developed a solution to implement ResNet, a well-known state-of-the-art deep learning architecture, which was originally built for analyzing daily photographic RGB images (thus limited to 3 channels), to be used in analysis of biological images at the multiplicity of unlimited number of fluorescent channels. This open-source solution can be robustly used by non-expert biologists to perform feature extraction, feature selection, and classification using supervised (random forest, support vector machine) and unsupervised learning (t-SNE, PCA, diffusion map) as well as deep learning on IFC imagery data. The complete workflow for IFC deep learning is published at http://github.com/broadinstitute/deepometry . The results encourage the adoption of IFC as a potential diagnostic tool in clinical practice.
 Blasi et al. Label-free cell cycle analysis for high-throughput imaging flow cytometry. Nat. Comm. 7, 10256 (2016)
 Hennig et al. An open-source solution for advanced imaging flow cytometry data analysis using machine learning. Methods 112, 201 (2017)
Minh Doan– Assay developer, Broad Institute of MIT and Harvard, Cambridge, MA
Broad Institute of MIT and Harvard
I've trained as a medical doctor (2004-2010) and did PhD in molecular cell biology (2010-2015) in University of Debrecen, Hungary. During this time, we developed a read-out procedure to monitor morphological and biochemical features of adipocytes during adipogenesis.
I then did my first postdoc in microbiology in Pasteur Institute, Paris, France for nearly 2 years, focusing on genomic interaction between the host cell and Chlamydial microbes. From 2016 until now, I join Dr. Anne Carpenter's Imaging Platform at the Broad Institute of MIT and Harvard, leading the effort to analyze Imaging flow cytometric data in clinical studies. Here we focus on developing machine learning methodology, especially deep learning, to automate the analytic processing of single cell images, from not only Imaging flow cytometry but may potentially applicable for general bioimaging. We recently exemplified this machine learning application in the studies of leukemia and eosinophil activation.