March 13th 2020
Topic: Deep-Learning based microscopy image analysis pipelines
Presenter: Dr. Conor Evans, Associate Professor, Wellman Center for Photomedicine, Massachusetts General Hospital, Harvard Medical School, Affiliate Faculty, Harvard University Biophysics Program
Abstract:
Quantitative microscopy experiments can generate mountains of data that must be organized, processed, and analyzed to extract quantitative and actionable information. Standard semi-automated approaches are often not up to the task as they require substantial hands-on involvement and do not scale to terabytes or more of data. Automated solutions are needed, but time-lapse multichannel microscopy data can be particular challenging as the acquired images often contain numerous heterogeneous structural and spectra features that can foil automated analysis routines. To overcome these challenges, we have developed a python-based microscopy data analysis pipeline that enables the automated analysis of medium-to-large size imaging datasets. This pipeline makes use of both deep learning-based methods as well as standard computer vision libraries that are packaged to provide completely hands-off image analysis. Quality control processes and logging are included in the pipeline to ensure that images have been processed correctly and accurately. We use Docker and python virtual environments to standardize the analysis across computers and operating systems, and the pipelines can be pushed to individual or multiple containers (e.g. Kubernetes) to separate, serialize, or accelerate processing. We will discuss how this generic approach can be applied to two entirely different sets of imaging data.