New Image-Based Data Set Transforms AI Label-Free Cell Segmentation
Artificial intelligence (AI) has been a game changer for automated processing of image content on a massive scale. Cell biologists are increasingly using it to analyze millions of cell images in a range of disciplines, including cancer biology. Advanced microscopy technologies are proving to be indispensable tools for expanding the applications of automated cell segmentation to a wider variety of cell culture models.
This article is posted on our Science Snippets Blog
A picture is worth a thousand insights
There is never a dull moment inside a living cell. Whether it is secreted growth factors, stress, or an oral therapeutic, our cells are always reacting to changes in their environment. Imaging cells and their subcellular organelles is crucial for helping us understand basic human biology, how diseases take hold, and how to combat them with effective treatments.
The scope of image-based cell research has always been limited by our capacity to analyze microscopy images at scale. Prior to advanced computational tools, this work was quite resource-intensive and heavily reliant on the subjective interpretation of well-trained experts. Unsurprisingly, progress was slow.
Man vs. machine
Recent breakthroughs in high-throughput microscopy and AI-enabled image analysis have empowered scientists to study biological phenomena on a whole other level. The insight such analysis provides is immense and it carries robust statistical power. In pharmaceutical research, such capabilities are a driving force for discovery by enabling fast in vitro drug screening and efficacy testing.
With automated cell segmentation, or cell identification, scientists can quantify discrete cellular features (e.g., cell type, division, shape) and characterize disease-associated phenotypes. Importantly, AI can reliably capture subtle changes over time or under varying conditions, that could otherwise be missed.
It takes training
Cell segmentation algorithms are “trained” using large, well-annotated imaging datasets. Most of the available algorithms, however, are designed for fluorescence-based cell imaging.
Growing evidence suggests that labels can interfere with the normal biology of cells. That’s why many researchers are ditching fluorescent probes for live-cell and label-free cell imaging approaches, which are more physiologically relevant and conducive to cell health. But there is one problem; well-annotated imaging datasets for this application remain scarce, or lack the morphological diversity seen in culture.
Go ahead, be label-free
The authors of a new study published in the journal Nature Methods created LIVECell (label-free in vitro image examples of cells) to help address this gap. LIVECell is a large, high-quality, expertly annotated dataset of phase-contrast cell images that can be used for training cell segmentation algorithms for biologically relevant cell imaging experiments.
They take advantage of the Incucyte® Live-Cell Analysis System for collecting 5,239 images, consisting of 1.6 million cells from diverse morphologies and culture densities. The Incucyte® Live-Cell Analysis System is designed specifically for performing continuous, non-perturbing live-cell analysis directly from the incubator, making it the ideal platform for this study.
The authors then used their dataset to train convolutional neural network-based models for single-cell applications and create a set of benchmarks to aid segmentation accuracy.
Read the study to learn about LIVECell and how it compares to other publicly available datasets for cell segmentation.