The Life Science company Sartorius open-sourced ‘LIVECell’, a deep learning dataset for label-free quantitative segmentation of live cell images. This was announced via a research paper published in Nature Methods magazine.
The dataset includes 5000 label-free phase-contrast microscopy images made up of 1.6 million cells of eight-cell types, all marked with distinct morphologies that an expert in this field has manually annotated. The images show a large variation in cell size and shape as the cells grow from initial seeding densities to fully confluent monolayers.
Neural networks are great at identifying cells, but they need training with high-quality datasets to learn how best to segment them.
Accurate segmentation is vital to downstream analysis, but this task can be daunting. Traditional image-based methods often require tedious customization and rigorous tuning for different types of cells with varying morphologies. The researchers believe that using a diverse set of cells and confluence conditions in the ‘LIVECell’ dataset can train deep-learning-based segmentation models more accurately. Therefore, researchers now have a robust and accurate way to train neural networks. Rather than being limited to one type of cell morphology, the neural networks used in this process can handle multiple classes. This will allow for more robust segmentation and ultimately minimize user-introduced biases.
Before the launch of the LIVECell dataset, researchers had access to a dataset of label-free images available to researchers consisting of only 4,600 images derived from 26,000 cells.
Sartorius has partnered with the German Research Center for Artificial Intelligence (DFKI) to demonstrate how this dataset can be used in deep learning, and they plan on continuing their work together.