Deep Learning Enables Cell Analysis Without Labeling

Fluorescently labeled cell micrographs are undeniably stunning, but they require intrusive and often destructive or deadly protocols to get their glow. In order to prevent such disruptions, researchers have created a computer program that can differentiate between cell types and recognize subcellular structures, among other features — all without the fluorescent probes on which our human eyes depend.

"This approach has the potential to revolutionize biomedical science," says Margaret Sutherland, program director at the National Institute of Neurological Disorders and Stroke, which partially supported the work.

Researchers who have published their work in Cell today (April 12) have developed their neural network, a brain modeled software, using an in-depth learning approach uses data to identify patterns, shape rules, and apply those rules to new knowledge."We trained the neural network by showing two sets of matching images of the same cells; one unmarked and one with fluorescent labels," says Eric Christiansen, a software engineer at Google Accelerated Research, in a press release. "This process has been replicated millions of times. Then, when we confronted the network with an unmarked picture that we had never seen before, we could foresee precisely where the fluorescent labels belong.

With high-quality photos, the program correctly identifies the nucleus inside the cells. It could also differentiate dead from living cells, spot neurons in groups of cells that included astrocytes and immature dividing cells, and even say a dendrite from an axon.

"Techniques like this appear to have a democratizing impact," Molly Maleckar, director of mathematical modeling at the Allen Institute for Cell Science, who was not involved in the research, told Wired, opening up possibilities for smaller groups.

The authors of the study claim that future studies would aim to refine the network and improve its output on specific tasks where it was less robust, such as selecting neural subtypes and the detection of axons in high-density cultures.