Leveraging Cross-Domain Synergies for Efficient Machine Learning of Nanoscale Tomogram Segmentation

Before you can get an artificial intelligence to detect certain things automatically, you have to teach it first. For example, if you want a computer to learn how to recognise specific structures inside cells, you first need to feed it with a few dozen images of cells taken under a microscope, in which you have marked, or annotated, those structures by hand. “But, creating these annotations for the computer is often incredibly laborious,” says Dr. Dagmar Kainmüller of the Max Delbrück Center for Molecular Medicine. Therefore, she and her colleagues are working on a method for teaching a machine using little or no annotated data at all.

Double image showing nanoparticles under electronic microscope in black and white and coloured

The team is currently working from images of cells taken at nanometre resolution under a cryo-electron microscope. Researchers use these images to identify specific membrane-like cell structures – and can give the job of finding them to an artificial intelligence. “There are several groups at a number of Helmholtz institutes working with images from cryo-electron microscopes,” Dagmar Kainmüller says.

Double image showing nanoparticles under electronic microscope, partly coloured

Yet, so far, they have to spend weeks or even months preparing these images before they can be used to train the artificial intelligence for each new structure. The aim is therefore to train an algorithm so that it can be used for this and other related search tasks as well – in other words, for capturing similar structures from different imaging techniques or for capturing different structures from similar imaging techniques.