AI microscopy catches fleeting movements inside living cells

Mitochondria, the energy-converting compartments inside cells, can send out small membrane extensions that retreat within seconds. Using AI-assisted imaging at two frames a second, researchers detected protrusions 80–200 nanometres long lasting 1.5–4 seconds that slower recordings missed. They analysed five cultured human U2OS cells at each of three imaging speeds.[1]

The work by Guillaume Minet, Anirban Ray and colleagues appeared in Nature Communications on 11 October.[1]

A microscope has to manage more than magnification. In fluorescent imaging, specially labelled structures emit the light used to make a picture. The MoNaLISA microscope, introduced in 2018, uses patterned illumination to switch fluorescent proteins on and off, restricting emission to very small regions. It scans these patterns to build a detailed image. The proteins supply a signal that can locate structures far smaller than an ordinary light microscope resolves.[2]

There is a physical price for collecting that information. Repeated illumination and switching can exhaust the fluorescent signal, a problem called photobleaching. Collecting an image also takes time, during which its subject keeps moving. The resulting choices concern how finely, how quickly and for how long a living system can be observed.[2]

Making a weaker signal useful

For longer recordings, the study applied Hierarchical DivNoising, or HDN. In experiments on vimentin, part of the cell’s internal scaffolding, roughly tenfold lower switching-light exposure supported up to fivefold longer imaging; the bleaching comparison covered five cells.[1]

HDN was developed as an image-restoration method that can represent several plausible clean images behind one noisy observation. Its neural network learns patterns at different scales, from broad structures to finer detail. The underlying problem is familiar from a grainy photograph: a faint feature may be part of the subject or part of the noise. Simply smoothing the picture can erase both. A model of the image and its noise offers a more selective reconstruction.[3]

Catching the movement between frames

The separate Fast-MoNaLISA approach sampled a quarter of the usual scan positions. A network reconstructed the images using five consecutive frames as context, reaching 2 Hz across about 35 by 35 micrometres, against the regular 0.5 Hz.[1]

That change addresses a different loss of information. A scanning instrument constructs a picture from measurements taken in sequence. If a thin extension grows while the scan proceeds and retreats before the next image, the record may miss its full excursion or distort its shape. Faster acquisition gives the scientist more opportunities to follow the action. Knowing where a structure is and seeing what it does are separate achievements.[2]

These gains came from separate acquisition strategies. Sparse reconstruction used bright signals; excessively dim inputs could lose faint details.[1]

For an imaging scientist, a useful reconstruction preserves the small features that make an event interpretable. A beautiful still can establish that a structure exists. A sufficiently detailed sequence can show its order of movement, its brief contacts and its return. The scientific reward is the action that becomes available to explain.[3]

Report an error or suggest a correction