Whereas traditional neuroscience relies on static snapshots of brain activity, the future of neuroimaging demands dynamic, high-fidelity monitoring of large-scale neuronal populations in freely behaving animals. Until recently, the main contender for correcting motion artifacts in live imaging has been rigid-body registration. Here, the assumption is that the entire brain moves as a single, solid block.
Researchers are also exploring the possibility of using optical flow algorithms to track pixel-by-pixel displacements: an approach known as non-rigid motion correction. This method holds great promise for capturing local deformations, particularly because biological samples undergo complex, multi-scale distortions due to breathing, heartbeat, and limb movements. The million-dollar question now is identifying the right computational framework to seamlessly integrate global and local corrections, a prerequisite for next-generation intelligent microscopy.
Now a research team led by scientists from Zhejiang University and Hangzhou City University has shown that DeepMoCo, an adaptive motion correction framework powered by Graph Neural Networks (GNNs), possesses properties that make it a highly practical choice for in vivo fluorescence imaging.
Graph Neural Networks: a straightforward solution for live imaging
In the search for robust motion correction, steps forward have already been made using conventional algorithms such as NoRMCorre. Yet since the advent of deep learning, GNN-based frameworks have become an intriguing contender. These architectures model image sequences as spatiotemporal graphs, which naturally endows them with the ability to aggregate local and global features, enabling robust prediction of neuronal displacements even in high-noise environments.
“This offers a significant advantage over traditional methods because you don’t need to rely on a single transformation model to handle all types of motion – the network intrinsically learns the spatiotemporal dynamics,” explains the research team. “This makes it easier to achieve pixel-level spatial fidelity and millisecond-level temporal consistency without needing ideal imaging conditions.”
The attractive but elusive prospect of multi-scale motion correction
There is one particular challenge in live imaging that has captivated researchers: multi-scale artifacts. Biological motion contains both global rigid shifts and local non-rigid deformations, which often cause traditional unified models to fail.
Why this dual-nature correction is so tantalizing is that it preserves the topological structure of neural circuits while eliminating jitter. “This is a very useful property as it means the corrected images retain their biological validity for downstream analysis,” says the team.
Yet despite the attraction of adaptive correction for neuroscience, until this latest study, achieving both global alignment and local deformation recovery simultaneously has remained a computational bottleneck.
Bridging the gap between complex motion and high-fidelity signals
To overcome this, hope has lain with deep learning architectures capable of distinguishing between different motion patterns. Yet a gap between theoretical models and noisy biological data has in the past hindered researchers from extracting accurate signals. “Researchers may have had the imaging data, but the true neural activity was buried in the motion artifacts,” explains the team.
In DeepMoCo, the system first classifies the motion mode, then applies targeted corrections. The team challenged every assumption with rigorous spatiotemporal tracking.
Then they made a crucial, adaptive step of employing a dual-modal correction strategy. “At first, balancing global and local corrections seemed challenging,” says the team.
Meticulously unpicking how different contributions complicated the calculations, they revealed that by fusing multi-scale features, the framework could maintain rigid registration accuracy while recovering subtle local deformations. In this way, they bridged the gap between theoretical deep learning and practical biological imaging, proving the efficacy of adaptive classification strategies.
Doors open to exploring neural circuits across species and modalities
Armed with the ability to accurately restore high-fidelity signals, the team went on to show that DeepMoCo’s performance is highly generalizable. Whether applied to the mushroom body and ellipsoid body in Drosophila, or to axons, dendrites, and the spinal cord in mice and zebrafish, the system consistently eliminated complex motion-induced distortions.
“This is a very useful property, as neuroscience research could potentially be conducted across different species and imaging modalities with a single, unified tool,” says the team.
Now, with theory and biological validation finally united, the wider research community is equipped with the means to explore neural circuit dynamics across a variety of models and optimise their applications for brain-computer interfaces and biodynamic research. Future iterations may further reduce sensitivity to region-of-interest masks and neuronal density, paving the way for even more complex physiological imaging.
PhotoniX
Computational simulation/modeling
Cells
DeepMoCo: graph neural network-based adaptive correction for motion artifacts in fluorescence microscopy
18-Aug-2026
The authors declare no competing interests.