Is that blip a signal or an artifact of misguided data analysis? That’s the question fueling a controversy surrounding “Direct Imaging of Neuronal Activity” (DIANA) fMRI,1 a potentially groundbreaking technique with temporal resolution fast enough to directly capture neuronal activity. A new preprint recently failed to replicate the original DIANA study and the authors reported that the only way that they could get a DIANA-like signal was by cherry-picking data.2
See also “New MRI Technique Tracks Brain Activity at Millisecond Timescales”
In the original study, researchers led by Jang-Yeon Park of Sungkyunkwan University submitted anesthetized mice to whisker shocks while taking small segments of fMRI images milliseconds apart. This burst of quick, partial images seemed to allow them to capture neuronal activity itself , rather than the blood flow changes usually measured by fMRI, although exactly how it translated into an MRI signal remained unclear.
The publication ignited the neuroimaging community. “Looking for an MRI signature of neuroactivity is the holy grail for fMRI field,” said Shella Keilholz, an MRI physicist from Georgia Tech and Emory University, who was neither involved with the original work nor the new preprint. “We were all very excited. They had done a lot of control experiments and really gone out of their way to convince people that this was real.”
Even so, some remained skeptical. “What they were claiming didn’t seem feasible,” said neurophysicist Ravi Menon from Western University, who decided to replicate the DIANA study with a team that included the director of Park’s own institute, the world-leading fMRI expert Seong-Gi Kim. They recently published their results in a preprint.2
The replication experiments were conducted in the same facility as Park’s study but had improved signal-to-noise ratios. “We should be able to see very small effects, and we don’t see them, period,” said Menon. Usually, researchers average the data collected in all of the whisker-shock trials, but Menon said that the only way his team could produce a signal was by removing certain trials from the overall average.
There was no mention of such data exclusion in the original DIANA paper, so in June, Menon asked Park about it at a talk he gave at the International Society for Magnetic Resonance in Medicine’s Annual Meeting in Toronto. “The answers weren’t satisfactory,” said Keilholz, who was also present. Although there was a language barrier, she said, “it sounded to me like he had thrown out data that didn’t match the expected template.”
Park sees it differently. According to him, there is much left to understand about DIANA signals and the data handling that they require. He said that neurons sometimes respond differently to the same stimulus,3 which means that the signal could vary between repeat trials . “Before [DIANA], we didn’t need to consider the behavior of neuroactivity per se in fMRI,” he said. “It’s time to think about the new implications of that.”
He said that averaging more trials didn’t necessarily lead to better DIANA results, and that it was better to average over earlier trials when the neuronal response is “fresh” and to take breaks between trials. He also encouraged averaging across animals to minimize baseline fluctuations that he believes come from spontaneous neuroactivity. Following this overall approach, Park said that he has replicated DIANA in both mice and humans, and he expects to publish that work soon.
That interests Timo van Kerkoerle at Radboud University Nijmegen and Martijn Cloos at the University of Queensland, coauthors of a preprint attempting to reproduce DIANA in humans.4 They couldn’t find the signal either, but they always expected it to be difficult, partly because it’s unclear what exactly produces the DIANA signal. In the original paper, it correlated well with electrophysiology, but no one could explain how that neuroactivity transformed into an MRI signal. “With not knowing that key part of the puzzle, we were already cautious,” said Cloos.
They also regret that Park didn’t specify how he selected and averaged trials in the original paper; whatever he did, they would like to learn from it. “These things are hard to do,” said van Kerkoerle. “The next step would be for [the researchers] to really talk to each other and say, 'okay, what did we do differently? How can we solve this?'”
While Cloos considered Menon’s preprint a strong statement in the field, he is not yet put off pursuing DIANA, although he is downscaling from humans. One of his next projects will look for DIANA in brain organoids. “[This way] we can have a very pure translation of the neuronal signals to MRI, hopefully, and try to understand how these signals may be formed,” he said.
Despite Cloos’ optimism, Keilholz senses a wind change. “People are maybe still playing around with it a little bit, but definitely the level of enthusiasm has dropped by an order of magnitude,” she said. “Hopefully in another year, we’ll have some sort of consensus.”
Note: The preprint findings reported in this article were published in Science Advances on March 27, 2024.
- Toi PT, et al. In vivo direct imaging of neuronal activity at high temporospatial resolution. Science. 2022;378:160-168.
- Choi SH, et al. No replication of direct neuronal activity-related (DIANA) fMRI in anesthetized mice. bioRxiv. 2023.
- Renart A, Machens CK. Variability in neural activity and behavior. Curr Opin Neurobiol. 2014;25:211-220.
- Hodono S, et al. Initial experiences with direct imaging of neuronal activity (DIANA) in humans. arXiv. 2023
