Release date: 2016-08-02
New research suggests that our functional magnetic resonance imaging (fMRI) algorithm may not be right—the researchers found an algorithmic error in fMRI-specific software that could lead to a deadly study of human brain activity over the past 15 years.
Currently, the best way to study brain activity is functional magnetic resonance imaging (fMRI). fMRI is so easy to use that we rely heavily on it – of course, as long as fMRI is accurate and reliable, this is not a bad thing. fMRI is used to observe changes in the minds of people in sports, games, love, and medication. Once they have problems, research results over the years will be questioned. In order to test the function of fMRI, researchers from Linköping University in Sweden conducted an unprecedented test of scale, but the results were not optimistic.
The researchers wrote: "Although fMRI is very common as a means of studying brain function, the statistical methods used are almost verifiable from real data."
The research team led by Anders Lund collected the results of the fMRI examination of 499 healthy people from all over the world at rest and divided them into 20 groups. Then they compared the groups to each other and got the results. Amazing 3 million random controls. They used these pairings to test the three most famous software for analyzing fMRI - SPM, FSL and AFNI.
The research team is looking forward to seeing the differences between the software (about 5%), but the results of the experiment surprised them: the software's false positive rate reached 70%. This means that software can sometimes be ridiculously wrong, even if the brain is at rest, it shows activity—that is, the displayed activity is the product of a software algorithm, rather than the brain being studied is really active.
"The results of this study have led to the questioning of the results of 40,000 fMRI studies and may have a huge impact on the interpretation of neuroimaging results."
They confirmed that one of the mistakes was already in the system 15 years ago, and the error was corrected when the research team started writing the paper in May 2015, but the study still questioned those previous fMRI-based studies. .
So, what is wrong with this research method? fMRI's work is based on a huge magnetic field that captures changes in the brain's local blood flow as it passes through the body. These tiny changes represent an increase or decrease in activity in a particular area of ​​the brain. Software is also based on this principle. But the problem is that when you look at the data, what scientists see is not the real brain—the images of the brain they see are made up of countless “stereopixels†that are then compiled by software, Richard Chaweng, the register Said the net.
"Unlike humans, when scanning stereo pixels, the software is looking for clusters," Chirgwin said. "If someone says 'scientists know that the next move will be: the image is evidence', they are actually telling the statistics software Their information is parsed."
Due to the higher price of fMRI, the sample size used in the study is usually small, and very few (if any) studies will conduct confirmatory tests to confirm the results. Currently, the means available for verification are also very limited.
Since the introduction of fMRI in the early 1990s, neuroscience and psychologists have faced enormous challenges in confirming the results of the trial. However, Eklund believes that as the results of the fMRI examinations are freely available online, and the validation methods continue to evolve, there will be more and more confirmatory tests, and people will find errors in the software early.
“It used to be 10 to 15 years to analyze the results with a single computer,†said Lund. “Now, due to the use of graphics cards, the time required for analysis has dropped from 10 to 20 days.â€
So what about 40,000 papers that may have problems? All we can do is repeat the experiment to see which ones are right and which are wrong.
Source: èŒèšª five-line spectrum
Shanghai J.Shine Co.,Ltd , https://www.jshinechem.com