Melanie Thernstrom lies motionless inside the large, noisy bore of a functional MRI scanner at Stanford University. She tries to ignore the machine's loud whirring as she trains her attention on a screen mounted inside the scanner, right in front of her eyes. An image of a flame bobs and flickers, shifting subtly in size. To her, the flame is a representation of the searing pain in her neck and shoulder, with its fluctuations reflecting the rise and fall of her discomfort. To the neuroscientists scrutinizing her through a window from the control room next door, the flame is a measure of the activity in a part of her brain.

As Thernstrom's pain increases, so does the amount of activity in part of this brain area, called the anterior cingulate cortex (ACC). This boomerang-shaped region, located in the frontal lobes, straddles the brain's midline between the ears and the forehead. Thernstrom's task is to will the flame to shrink, thereby reducing the neuronal hubbub in that region and the sensation of pain. With software rapidly parsing the machine's data to update the image of her ACC, Thernstrom can peer inside her own mind. She can observe, fuzzily, her brain's inner workings almost in time with the conscious manifestation of her discomfort.

Pain is your brain's way of telling you that your body is facing impending or actual damage to your tissue. To send that signal, the brain constructs an unpleasant sensory and emotional experience. When you get a paper cut, for example, the nerves in your finger shuttle a message to your brain, which interprets that incoming missive to beget the experience of pain. The ACC helps to modulate the pain response. The prickling sting of a sliced fingertip serves as an internal red flag alerting you to an attack or threat, and it vies for your attention with your other perceptions and cognitive states—musings about lunch, the ping of an incoming e-mail, a co-worker's pungent cologne.

The brain is not an unerring interpreter of the body's maladies. For Thernstrom, chronic pain was impeding her ability to get on with life, and she often felt as if her pain was draped like a veil over her thoughts. If she could lift the veil, she would be able to resume the daily activities she had relinquished. To do so, we and our collaborators hypothesized, she had to learn how to gain conscious control over her ACC.

Marshaling one's neurons to behave in a certain manner is no easy feat. Much as a baby might learn how to manipulate her fingers and toes through trial and error, Thernstrom has to discover what patterns of thoughts stoke the fire versus snuffing it out. First, she tries to convince herself that the burning sensation is soothing warmth, as if she is on a beach basking in the sun or relaxing in a Jacuzzi. The flame amplifies. The flickering image reminds her of scenes in a recent theological book she had read, in which religious martyrs were being burned at the stake. This gives her an idea.

She calls to mind the story of Akiba ben Joseph, who was said to have joyfully surrendered to his tortuous fate as a way of asserting his devotion to God. As she imagines herself taking on a similar role, she notices the flame begin to dwindle, along with her pain. As long as she concentrates on the feeling of surrender, the aches in her neck and shoulders lose their edge.

Thernstrom is one of several chronic pain sufferers who have volunteered to help our laboratory investigate an emerging technology called real-time fMRI (rtfMRI) neurofeedback. Imaging technologies, widely used to produce snapshots of the brain in action, are now gaining traction as tools for rehabilitation, therapy and brain training more broadly. This neurofeedback technique builds on the idea that exposing a person to his or her own patterns of brain activity could help that individual modify harmful or undesirable cognitive processes. Rather than training someone's brain to adopt new habits by teaching a new task—say, learning to juggle to improve hand-eye coordination—this approach aims to alter brain activity directly, through a person's own process of discovery.

For certain disorders, among them chronic pain and the loss of movement in Parkinson's disease, scientists have identified key brain regions implicated in the ailment. By showing a person how that target area is behaving in the present moment, a recipient of neurofeedback can experiment with mental strategies to alter the brain's behavior. As the technology develops, rtfMRI has the potential to help sufferers of numerous brain-based disorders exert greater control over their disease process. Our latest work has investigated using it to modulate the brain in depression, and it may also assist with anxiety, phobias and physical rehabilitation after a stroke.

It might enhance cognition for healthy people, too, by identifying when key brain regions are primed to learn or by helping individuals internalize the thought patterns associated with a creative mind-set. Brain imaging has revolutionized how neuroscientists and psychologists view the human experience. Now it is poised to help the rest of us alter that experience, from the inside out.

Seeing Inside the Mind

Researchers began exploring the use of neurofeedback in the early 1970s. At the time they focused on electroencephalography (EEG). An EEG reading uses electrodes placed on a person's scalp to pick up patterns of electrical activity near the surface. Typically a study participant might don the electrodes and perform some cognitive task, such as imagining moving an arm. Researchers would record the signals, translate them into a picture for the participant to see, and ask that subject to try to regulate their brain activity by mentally altering the picture in some way.

The clinical potential of neurofeedback soon caught researchers’ attention. Among the disturbances that neuroscientists sought to alleviate were seizures, anxiety, depression, addiction and chronic pain. They saw some success using EEG neurofeedback to train patients suffering from epilepsy to normalize the neuronal rhythms underlying seizures. Yet the technology of the time had major limitations. Mapping the signals picked up by the EEG's electrodes to a specific brain area is exceedingly difficult and often impossible. The electrodes, which most clearly register neural activity near the surface of the brain, cannot listen in on the deep-brain structures implicated in many disorders.

In the 1990s fMRI revolutionized neuroimaging research. This now widespread technique works by measuring the oxygen content of blood. When the neurons in a certain region of the brain become highly active, they draw more oxygen as fuel from nearby blood vessels, thus triggering an increase in blood flow to that area. Blood that is rich in oxygen and blood that is oxygen-impoverished differ in their magnetism, and these distinctions serve as a proxy for levels of brain activation. So when the powerful magnet of an fMRI machine releases its bursts of radio waves, brain regions that are more or less engaged will produce a correspondingly strong or weak signal. By comparing the resulting maps of variations in the oxygen content of blood under different conditions, neuroscientists and psychologists can gain insight into how the brain carries out a given task.

The first serious efforts to adapt fMRI to neurofeedback occurred in 1995, when biophysicist Robert Cox, then at the Medical College of Wisconsin, and his colleagues found a way to process data from brain scans in real time, as opposed to after an experiment was already over—a crucial initial step. Seven years later several laboratories showed they could share that continuous stream of data with the person being scanned and coach that individual into altering brain activity in specific areas.

In subsequent years, our lab at Stanford, along with neuroscientist R. Christopher deCharms, turned its attention to training people with pain to use neurofeedback to alleviate their suffering. We asked eight healthy participants and eight people with chronic pain to undergo fMRI scanning. The healthy subjects held a heat probe in their left palm as their head lay inside the machine. The temperature of each probe was set to the maximum level that its holder could endure without squirming, equivalent to a seven out of 10 on a pain scale.

We then described some strategies to both the healthy participants and the chronic pain sufferers for either increasing or tamping down their hurt. To amplify pain, for example, we suggested to our subjects that they attend to their discomfort, consider the sensation a threat—perhaps by focusing on its frightening aspects—or allow the painful sensation to wash over them. To diminish pain, we offered techniques such as shifting focus away from the pain, interpreting the sensation as nothing special or attempting to gain control over the experience. Participants were also encouraged to come up with their own strategies. This freedom quickly became an important part of the treatment's success. None of the researchers could have guessed that Thernstrom imagining herself as a martyr being burned at the stake would have been most effective for her.

As our subjects tried out their cognitive strategies, software analyzed the fMRI signal corresponding to their ACC and presented it back to them as a growing or shrinking virtual flame. The participants adjusted their thought patterns until they found one that drove the flame in the desired direction. Afterward, they all rated their pain.

We found that both groups could change their experience of pain; in fact, the chronic pain sufferers cut their pain ratings by half. The greater the participant's ability to control ACC function, the more that person's pain diminished. Both groups were also able to maintain control over their ACC activity and their experience of pain even when they were no longer receiving visual feedback. Our control groups—composed of subjects receiving no feedback, sham feedback, or biometrics such as their heart rate and perspiration—did not show the same degree of control over their pain and ACC activation after practicing their cognitive strategies.

Other scientists have applied this neurofeedback approach to combating the symptoms of Parkinson's. In 2011 neuroscientist Leena Subramanian of Cardiff University in Wales and her colleagues tested rtfMRI methods on 10 individuals with early-stage Parkinson's by scanning them twice in sessions spanning two to six months. During the first visit, half the participants observed the activity in their supplementary motor area (SMA), a motion-control region that is hypoactive in Parkinson's patients, while lying in a scanner. These individuals were given free rein to imagine any kind of movement in an effort to engage more of the SMA. The other five subjects made up the control group. They also imagined moving while in the scanner but did not see their brain activity. In the intervening months, all 10 participants devoted time at home to picturing themselves executing complex movements, such as playing a sport.

When the researchers scanned the participants’ the second time, the patients who had received neurofeedback showed more activation in their SMA, performed faster on a finger-tapping task and improved on clinical symptoms of Parkinson's 37 percent more than the control subjects. Neurofeedback appears to have helped these patients develop a more effective mental imagery strategy than those who lacked that information, which gave the former group an advantage in their home practice.

Focusing on single brain areas, as these two studies did, has produced exciting results, but this approach has its limits. Any thought or feeling invokes complex networks in the brain. Even simple acts, such as bending down to sniff a flower or contemplating shapes in a cloud, emerge from a precise choreography of chemical and neuronal ensembles. As our grasp of the dynamics underlying our mental states improves, we can unlock the true potential of rtfMRI.

A Tune-up for Brain Networks

A big step forward for this technology will come from matching specific mental states to activation patterns that encompass the entire brain, so that people learn how to alter broad patterns rather than particular regions. Already scientists have been able to map the complex activation patterns seen in fMRI images to subjective reports of what a person is thinking, allowing neuroscientists to pull off a rudimentary form of mind reading.

To intrude so deeply into a person's thoughts, that individual first must look at thousands of images while lying in a scanner. After building up a database of pairs of activation patterns and the images that triggered them, a computer can decode what a person might be picturing at a given time [see “Movies in the Cortical Theater,” by Christof Koch; Scientific American Mind, January/February 2012]. We can gain further resolution, too, by having a computer learn to distinguish between different brain states associated with a certain stimulus or experience—say, a happy thought versus a sad reaction in response to a picture of a pony.

One way to improve rtfMRI is to conduct this type of pattern matching along with neurofeedback. We would need to assemble the pairs of brain states and stimuli anew for each participant, as the encodings of thoughts and memories differ from person to person. Part of the challenge here is that the software can err when classifying a volunteer's activation information as a particular brain state. With brain activity shifting subtly in fractions of a second, the desired state can end up labeled incorrectly or muddied by overlapping cognitive states. Sharing with a volunteer how his or her brain states are classified could expose errors and encourage that person to conjure up clearer brain states that are more representative of a certain thought or feeling. Ideally, this collaborative process would yield perfect accuracy so that our software could always tell if you are, say, happy versus sad.

Preliminary work by Stephen LaConte, now at the Virginia Tech Carilion Research Institute, and his colleagues suggests that this new technique may be useful for training brain states related to reducing cravings in addicts. In 2009 the researchers found that they could predict, based on brain-wide patterns of data, whether a chronic smoker was in a state of craving. Therapies targeted at diminishing the intensity of that brain pattern could help substance abusers overcome their harmful urges. More recently, our lab has used this pattern-matching method to detect the presence or absence of acute or chronic pain.

Even the therapies themselves could become tailored to individual use. As we learn more about what brain processes support specific cognitive techniques, rtfMRI neurofeedback could strengthen the relevant networks. A broader range of people could end up benefiting from strategies such as mindfulness meditation or cognitive-behavior therapy, which are already used to improve emotional, cognitive and physiological dysfunction.

The potential of rtfMRI is not limited to disease. Neurofeedback could be used to train people to develop subtle mental strategies that alter their neural function to promote creativity, for example. Once they have learned the techniques for inducing a more creative brain state, they can rehearse this frame of mind in their day-to-day activities, similar to the approach used in the Parkinson's study mentioned earlier. Practitioners would likely return to the scanner for an occasional mental tune-up to update their strategies as their brains adapt.

Neuroscientists have made initial strides toward applying rtfMRI to enhancing learning, perception, performance and wellness. Our ability to pick up new information and skills fluctuates—at times, we are either more or less prepared to learn, and neuroimaging has revealed the underlying brain states that correlate with this readiness. In one experiment published in 2012, for example, a team led by John Gabrieli of the Massachusetts Institute of Technology showed its subjects pictures of scenes and parsed their brain data in real time to monitor the performance of the parahippocampal place area (PPA), a region involved in remembering and recognizing scenes. The scientists discovered that their subjects formed more accurate memories of the pictures they viewed when the PPA was in a prepared state than when it was in a less optimal condition. Such studies suggest that we can accelerate learning by adapting a training program to the brain's present condition.

More generally, rtfMRI neurofeedback can also be used as a novel tool for probing brain function. Traditional fMRI research involves asking participants to engage in a task and measuring the effect on the brain. The results give us associations, but we cannot know definitively whether the task caused the brain changes. With rtfMRI, we are able to test our assumptions about how the brain works by selectively manipulating specific brain areas or networks and observing the outcome. Neuroscientist Mitsuo Kawato, director of the ATR Computational Brain Information Communication Research Group in Japan, has coined the term “manipulative neuroscience” for this burgeoning field.

In work published in 2012 Kawato and his colleagues used rtfMRI techniques to test whether they could improve one small aspect of their subjects’ visual perception without them ever becoming consciously aware of what they were learning. First, they had a computer learn the activation patterns in the visual cortex associated with specific orientations of lines—30, 70 or 120 degrees. They then gave their subjects feedback on how closely the activation in their visual cortex resembled one of these patterns and tried to coax them into matching their brain activity to that associated with seeing a particular diagonal line. The experimenters did so without showing their subjects any lines, explaining the meaning of the neural patterns they were pursuing or revealing the intention of the experiment.

Afterward, the participants were significantly better at detecting the diagonal line they had just been trained on than they had been at the beginning of the experiment. This study revealed an elegant way to test that the activation patterns we suspect are associated with a given stimulus or behavior—in this case, observing a particular diagonal line—are indeed linked. The vague relations hinted at by traditional brain scanning are finally giving way to more concrete results.

Before rtfMRI neurofeedback can become a widely used therapeutic tool, however, we will have to address the exorbitant cost of an fMRI scanner. Its initial adoption, then, will likely be in assuaging conditions that are notoriously difficult or expensive to treat long term, such as chronic pain and addiction. Other opportunities lie in blending rtfMRI with less expensive, more mobile imaging technologies, such as EEG or near-infrared spectroscopy (NIRS). NIRS is similar to fMRI, but it uses light rather than a magnet to measure brain function. Although EEG and NIRS do not offer the same whole-brain access as fMRI, researchers might be able to translate the portrait of brain activity achieved through rtfMRI neurofeedback into an EEG or NIRS signature.

With rtfMRI neurofeedback, we have the opportunity to peek under the hood—to access the origins of our conscious and unconscious thought processes. It allows therapists to offer treatment and simultaneously monitor the brain's response to that treatment. And it opens up the possibility of having therapies and training regimens evolve in step with an adapting brain. In exploring this new landscape, both to aid research and to accelerate healing, we are only beginning to learn of our own capacity for self-directed growth.