authors_brain

by Kevin Sitek
figures by Daniel Utter

Over the years, scientists have developed many techniques to observe what’s going on in the human brain as we think or move. Unfortunately, few of the insights we have made so far have resulted in any improvements in standard clinical mental care. Recent advances in neuroimaging may be changing this. Studies from the past few years have shown that various measures of brain activity and structure can predict who will develop certain mental disorders as well as how individuals will respond to particular treatments. In fact, measurements of brain activity can be better at diagnosing depression than standard clinical assessments. This is just one of many examples of modern neuroimaging techniques beginning to contribute to how we diagnose and treat diseases.

Developing tools to look at what’s inside our heads

Despite centuries—or perhaps millennia—of questions about how the brain works, there weren’t tools for looking at the living brain until the second half of the twentieth century. Before that, our knowledge of the brain’s organization and function largely came from dissecting patients’ brains after death and matching up any physical abnormalities with their reported symptoms during life.

Starting in the 1960s with the development of computerized axial tomography (CT or CAT scans) and magnetic resonance imaging (MRI), it became possible to take images of the structure of a living person’s brain. Next, in the 1980s, scientists figured out how to produce functional brain images using positron emission tomography, better known as PET. Instead of one image of the brain’s structure, they could now take snapshots of the brain over time to observe how brain activity changed.

However, a downside to PET is that it requires injecting the participant with a radioactive isotope, which is then tracked by the scanner while the participant performs a given task. The radioactivity is equivalent to what you’d receive normally just by living on Earth for about 5 years. Inspired by functional PET imaging, researchers figured out how to use an MRI scanner to create functional images without any radioactive isotopes.

Figure 1. (A) Computerized axial tomography (CT or CAT scan) combines many X-ray images, collected using small doses of radiation, to create 3-D images of the human brain. (B) Magnetic resonance imaging (MRI; right) measures the magnetic properties of various tissues in the human brain. Functional brain activation can be measured using (C) positron emission tomography (PET) and (D) functional magnetic resonance imaging (fMRI; here, laid on top of a structural MRI for reference). In both functional images, warmer colors represent brain areas with more activity. In PET this is inferred by measuring metabolism, whereas in fMRI it is inferred based on where oxygenated blood is. While each imaging method has its own strengths, one benefit of MRI (both structural and functional) is that it does not expose the participant to radiation. (Wikimedia)
Figure 1: (A) Computerized axial tomography (CT or CAT scan) combines many X-ray images, collected using small doses of radiation, to create 3-D images of the human brain. (B) Magnetic resonance imaging (MRI; right) measures the magnetic properties of various tissues in the human brain. Functional brain activation can be measured using (C) positron emission tomography (PET) and (D) functional magnetic resonance imaging (fMRI; here, laid on top of a structural MRI for reference). In both functional images, warmer colors represent brain areas with more activity. In PET this is inferred by measuring metabolism, whereas in fMRI it is inferred based on where oxygenated blood is. While each imaging method has its own strengths, one benefit of MRI (both structural and functional) is that it does not expose the participant to radiation. (Adapted from Wikimedia)

Instead, the MRI scanner uses a large magnet to measure differences in various tissues in the body (Figure 2). During functional MRI, or fMRI, the scanner is specifically sensitive to oxygenated blood. Since active parts of the brain use up a lot of oxygen, new oxygen must be delivered through the blood stream. As a result, fMRI infers brain activity by tracking where oxygenated blood is increasing in the brain (see the image in Figure 1D). With this information we can look at brain activity differences between people with different mental states and disorders.

Figure 2. Schematic of how functional magnetic resonance imaging (fMRI) works. Protons inside atoms in the brain are always spinning, but each atom’s spin is in a different direction. Using a powerful magnet, we can align all the spins so they are pointing in the same direction. Next, using a radiofrequency pulse, the spins of all the atoms are tilted over, but they’re still all pointing in the same direction. However, over time, their spins will move back to the original position. Depending on the characteristics of the atoms, some will return more quickly than others. Importantly for fMRI, oxygenated blood takes longer to return to normal than deoxygenated blood. FMRI measures this difference as the blood oxygenation level dependent, or BOLD, signal. Active parts of the brain need to have their oxygen stores replenished, so the BOLD signal tells us which brain areas have been more active than others.
Figure 2: Schematic of how functional magnetic resonance imaging (fMRI) works. Protons inside atoms in the brain are always spinning, but each atom’s spin is in a different direction. Using a powerful magnet, we can align all the spins so they are pointing in the same direction. Next, using a radiofrequency pulse, the spins of all the atoms are tilted over, but they’re still all pointing in the same direction. However, over time, their spins will move back to the original position. Depending on the characteristics of the atoms, some will return more quickly than others. Importantly for fMRI, oxygenated blood takes longer to return to normal than deoxygenated blood. FMRI measures this difference as the blood oxygenation level dependent, or BOLD, signal. Active parts of the brain need to have their oxygen stores replenished, so the BOLD signal tells us which brain areas have been more active than others.

From voxels to discoveries: Analyzing brain imaging data

An fMRI scan results in a three-dimensional image made up of “voxels,” or 3-D pixels. Each voxel represents a volume in the brain about 2-3 cubic millimeters in size—pretty small, but big enough to contain hundreds of thousands of neurons. With fMRI, the scanner takes one image every two seconds or so, resulting in a time series of brain images, each with hundreds of thousands of voxels. Each voxel for each time point has a particular level of activity. This activity is “Blood Oxygenation-Level Dependent, ” meaning more oxygen leads to greater signal, so it goes by the name BOLD.

However, this time series isn’t ready to be published as a pretty picture yet. There are differences from one image in the time series to the next that aren’t caused just by changes in brain activity. This can be due to the participant moving in the scanner or even just due to normal breathing. If this isn’t fixed before the images are combined, then a voxel that’s in brain area A in one image could be in brain area B in another image! A computer program aligns all the images so they are lined up and can be compared across time.

After the fMRI data have been processed, they next go through rigorous statistical tests to figure out which parts of the brain were active during the scan. A common misconception is that only about 10% of your brain is doing anything at any particular time. In fact, your whole brain is always active! What this fMRI statistical analysis does is look for which brain areas are more or less active than the rest of the brain during a particular task. These differences are very small—BOLD activity might increase only 1% during a task—but historically, the differences have been noticeable if we combine results from many people, each of whom does the task a number of times.

Figure 3. Results of a statistical fMRI analysis. Participants in this task were asked to actally move their right hand or to mentally think about moving their right hand. Although the whole brain is active all the time, the colored voxels in this image show which areas are more active during this particular task. The results show that thinking about doing a motion and actually doing that motion have similar brain activity. Interestingly, the left side of the brain is more involved in controlling the right side of your body. (Wikimedia)
Figure 3: Results of a statistical fMRI analysis. Participants in this task were asked to actually move their right hand or to mentally think about moving their right hand. Although the whole brain is active all the time, the colored voxels in this image show which areas are more active during this particular task. The results show that thinking about doing a motion and actually doing that motion have similar brain activity. Interestingly, the left side of the brain is more involved in controlling the right side of your body. (Wikimedia)

Making new discoveries with brain imaging

Functional human brain imaging makes it possible for scientists and doctors to “see inside” the working brain. As a result, brain imaging can be used to investigate how the brain functions differently in people with various psychiatric disorders. Studies showed that during tasks involving emotions, people with depression had different activity in a region in the middle of the front of the brain than people without depression. Other studies showed that other frontal areas involved in attentional control are less active in people with attention deficit hyperactivity disorder.

Not all brain imaging discoveries had to do with how participants performed tasks. In fact, one of the biggest insights came from observing brain activity when people were asked to do nothing at all! When researchers looked at how similar activity across different brain areas was, they found fascinating networks of “resting state” activity.

Figure 4. When participants aren't given a task and are just asked to lie still in the MRI scanner, the "resting state network" becomes active. Areas in blue and green are highly correlated, or synchronized, with each other, meaning they are active at the same time. Since the participant wasn’t asked to do any particular task, this pattern of activity has come to be known as the resting state network. By looking at how strong the resting state network is, or at which other areas are also synchronized with the resting state network, we can make inferences about how brain activity is organized in people with various psychiatric conditions. (Wikimedia)
Figure 4: When participants aren’t given a task and are just asked to lie still in the MRI scanner, the “resting state network” becomes active. Areas in blue and green are highly correlated, or synchronized, with each other, meaning they are active at the same time. Since the participant wasn’t asked to do any particular task, this pattern of activity has come to be known as the resting state network. By looking at how strong the resting state network is, or at which other areas are also synchronized with the resting state network, we can make inferences about how brain activity is organized in people with various psychiatric conditions. (Wikimedia)

What can brain imaging tell us about a single person?

New discoveries about brain functions and disorders are made using brain imaging every day. These findings tell us how most people use different brain areas, or how people with a particular disorder differ from people without that disorder as a group. However, it has been extremely difficult to use brain imaging to tell us information about a single person. So as we mentioned in the previous paragraph, people with depression process emotions differently than non-depressed people, on average. But we haven’t been able to pinpoint whether one specific person processes emotions differently enough that they’re probably depressed.

As mentioned earlier, some of the difficulty is due to BOLD activations being relatively small, so researchers need to collect data from 20 or 30 participants and average their results together in order to see any effects. Other issues arise from the computational challenges of brain imaging—each image can be made up of 100,000 voxels, so doing statistical analyses on these data have historically taken powerful computers a long time.

However, recent advances in computational power and machine learning algorithms have opened the door to single-subject data analysis. Faster, more powerful computers means that we can use higher resolution brain scans with more sophisticated processing steps. Improved data processing steps result in “cleaner” data and make BOLD activity easier to see. This means participants don’t have to do a task for quite as long, and we don’t need to combine results from as many participants in order to see brain activity differences.

Machine learning algorithms—along the same lines as what Netflix uses to recommend you movies based on what you’ve enjoyed previously—make predictions about new data based on earlier data. (See the special edition article on computational neuroscience for more.) These same principles can be applied to neuroimaging data. An algorithm can look at the fMRI activity of some participants, along with important information about the participants (such as their age, gender, or intelligence). Then, with only the fMRI activity of a new group of participants, the algorithm can “predict” information about this new group (like which ones are male, old, or average IQ).

Figure 5. A machine learning algorithm known as a "random forest classifier" takes in a lot of data (A). It looks at a subset of those data (B) and figures out characteristics of the data (C) in a “decision tree.” It does this a number of times, constructing many trees (hence the name “forest”). Next, it looks at the subset of the original data that it didn’t consider initially (D), which weren't used earlier to figure out the important characteristics in (C). The algorithm then classifies the “new” data in (D) based on the characteristics it learned from the random forest (C) determined by the previous subsets (D).
Figure 5: A machine learning algorithm known as a “random forest classifier” takes in a lot of data (A). It looks at a subset of those data (B) and figures out characteristics of the data (C) in a “decision tree.” It does this a number of times, constructing many trees (hence the name “forest”). Next, it looks at the subset of the original data that it didn’t consider initially (D), which weren’t used earlier to figure out the important characteristics in (C). The algorithm then classifies the “new” data in (D) based on the characteristics it learned from the random forest (C) determined by the previous subsets (D).

In the past few years, neuroscientists have used these computational advances to investigate various mental health disorders. For instance, a study of patients with social anxiety disorder took brain scans before the patients underwent cognitive behavioral therapy. The patients looked at pictures of faces with various emotions. The study found that the patients who went on to respond best to the therapy had the greatest brain responses to angry faces before the therapy.

A recent study from MIT and Harvard Medical School scanned children whose parents have major depressive disorder. Children with a parent who is depressed are three times more likely to develop depression themselves. The study also scanned a group of children whose parents were not depressed. By looking at brain activity when there was no task, the researchers found that the resting state network was more synchronized with areas that process emotions in the brain in the at-risk children. That is, when the fMRI participants were lying in the scanner without an explicit task, the emotion-processing regions were active in similar patterns as the resting state network (which is shown in Figure 4),. This synchronicity could represent overactive emotion processing and could explain why children whose parents have depression are more susceptible to developing depression themselves.

Beyond the findings that compared the two groups directly, the researchers also trained a computer classification model to guess whether each child’s brain activity makes them more likely to be in the at-risk group or the control group. The computer model based on brain activity ended up classifying the children correctly 80% of the time with very few misclassifications of control children as at-risk (20%). In contrast, a computer model that was built on behavioral tests that clinicians use to diagnose mental disorders was correct only 64% of the time, with many more misclassifications of controls as at-risk (73%). These results suggest that brain activity is actually a better predictor of whether someone is at risk for a psychiatric disorder than the behavioral data used by clinicians!

So are my next doctors an MRI and a computer?

Despite promising results using computer algorithms to predict whether someone has a disease or not, we won’t be replacing MDs with MRIs anytime soon. For one, most studies have been successful at picking out who has a disorder from the pool of study participants that were included because they had the disorder (the test participants) or because they didn’t (they were control participants). It is much more difficult to pick someone out of the general population and accurately predict which disorder they may have—the resting state networks of two very different disorders could actually look quite similar.

However, it is extremely encouraging that researchers may soon be able to make accurate medical predictions with a “data-driven” approach. Diagnosis of mental health disorders is notoriously subjective, with clinicians diagnosing patients differently. Recent advances in human brain imaging show that diagnosing medical disorders will be more consistent and more accurate in the near future.

Kevin Sitek is a graduate student in the Harvard Program in Speech and Hearing Bioscience and Technology and conducts research at the McGovern Institute for Brain Research at MIT.

This article is part of the April 2016 Special Edition on Neurotechnology.

For more information:

“Diagnosing depression before it starts”

“Twenty years of fMRI”

“Idle minds: Neuroscientists are trying to work out why the brain does so much when it seems to be doing nothing at all”

Featured image (top) created by Kevin Sitek.

4 thoughts on “Can Computers Use Brain Scans to Diagnose Psychiatric Disorders?

  1. Autism – There is a “Brain Balance” theory that suggests autism may occur when the two hemispheres of the brain do not ‘cross over’. The patient is given exercises to ‘teach’ and help this connection in the brain. In the clinic, children have been said to be much improved and especially those with severe autism. Wondering how this might be a great breakthrough to help if the study above can show this premise to be true.

  2. Kevin, I thought this was kind of interesting. I’ve heard that in the past that the ridges and grooves on your skull were actually used to help diagnose things. I think it’s amazing how advances in technology allow us to further understand the brain and neurological structures.

  3. Dr. Melillo, Founder of Brain Balance, is said to be releasing a clinical study on his theory this Fall. Hopefully, it will provide a strong case against the nay-sayers.

    On another note, Dr. Daniel Amen has done clinical studies using SPECT with great results.

Leave a Reply

Your email address will not be published. Required fields are marked *