ADVERTISEMENT
What Does Imaging Data Show Us About Schizophrenia?
In this video, Sanjai Rao, MD, discusses whether imaging data could play a role in diagnosing schizophrenia. Dr. Rao made his comments during a Psych Congress 2020 session titled "Solving Clinical Challenges in Schizophrenia," in which he answered questions related to schizophrenia that were submitted by clinicians.
Dr. Rao is Associate Clinical Professor of Psychiatry and Associate Residency Training Director, University of California San Diego School of Medicine and VA San Diego Healthcare System.
Read the transcript:
We have also known for a number of years that there are various abnormalities that you can see in groups of people with schizophrenia, compared to groups of people without schizophrenia. You can use structural MRI techniques that demonstrate gray matter loss and increased ventricle size and decreased volume in various places.
You can also study connectivity between different regions of the brain using techniques like diffusion tensor imaging. This allows you to see that there are deficiencies in connectivity, including in really big areas, like the corpus callosum, but also a number of other areas as well.
There are some small studies suggesting that you can use techniques like this to classify whether a group of patients has schizophrenia or doesn't have schizophrenia and they're reasonably sensitive and specific.
Of course, you've got things like PET and SPECT that can be used to characterize receptors and characterize things like dopamine synthesis capacity. This is actually the source of our information that patients with schizophrenia have increased striatal dopamine synthesis capacity.
Because you can characterize receptor targets, this may be the thing—or one of the things—that helps us guide targets and dose ranges for new treatments. With all of this wealth of data that we have, you're probably wondering, "Well, why can't we just stick somebody in a scanner and figure out if they have schizophrenia?"
The answer, of course, is that, for all of their promise, there are a number of limitations in imaging studies. The sample sizes are usually pretty small. These are expensive studies to do.
The population is usually pretty heterogeneous. People with first-episode, people with later-episode, people with comorbidities, and when you put it all together, there's a big overlap between what's considered normal and abnormal.
As an example, if I gave you two buckets of data, and one was patients with schizophrenia and one was patients without, and the data I gave you for both groups was ventricle size, if you take average ventricle size between the schizophrenia group and the non-schizophrenia group, you'd see a difference.
On the other hand, if I gave you a single person from each group, you may not be able to tell the difference, because there's a big overlap between what's considered normal and abnormal in these studies.
How do you get around this? One tried and true way of getting around things like sample size limitations is, of course, by doing a large meta-analysis. That can help. In this case, it's definitely not the cure-all, because if you have a lot of groups that have a lot of overlap between normal and abnormal, then you just end up with a really large group that has a lot of overlap between normal and abnormal.
Then the thinking goes, "Well, what if it's just not one thing?" What if looking at just something like ventricle size is not the way to go, but really, it's a pattern of abnormalities. Maybe the detection of this is in the patterns.
You can do that. You can do what's called multivariate pattern analysis, where you look at multiple findings, put them together, and try to spot patterns. This shows some promise. Then, if you want to really make this efficient and automate this, you train a computer algorithm to do it. That's where machine learning comes in.
Here's where you essentially take a computer algorithm, and you feed it a big bucket of data where you already know the answer—a known, labeled dataset. Then you try to get it to learn and refine itself based on that dataset. Then you give it another dataset and see if it can pick up on the differences there.
The best example of this is any of you out there—and this should be all of you—who've used a modern smartphone have witnessed this in the photos app of your smartphone. You can go into your photos app on your iPhone, your Android phone, or whatever. Type in cat, beach, train, cable, or whatever into the search field, and your photos app will magically find photos that you've taken that have those elements in it. That's not because some person has looked through all your photos and figured that out for you. It's because the computer algorithm in the phone itself has been trained on a big set of data to pick up things like that. If you can do that in a phone, maybe you can do that in neuroimaging.
Bottom line, you still can't use this in clinical practice. We can't just take our patients, put them in scanners, and figure this out, but there is some promise here. There are newer techniques and analysis that are showing higher sensitivity and specificity. Perhaps if machine learning and multivariate analysis advance far enough, this might be someplace where we could go.
