If, how, and when fMRI goes clinical

This blog post was inspired by the twitter conversation that ensued after Chris Gorgoleski’s provocative tweet shown below. The link to the entire thread is  provided here.
Before I begin, I have to emphasize that while I am an NIH employee, my opinions in this blog are completely my own based on my own admittedly fMRI-biased perspective as an fMRI scientist for the past 28 years, and not in any way associated with my employer. I don’t have any official or unofficial influence on, or representation of, NIH policies. 
Back in 1991, the first fMRI signal changes were observed, ushering in a new era in human brain imaging that has reaped the benefits from its relatively high resolution, sensitive, fast, whole brain, and non-invasive assessment of brain activation at the systems level. With layer and columnar resolution fMRI currently producing promising results, it is starting to approach circuits level.  Functional MRI has filled a large temporal/spatial gap in our ability to non-invasively map human brain activity.  The appeal of fMRI has cut across disciplines – physics, engineering, physiology, psychology, statistics, computer science, and neuroscience to name a few, as the contrast needs to be better understood, the processing methods need to be developed, the pulse sequences need to be refined, the reliability needs to be improved, and ultimately the applications need to be realized. Neuroscientists and clinicians have applied fMRI to a wide range of questions regarding the functional organization and physiology of the brain and how they vary across clinical populations.
Because meaningful activation maps could be obtained from individual subjects (tap your fingers or shine a flickering checkerboard in your eyes, and the fMRI signal changes in the appropriate area in seconds – easily visible to the eye), the hope arose early on that this was a method that could be used clinically to complement prediction, diagnosis, and treatment of a wide range of neurologic and psychiatric pathologies. Sure, we can see motor cortex activation but can we differentiate, on an individual level, say, who is left handed vs right handed by comparing this activation? Perhaps group statistics might pull out a difference, but to assign an individual to one group (left handers) versus the other (right handers) with a level of certainty above 90% is a much more difficult problem. This type of problem encapsulates the essence of the difficulty associated with many hoped for clinical implementations of fMRI. Nevertheless, funding agencies embraced fMRI as it was generally accepted that its potential was high for shedding light on understanding the human brain and enhancing clinical treatment. Even with no clear clinical application, NIH embraced fMRI for its research potential. A sentence taken out of NIH’s mission statement is as follows:
“The mission of NIH is to seek fundamental knowledge about the nature and behavior of living systems and the application of that knowledge to enhance health, lengthen life, and reduce illness and disability.”
This clearly states the position that fundamental knowledge is important for clinical applications, even if the applications are not clearly defined. Functional MRI has certainly contributed to fundamental knowledge.
Over the years fMRI has grown in maturity as a tool for neuroscience research, substantially impacting the field, however, the clinical applications have not quite panned out. Pre-surgical mapping emerged as the only billable clinical application, obtaining a CPT (Current Procedural Terminology) code in 1997 – and even here it has not become the standard approach as it is being carried out in a relatively small number of hospitals worldwide.
There are several techniques that are being tested in the clinic. One promising example is a novel application and analysis of resting state fMRI that extracts the relative time shift of the fluctuations across the brain, is being tested and used in clinics in Germany and China. The basic idea is that in regions with compromised flow due to stroke, the temporal delay in a component of the BOLD based resting stater fluctuations is clearly visible. This is a method that may obviate the need for the current clinical practice of using Gd contrast agents in these patients as not only is the specificity outstanding but the sensitivity is comparable. 
Why the stalled clinical implementation of fMRI?
What are the reasons for this stalled clinical implementation? Let’s take a step back to look at why MRI, the precursor to fMRI by about a decade, has been so successful for clinical use. Using an array of available pulse sequences and corresponding structural contrasts, MRI can effectively be used to detect most tumors as well as most lesions associated with stroke and other types of trauma.  The lesions that are detectable are visible with minimal processing, allowing the Radiologist to simply view the image and make a diagnosis. The effective lesion or tumor contrast to noise ratio is high enough (at least above 10) such that detection is routine on a single subject basis by a trained Radiologist. 
Functional MRI on the other hand, requires several processing steps – all of which may influence the final result, as well as for the subject to either perform a task – or not (with resting state fMRI), and remain completely motionless as the threshold for motion is much more strict for fMRI. After processing, a map of activity or connectivity is created. These maps, typically color coded and superimposed on high resolution anatomic scans, show individual results with relatively high fidelity, but unfortunately, the difference between a functional map (from either a task or from resting state) of an individual with a pathology and that of a healthy volunteer, relative to the noise and variance among subjects, is too low for a visual assessment of a Radiologist or even for statistical reliability. There has also been the question of what task to use to highlight differences between normal controls and individuals with pathology. In resting state, there’s the issue of not being able to really know what the subject is doing – introducing further uncertainty.
In the case of presurgical mapping however, the fidelity of mapping the location of some functional regions (motor, somatosensory, visual, auditory, language) is high enough to allow the surgeon to identify and avoid these areas in individual subjects. However, even with presurgical mapping, the method is potentially confounded by compromised neurovascular coupling in the lesioned area, up to an hour of additional scanning, extreme sensitivity to motion (as mentioned, more than typical MRI scans), unique warping of echo planar images relative to structural scans causing misregistration, and, again, additional offline processing steps that add a degree of difficulty and uncertainty in functional localization. 
For the above reasons, fMRI has not caught on clinically even with presurgical mapping, as other more invasive approaches are arguably more precise, straightforward to implement, and less expensive. 
Now the question starts to loom, how much longer should clinically focused funding agencies need wait to see fruition before looking elsewhere? A large fraction of fMRI researchers including both those who develop the methods and those that apply them towards some neuroscience or clinical question generally maintain a belief that fMRI will become more clinically useful in the near or intermediate future. This position is not just a bluff or a vacuous promissory note by researchers willing to give proper lip service to a distant goal over the horizon. I think most of us get it – that we really want this all to pay off. It would be beneficial for many grants to include careful thinking on the steps that would be necessary to take the research to clinical practice.  Others think that health-focused funding agencies should start to actively look elsewhere for potential techniques that are more likely to achieve clinical traction in the near future. 
A current growth phase of fMRI
My own sense is that fMRI is in or rapidly approaching another major growth phase. New insights into brain organization are emerging at an increasing rate due to new and more sophisticated paradigms (real time fMRI, resting state fMRI, naturalistic viewing, fMRI adaptation), higher field strengths, better RF coils, and more specific and sensitive pulse sequences (blood volume sensitive imaging for layer specific fMRI), large multi-modal pooled data sets that allow world-wide access for data mining (Connectome project, UK Biobank, etc..), and perhaps most importantly, more sophisticated processing approaches (dynamic connectivity measures, cross subject correlation, machine learning, etc..). These advances have also enabled deeper insights into the functional organization of brains from individuals with psychiatric or neurologic disorders. Specifically, the use of Big Data in combination with machine learning or multivariate analysis in general, in combination with other modalities (genetics, EEG),  have started to generate potentially useful biomarkers that could be applied on individual subjects for disease diagnosis, prediction, and treatment.
Just one clinical application away
An second growth phase may be precipitated by one major clinical application that is more effective and perhaps even less expensive than the clinical practice that it replaces. Once this happens, I believe that the big scanner vendors (Siemens, GE, and Philips), and perhaps new companies will direct more attention to streamlining the basic implementation of fMRI in the clinic. Better hardware, pulse-sequence, subject interface devices, and processing methods will rapidly advance as economic incentives will supersede the influence of grant money in this context. Of the potential clinical applications mentioned below, it’s not clear which one will emerge first to break into clinical practice. 
For the past two decades, fMRI has benefited substantially from the success of MRI, as this has caused a proliferation of fMRI-ready scanners worldwide and has kept many costs down. Can you imagine how anemic the field of fMRI would be if MRI were not clinically useful? The substantially smaller research market of fMRI would have consisted of substandard and much more expensive scanners resulting in much slower advancement. Likewise, imagine what the field could look like if the fMRI market moved from research to clinical? The field would experience a transformation. Researchers would have immediate access to a wider variety of state of the art sequences that exist on only a handful of scanners today. Methodology including subject interface devices, processing pipelines would not only advance more rapidly, but be more standardized and quality-controlled across centers. The on-ramp to further clinical implementation would be much smoother. 
How long to wait?
So the question remains, how long should funding agencies wait to determine if fMRI will catch on clinically? Some feel that they’ve waited long enough. Others feel, as I do, that the increased focus of the field on fMRI towards individual assessment as well as layer specific fMRI will likely make clinical inroads and is really just getting started. I also believe that fMRI – in synergy with other modalities – is not anywhere close to realizing its full potential in revealing fundamental new insights into functional organization useful to both basic neuroscience and clinical practice. To stop or even reduce support of fMRI now would be tragic. 
Potential Clinical Applications of fMRI in the Immediate Future. 
What are the potential clinical applications and what specifically would be necessary to allow fMRI to be used on a day to day basic with patients? 
  1. Disorder/Disease Biomarkers: Large pooled data sets that also contain structural data, genetic data, and a slew of behavioral data are just starting to be mined with advanced processing methods. Already specific networks related to behavior, lifestyle, and genetic disorders have been discovered. The long term goal here is the creation of multivariate biomarkers that can be applied to individuals either to screen, diagnose, or guide treatment with an acceptable degree of certainty. There are perhaps hard limits to fMRI sensitivity and reliability, but if the number of meaningful dimensions of information from fMRI are increased, then the hope is that this massively multivariate data may allow highly sensitive and specific individual subject and/or patient differentiation based on resting state or activation information. 
  2. Biofeedback: It has been demonstrated that when presented in real time with useful fMRI activation-based feedback in real time on a specific aspect of their dynamic brain activity, subjects were able to alter and tune their activity. In many studies, this led to a change in an aspect of their behavior – touching on depression, phobias, and pain perception. The fMRI signal is still slow and noisy, however, of higher fidelity than other real time neuronal measures. Recently, simultaneous use of EEG has been proposed to enhance the effectiveness of real time fMRI feedback. This is still in its early stages, however, clinical trials are underway. 
  3. Localization for Neuromodulation: An emerging area of clinical treatment is that of neuromodulation by the use of methods to stimulate or interfere with brain activity in a targeted manner either invasively or non-invasively. Deep brain stimulation, TMS, tDCS, focused ultrasound, and more are currently being developed for clinical applications – alleviating depression, Parkinson’s disease, and other disorders. The placement and targeting of these interventions is critical to their success. I see fMRI a playing a significant role in providing functional localizers so that the efficacy of these neuromodulation approaches may be fully realized.
  4. Assessment of locked in patients:Recent studies have shown that fMRI is superior to EEG in assessing the brain health, activity, and function of locked in patients. In some instances fMRI activity was used as a means for communication. This approach has considerable potential to be used on a regular basis in a clinical setting as no other methods compare  – even in its early stages of implementation.
  5. Brain Metabolism/Neurovascular Coupling/Blood Oxygenation Assessment: While activation and connectivity studies dominate potential fMRI clinical applications, more fundamental physiologic  information obtained using simultaneous fMRI measures with the appropriate pulse sequence, such as a combined arterial spin-labelling (ASL) for perfusion, blood oxygenation level dependent (BOLD), and/or Vascular Space Occupancy (VASO) contrast for blood volume, during a stress such as breath-hold or CO2 inhalation – or even during normal breathing variations at rest, can provide insights into baseline blood oxygenation, neurovascular coupling, and even resting and activation-induced changes in Cerebral Metabolic Rate (CMRO2). All these provide potentially unique and useful information related to vascular patency and metabolic health of brain tissue – with potentially immediate clinical applications that may fill a niche between CT angiography, ultrasound, and positron emission tomography (PET). 
  6. Perfusion Deficit Detection using ASL: has been in existence as long as BOLD contrast and significant effort has been made to test it clinically. While the baseline perfusion information that it provides is comparable to that obtained with injected Gd contrast, its sensitivity is significantly lower, requiring a much longer acquisition time for averaging. This has slowed widespread clinical implementation.
  7. Perfusion Deficit Detection using resting state BOLD: This is perhaps the most promising of the possible clinical implementations of fMRI in the broadest interpretation of the name. Mapping the relative latencies of resting state BOLD fluctuations clearly reveals regions of flow deficit. This approach compares well to the clinically used approach of Gd contrast in terms of sensitivity and specificity. Creation of latency maps from BOLD fluctuations is also relatively straightforward and could be performed seamlessly and quickly in an automated manner. This approach is currently being implemented in a limited manner in hospitals in Germany and China. 
  8. Localization of seizure foci: The flip side of mapping regions for surgeons NOT to remove for pre surgical mapping applications is the mapping  of seizure generating tissue to provide surgeons with a target for removal. For certain types of seizure activity, the brain is constantly generating uniquely unusual activity, which translates into unique temporal signatures recorded with either EEG or resting state fMRI. Detection with EEG is much more easily and cheaply performed, but has less spatial precision fMRI. 
  9. Clinical Importance of Basic Neuroscience: Many would argue that the clinical importance of basic and cognitive neuroscience research, while not having a direct clinical application, has so many secondary and tertiary influences on the state of the art of clinical practice that this is in itself a sufficient justification for continued fMRI research funding by both basic science funding agencies as well as more clinically focused agencies.
Success? How to measure it – and on what time scale?
Getting back to the issue of funding. From my  perspective, there are two primary issues: 1. How to achieve a balance of short term and long term success. 2. How to even gauge the effectiveness of a funding initiative or of a specific funded project.
Clinical funding agencies generally fund basic research with the idea that clinical implementation is a long term goal that requires basic science groundwork to be established. If funding were only short term, many discoveries and new fruitful directions and opportunities would be missed. About 30 years ago, several notable large companies supported research of select employee/scientists that was more open ended. Examples are Varian (my Ph.D. co-advisor, Jim Hyde emerged from this renowned group) and famously, Bell Labs who allowed one of their scientists by the name of Seiji Ogawa to dabble in high field MRI – using hemoglobin as a potential contrast. Back then, companies seemed to have more latitude for open ended creative work but the culture seems to have shifted (with perhaps the exception of Google and the like). Today, MRI research by vendor employees has become more product focused and usually on short term problems. While this is an effective approach in many contexts, in my opinion, much of the creative potential of these employee scientists is lost on product development and troubleshooting.
Regarding the second issue of measures of success, this is an open problem that I believe vexes funding agencies and program officers around the word. Measures such as papers published or citations don’t really capture the essence of a successful new research direction. One has to gauge the entire field to determine the success of a new method. One also has to wait potentially decades to determine the true payoff. To the best of my knowledge, there are no clear objective or quantitative measures of funding success. Those deciding on the funding typically base their decisions on their own broad and deep knowledge of the field and advice from experts doing the research. Grant reviewers assess the quality of the proposals but the directors and program officers set the initiatives. It would be interesting and useful to develop more of a science for what general directions and what grants would be best to fund, looking back on what was funded and coming up with measures that can effectively predict “success.” This task might be a problem for the machine learning community.
What will it take for fMRI to be a clinical method? 
What it will take for fMRI to become a sought-after clinical method? To begin, a foundation of streamlined clinical testing needs to be established. At minimum, this will require a highly streamlined, patient/clinician – friendly protocol that collects fMRI data in real time (to allow for immediate identification of unacceptable motion, etc. so that the scans can be quickly cancelled and redone), an agreed upon processing pipeline that then collapses the salient information into a map or even a set of numbers that are both meaningful and easily understood by those making clinical decisions. Functional MRI subject interface devices need minimal setup time, and the protocol itself should take no longer than any other structural scan that is performed. Currently, no such highly integrated systems exist. With increased focus on better extraction and differentiation of individual information, clinical implementation will be a natural next step. I believe we just have to wait a bit, and no one really has a solid sense of whether or not fMRI will successfully penetrate clinical practice, but there’s a few things that can be done. 
Regarding utility and reliability, I think that currently, with our hardware, acquisition methods, noise reduction approaches, and other post processing methods, fMRI is not quite reliable or sensitive enough. One example is of how physiologic noise reduction can immensely improve the state of the art. Currently, physiologic noise sets an upper limit of about 120/1 on fMRI time series, no matter what the coil sensitivity or field strength is. If we were able to remove this physiologic noise, then the time series signal to noise ratio would be limited only by coil sensitivity – potentially increasing the time series signal to noise ratio by an order of magnitude. 
There are the large obstacles of cost effectiveness and clinical uniqueness. The cost/benefit has to place it above the competing clinical methods. Given the current rate that the field is making progress on individual assessment methods, my sense is that it will become reliable enough for a small but growing number of clinical applications. Which ones and when, I don’t think anyone knows, but I think at least one of the applications that I mentioned above will emerge within the next decade. Specifically, it appears that applications 5,6, and 7 which use fMRI to map physiology rather than function, and application 3 which is the use of fMRI activation as a functional localizer for neuromodulation, have the highest likelihood for clinical penetration. Approach 7, that of mapping resting state latencies and using these maps for perfusion deficit assessment, has the necessary ingredients for success: similar ease of implementation, sensitivity, and specificity to current approaches, and an added benefit of being less invasive than current clinical practice involving Gd injection. 
Funding the vendors
A ripe target for funding might be to the major scanner vendors or small businesses to create such a clinically viable platform that would be able to immediately implement and test the most promising basic science findings. At the moment, I feel that vendors are not devoting enough man hours to any major fMRI platform development, as there are no clearly profitable applications that exist in the short term. Catalyzing development along these lines by grants would enable more rapid clinical implementation and testing. As mentioned, once a clear clinical application is established, more vendor-funded fMRI development would then allocated by the vendors as it would translate into profit.
Other Suggestions
In the twitter conversation there were a few suggestions that emerged. One that is generally practiced but perhaps should be emphasized further, is, for those applying for grants from agencies where the mission is human health, more detail regarding how their research will lead to better clinical practice should be included. What are the steps needed? What clinical practice will be improved and how? What might be the timeline? I think that this approach should apply to a large fraction of these grant applications but for many, I don’t think that this should be a requirement as its generally accepted that the fallout of better understanding brain organization in health and disease can inform unexpected new avenues of clinical practice. One cannot and sometimes should not always connect the dots. There is a significant role of basic research – without an obvious or immediate clinical application – that is still beneficial to clinical practice in the long run. 
Fund more tool development, implementation, and streamlining. One gap that I see in some of the funding opportunities is that of taking a potentially useful tool and making it work in regular clinical practice. This could be either before or after the clinical trials stage. I think that funding more nuts and bolts research and development – scaling up a tool from concept to general practice – should have a larger role as often this gap is prohibitively wide.
Fund infrastructure creation for data, tool, and model sharing and testing. In recent years, the creation of large, curated, mine-able databases has shown to be effective in accelerating, among other things, methods development research and discovery science as well as transparency and reproducibility. One can imagine other useful infrastructures created for computational model sharing, cross modality data pooling, tool testing and development, and generally integrating the vast disconnected body of scientific literature in neuroscience. As a concrete example, I’m often struck with how disconnected the information is at a typical Society for Neuroscience Meeting. Attendees are quickly overwhelmed with the information. If there was some structure, perhaps organized by high priority open questions or models that need to be tested, that the diverse findings could be linked with, this would go a long way towards increasing the focus of the community, identifying research opportunities, and pointing out clear gaps in our understanding. 
Funding for fMRI is well worth it. 
My response to those who feel that fMRI funding should be cut is to of course welcome them to provide viable alternatives. Perhaps there are new directions out there that need more focus. I think that most in the field of neuroimaging – as well as those outside – would agree however that fMRI has not only established its place as a formidable tool in neuroscience and clinically directed research, it is a technique that has revolutionized much of cognitive neuroscience. It’s also clear that we are currently in the midst of a wave of innovation in everything from pulse sequence design to multi-modal integration to processing methods. The field is advancing surprisingly well. It is making a growing number of clear contributions to neuroscience research and will eventually make inroads, one way or another, into clinical practice.