Serving the fNIRS community
Every limit of fNIRS that we have discussed in this course is also a target for someone working to overcome it. Coarse spatial resolution motivated high-density imaging; the inability of continuous-wave systems to measure absolute values motivated frequency- and time-domain instruments; the ambiguity of oxygenation motivated measurements of blood flow and of cellular metabolism; and the slowness and noise of the hemodynamic signal have shaped how it is used in real time. In this lesson, we look at where these efforts stand today and what they still have to solve.
The field moves quickly, so this lesson does not try to list every new system. Instead, for each frontier, we ask which limit it addresses, what has been demonstrated, and what remains to be done. Several of these technologies were introduced earlier in the course, and here we connect them to the applications of this unit. The last section asks the opposite question: which limits no technology is likely to remove, because they come from the physics of light in tissue.
Table 1 maps the main technological frontiers to the limits they address. It is a useful habit to read any new technology this way: what problem does it solve, what does it cost, and what problem does it leave untouched? A technology that improves resolution does nothing for scalp contamination unless it is designed to, and a classifier with high accuracy says nothing about where its information comes from unless that is tested.
| Frontier | Limit it addresses | What remains |
|---|---|---|
| Wearable HD-DOT | The trade-off between resolution and wearability (Lessons 2.4 and 5.4) | Whole-head coverage, weight, power, heat, coupling through hair, and the need for head models |
| Compact time-domain systems | Absolute values and depth discrimination outside the laboratory (Lesson 4.3) | Cost and number of channels, calibration, and common performance standards |
| Speckle-based blood flow (SCOS, multi-speckle DCS) | The photon budget of DCS, which limits flow measurements to few channels and short separations (Lessons 4.4 and 6.3) | Correction for detector noise, validation in patients, comparison between people, and wearable formats |
| Broadband spectroscopy | Oxygen use inside the cell, through cytochrome c oxidase (Lesson 1.2) | A small signal, cross-talk with hemoglobin, and few available systems |
| Multimodal systems | Timing, anatomy, physiology, and behavior that fNIRS alone does not provide | Integration on the head, synchronization, and methods for joint analysis |
| Brain-computer interfaces and machine learning | Real-time use and decisions about individuals | Speed, confounds, small samples, and generalization to new sessions and people |
Table 1. Technological frontiers of fNIRS, the limits they address, and what they still have to solve.
In Lesson 2.4, we saw that high-density diffuse optical tomography (HD-DOT), with many overlapping channels at several distances, can approach the localization accuracy of fMRI for the cortical surface. For most of its history, however, HD-DOT required hundreds of optical fibers connecting the cap to a large instrument, which made the cap heavy, limited the participant’s movement, and gave up the very advantages that make fNIRS attractive. fNIRS users had to choose between resolution and wearability. The modular, fiberless designs introduced in Lesson 4.1, in which small tiles carrying their own sources, detectors, and electronics are placed directly on the head, are removing that choice.
Vidal-Rosas et al. (2021) tested such a wearable system against the classic visual mapping experiments that fiber-based HD-DOT had established, recording one participant over 15 sessions. The wearable system reproduced the focal activations and response shapes of the larger systems, with high repeatability across sessions, and a simulated sparse array recorded from the same data showed how much localization and resolution would be lost without the high density. Frijia et al. (2021) then performed the first study of the infant brain with wearable HD-DOT, using a social stimulus paradigm established with sparse fNIRS. The images agreed with the earlier sparse results but showed better localization, better depth specificity, higher signal-to-noise ratio, and much more consistent responses across infants, and the system was well tolerated. Zhao et al. (2021) designed an ultralight, flexible version for newborns, weighing about 70 grams, aimed at repeated imaging at the cot side.
The remaining challenges are practical and computational. Covering the whole head with high-density tiles multiplies the weight, the power consumption, and the heat produced on the scalp, and each tile must still make good contact through hair. The number of channels, and with it the volume of data, grows quickly with the number of optodes, and turning these data into images still requires the forward models, head models, and regularization choices of Lesson 5.4. The question raised in that lesson, whether reconstruction is worth it, has a clearer answer for these systems than for sparse probes: with dense, overlapping measurements, image reconstruction is no longer optional but the natural way to use the data.
Time-domain (TD) systems measure the time of flight of photons and, as we saw in Lesson 4.3, can recover absolute optical properties and separate early photons, which mostly sample superficial tissue, from late photons, which reach deeper. For decades, these advantages came with large, expensive, and delicate laboratory instruments built around pulsed lasers and photon-counting electronics. Pifferi et al. (2016) reviewed how new photonic components are changing this: compact pulsed lasers, single-photon avalanche diodes and silicon photomultipliers as detectors, fast gating circuits, and time-to-digital converters integrated on chips. These components point in two directions at once: toward compact and wearable TD devices, such as the backpack system described in Lesson 4.3, and toward powerful systems with greater depth sensitivity than was previously possible.
If TD systems become as practical as CW systems, several problems discussed in this course would become easier: absolute values that can be compared between people and sessions (Lessons 4.3 and 6.2), depth discrimination built into each measurement rather than added through short channels (Lesson 3.5), and clinical oximetry that separates the brain from the scalp (Lesson 6.3). What remains is cost per channel, the stability and calibration of the instrument’s timing response, and agreed protocols for testing performance, so that values from different systems can be compared.
In Lesson 4.4, we saw that diffuse correlation spectroscopy (DCS) measures blood flow, and that its main weakness is the number of photons: each detector sees a single speckle, so DCS systems have few channels, short separations, and noisy signals. We also saw the new approaches that attack this limit, and they are now among the most active frontiers of the field. Speckle contrast methods, such as speckle contrast optical spectroscopy (SCOS), record many speckles at once with a camera and infer flow from how much the pattern blurs during the exposure. Multi-speckle DCS does the same with arrays of single-photon avalanche diodes (SPADs), each pixel computing its own correlation. Interferometric DCS amplifies the weak signal from the tissue with a reference beam, and longer wavelengths increase the number of photons that can be safely used. Lesson 4.4 describes how each of these works; here we ask what they make possible.
The most important change is one of scale. With signal-to-noise ratios one or more orders of magnitude higher than conventional DCS, blood flow can be measured at the source-detector separations of fNIRS, where the brain contributes more of the signal, and with many more channels. In Lesson 4.4, we saw that a speckle contrast system detected brain activation at a separation of 33 mm with more than ten times the signal-to-noise ratio of DCS. If this progress continues, flow could become a routine companion of oxygenation rather than a specialized measurement, and with both, the oxygen metabolism of Lesson 4.4 could be mapped across the cortex during a task, in the same way that fNIRS now maps oxygenation. At the bedside, cheaper detectors and longer separations would address several of the limits of DCS discussed in Lesson 6.3, bringing continuous monitoring of cerebral blood flow and autoregulation closer to clinical use.
These developments also connect to the other frontiers of this lesson. SPAD arrays are the same family of detectors that are making time-domain systems compact (Section 3), and time-domain DCS combines the two ideas, selecting late photons to make flow measurements more sensitive to depth. The technology is still young. Speckle contrast measurements must be carefully corrected for the noise of the camera itself, which can mimic the blurring caused by flow. Like DCS, these methods report a blood flow index that is easier to compare within a person than between people, and the scalp still contributes to the signal. Most demonstrations so far come from a small number of laboratories, in healthy adults, with laboratory systems. Validation against established flow measurements, in patients, and in wearable formats, is the work that remains.
Hemoglobin tells us about the delivery of oxygen, not about its use. In Lesson 1.2, we met cytochrome c oxidase (CCO), the last enzyme of the mitochondrial respiratory chain, whose oxidation state changes with the cell’s use of oxygen and whose absorption in the near infrared was part of Jöbsis’s original motivation. Measuring CCO is difficult because its concentration in tissue is much lower than that of hemoglobin, so its signal is small and easily contaminated by cross-talk from the much larger hemoglobin changes. Broadband systems, which measure attenuation at many wavelengths across the near infrared, provide enough spectral information to separate it (Lesson 2.3).
In healthy adults, Kolyva et al. (2012) measured changes in oxidized CCO during a frontal problem-solving task with a hybrid broadband and frequency-domain system, and systematically tested possible confounds, such as the wavelength range used in the fit, the path length, systemic changes, and changes in scattering. None explained the CCO response, which was heterogeneous: most participants showed increased oxidation, but some showed a decrease. In newborns with hypoxic-ischemic brain injury, Bale et al. (2019) monitored 50 infants during therapeutic hypothermia and found that a strong coupling between CCO and brain oxygenation during spontaneous drops in oxygen saturation was associated with an unfavorable outcome, as predicted by magnetic resonance spectroscopy, with 64% sensitivity and 79% specificity. These are promising first steps toward a bedside measure of metabolism at the level of the mitochondria. The technique is still limited by the small size of the signal, by the dependence of path length on wavelength, and by the small number of broadband systems in use.
No single technique measures everything, and fNIRS is unusually easy to combine with others, because optical measurements interfere little with electrical recordings and, with optical fibers or suitable electronics, can be made inside an MRI scanner. Each partner fills a specific gap. EEG adds millisecond timing of the neural activity that the hemodynamic response follows (Lesson 3.4). fMRI adds anatomy and depth, and simultaneous recordings were essential for validating fNIRS (Lesson 3.4). DCS adds blood flow and, with it, oxygen metabolism (Lessons 4.4 and 6.3). Physiological recordings, accelerometers, eye tracking, and video add the information needed to model systemic signals, motion, and natural behavior (Lessons 5.3 and 6.1). fNIRS can also be combined with transcranial stimulation, to measure the brain’s response to the stimulation itself.
Combining modalities is most convincing when each contributes information the other cannot. Fazli et al. (2012) recorded EEG and fNIRS simultaneously in a real-time brain-computer interface based on imagined movements and found that adding fNIRS improved the classification accuracy of EEG in more than 90% of participants, by 5% on average, because the two signals carried complementary information. They also noted that the delay of the hemodynamic response limits how fast such a system can operate. The challenges of multimodal systems are partly physical (fitting optodes and electrodes on the same scalp, often in the same positions) and partly analytical, since signals with different time scales, spatial sensitivities, and noise must be synchronized (Lesson 4.1) and modeled together.
A brain-computer interface (BCI) translates brain signals into commands, without the need for muscles. Its most important users are people who cannot move or speak, such as patients with advanced amyotrophic lateral sclerosis or brainstem stroke. fNIRS was proposed as a BCI signal because it is safe, easy to set up, and tolerant of the environment. Coyle et al. (2004) proposed an optical BCI based on hemodynamic responses to imagined movements, and Coyle et al. (2007) built a simple system that allowed a user to make an on/off selection, with each selection taking about a minute. Sitaram et al. (2007) recorded 20 channels over the motor cortex while healthy volunteers imagined moving the left or right hand, and showed that the two could be classified offline with accuracies of 73% to 89%, depending on the algorithm. The fundamental limit of these systems is speed: because the hemodynamic response takes seconds to develop, an fNIRS BCI makes decisions far more slowly than an EEG BCI. This is why hybrid systems, neurofeedback (Lesson 6.3), and passive BCIs, which monitor states such as mental workload rather than issuing commands (Lesson 6.1), are where fNIRS has been most useful.
The same classification methods, often called machine learning, are now used much more widely: to decode cognitive states, to predict performance, and to classify patients and controls, as we saw in Lesson 6.3. These methods are powerful, and they are also very good at finding whatever separates the classes in the data, whether or not it comes from the brain. If one condition makes participants breathe differently, raise their blood pressure, move more, or flush, a classifier will use that difference, and its accuracy will be high for the wrong reason. von Lühmann et al. (2020) showed that the analysis principles of Lesson 5.3 help here too: estimating single-trial responses with a GLM that includes short-separation and other physiological regressors gave features that better reflected brain activity and improved classification accuracy by 7.4% on average, using data with a known ground truth. Figure 1 shows both problems in simulated data: a classifier that owes most of its accuracy to the body, and a test that overestimates accuracy because neighboring trials share the same slow drift.
Figure 1. What a classifier learns (simulated data). (a) When a task changes breathing and scalp blood flow, a classifier trained on uncorrected channels reaches 81% accuracy, almost as much as one trained on short channels that see no brain at all (77%); after short-channel regression, the brain-based accuracy is 61%. Tested on participants left out of training; error bars show the spread across 30 simulated datasets. (b) With slow drifts and block designs, splitting trials at random puts neighboring trials, which share the same drift, in both training and test sets (77%); testing on a new session gives the honest estimate (56%). Error bars show the spread across participants.
💭 Pause and Think
A study reports that a classifier distinguishes mental arithmetic from rest with 92% accuracy, using HbO₂ from 8 prefrontal channels, in 12 participants. The data were filtered and normalized using the whole recording, and the classifier was tested with cross-validation over trials pooled from all participants. What would you want to check before believing that the classifier decodes brain activity, and that it would work on a new person?
Think about breathing, heart rate, and scalp blood flow during mental effort (Lesson 3.5), whether short channels and HbR were used, whether information from the test trials leaked into the training through the normalization or the cross-validation, and whether testing on participants left out of training would give the same accuracy.
⚠ A Common Mistake to Avoid
It is tempting to read a high classification accuracy as evidence that the brain signal contains the information being decoded. Accuracy measures how well the classes were separated, not where the separating information came from, and it can be inflated by systemic physiology, motion, and analysis choices that let information from the test data reach the training data.
Keep test data completely separate from every step of training, including filtering choices, normalization, and feature selection. Test on new sessions and new participants, not only on new trials. Remove systemic components with short channels and auxiliary measurements, compare the classifier with one trained on those auxiliary signals alone, and report HbR as well as HbO₂. If a classifier trained on breathing performs as well as one trained on the brain, you have learned something important about your data.
🎍 MUST-READ PAPER
Ayaz, H., Baker, W.B., Blaney, G., Boas, D.A., et al. (2022). “Optical imaging and spectroscopy for the study of the human brain: status report.” Neurophotonics, 9(Suppl 2), S24001.
A status report written by a large group of researchers across the field, covering the optical technologies and software for noninvasive studies of the human brain, their applications in neuroscience and the clinic, the areas where innovation is most needed, and the outlook for the coming years. The best single entry point to everything in this unit, and a map of where you might contribute.
It is worth ending a lesson about new technology with the limits that new technology is unlikely to remove. Light in tissue diffuses, and the number of photons that return from a given depth falls steeply with that depth (Lessons 2.1 and 2.5). Better detectors, more channels, and time-of-flight information can push the measurement somewhat deeper and sharpen it, but they will not make the adult hippocampus or amygdala accessible to light delivered at the scalp. Every photon that reaches the brain from the scalp must cross the scalp twice, so extracerebral signals can be reduced and modeled but never avoided. And the hemodynamic response will always follow neural activity by seconds, because that delay is set by physiology, not by the instrument.
What technology can change is how well we measure within these limits: with more coverage, finer resolution, absolute values, flow and metabolism alongside oxygenation, and fewer constraints on where and with whom we measure. Using these advances well will still require the reasoning of this course: knowing what the signal is, where it comes from, and what else could explain it. The last lesson turns from instruments to the field itself, and to the open problems that are not technological at all.
📌 Key Takeaways
📚 Further Reading & Key References
Society for Functional Near-Infrared Spectroscopy.
Content is for educational purposes.