nueva página del texto (beta)
Inglés (pdf)
Artículo en XML
Referencias del artículo
Traducción automática
Enviar artículo por email
Citado por SciELO 
Similares en
SciELO 
Permalink
Introduction
Optical imaging modalities have significantly advanced over the past decade, becoming valuable tools for non-invasive human brain mapping. Techniques such as functional near-infrared spectroscopy (fNIRS), diffuse optical tomography (DOT), diffuse correlation spectroscopy (DCS), hyperspectral imaging (HSI), and photoacoustic imaging (PAI) measure neural activity and hemodynamics through light-tissue interactions. These methods are portable, safe, and cost-effective compared to magnetic resonance imaging (MRI) or positron emission tomography (PET), allowing bedside or naturalistic use, especially suitable for infants or active subjects1. fNIRS and DOT use near-infrared (NIR) light (650-900 nm) to detect oxyhemoglobin (HbO) and deoxyhemoglobin (HbR) changes linked to neural activation1. DCS estimates real-time cerebral blood flow (CBF) by analyzing fluctuations of coherent NIR light2. HSI differentiates cortical tissues based on spectral reflectance, aiding tumor identification and assessing tissue oxygenation3,4. PAI combines optical contrast with ultrasonic detection, achieving deeper tissue penetration and high-resolution imaging despite skull-induced acoustic challenges5. From 2015 to 2025, these modalities have increasingly contributed to cognitive neuroscience, neurodevelopmental studies, and clinical neurology. A representative schematic of the modalities covered in this work is depicted in figure 1A. This review summarizes recent advancements, methodological trends, limitations, and future directions in optical brain mapping, following Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) 2020 guidelines6.
Methodology
Search strategy
We performed a systematic literature search to identify relevant studies from January 2015 up to March 1, 2025. The search encompassed multiple databases: PubMed/MEDLINE, Web of Science, Scopus, and IEEE Xplore, to capture both biomedical and engineering literature. We used Boolean search queries combining terms for optical imaging modalities and brain mapping. For example, the PubMed query is illustrated in figure 1B:
For reproducibility, the full Boolean search strings for each database are provided in table 1 of the Supplementary Material. We also hand-searched reference lists of relevant papers and prior reviews to ensure inclusion of any studies missed by the database search. All results were imported into a reference manager, and duplicates were removed7. This review was conducted following the PRISMA 2020 reporting guidelines6, aiming for a transparent and reproducible selection process.
Table 1 Summary of 118 optical brain-imaging studies published 2015-2025, grouped by modality, typical human populations investigated, and representative key findings
| Modality | Studies (n) 2015-2025 | Typical human populations | Representative key findings |
|---|---|---|---|
| fNIRS | 60 | Healthy adults, infants, preschoolers, stroke, TBI, neuro-psychiatric cohorts | Portable systems enable naturalistic tasks, hyperscanning, and neuro-feedback; wearable caps combined with VR for social-development research |
| DOT | 20 | Adults, toddlers (1-7 year), epilepsy, peri-operative language mapping | High-density arrays provide fMRI-comparable cortical maps; feasible in awake children watching movies |
| DCS | 15 | Acute ischemic stroke, neuro-ICU, cardiac arrest survivors, healthy volunteers | Real-time bedside CBF monitoring; detects reperfusion after thrombolysis; reviews highlight no RCTs to date |
| HIS | 13 | Intra-operative brain-tumor patients (glioma, metastasis) | Spectral classifiers delineate tumor margins; first multi-site benchmark dataset released |
| PAI | 10 | Adults with hemicraniectomy, neonates (fontanelle), intra-operative vascular mapping | First transcranial functional PAI in a human; reviews outline roadmap and skull-mitigation strategies |
fNIRS: functional near-infrared spectroscopy; DOT: diffuse optical tomography; DCS: diffuse correlation spectroscopy; HIS: hyperspectral imaging; PAI: photoacoustic imaging, CBF: cerebral blood flow, TBI: traumatic brain injury; ICU: intensive care unit; RCTs: randomized controlled trials.
Prescreening and eligibility criteria
We included primary research articles meeting the following criteria: (1) published in English in peer-reviewed journals between 2015 and 2025; (2) involved in vivo human participants (either healthy individuals or patients) undergoing brain mapping with one or more of the optical modalities of interest (fNIRS, DOT, DCS, HSI, PAI); (3) reported functional brain data (e.g. neural activation, hemodynamic changes) or diagnostic imaging relevant to brain structure/function. We excluded animal studies, cadaver or phantom studies, technical papers without in vivo human data, editorials/commentaries, and conference abstracts.
The selection process proceeded in two stages. First, one reviewer screened titles and abstracts of all retrieved records for relevance to human optical brain imaging. Obviously, irrelevant records were excluded at this stage. Next, full-text of potentially eligible articles was obtained and assessed against the inclusion criteria. Studies that did not explicitly report results from human brains (e.g., purely simulation or algorithm development papers) were excluded. In total, the search identified n = 1,247 records after duplicate removal, of which n = 220 articles underwent full-text screening. Finally, n = 118 studies were deemed eligible and included in this review. A PRISMA 2020 flow diagram (Fig. 2) illustrates the study identification and selection process.
Risk of bias assessment
A formal risk of bias assessment was performed on the included studies to gauge the reliability of their findings. Given that many studies evaluated imaging modalities against clinical or neuroimaging benchmarks (diagnostic-type studies) or reported observational results, we employed the quality assessment of diagnostic accuracy studies-2 (QUADAS-2) tool for quality assessment8. Each study was assessed in four domains: patient selection, index test, reference standard, and flow and timing. Signaling questions for each domain were answered based on information in the papers, and each domain was rated as "Low," "High," or "Unclear" risk of bias per guidelines8. A single reviewer (the first author) performed the risk of bias assessment for all studies. We acknowledge that having a single reviewer conduct the study screening and bias assessments is a methodological limitation; ideally, two or more independent reviewers would carry out these steps to strengthen rigor and reduce potential bias. Nevertheless, the reviewer adhered to the standardized QUADAS-2 protocol and consulted with co-authors regarding any uncertainties to mitigate subjectivity.
Results
A total of 118 studies met our inclusion criteria, encompassing applications of fNIRS, DOT, DCS, HSI, and PAI in human brain mapping from 2015 to 2025. We organized the findings by modality, highlighting key technical developments, typical applications, and overarching trends for each. These 118 primary studies form the basis of the results summarized by technique as shown in table 1.
Across optical brain-imaging methods, fNIRS and high-density DOT provide centimeter-deep cortical coverage with spatial resolutions of roughly 5-30 mm and the highest portability, while DCS offers similar depth but adds fast blood-flow metrics at moderate portability; HSI delivers sub-millimeter surface detail yet is limited to exposed cortex and remains less portable, and PAI uniquely penetrates up to several centimeters with sub-millimeter resolution but still faces skull-related constraints and lower portability. These results are summarized in figure 3.
Figure 3 Comparison of the reviewed optical imaging modalities according to their temporal and spatial resolution, penetration depth, and portability.
fNIRS
fNIRS has become a widely used modality in human neuroscience over the past decade, reflecting its maturation into a robust functional imaging tool1,9. fNIRS uses sources and detectors placed on the scalp to measure changes in NIR light absorption associated with CBF and oxygenation changes in cortical regions1. From 2015 to 2025, there was an explosion of fNIRS research in diverse fields, including cognitive psychology, social neuroscience, developmental neuroscience, and clinical neurology9-12. This surge is evidenced by a large number of publications and review articles, which "testifies to the maturity achieved by this non-invasive optical, vascular-based functional neuroimaging technique"1.
One important trend has been the deployment of fNIRS in naturalistic and mobile settings. Whereas early fNIRS studies (pre-2015) often occurred in laboratory environments with participants seated still, newer studies have leveraged advances in portable and wireless fNIRS devices. By around 2018, truly wearable fNIRS systems enabled recording brain activity in freely moving subjects during realistic tasks13,14. For example, researchers have used fNIRS to monitor prefrontal cortex activity while people navigate real-world environments or engage in social interactions, which is infeasible with functional magnetic resonance imaging (fMRI). Pioneering studies even conducted hyperscanning (simultaneous fNIRS on multiple people) to study interactive brain dynamics (e.g., teacher-student or parent-child interactions). These developments were facilitated by compact wireless emitters/detectors and Bluetooth data transmission, as well as improved signal processing to filter out motion and ambient light artifacts. Today, fNIRS is utilized on healthy subjects moving freely in different naturalistic settings, an achievement anticipated to grow with further instrument development1.
In parallel with mobility, the spatial coverage and density of fNIRS arrays have increased. Traditional fNIRS used sparse optode arrangements, yielding only rough topographical maps. Recent high-density configurations place many optodes in a grid, with overlapping measurements at multiple source-detector distances. This approach, often termed high-density DOT, when paired with image reconstruction, dramatically improves image quality over sparse fNIRS15. For instance, Tripathy et al. built a high-performance high-density DOT system with an array of sources and detectors specifically designed to be child-friendly15. They demonstrated that such high-density fNIRS can produce image quality benchmarked against fMRI - mapping functional responses in 1-7 year-old children watching movies, a scenario where fMRI is challenging due to motion and compliance issues15. The ability to obtain spatially detailed maps of cortical activation (e.g., distinguishing functional responses in Brocas area vs. adjacent cortex during speech stimuli) has garnered significant interest in both cognitive and clinical research16. Indeed, high-density fNIRS/DOT approaches have achieved enough spatial resolution to resolve localized activation comparable to fMRI findings, while retaining fNIRSs logistical advantages (silent, portable, tolerant to motion)15,16.
Applications of fNIRS in healthy populations have expanded to a wide array of cognitive functions - from language processing and auditory perception to decision-making, attention, and motor control. In clinical populations, fNIRS has been explored for bedside functional monitoring and neurorehabilitation progress. For example, multiple studies used fNIRS to assess motor cortex activation in stroke patients during rehabilitation tasks, finding that fNIRS activation magnitudes correlate with motor recovery stages (references in included studies). fNIRS has also been applied to monitor frontal lobe activity in disorders of consciousness, to map language or memory function in epilepsy patients, and even to assist brain-computer interface (BCI) development for patients with paralysis17. The data suggest that fNIRS signals (changes in HbO and HbR concentrations) can serve as biomarkers for brain function in situations where MRI is impractical or unsafe (e.g., early after stroke, in the intensive care unit [ICU], or in patients with motor deficits12,18,19). Furthermore, interactive neuroscience paradigms have emerged, such as neurofeedback training using real-time fNIRS signals and studies of how physical exercise modulates cortical activation1.
Despite these advances, limitations of fNIRS remain. The technique is fundamentally limited to sensing cortical surface activity (typically ≤ 1.5 cm deep) because NIR light is strongly attenuated in tissue9-11,20-22. Signals are also contaminated by superficial blood flow in the scalp/skull (systemic physiology, skin blood vessels), which can confound true cerebral signals. A pressing need for standardized protocols to remove or account for these extracranial signal components has been identified1,22. Short-separation detectors (to measure and subtract scalp-only signals) and improved algorithms (e.g., regression of global trends, machine learning classifiers) are being incorporated to address this. In addition, because many fNIRS setups are custom, results can vary across labs. Efforts like near-infrared spectroscopy (NIRS)-brain imaging data structure (BIDS) (a BIDS extension) have begun standardizing data formats to improve reproducibility23,24. Before fNIRS can become a clinical neurodiagnostic tool, consensus on systemic interference correction, probe placements, and quantification methods is needed1,9. Nonetheless, the trend over 2015-2025 shows fNIRS evolving from a niche laboratory technique to a mainstream neuroimaging modality with growing real-world and clinical applications, thanks to its non-invasiveness and flexibility.
DOT
DOT extends fNIRS using high-density source-detector arrays and image-reconstruction algorithms to generate 3D maps of cortical HbO and HbR. Between 2015 and 2025, it matured into a tool that can approach fMRI-level localization when optode spacing is ≤ 1 cm and measurements overlap densely, driving widespread adoption in cognitive and clinical research15,25. Continuous-wave (CW) DOT dominates because the hardware is inexpensive and portable; a 2024 systematic review tallied 83 CW -DOT human studies, underscoring rapid growth16. High-density systems have mapped language in toddlers, neonatal strokes, and peri-operative cortex, tasks previously impractical for sparse optics or fMRI26.
Algorithmic advances, regularized inverse models, MRI-based head priors, and multi-wavelength fitting now resolve activity in individual gyri and yield quantitative hemoglobin changes; DOT activation patterns routinely overlap with fMRI benchmarks25,27. Pilot clinical work shows promise for stroke, epilepsy, and intraoperative mapping, where neuronavigation-guided DOT offers an optical analog to fMRI16.
Challenges remain
Image quality depends on head-model assumptions, depth sensitivity is predominantly cortical, and high-density caps can be cumbersome. Standardized protocols and deeper-penetrating time-domain or frequency-domain systems are needed to improve quantification and broaden adoption16,28. Nevertheless, DOT as of 2025 remains primarily a research tool. No DOT system is yet widely approved for clinical use, and further validation in multicenter studies is needed to establish its accuracy and reliability compared to gold standards.
DCS
DCS probes CBF by tracking speckle-intensity fluctuations produced by moving red blood cells under coherent NIR illumination, yielding a continuous index of microvascular perfusion that complements fNIRS/DOT oxygenation data2,29.
Bedside utility is illustrated by Delgado-Mederos et al., who combined DCS and NIRS in five acute-stroke patients and detected reperfusion within 2.5 h of thrombolysis, ahead of repeat imaging30. Subsequent clinical reports extend monitoring to hypothermic circulatory arrest, traumatic brain injury, and post-cardiac-arrest care, where continuous CBF trends inform management when ultrasound or MRI are impractical31,32.
Technical advances from 2015 to 2025 focus on signal-to-noise, depth, and coverage. Multi-speckle detection and parallelized cameras reduce variance and expand the field-of-view2. Longer wavelengths (1064 nm) improve penetration, while time-gated and time-domain DCS isolate late-arriving photons for deeper cortical sensitivity33. Hybrid systems integrate DCS with NIRS to estimate cerebral metabolic rate of oxygen or with DOT to create emerging 3D flow tomograms29.
Limitations persist
Probes must remain immobile; absolute calibration requires external standards; adult measurements are confined to superficial cortex; and spatial localization is coarse unless high-density grids or MRI co-registration are used2,29. Nonetheless, progressive hardware miniaturization, depth-discrimination schemes, and early commercial devices position DCS as a valuable, real-time perfusion monitor poised for wider neurocritical and functional applications33.
HSI
HSI records a complete reflectance spectrum for every pixel, yielding rich optical fingerprints that differentiate brain tissues without dyes. A systematic survey of clinical literature tallied > 25 operative HSI systems and 45 human studies, confirming rapid growth in neurosurgical use (though typically sample sizes were small)34,35. The foremost application is intraoperative tumor delineation: the HELICoiD consortium built an open VNIR database of 36 cubes from 22 glioma patients (> 300,000 labelled spectra) and showed that spatial-spectral classifiers could highlight tumor margins with ~ 80% accuracy in real time, guiding resection decisions3. A subsequent multicenter benchmark enlarged the dataset to 61 images from 34 patients and reported a median macro-F1 of 70% for tumor detection, establishing a common reference for algorithm development36.
HSI also exploits hemoglobin absorbance to map cortical oxygenation and vascular anatomy during aneurysm repair, bypass grafting, or decompressive craniectomy, offering wide-field perfusion assessment that complements point probes35. Early reports indicate that task-evoked spectral changes can localize functional cortex in awake craniotomy, suggesting a contact-free adjunct to electrical stimulation. Hardware advances have catalyzed translation: snapshot cameras using birefringent spectral demultiplexers now capture ≥ 64 spectral bands in a single exposure, delivering video-rate cubes compatible with the pulsating brain37. Compact modules mount on operative microscopes or exoscopes, allowing surgeons to toggle between white-light and hyperspectral views without workflow disruption35. Concurrently, deep-learning pipelines fed by public datasets generate color-coded overlays of tumor probability directly in the eyepiece, standardizing spectral interpretation and accelerating decision-making3,4,36.
Despite these advances, HSI remains confined to the exposed cortex: reflected photons interrogate only the first few millimeters, leaving subsurface tumor or hemorrhage invisible. Image quality can be degraded by surface blood, irrigation fluid, or variable illumination, necessitating stringent calibration and spectral normalization35. Most clinical reports are single-center case series with modest cohorts; robust multicenter trials are still scarce, although the Leon benchmark represents a critical step toward standardization and external validation36. Active research goals include depth-resolved extensions (e.g., endoscopic or hybrid HSI-fluorescence systems) and harmonized acquisition protocols.
In sum, 2015-2025 saw HSI progress from proof-of-concept to a nascent operating-room adjunct that can delineate tumors, assess perfusion, and potentially map function in real time. While feasibility has been demonstrated in operative settings, HSI remains an experimental adjunct; no HSI system is yet cleared for routine neurosurgical use. Further, inter-patient variability in spectra means algorithms need robust calibration and possibly large training datasets to be reliable. As of 2025, HSI shows great promise for intraoperative guidance, but it will require larger trials (and possibly regulatory approval) to establish clinical utility.
PAI
PAI combines optical contrast with ultrasound detection to achieve deeper imaging than purely optical methods. In transcranial applications, PAI has been investigated in specialized scenarios - notably in patients with skull openings (e.g., hemicraniectomy or fontanelle in neonates) where acoustic coupling is feasible. About 10 studies from 2015 to 2025 applied PAI or photoacoustic tomography (PAT) to human brain imaging. One milestone was the first report of transcranial functional PAT in an adult human, achieved using a patient with a large skull defect to bypass skull attenuation38. Other studies imaged neonatal brains through the fontanelle, visualizing blood oxygenation changes during interventions. PAI has also been explored intraoperatively for vessel mapping or cerebral perfusion monitoring by applying an ultrasound detector over exposed brain tissue.
Technical innovations for PAI include using 1064 nm lasers (which penetrate bone somewhat better) and full-ring ultrasound detector arrays to capture signals from multiple angles5. Multiwavelength PAI has been employed in preclinical models to differentiate HbO versus HbR signals, yielding functional maps of oxygenation in the brain5. The primary challenge for human PAI is the skull: the bone greatly attenuates and distorts acoustic signals. Strategies such as skull thinning, cooling, or acoustic coupling media have been proposed to mitigate this, but remain experimental. Current PAI setups are also relatively bulky and require careful alignment of lasers and ultrasound sensors, limiting clinical practicality.
To date, PAI in humans is limited to feasibility demonstrations, and its clinical readiness is the most nascent among the modalities discussed. No large-scale human PAI studies exist yet, and safety considerations (ensuring laser exposure is within ANSI limits for skin/eye safety) must be managed as higher energies are considered for deeper imaging. Still, PAIs unique ability to provide high-resolution optical contrast at depth (by listening for photo-induced ultrasound) makes it a compelling future modality if technical barriers can be overcome. Reviews in the field have outlined a roadmap and remaining challenges for translating PAI into clinical neuroimaging39,40.
Risk of bias and study quality
The overall quality of evidence across the included optical brain imaging studies is moderate, as assessed by our QUADAS-2-based analysis. Most studies were early-phase explorations, often lacking control groups or reference standards, which introduces various biases. Common issues included non-random participant selection (e.g., enrolling only young healthy adults for feasibility, or selecting patients with favorable conditions such as thin skulls or superficial tumors), which may overestimate the success of the modality in the general population. Blinding was seldom reported; for instance, investigators conducting fNIRS or DOT data analysis often knew the task conditions, and surgeons knew the ground truth of tumor locations when evaluating HSI or PAI results, potentially biasing outcome assessment. Supplementary figure S1 provides an overview of the risk-of-bias ratings (low, high, unclear) across all included studies.
In studies where a reference like fMRI was used, some did not clearly state if the fMRI analysis was blinded to the optical results, raising the possibility of confirmation bias. Several patient studies (especially in stroke or ICU monitoring with NIRS/DCS) had no independent gold standard for the measured outcome, meaning they inferred clinical relevance indirectly (e.g., assuming that increased flow per DCS is good because the patient improved clinically, but without imaging confirmation). Flow and timing biases were present in longitudinal studies that had dropouts or where the optical measurement and reference test were not perfectly aligned in time. For example, in stroke monitoring, optical data were continuous, but the "ground truth" (such as an MRI scan) might have been done hours later, during which the patients condition could have changed.
Another notable aspect is that many studies were small (median sample size 15 in our included set). This raises concerns about statistical power and selective reporting. Positive findings (e.g., successful mapping of a function or detection of a pathology) are likely over-represented in published studies, while negative or inconclusive studies may be underreported - a form of publication bias. This review cannot fully account for that, but it should be considered when interpreting the generally positive tone of results reported for each modality.
Encouragingly, some studies did incorporate measures to reduce bias: for instance, a few DOT versus fMRI comparison studies had independent analysts process each modalitys data15,25; some HSI tumor studies used a third-party algorithm to classify spectra without knowing histology results4,36, and then compared objectively. However, these were exceptions. The risk-of-bias assessment suggests that while the field demonstrates clear feasibility of each technology, the level of evidence is still preliminary, often in the form of case series or unblinded comparisons. There were no randomized controlled trials16 and only a handful of multicenter studies2,3,39. Therefore, any quantitative performance metrics (tumor detection accuracy, stroke monitoring sensitivity, etc.,) should be considered provisional.
In summary, across modalities, there is a need for future studies with more rigorous designs: larger sample sizes, appropriate control conditions or reference standards, blinding of outcome assessments, and, if possible, multicenter collaboration to improve generalizability. The literature provides a strong proof-of-concept foundation, but the actual clinical utility and comparative performance of these optical techniques will require the next level of evidence.
Discussion and future directions
Over the last decade, optical imaging methods for human brain mapping have progressed from niche experimental techniques toward more mainstream neuroimaging and clinical research tools. Functional NIRS and DOT have seen the most widespread use, moving into real-world applications and complex populations (like infants and those who cannot undergo fMRI) while improving image quality. These modalities have clearly demonstrated feasibility for human brain mapping in research and even some translational clinical studies; however, they remain primarily research tools at present and have not yet achieved routine clinical adoption. DCS has added the capability to measure CBF, addressing a gap left by hemodynamics-only tools and opening new avenues in bedside monitoring. At the same time, DCS is still in an exploratory phase and requires further validation before it can be considered for standard clinical use. HSI and PAI, though still largely experimental in human brain applications, have demonstrated promising potential to provide rich spectral and high-resolution insights, respectively, although these results are still preliminary, especially in surgical or acute settings. At present, both HSI and PAI remain at a proof-of-concept stage in humans and are far from any routine clinical use.
The collective evidence from 2015 to 2025 indicates that these modalities can successfully detect functional activation and pathological changes in the human brain, but each comes with limitations that temper its current utility. Spatial resolution and depth remain key challenges: even with high-density DOT, optical methods cannot yet match MRIs sub-millimeter resolution or whole-brain coverage. Most fNIRS/DOT measurements are confined to cortical convexities, and deep structures (basal ganglia, hippocampus) remain noninvasively out of reach. PAI promises depth, but the skulls acoustic impedance is a formidable barrier. Efforts like using 1064 nm light and full-ring detectors5 are partially mitigating this, and further innovations (e.g., skull cooling or coupling media to improve ultrasound transmission) are being considered. In parallel, the temporal resolution of optical techniques is high for detecting hemodynamic oscillations (on the order of milliseconds for DCS speckle or instrument sampling, though physiological signals are slower). Still, they measure slower signals than neural electrophysiology. Integration with electroencephalography (EEG) or magnetoencephalography could provide a more complete picture of brain activity, combining fast neural and slower hemodynamic information.
A critical issue across modalities is signal quantification and specificity. fNIRS and DOT signals can be contaminated by scalp blood flow; standardizing short-separation regression and other preprocessing is a priority before clinical adoption1,9. DOT reconstructions need to become more quantitatively accurate (current images often require assuming bulk optical properties and use relative changes). DCS provides an index of flow, but calibrating it to absolute CBF is an area of ongoing research - one future direction is combining time-domain diffuse optics with DCS to independently measure optical path lengths and make DCS quantitative in absolute terms. For HSI, isolating the effect of one variable (say oxygenation) from others (such as blood volume or light scattering changes) is complex; advanced spectral unmixing algorithms and perhaps complementary measurements (like simultaneous NIRS for deeper layers) might help. PAIs specificity could be enhanced by multispectral PAT, taking images at multiple wavelengths to differentiate HbO versus HbR, lipid, and other tissue constituents. Indeed, multi-wavelength PAI has been used in preclinical studies to produce functional oxygenation maps of the brain and could be translated to humans to provide both structure and oxygenation in one modality41.
From a technology standpoint, the next decade will likely see further miniaturization and integration. For fNIRS, one exciting avenue is the development of fully wearable, cap-based systems using flexible electronics and high-density grids that can record from most of the cortical surface. Some prototypes involve dozens of LED sources and silicon photodiode detectors embedded in a soft cap, powered by a battery pack or even operating wirelessly. These could turn fNIRS into a "wear it and forget it" technology for continuous brain monitoring in daily life or rehabilitation training. Similarly, integrating fNIRS and DCS in one device (as some research teams have done) provides a more complete hemodynamic picture (both oxygenation and flow). A logical next step would be commercial development of hybrid NIRS-DCS monitors for neonatal ICU or stroke units, which could noninvasively track cerebral oxygen delivery and utilization. The combination of flow and oxygenation can yield indices of metabolic rate (via Ficks principle), giving a window into brain metabolism at the bedside - a long-sought goal in neurocritical care.
Multimodal integration is another promising direction. We already see fNIRS being combined with fMRI (for cross-validation and to leverage fMRIs deep reach), with EEG (to support BCI designs where fNIRS provides complementary information to EEG signals), and even with transcranial electrical or magnetic stimulation (to measure brain hemodynamic response to neuromodulation). For example, concurrent EEG-fNIRS is used to study neurovascular coupling in epilepsy or to improve the accuracy of BCIs using two modalities. As we advance, one could envision trimodal systems (EEG + fNIRS + DCS) to monitor neural, hemodynamic, and perfusion changes simultaneously - this could be powerful in scenarios such as brain-computer interfaces for locked-in patients, where EEG might reveal intent and NIRS/DCS ensure adequate cerebral perfusion or track mental workload.
On the analytical front, the growing datasets from optical neuroimaging call for sophisticated analysis methods. Machine learning and deep learning approaches have started appearing for DOT image reconstruction and HSI spectral classification. For instance, deep neural networks can learn from examples of optical measurements paired with known brain activity (from fMRI or simulations) to directly reconstruct brain activation patterns, potentially outperforming traditional analytical models42. In HSI, deep learning is already used to improve the accuracy of tumor detection in the spectral images43. One challenge that ML can help with is denoising and artifact removal - for example, using convolutional neural networks to filter motion artifacts from fNIRS time series, or to correct for skull-induced distortions in photoacoustic data by learning the distortion patterns5,42. Caution is needed to ensure these models generalize and do not introduce bias, but early results are promising.
The field is poised for larger validation studies in terms of clinical translation and trials. For fNIRS/DOT, that might mean multi-site trials to test fNIRS as a tool for presurgical mapping of the language cortex (comparing it to the gold standard of direct cortical stimulation or fMRI). For DCS, a logical trial would be to validate DCS against established CBF measures such as PET or MRI-ASL (MRI-arterial spin labelling) in patients with cerebrovascular disease, to see how well DCS can track perfusion changes. If successful, DCS could become an accepted monitor for conditions such as carotid stenosis or head-of-bed positioning effects in the ICU. HSI will likely see further clinical studies in neurosurgery: perhaps a Phase II trial of HSI-guided tumor resection, measuring if surgeons achieve more complete resection or better outcomes with HSI assistance (compared to standard white-light or 5-ALA (5-aminolevulinic acid) fluorescence guidance). PAI might initially find a niche in specific procedures - for example, monitoring cortical perfusion during cardiac surgery or neuroendovascular procedures, where a PAI probe could be placed on the head to watch for ischemia in real-time (something currently not done, but technically feasible with a craniotomy or thin bone window).
Regulatory and practical considerations will also shape future directions. Several fNIRS devices have already been approved for clinical use (e.g., for neonatal cerebral oximetry or functional brain imaging), but DOT systems and DCS are mostly confined to research. To translate, devices must be user-friendly and have cleared safety profiles. PAI systems will have to ensure laser exposure stays within ANSI (American National Standards Institute) limits for skin and eye safety; as energies increase for deeper imaging, ensuring no damage (especially to the eye from scattered light) is paramount. HSI devices should deal with issues such as sterilization (since they may be used in sterile fields) and real-time performance.
Finally, standardization and data sharing will significantly benefit the field. The past decade taught us that different groups sometimes reported divergent results for similar applications, likely due to differences in hardware or analytical choices. Initiatives such as open-source platforms for fNIRS (e.g. HOMER44, AtlasViewer45) and shared datasets (like the DOI for the brain tumor HSI dataset4) should be expanded. A community consensus on reporting and comparing optical imaging findings (akin to the reporting standards in fMRI or EEG research) would help pool data and conduct meta-analyses. Applying the PRISMA framework in this review highlighted the need for more systematic reporting in optical neuroimaging studies themselves, such as more precise descriptions of participant selection and blinding.
Despite our focus on technological advances, it is important to acknowledge the limitations of our own review methodology. In particular, the study screening and quality appraisal for this review were conducted by a single reviewer. While this ensured consistency in applying inclusion criteria and bias assessments, it may also introduce subjective bias. Ideally, future systematic reviews in this area will involve multiple independent reviewers at each stage to enhance the reliability and reproducibility of the conclusions.
Conclusion
Optical imaging techniques have emerged as powerful tools for non-invasive human brain mapping, bridging gaps in traditional imaging modalities. Their versatility in clinical and naturalistic environments offers significant potential for advancing neuroscience and patient care. Future research should focus on standardization, quantification, multimodal integration, and rigorous validation through larger, multicenter studies. Enhancing depth penetration, spatial resolution, and real-time capability will further accelerate their translation from research to clinical applications, ultimately transforming how we understand and monitor the human brain.
