From Channels to Images: Spatial Coverage, High-Density fNIRS, and Image Reconstruction

We now understand how the MBLL turns two optical density changes into hemoglobin concentration time series. That calculation happens at a single source-detector pair, producing a single channel of data.

The transition from a single measurement to a spatial map of brain activity is one of the most exciting recent developments in fNIRS. Understanding how it works conceptually, and what it requires, is the main goal of this Lesson.

1. What Is a Channel?

In fNIRS, a channel is a single source-detector pair. Each channel produces one time series of ΔHbO₂ and one of ΔHbR. One important thing to note is that the channel does not measure a point; it measures the tissue volume sampled by photons that travel between that source and detector (the banana-shaped sensitivity region whose depth and lateral extent depend primarily on the source-detector separation that we discussed in Lesson 2.1).

Two properties of this sensitivity region are particularly important for understanding what a channel measurement implies:

• Depth sensitivity increases with source-detector separation. A pair separated by ~3 cm samples tissue at depths of roughly 1-2 cm from the interface (e.g., the scalp). On the other hand, a pair separated by only 1 cm samples primarily superficial tissue (e.g., in the brain 1 cm separation would sample the scalp and skull, and it barely reaches the cortex at all). This depth-separation relationship is important to how multi-distance arrays gain spatial information.
• The channel is not spatially specific. Within the banana-shaped sensitivity region, all tissue contributes to the measured signal. You cannot tell from a single channel whether a signal change came from a focal cortical region, from diffuse extracerebral tissue, or from some combination. This is due to the partial volume and extracerebral contamination problem we discussed in Lesson 2.2.

Short-separation channels: sampling the scalp selectively

One of the most practically important innovations in fNIRS methodology in the late 2000s is the deliberate use of short-separation channels, i.e.,  source-detector pairs with a very small separation (typically 8-15 mm). At this distance, photons barely penetrate below the scalp. The channel therefore measures almost exclusively the extracerebral signal: blood volume and oxygenation changes in the scalp and skull driven by systemic physiology (heartbeat, respiration, blood pressure fluctuations).

This is useful because, as we saw in the previous lesson, the extracerebral signal is the dominant source of physiological noise in standard long-separation fNIRS channels. If you can measure the extracerebral signal separately with a short-separation channel, you can regress it out of the nearby long-separation channel and isolate the cortical signal with much greater specificity. Short-separation channels are now widely considered best practice in fNIRS signal processing and are built into many modern systems.

2. Multiple Channels Give Spatial Coverage and Montage Design

A single channel tells you about one region of the brain. Place multiple sources and detectors on the scalp, and each source can pair with multiple nearby detectors and vice versa. The result is an optical probe montage: a collection of channels that together cover a region of the scalp and the cortex beneath it.

The efficiency of a montage is notable: with N sources and M detectors, you can obtain many more than N + M channels, because each source can pair with several detectors and each detector can receive from several sources. A system with 8 sources and 8 detectors, for example, could yield up to 64 channels – although, in practice, it will realistically yield approximately 20-30 channels depending on geometry, each sampling a distinct (though overlapping) cortical region. This happens because data at separations beyond 40-45 mm will be completely dominated by noise.

The spatial arrangement of sources and detectors is designed by the researcher to match the brain regions relevant to their question, a process generally referred to as  montage design. In the simplest way, channels are typically registered to standard anatomical coordinates using the 10-20 system or probabilistic atlases, allowing the spatial location of each channel to be mapped approximately to a brain region. Several tools exist to support this: AtlasViewer, fOLD, devfOLD, and others allow virtual optode placement and spatial coverage visualisation before any data collection begins.

Figure 1. A multi-channel fNIRS montage. Each source-detector pair constitutes a channel with its own banana-shaped sensitivity region. Together, multiple channels provide spatial coverage of a cortical area. Source: Wikimedia Commons.

The result of a multi-channel fNIRS experiment is a set of spatially distributed hemoglobin time series (one ΔHbO₂ and one ΔHbR per channel). These can be  readily displayed as a topographic map, in which channel values are projected onto a 2D representation of the scalp surface, colour-coded by amplitude. This is the standard output of most fNIRS experiments, and it gives an immediate visual sense of which scalp regions were more or less oxygenated during a task.

Topographic maps are informative, but they are representations of channel values projected onto the surface and not images of the brain itself. (Note, in some cases, you may see projections onto a 3D representation of the brain or scalp; even in these cases, what you see is not an image.) Regardless of 2D or 3D projects, in these maps each channel value reflects a weighted average of haemoglobin changes across the entire banana-shaped sensitivity volume, including scalp and skull contributions. The spatial resolution is limited by channel density and the broad shape of the sensitivity profile (typically ~2-3 cm for standard systems). Two activations closer than this distance cannot be distinguished.

3. High-Density fNIRS Arrays

Standard fNIRS systems typically use sparse arrays comprising 20 to 100 channels, with (long) source-detector separations of 2.5-4 cm and inter-optode distances that leave large gaps in spatial sampling. These systems are practical, portable, and well suited to many research questions. But their spatial resolution and depth specificity are limited by the sparsity of the sampling.

High-density diffuse optical tomography (HD-DOT) takes a fundamentally different approach. By placing sources and detectors much closer together (typically with first nearest-neighbour separations of 1-1.5 cm) and using many more of them, HD-DOT systems generate arrays with hundreds or thousands of channels covering large areas of the scalp. Importantly, the channels in these arrays have overlapping sensitivity regions at multiple source-detector separations simultaneously. The same volume of cortex is sampled from many different angles and depths.

This is a critical difference. In a sparse system, each cortical location is sampled by at most one or two channels. In an HD array, each cortical location is sampled by many channels simultaneously, each sensitive to a slightly different depth and angle. This multiplicity of overlapping measurements is what makes image reconstruction in fNIRS possible (which has historically been termed “Diffuse Optical Tomography”).

4. Diffuse Optical Tomography (DOT)

Diffuse optical tomography (DOT) is the process of using the full array of channel measurements simultaneously, together with a mathematical model of how light travels through the head, to reconstruct a 3D image of where and how much absorption changed. It is, mathematically, an inverse problem: you know the outputs (all the detector readings), you know the model (how light propagates through the head), and you want to find the inputs — the spatial distribution of absorption changes that best explains what the detectors measured.

This is conceptually different from what the MBLL does. The MBLL converts each channel’s optical density change independently into a haemoglobin estimate, treating each channel in isolation. DOT reconstruction uses all channels together: the spatial overlap between channels means that the measurements from one channel provide information about the same tissue volume as adjacent channels, and this redundancy can be exploited to localise the source of changes more precisely.

The forward model: why reconstruction needs physics

Reconstruction requires knowing, for each possible location in the head, how much a unit absorption change at that location would affect each detector reading. This set of relationships — the sensitivity matrix or Jacobian — is computed from a forward model of light transport through a head geometry.

Typically this involves a segmented anatomical model of the head — derived from an MRI scan, an age-appropriate atlas, or a standard template — with different optical properties assigned to each tissue layer (scalp, skull, CSF, grey matter, white matter). The photon diffusion equation is then solved numerically through this layered head model to compute how much light from each source reaches each detector, and how sensitive each detector is to changes at each voxel in the brain. This is the calculation that the MBLL’s semi-infinite homogeneous approximation replaces with a much simpler formula — which is precisely why the MBLL cannot support genuine tomographic reconstruction, as we will discuss in Lesson 2.5.

5. What HD-DOT Offers — and What It Requires

HD-DOT represents a genuine qualitative advance over sparse fNIRS in several dimensions. But it also comes with requirements that standard fNIRS does not. Understanding both sides of this trade-off is essential for choosing the right approach for a given research question.

What HD-DOT offers

• Better spatial resolution. HD-DOT can resolve functional responses of ~1 cm — comparable to fMRI at the cortical surface — compared to the ~2–3 cm resolution of sparse systems. Zeff et al. (2007) demonstrated shifts of <1 cm between adjacent visual field representations, which sparse fNIRS cannot resolve.
• Depth discrimination. Because HD arrays include channels at multiple source-detector separations simultaneously, they can in principle distinguish signals from different depths. This means the reconstruction can separate cortical haemodynamics from extracerebral contamination more effectively than sparse systems or short-separation regression alone.
• Anatomically registered images. Rather than channel values mapped to scalp positions, HD-DOT reconstruction produces voxel-level images registered to standard brain atlases. This facilitates direct comparison with fMRI, group-level analyses in common atlas space, and more precise localisation claims.
• Sensitivity to distributed networks. The wide field of view of whole-head HD-DOT systems enables mapping of distributed resting-state networks and bilateral activations — capabilities that sparse, region-specific fNIRS montages cannot achieve without multiple probe relocations.

What HD-DOT requires

• A forward model of light transport. Reconstruction cannot be performed without computing how light travels through the specific head being imaged. This requires either a subject-specific MRI (not always available), an age-appropriate atlas model, or a generic template. The accuracy of the reconstruction depends directly on the accuracy of this model — errors in the head geometry or assumed optical properties translate directly into errors in the reconstructed image.
• More hardware and setup time. Placing and coupling hundreds of optodes requires more time and care than a sparse system. Hair coupling is a particular challenge with high-density arrays. The systems are typically larger and less portable than sparse fNIRS, though this is changing rapidly with new wearable HD designs.
• More complex analysis pipelines. DOT reconstruction requires forward modelling, inversion, regularisation, and atlas registration — a substantially more involved pipeline than channel-by-channel MBLL application. Dedicated software tools exist (NeuroDOT, NIRFAST, AtlasViewer) but they require training.
• Not always necessary. For many research questions — lateralisation of a frontal cognitive task, monitoring cerebral oxygenation during exercise, studying a well-localised response — a sparse system provides sufficient information at far lower cost and complexity. HD-DOT is the right tool when spatial precision matters; it is not the right tool by default.

6. Why This Sets Up Lesson 2.5

This lesson has shown that fNIRS can be much more than a collection of independent channel measurements. With high-density arrays and reconstruction, it can approach MRI-like spatial resolution while retaining its unique advantages of portability, silence, and compatibility with diverse populations.

But this progression also makes something clear: the more seriously you take spatial imaging, the more the limitations of simple models become apparent. The MBLL treats tissue as a homogeneous half-space and each channel independently. HD-DOT reconstruction requires a layered, anatomically realistic model and treats all channels simultaneously. The gap between these two approaches is not just computational — it reflects a genuine difference in the physical accuracy of the underlying model.

Lesson 2.5 confronts this directly. It will explain what the MBLL is, formally, as a physical model — and what it assumes. It will name the more complete framework that the MBLL approximates. And it will give you the honest, consolidated picture of what CW fNIRS can and cannot measure, regardless of whether you are using a sparse system or an HD array. That lesson closes Unit 2 — and opens the door to Unit 3, where the question shifts from physics to neuroscience.

Society for Functional Near-Infrared Spectroscopy.
Content is for educational purposes.