🔢 mathematics

Projections for handling uncertainties and enabling domain truncation in diffuse optical tomography

This paper introduces a projection-based Bayesian inversion technique for diffuse optical tomography that mitigates modeling errors caused by domain truncation, optode coupling variations, and misspecified tissue parameters by projecting the primary Jacobian onto the orthogonal complement of nuisance Jacobians or their singular vector spans, thereby improving reconstruction accuracy in neonatal brain imaging.

Original authors: Aada Hakula, Pauliina Hirvi, Nuutti Hyvönen, Altti Jääskeläinen, Ville Kolehmainen

Published 2026-04-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Aada Hakula, Pauliina Hirvi, Nuutti Hyvönen, Altti Jääskeläinen, Ville Kolehmainen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to take a clear photograph of a specific room inside a house (the brain) using a flashlight that shines through the walls. This is essentially what Diffuse Optical Tomography (DOT) does: it uses light to see changes in blood flow inside the brain. However, just like taking a photo in a messy room, the picture often comes out blurry or distorted because of "noise" and "mistakes" in how the camera is set up.

This paper presents a clever mathematical trick to clean up these blurry pictures without needing a perfect camera or a perfectly known house.

The Three Main Problems (The "Mess")

The authors identify three specific things that ruin the clarity of the brain images:

The "Loose Connection" Problem: Imagine your flashlight and your camera lens aren't perfectly pressed against the skin. There might be a tiny gap, or maybe the hair is in the way. This changes how much light gets in and out. In the paper, they call this a change in "coupling coefficients." It's like the volume knob on your radio being turned up or down unexpectedly, making the signal sound wrong.
The "Too Big Room" Problem: To get a clear picture of the brain, you usually have to model the entire head (skin, skull, brain, etc.). But modeling the whole head takes a huge amount of computer power. The authors want to just focus on the brain (the "Region of Interest") and ignore the rest (the "Region of Non-Interest"). However, if you ignore the rest of the head, the light bouncing off the skull can still mess up your brain picture.
The "Wrong Map" Problem: To calculate the image, the computer needs a map of how light travels through different tissues (like gray matter or white matter). But scientists don't always agree on the exact numbers for these tissues. If the computer uses a slightly wrong map, the resulting picture will be distorted.

The Solution: The "Noise-Canceling" Filter

The paper proposes a method called Projection. Think of this as a sophisticated "noise-canceling" filter for your data.

Instead of trying to fix the messy data after the fact, the authors change the rules of the game before they try to solve the puzzle. They mathematically "project" the data onto a new plane where the annoying errors simply don't exist.

Here is how they do it for each problem:

Fixing the Loose Connections: They calculate exactly how a loose connection would change the light signal. Then, they create a filter that removes any part of the signal that looks like a loose connection. It's like wearing noise-canceling headphones that specifically silence the sound of a door creaking, so you can hear the music clearly.
Truncating the Room (Ignoring the Rest): Since they can't ignore the whole head without losing information, they use a "smart filter." They figure out which parts of the "ignored" head (the skull and skin) are most likely to mess up the brain signal. They then build a filter that removes only those specific messy parts, leaving the useful brain data intact. It's like using a sieve that lets the gold (brain data) through but catches the biggest rocks (skull interference) without catching the tiny grains of sand (useful details).
Fixing the Wrong Map: This is their most unique idea. If the computer is using a slightly wrong map of the tissue, they don't try to guess the right map. Instead, they look at the difference between two slightly different maps. They use this difference to build a filter that cancels out the errors caused by not knowing the exact map. It's like realizing your compass is slightly off; instead of buying a new one, you adjust your path based on exactly how much it's off, so you still end up at the right destination.

The Results: Clearer Pictures

The authors tested this method on a simulated model of a newborn baby's head (because their heads are smaller and the light travels differently). They created fake "brain activity" (like a light turning on in the brain) and then added the three types of messiness described above.

Without the filter: The pictures were a disaster. The brain activity was either invisible or looked like it was in the wrong place, covered in strange artifacts (ghostly shapes).
With the filter: The pictures became remarkably clear. The brain activity appeared exactly where it should be, and the "ghosts" caused by loose connections or wrong maps disappeared.

The Bottom Line

The paper claims that by using these mathematical "projections," researchers can:

Ignore messy connections between the device and the skin.
Focus their computer power only on the brain, ignoring the rest of the head, without losing image quality.
Get good results even if they don't know the exact optical properties of the brain tissue.

The authors emphasize that this method allows them to get high-quality images faster and more reliably, even when the setup isn't perfect. They tested this specifically on simulated data for a neonate (newborn) and found that their "projection" technique successfully cleaned up the images, making them look almost as good as if everything had been perfect to begin with.

1. Problem Statement

Diffuse Optical Tomography (DOT) is a functional neuroimaging technique used to map hemodynamic activity (changes in oxygenated and deoxygenated hemoglobin) by measuring light propagation through tissue. The reconstruction of absorption coefficient changes relies on an inverse problem that is ill-posed and highly sensitive to modeling errors.

The paper addresses three specific sources of uncertainty in the forward model that degrade reconstruction quality:

Optode Coupling Coefficients: Uncertainties in the amplitude losses and phase shifts at the interface between the optical fibers (sources/detectors) and the skin (e.g., due to air gaps, hair, or movement).
Domain Truncation (Region of Non-Interest - RONI): The desire to reconstruct only a specific Region of Interest (ROI), such as the cerebral cortex, while ignoring activity in the rest of the head (RONI). Standard inversion attempts to explain all data via the ROI, leading to artifacts at the ROI boundary.
Misspecified Baseline Optical Parameters: Inaccuracies in the assumed baseline absorption coefficients for specific tissue types (e.g., Gray Matter or White Matter) used to compute the Jacobian matrix, often due to variability in literature values.

The goal is to enable accurate reconstruction of the ROI even when these uncertainties exist, potentially allowing for computational domain truncation to speed up processing and improve stability.

2. Methodology

The authors propose a projection-based technique integrated into a Bayesian inversion framework. The core idea is to project the linearized forward model onto the orthogonal complement of the subspace spanned by the "nuisance" parameters (uncertainties), effectively filtering out their influence before solving the inverse problem.

A. Mathematical Framework

The linearized measurement model is $y = Jx + \tilde{J}\tilde{x} + e$ , where:

$y$ : Difference measurements (log-amplitude and phase).
$x$ : Unknown absorption changes in the ROI (primary unknown).
$\tilde{x}$ : Nuisance parameters (coupling coefficients, RONI absorption, or baseline errors).
$J, \tilde{J}$ : Primary and nuisance Jacobian matrices.
$e$ : Measurement noise.

The method applies an orthogonal projection matrix $P$ to the equation, yielding $Py = PJx + Pe$, and then performs Bayesian inversion on this projected system.

B. Construction of Projections

The construction of $P$ varies depending on the type of uncertainty:

Low-Dimensional Uncertainties (Coupling Coefficients):
- The nuisance Jacobian $J_c$ (relating measurements to coupling coefficients) is low-dimensional.
- $P$ is constructed as the projection onto the orthogonal complement of the range of $J_c$ : $P = I - J_c(J_c^\top J_c)^{-1}J_c^\top$ .
- This effectively removes the influence of coupling errors entirely.
High-Dimensional Uncertainties (Domain Truncation/RONI):
- The nuisance Jacobian $\tilde{J}$ (relating measurements to absorption changes in the RONI) is high-dimensional (spanning nearly the whole image space). A full projection would eliminate almost all data.
- Partial Projection: The authors select a subspace of the range of $\tilde{J}$ expected to have the strongest impact on the data. This is done using the first $k$ left singular vectors of a weighted nuisance Jacobian ( $\tilde{J}\tilde{A}$ ), where $\tilde{A}$ incorporates prior information (covariance) about the RONI.
- $P$ projects onto the orthogonal complement of this selected subspace.
Misspecified Baseline Parameters:
- This is a novel approach. Instead of a Jacobian with respect to the baseline parameter, the authors compute a weighted difference matrix between two Jacobians evaluated at two different baseline absorption levels (e.g., literature value vs. assumed value).
- The nullspace of the projection is formed by the eigenvectors of this difference matrix (weighted by the prior covariance of the ROI).
- This approximates the sensitivity of the forward model to errors in the baseline parameter.

C. Bayesian Inversion

The projected linear equation is solved using a Gaussian prior and noise model. The posterior mean and covariance are derived analytically for the projected system, allowing for efficient computation.

D. Simulation Setup

Forward Model: Monte Carlo simulations of the Radiative Transfer Equation (RTE) using MCX (Monte Carlo eXtreme) on GPUs.
Phantom: A voxelized neonatal head model (41.7 weeks gestational/chronological age) with 15 sources and 21 detectors.
Data: Frequency-domain (FD) measurements of log-amplitude and phase delay.

3. Key Contributions

Novel Projection for Baseline Errors: The paper introduces a method to handle misspecified baseline optical parameters by constructing the projection nullspace from the eigenvectors of a difference of Jacobians, a technique not previously tested in imaging modalities.
Enabling Domain Truncation: The method successfully allows for the reconstruction of the ROI (brain cortex) while ignoring the RONI (scalp/skull/other brain regions) without introducing severe boundary artifacts, facilitating faster online processing.
Unified Framework: It demonstrates a unified projection strategy that can handle coupling errors, domain truncation, and parameter misspecification individually or simultaneously.
Avoidance of Approximation Error (AE) Sampling: Unlike the Approximation Error method which requires extensive sampling of priors and forward solves to estimate error statistics, this projection method avoids additional sampling, making it computationally more efficient.

4. Results

The method was evaluated on simulated data with two targets (localized absorption perturbations) under four scenarios:

Case 1 (Coupling Coefficients): Without projection, mismatched coupling coefficients resulted in uninformative reconstructions ( $L_2$ error ~9.9). With projection, the error dropped to ~0.195, matching the quality of the ideal baseline case.
Case 2 (Domain Truncation): Reconstructing only the brain while ignoring the scalp (RONI) caused strong artifacts at the boundary in naive reconstructions ( $L_2$ error ~1.5). The projection method reduced the error to ~0.183, producing results nearly identical to the full-domain reconstruction.
Case 3 (Misspecified Baseline): Using incorrect baseline absorption values for Gray/White matter caused significant artifacts and contrast errors ( $L_2$ error ~1.2). The projection method improved localization and reduced the error to ~0.39, though not as perfectly as the other cases.
Case 4 (Combined Uncertainties): When coupling errors and domain truncation were combined, the naive approach failed completely ( $L_2$ error > 20). The combined projection strategy successfully recovered the targets with an error of ~0.16, comparable to the ideal cases.

5. Significance

This work significantly advances the robustness of Diffuse Optical Tomography for functional brain imaging. By mathematically decoupling nuisance parameters from the primary reconstruction, the method:

Improves Clinical Utility: Makes DOT more viable for real-world applications where optode coupling is unstable and tissue properties vary between subjects.
Enables Real-Time Processing: Allows for computational domain truncation (focusing only on the cortex), which drastically reduces the dimensionality of the inverse problem and speeds up reconstruction, crucial for monitoring long time-series data.
Reduces Dependency on Precise Calibration: Reduces the need for perfectly accurate prior knowledge of tissue optical properties, which is often difficult to obtain.

The study validates that projection-based marginalization is a powerful alternative to approximation error methods, particularly for handling structural and parametric uncertainties in complex, stochastic forward models like Monte Carlo simulations.