High-Fidelity 3D Simulator for Synthetic fNIRS Data Generation

arXiv:2605.30552 · q-bio.NC · Submitted 2026-05-28 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: We're looking at the paper: "High-Fidelity 3D Simulator for Synthetic fNIRS Data Generation".

Marcus: Here's the gist of the paper. A high-fidelity 3D simulator for synthetic fNIRS data generation introduces a mesh-based Monte Carlo simulation approach to create physiologically realistic, full-head synthetic recordings with high spatiotemporal fidelity This simulator allows users to generate virtually unlimited labeled datasets for testing denoising algorithms, data augmentation, mechanistic modeling, or in silico experimentation Functional near-infrared spectroscopy (fNIRS) provides a noninvasive window into brain activity by measuring task-related changes in oxygenated and deoxygenated hemoglobin in the cortex Analyzing fNIRS data is challenging because of motion artifacts, physiological noise, and other confounding factors The primary goal is to extract cortical activation profiles from optical time series measurements that often contain features generated by numerous confounding factors, including motion artifacts, superficial signal contamination, and systemic physiological signals This fragmentation has also contributed to a scarcity of large, annotated datasets for cross-validation and for supporting the development of next-generation analysis pipelines Model Architecture The simulator introduces a multi-component forward model for generating synthetic fNIRS time series and establishes the relationship between the forward model’s parameters, denoted collectively as θ ∈ Θ (for an overview see Table 1) and the time-dependent vector of intensities, denoted as ϕ1(t; θ) ∈ R C The approach starts by introducing the Rytov approximation to generate synthetic intensities ϕ1(t; θ) (see Section 2.1), using both a baseline intensity ϕ0 and the sensitivity matrix J, which are computed via MMC simulations (see Section 2.2) The total change in absorption coefficient for wavelength λ across all M mesh points is modeled as a weighted sum of cortical and systemic contributions that are parameterized via θ (see Table 1) Modeling Cortical Activity The cortical contribution is modeled as a linear combination of spatial activity maps and time-varying cortical responses This involves modeling the temporal response c(t; θC) by convolving the experimental design vector x(t) with a parametric hemodynamic response function (HRF) h(t, θC,l) The HRF hl(t) models the stereotypical shape of the cortical response using the canonical double-gamma HRF The design vector x(t) indicates when a subject is presented with an experimental event (1 = ON, 0 = OFF) based on the experimental task Modeling Systemic Physiology Changes in absorption due to systemic physiology δµP (see Figure 2) are modeled in the frequency domain and transformed into the time domain via the inverse Fourier transform (McSharry et al., 2003) The power spectral density (PSD) is modeled as the sum of an aperiodic 1/f component and periodic Gaussian peaks corresponding to physiological oscillations The aperiodic component accounts for broadband background fluctuations and is modeled as a power-law (1/f) process with a low-frequency stabilizer External Artifact Generation To generate synthetic data representative of artifacts commonly observed in fNIRS, motion artifacts are explicitly modeled as exponentially decaying pulses centered at randomly selected time points, independently for each channel Baseline shifts are modeled as step-like offsets with randomly selected onset times, independently for each channel Channel dropout is modeled by assuming a random probability Pd that any given channel is replaced by Gaussian noise with mean equal to that channel’s mean and amplitude Ad Validation and Comparison The proposed simulator is compared against an autoregressive (AR) simulator implemented in existing fNIRS analysis toolboxes It is also compared against a Jacobian-free version that omits the Jacobian-based spatial projection The ablation study shows that the full simulator best reproduces the heterogeneous cross-channel structure observed in the experimental data, with stronger correlations among channels over the motor ROI The full simulator combines a physics-aware measurement model with interpretable signal components, making it suitable for generating controlled synthetic fNIRS data with both physiological and spatial structure The Jacobian-free simulator provides a cheaper, low-fidelity alternative that can reproduce some global and temporal signal properties, but it is largely unsuitable for preserving spatially localized channel structure The main strength of the proposed model is its ability to jointly reproduce global signal scale, physiological spectral structure, and localized channel relationships The Jacobian maps localized absorption changes in the head volume to probe-specific intensity measurements, preserving how cortical activity is expressed across nearby and distant channels Limitations If the goal is to reproduce the characteristics of a particular experimental dataset, the simulator’s parameters currently require manual tuning A natural next step is to estimate simulator parameters directly from experimental recordings using likelihood-free approaches such as simulation-based inference (SBI; Cranmer et al., 2020) Another limitation on the fidelity of this work is the current reliance on a single head geometry Incorporating landmark-based constraints and simple mechanical priors would improve physical plausibility and reproducibility of registration across heads The simulator can incorporate alternative HRF parameterizations, including finite impulse response models and other flexible bases that can better capture sustained or non-canonical dynamics The current registration procedure aligns the montage to the head mesh without explicitly enforcing anatomical landmarks (fiducials) or physical constraints of the probe geometry Acknowledgments This work was supported in part by NIH Grant No.

Ines: F31EB035931 and NSF grant No.

Marcus: 2448380 The code associated with this manuscript will be made public upon publication The paper is supported in part by NIH Grant No.

Ines: F31EB035931 and NSF grant No.

Paper summary: Ines: We’ve covered the basics of the "High-Fidelity three dee Simulator for Synthetic fNIRS Data Generation" paper, focusing on how they build a system that creates realistic full-head synthetic recordings using mesh-based Monte Carlo simulations <ref:2605.30552#pg1,High-Fidelity 3D Simulator for Synthetic fNIRS Data Generation>. We touched on how this helps overcome issues with data scarcity and enables testing new analysis pipelines.

Marcus: To build on that, the core thesis is that a high-fidelity simulator can provide access to flexible and controllable ground truth data across different experimental setups. They argue this allows for systematic benchmarking of signal processing algorithms while also providing a data augmentation pipeline for AI models, like those using Generative Adversarial Networks.

Yuki: It’s interesting how they set up the simulation framework, starting from an experimental design and optode montage, combining task-evoked responses, global physiological fluctuations, and external artifacts with a spatial sensitivity matrix derived from light-transport simulations. They show this projects to channel-level changes in optical density to produce synthetic long- and short-channel time series.

Ines: And the model components are detailed: they use the Rytov approximation for intensities, and they parameterize the cortical response using a parametric HRF, which is then convolved with the experimental design vector. They also explicitly model systemic physiology in frequency domain before transforming it back to time.

Marcus: The external artifact generation is specific too: motion artifacts are modeled as exponentially decaying pulses at random times, baseline shifts as step-like offsets, and channel dropout based on a probability Pd where channels get replaced by Gaussian noise with the channel's mean amplitude.

Yuki: What stands out to me from this overview is how they compare their output against empirical recordings using metrics like timefrequency maps or inter-channel correlation maps to see if they actually capture the structure well, not just some global average.

Ines: That comparison is crucial because it shows whether the simulated data truly reflects the real complexity of fNIRS signals, which is what drives the claim about reproducing that heterogeneous cross-channel structure.

Marcus: So, in short, this paper introduces a simulator that combines physics with interpretable signal components to generate synthetic data with both physiological and spatial structure for testing and augmenting AI. That’s where we are now before we get into the broader implications.

Conclusion: Ines: When you look at the title, "High-Fidelity three dee Simulator for Synthetic fNIRS Data Generation," it tells you exactly what they built: a tool that creates high-fidelity synthetic data for fNIRS recordings across a full head <ref:2605.30552#pg1,High-Fidelity 3D Simulator for Synthetic fNIRS Data Generation>. They are focused on achieving that realistic representation of the underlying physics and biology.

Marcus: And when we look at the authors, Condell Eastmond, Niels Bracher, Xavier Intes, and Stefan T. Radev, they’re the team responsible for developing this mesh-based Monte Carlo simulation approach to model this complex system. They are putting out open-source software so other researchers can reproduce the results and use it for their own customized simulations.

Yuki: For someone listening who is only focused on the application, what this means is that they’ve provided a way to generate controlled synthetic data that has both physiological and spatial structure, which is something that was hard to get before. It’s a resource for training better AI models without relying solely on noisy real datasets.

Ines: Exactly. The main point here is the simulator combines a physics-aware measurement model with interpretable signal components, making it suitable for generating controlled synthetic fNIRS data with both physiological and spatial structure. That capability is what makes it so powerful for testing new methods in this field.

Marcus: They also explicitly mentioned that they developed a Jacobian-free version as a cheaper alternative, but they admitted that the full simulator is largely unsuitable for preserving spatially localized channel structure, though it still reproduces some global and temporal signal properties.

Yuki: So, the real takeaway is that while the full simulator is powerful for structure, even if it’s complex to run because of all those physics components, its ability to map localized absorption changes in the head volume to probe-specific intensity measurements is what preserves how cortical activity is expressed across nearby and distant channels.

Ines: That capability allows them to jointly reproduce global signal scale, physiological spectral structure, and those localized channel relationships simultaneously. It’s a sophisticated way to generate data that respects how fNIRS works physically.

Marcus: And the limitation they state plainly is that if the goal is to reproduce the characteristics of a particular experimental dataset, the simulator’s parameters currently require manual tuning. A natural next step, which they suggest, is using simulation-based inference approaches to estimate those simulator parameters directly from experimental recordings.

Yuki: Incorporating landmark-based constraints and simple mechanical priors would also improve physical plausibility when aligning the montage to the head mesh without explicitly enforcing anatomical landmarks or physical constraints of the probe geometry. It’s a clear path for making it more physically sound over time.

q-bio.NC

Submitted: 2026-05-28

Updated: 2026-11-16

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 88/100

The gist: A high-fidelity 3D simulator for synthetic fNIRS data generation introduces a mesh-based Monte Carlo simulation approach to create physiologically realistic…

Key concepts

Mesh-based Monte Carlo Simulation
This is the core method used to simulate the physical measurements. It involves creating a 3D mesh representing the head and using Monte Carlo simulations to determine how light travels through this structure, which then generates synthetic intensity data that mimics real fNIRS recordings.
Hemodynamic Response Function (HRF)
The HRF models how brain activity changes the absorption of light over time. This function describes the stereotypical shape of a cortical response, typically using a double-gamma model to capture the temporal dynamics of blood oxygenation changes during a task.
Power Spectral Density (PSD) Modeling
This concept is used to simulate systemic physiological noise. The PSD models how different frequencies contribute to the background signal, combining broadband fluctuations with periodic peaks that represent known physiological oscillations in the system.

Terminology

Summary

A high-fidelity 3D simulator for synthetic fNIRS data generation introduces a mesh-based Monte Carlo simulation approach to create physiologically realistic, full-head synthetic recordings with high spatiotemporal fidelity This simulator allows users to generate virtually unlimited labeled datasets for testing denoising algorithms, data augmentation, mechanistic modeling, or in silico experimentation

Introduction to the Problem

Functional near-infrared spectroscopy (fNIRS) provides a noninvasive window into brain activity by measuring task-related changes in oxygenated and deoxygenated hemoglobin in the cortex Analyzing fNIRS data is challenging because of motion artifacts, physiological noise, and other confounding factors The primary goal is to extract cortical activation profiles from optical time series measurements that often contain features generated by numerous confounding factors, including motion artifacts, superficial signal contamination, and systemic physiological signals This fragmentation has also contributed to a scarcity of large, annotated datasets for cross-validation and for supporting the development of next-generation analysis pipelines

Model Architecture

The simulator introduces a multi-component forward model for generating synthetic fNIRS time series and establishes the relationship between the forward model’s parameters, denoted collectively as θ ∈ Θ (for an overview see Table 1) and the time-dependent vector of intensities, denoted as ϕ1(t; θ) ∈ R C The approach starts by introducing the Rytov approximation to generate synthetic intensities ϕ1(t; θ) (see Section 2.1), using both a baseline intensity ϕ0 and the sensitivity matrix J, which are computed via MMC simulations (see Section 2. p. 2

Where the paper says it

We start by introducing the Rytov approximation (O’Leary, 1996) to generate synthetic intensities ϕ1(t; θ) (see Section 2. 1), using both a baseline intensity ϕ0 and the sensitivity matrix J, which are computed via MMC simulations (see Section 2.

2) The total change in absorption coefficient for wavelength λ across all M mesh points is modeled as a weighted sum of cortical and systemic contributions that are parameterized via θ (see Table 1)

Modeling Cortical Activity

The cortical contribution is modeled as a linear combination of spatial activity maps and time-varying cortical responses This involves modeling the temporal response c(t; θC) by convolving the experimental design vector x(t) with a parametric hemodynamic response function (HRF) h(t, θC,l) The HRF hl(t) models the stereotypical shape of the cortical response using the canonical double-gamma HRF The design vector x(t) indicates when a subject is presented with an experimental event (1 = ON, 0 = OFF) based on the experimental task

Modeling Systemic Physiology

Changes in absorption due to systemic physiology δµP (see Figure 2) are modeled in the frequency domain and transformed into the time domain via the inverse Fourier transform (McSharry et al., 2003) The power spectral density (PSD) is modeled as the sum of an aperiodic 1/f component and periodic Gaussian peaks corresponding to physiological oscillations The aperiodic component accounts for broadband background fluctuations and is modeled as a power-law (1/f) process with a low-frequency stabilizer

External Artifact Generation

To generate synthetic data representative of artifacts commonly observed in fNIRS, motion artifacts are explicitly modeled as exponentially decaying pulses centered at randomly selected time points, independently for each channel Baseline shifts are modeled as step-like offsets with randomly selected onset times, independently for each channel Channel dropout is modeled by assuming a random probability Pd that any given channel is replaced by Gaussian noise with mean equal to that channel’s mean and amplitude Ad

Validation and Comparison

The proposed simulator is compared against an autoregressive (AR) simulator implemented in existing fNIRS analysis toolboxes It is also compared against a Jacobian-free version that omits the Jacobian-based spatial projection The ablation study shows that the full simulator best reproduces the heterogeneous cross-channel structure observed in the experimental data, with stronger correlations among channels over the motor ROI

Conclusion

The full simulator combines a physics-aware measurement model with interpretable signal components, making it suitable for generating controlled synthetic fNIRS data with both physiological and spatial structure The Jacobian-free simulator provides a cheaper, low-fidelity alternative that can reproduce some global and temporal signal properties, but it is largely unsuitable for preserving spatially localized channel structure The main strength of the proposed model is its ability to jointly reproduce global signal scale, physiological spectral structure, and localized channel relationships The Jacobian maps localized absorption changes in the head volume to probe-specific intensity measurements, preserving how cortical activity is expressed across nearby and distant channels

Limitations

If the goal is to reproduce the characteristics of a particular experimental dataset, the simulator’s parameters currently require manual tuning A natural next step is to estimate simulator parameters directly from experimental recordings using likelihood-free approaches such as simulation-based inference (SBI; Cranmer et al., 2020) Another limitation on the fidelity of this work is the current reliance on a single head geometry Incorporating landmark-based constraints and simple mechanical priors would improve physical plausibility and reproducibility of registration across heads The simulator can incorporate alternative HRF parameterizations, including finite impulse response models and other flexible bases that can better capture sustained or non-canonical dynamics The current registration procedure aligns the montage to the head mesh without explicitly enforcing anatomical landmarks (fiducials) or physical constraints of the probe geometry

Acknowledgments

This work was supported in part by NIH Grant No. F31EB035931 and NSF grant No. 2448380 The code associated with this manuscript will be made public upon publication The paper is supported in part by NIH Grant No. F31EB035931 and NSF grant No.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. A high fidelity 3D simulator can generate virtually unlimited labeled datasets for testing denoising algorithms, data augmentation, mechanistic modeling, or in silico experimentation by combining anatomically accurate sensitivity profiles with parameterized models of hemodynamic responses and systemic physiology.

  2. The improved AI system can systematically benchmark signal processing algorithms using realistic ground truths and provide a data augmentation pipeline for data hungry AI models to learn the statistical structure of observed data.

  3. The simulator can support simulation based inference approaches that link neural time series to mechanistic models of brain activity, accelerating experimental design by allowing researchers to virtually prototype optode configurations and predict signal quality prior to conducting costly in vivo experiments.

  4. By incorporating a Jacobian-based spatial projection, the simulator preserves probe specific and wavelength dependent spatiotemporal fidelity, which is crucial for accurately representing how cortical activity is expressed across nearby and distant channels.

Abstract

Functional near-infrared spectroscopy (fNIRS) provides a noninvasive window into brain activity by measuring task-related changes in oxygenated and deoxygenated hemoglobin in the cortex. However, analyzing fNIRS data is challenging because of motion artifacts, physiological noise, and other confounding factors. This challenge is further compounded by the limited availability of annotated datasets, which hinders the development and validation of new analysis pipelines, particularly given the growing use of AI methods. Recognizing these challenges, we introduce a 3D fNIRS simulator that uses mesh-based Monte Carlo simulations to generate physiologically informed, full-head synthetic recordings with controllable spatiotemporal characteristics. Our simulator combines anatomically informed sensitivity profiles with parameterized models of hemodynamic responses, systemic physiology, and measurement artifacts. As a result, users can generate large labeled datasets with known cortical ground truth for developing and evaluating denoising algorithms, data augmentation strategies, mechanistic models, and in silico experiments. We validate key spatiotemporal and physiological properties of the simulator against experimental fNIRS data from open-source finger-tapping, pain-assessment, and surgical-skill datasets.

Sources

Related papers