fNIRS Measurements and High Density Probes

fNIRS Measurements and High Density Probes

Introduction

• The discussion focuses on fNIRS (functional Near-Infrared Spectroscopy) measurements, particularly the transition from low-density to high-density probes.

Low Density fNIRS Measurements

• Early fNIRS systems (4th generation) used frequency-encoded lasers and avalanche photodiodes.

• Signals were measured simultaneously from all lasers and detectors using band-pass filtering.

• Initial setups involved low-density arrays of sources and detectors (e.g., four sources in a row with detectors 3 cm apart).

• Measurements were typically limited to these 3 cm channels without overlap.

• Despite being low density, these setups could effectively measure brain activity.

• Example: Manager Franceschini's work demonstrated localized changes in oxyhemoglobin and deoxyhemoglobin during finger opposition, motor tasks, sensory tasks, and electrical stimulation.

The Need for High Density Probes

• The primary motivation for high-density measurements is to improve spatial resolution.

How High Density Probes Improve Spatial Resolution

• Overlap Measurements: High-density arrays create overlapping measurements, where each brain location is measured by multiple source-detector pairs.

• Imagery Construction: This overlap, combined with advanced imagery construction techniques, enhances spatial resolution and uniformity.

• Spatial resolution can potentially be improved by a factor of two or more.

  • Low-density systems with 30 mm source-detector separation have a spatial resolution of approximately 30 mm.

  • High-density systems can reduce this to 15 mm or even 10 mm in the adult cortex near the skull.

Early Attempts at High Density Measurements (2005-2006)

• A 5th generation imaging system was used with 16 sources and 32 detectors in a hexagonal pattern.

• Time multiplexing and frequency encoding were used for the sources.

• Time multiplexing was necessary because detectors close to a source received a significantly large signal and were unable to detect signals from long separation source.

• Long separation detectors only saw long separation sources at a given time, and vice versa.

• A hexagonal pattern was chosen to maximize the duty cycle, requiring only three states for multiplexing, unlike grid patterns which require eight or nine.

Phantom Tests

• Phantom tests were conducted using absorbing targets placed under the source-detector array.

• Image reconstruction was performed using:

  • Nearest neighbor measurements.

  • Second nearest neighbor measurements alone.

  • Combined nearest and second nearest neighbor measurements.

• Using both nearest and second nearest neighbors yielded the best high-resolution images of the absorbent target.

• Moving the target demonstrated the benefits of overlapping, multi-distance measurements for spatial sensitivity.

  • Nearest neighbor measurements alone lacked uniform spatial sensitivity.

  • Overlapping second nearest neighbor measurements improved visibility and spatial resolution.

Human Measurements (Finger Tapping)

• Brief finger-tapping experiments were conducted.

• Image reconstruction was performed using:

  • Shortest separation measurements.

  • Nearest neighbor measurements.

  • Second nearest neighbor measurements only.

  • Combined measurements.

• The combined measurements provided better spatial resolution.

Challenges and Limitations

• The duty cycle between different states took about 0.5 to 1 second, limiting the ability to measure the cardiac cycle and assess signal quality.

• Detector fibers were large, preventing close placement, making the SNR poor for larger separations (ideally 30 mm, but achieved 36 mm).

Advancements by Joe Culver's Group

• Joe Culver and his group developed a special-purpose system with a grid geometry for high-density measurements.

• They demonstrated retinotopy in 2007.

• In 2014, they published a Nature Photonics paper showcasing a high-density probe wrapped around the whole head.

Image Reconstruction

• Basic Principle: Measures the perturbation in the optical signal (photon fluence) between a source and detector, which is caused by changes in absorption due to brain activity.

• Equation:

  • $\text{Perturbation in signal} = \int \text{Sensitivity profile} \cdot \text{Change in absorption} \, dV$

  • Where the sensitivity profile represents the light path from the source through the tissue to the detector (banana pattern).

• Linearization: The continuous integral can be linearized by summing over voxels in the brain:

  • $\Delta \text{Intensity} = \sum<em>{i} \text{Sensitivity}</em>{i} \cdot \Delta \text{Absorption}_{i}$

• Matrix Equation: Measurements from multiple sources and detectors are represented as a vector y, related to a vector x (change in absorption coefficient at each point) through a matrix A:

  • $y = Ax$

  • Where matrix A is derived from photon fluence profiles.

Solving for the Image

• A least squares solution can be used:

  • $x = (A^T A)^{-1} A^T y$

• However, diffuse optical tomography typically has more unknowns (voxels) than measurements, leading to an underdetermined inverse problem.

• The problem is also ill-conditioned, making it highly sensitive to noise.

• Regularization is necessary to address these issues.

Regularization Techniques

• Truncated Singular Value Decomposition (SVD): A linear algebra technique (not detailed).

• Spatial Regularization:

  • Introduce prior information about the expected image location and size.

  • $x = (A^T A + R)^{-1} A^T y$

  • Incorporates spatial regularization (R) and measurement covariance (C) matrices.

  • $x = (A^T C^{-1} A + R)^{-1} A^T C^{-1} y$

Impact of Subject-Specific Anatomy

• Studies were conducted to investigate the impact of using a subject's true brain anatomy (from MRI) versus a brain atlas for image reconstruction.

• Registration of probe positions on the subject's head was performed to allow reconstruction on both the subject's anatomy and a registered brain atlas.

GPU-Accelerated Monte Carlo Code (tMCimg)

• A GPU-accelerated Monte Carlo code was developed to calculate the sensitivity matrix.

• The code is significantly faster than CPU-based versions.

Simultaneous fMRI and fNIRS Measurements

• Simultaneous fMRI and fNIRS measurements were taken during median nerve stimulation to validate fNIRS image reconstruction.

• fMRI data provided a benchmark for brain activation location.

• Images were reconstructed on the subject's true anatomy and the brain atlas.

• There was generally good overlap between the reconstructions.

Validation with Simulated Data

• A database of 30 anatomical MRIs was used to simulate brain activity measurements with a whole-head probe.

• A hexagonal high-density probe design was used.

• Monte Carlo simulations were run to obtain photon sensitivity maps for each subject.

• Simulated brain activations were placed at different locations in each brain using FreeSurfer to transform locations between brains based on gyro folding patterns.

• Images were reconstructed in the subject space and the registered atlas space.

Localization Error Analysis

• Localization error was calculated as the Euclidean distance between the true and reconstructed activation locations.

• Geodesic distance, which follows the surface of the brain, was also used.

• Reconstruction on the subject's true anatomy resulted in lower localization errors compared to using the brain atlas.

• The average Euclidean localization error was about 18 mm when using the atlas and improved with the subject's MRI.

Atlas Viewer Software

• All image analysis tools were integrated into Atlas Viewer.

• The software is used for reconstructing images from fNIRS data, starting with temporal processing in HOMER.

• Atlas Viewer is also used to design probes targeting specific brain regions.

Recent Success with High Density fNIRS Measurements

• After years of attempts, high-density fNIRS measurements in adults were finally achieved with the help of Alex.

• The Mira Sport 2 system was used.

• A high-density probe was designed in Atlas Viewer and 3D-printed using a Ninja cap.

• Optodes were placed using optode holders from Miraix with spring tops.

• The first brain activation image was successfully obtained using this system during finger-tapping tasks.

Future Directions

• Increasing the number of sources and detectors is a goal.

• Combining multiple MiraX systems could provide 64 or 80 sources and detectors.

• A new system with 36 sources and 108 detectors to cover the whole head is being developed, using a hexagonal pattern.

• Hexagonal pattern allows coverage of the whole head with fewer sources/detectors, better spatial resolution & better duty cycle.

Probe Design in Atlas Viewer

• The process of designing a probe in Atlas Viewer was demonstrated.

• Key steps include:

  • Displaying reference points (e.g., 10-20 locations).

  • Positioning the probe around a target area (e.g., C3 for motor cortex).

  • Adding sources and detectors by clicking on the brain surface.

  • Placing detectors roughly and correcting distances later.

  • Adding dummy optodes as anchors, securing it in three locations at Cp5, Cz, and F3

  • Adding springs to the dummies and other Optos

  • Setting target disances for shor versus long separation channels

  • Registering the probe.

Integration with NearSite and Aurora

• The probe geometry can be exported as an SNRF file.

• This file can be imported into NearSite software.

• NearSite then uses the information for data acquisition in Aurora.

Tips for good signal quality

• Using ninja mirrors cap and spring top optodes

• Wiggling spring tops to get good contact

• Using an light blocking shower cap to eliminated ambient light.