🧠 neuroscience

Modeling behavior to disentangle motion-related effects in functional ultrasound imaging in awake, head-fixed mice

This paper introduces a behavior-informed modeling framework that integrates continuous measurements of running speed and head motion into functional ultrasound imaging analysis, effectively disentangling motion-related confounds from neural signals to enable reliable brain-wide mapping during naturalistic, high-motion behaviors in awake mice.

Original authors: Qin, C., Nelissen, F., Waasdorp, R., Lotfi, A., Rojas, C., De Angelis, L., Heemskerk, M., Maresca, D., Keysers, C., Heiles, B., Gazzola, V.

Published 2026-02-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Qin, C., Nelissen, F., Waasdorp, R., Lotfi, A., Rojas, C., De Angelis, L., Heemskerk, M., Maresca, D., Keysers, C., Heiles, B., Gazzola, V.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Picture: Listening to the Brain While the Mouse is Dancing

Imagine you are trying to listen to a quiet conversation in a room, but suddenly, someone starts jumping up and down on a trampoline. The floor shakes, the furniture rattles, and the conversation becomes impossible to hear.

This is exactly the problem scientists face when they try to study the brains of awake mice using Functional Ultrasound Imaging (fUSI).

The Goal: They want to see which parts of the mouse's brain "light up" (get active) when the mouse sees a picture or feels a tiny shock.
The Problem: When the mouse runs, gets excited, or feels pain, its brain moves slightly inside its skull. Because the ultrasound probe is sitting right on top of the skull, even a tiny wobble creates a massive "rattle" in the data. This rattle looks like brain activity, but it's actually just motion noise.

For a long time, scientists had to either:

Freeze the mouse: Only study it when it was perfectly still (boring, not natural).
Throw away data: Delete any moment the mouse moved (wasting a lot of information).
Use a "Blind Filter": Try to guess what is noise and what is signal without knowing why the mouse moved (like trying to clean a muddy window without knowing where the mud came from).

The New Solution: The "Motion Detective"

This paper introduces a clever new method. Instead of guessing or throwing away data, the scientists decided to track the mouse's movement explicitly and tell the computer, "Hey, this part of the signal is just because the mouse ran, not because it was thinking."

Think of it like this:

The Old Way (Blind Filter): You hear a loud thump in the house. You assume it's a ghost and try to filter it out. But what if it was just the dog jumping? You might accidentally filter out the dog's bark, which was the important sound you wanted to hear.
The New Way (Behavioral Modeling): You have a camera on the dog. When you hear a thump, you look at the camera, see the dog jumped, and say, "Okay, that thump is the dog. I will ignore it." Now, you can clearly hear the conversation happening in the next room.

How They Did It

The researchers set up two experiments with head-fixed mice (mice whose heads were held still, but whose bodies could move on a treadmill):

The Visual Test (Low Motion): They showed the mice moving stripes on a screen. The mice sometimes ran a little.
- Result: The new method successfully separated the "seeing" signal from the "running" signal. It found the visual brain areas much more clearly than the old "blind filter" methods.
The Pain Test (High Motion): They gave the mice tiny, harmless electric shocks on their tails. This made the mice run fast and get very excited.
- Result: This was the hard test. The running and the pain happened at the exact same time. The old methods got confused and thought the "running noise" was the "pain signal."
- The Breakthrough: The new method used the treadmill speed as a "noise map." It subtracted the running noise from the data. Amazingly, it revealed that the Primary Somatosensory Cortex (the pain center) was actually reacting differently depending on how strong the shock was. The old "blind filter" had accidentally erased this important detail!

The Analogy: The Noisy Concert

Imagine you are at a concert (the brain activity) while a friend is shaking the stage (the mouse running).

The Microphone (fUSI) picks up both the music and the shaking.
The Old Method (PCA/Blind Filter): The engineer says, "That shaking is too loud, I'll just turn down the volume on everything that shakes." But the music also shakes a little bit! So, they accidentally turn down the music too.
The New Method (GLM Modeling): The engineer has a sensor on the stage shaking. They say, "I know exactly how much the stage shook. I will mathematically subtract only the shaking sound." Now, the music comes through crystal clear, even though the stage is still shaking.

Why This Matters

This paper is a game-changer for neuroscience because:

It saves data: We don't have to throw away minutes of recording just because the mouse moved.
It allows for "Real Life" studies: We can finally study complex behaviors like decision-making, fear, or social interaction in mice, not just simple, boring tasks where the mouse sits still.
It bridges the gap: It makes studying mice more like studying humans. Humans can't sit perfectly still in an MRI scanner while thinking complex thoughts; they move. This method helps us understand how brains work when they are actually doing things, not just sitting there.

In short: The authors built a "motion detective" for brain scans. By teaching the computer to recognize the difference between "brain thinking" and "body moving," they can finally hear the brain's true voice, even when the mouse is running for its life.

1. Problem Statement

Functional Ultrasound Imaging (fUSI) is a powerful tool for mapping brain-wide hemodynamic activity in awake rodents, offering high spatiotemporal resolution and portability. However, its application to naturalistic, complex behaviors (beyond tightly controlled sensory paradigms) is severely limited by motion-related confounds.

The authors identify three distinct mechanisms by which motion corrupts fUSI signals:

Mechanical Artifacts: Physical displacement of the brain relative to the external probe causes acoustic field perturbations and backscatter artifacts.
Systemic Physiology: Locomotion induces systemic hemodynamic changes (e.g., increased cardiac output, arterial pulsatility) that modulate cerebral blood volume independently of local neural activity.
Neural Co-activation: Motor and premotor circuits are genuinely activated during movement, generating neural signals that correlate with motion but may obscure task-specific cognitive or affective signals.

Existing mitigation strategies (e.g., spatiotemporal clutter filtering, frame censoring, and removing the first principal component [PC1]) are largely data-driven and "blind." They rely on statistical properties of the signal rather than explicit knowledge of the animal's behavior. Consequently, these methods often fail to distinguish between motion artifacts and genuine neural activity, potentially attenuating or removing task-relevant signals, especially in high-motion paradigms like aversive stimulation.

2. Methodology

The authors propose a behavior-informed General Linear Model (GLM) framework that explicitly incorporates continuous behavioral measurements as nuisance regressors.

Experimental Setup:

Subjects: Awake, head-fixed C57BL/6JRj mice on a running wheel.
Imaging: fUSI acquired via a cranial window using a high-frequency linear probe (18.5 MHz).
Behavioral Monitoring:
- Running Speed: Measured via a wheel-mounted encoder.
- Head Motion: Measured via a 3-axis accelerometer attached to the headplate.
Paradigms:
1. Visual Stimulation: Moving gratings (low-motion condition). Used to validate the method against a "ground truth" of stationary trials.
2. Noxious Stimulation: Tail shocks of increasing intensity (high-motion condition). This induces robust escape running and arousal, creating a scenario where motion and stimulus are highly correlated.

Analytical Framework:
The core innovation is the integration of behavioral covariates into the GLM:

Regressors: The model includes:
- Task predictors (Visual or Shock).
- Running Speed (Raw): Models immediate mechanical artifacts.
- Running Speed (Convolved with HRF): Models hemodynamic changes driven by neural activity associated with running.
- Interaction Terms: To capture state-dependent modulation (e.g., how running alters visual processing).
Comparison: The authors compared their behavior-informed GLM against:
- Models ignoring motion.
- Blind denoising methods (removing PC1).
- Simple correlation analyses.
Validation:
- In the visual task, "stationary trials" served as the ground truth reference.
- In the noxious task, baseline running patterns served as a reference for motion-related variance.
- Metrics included 2D spatial correlation of effect size maps, variance explained ( $R^2$ ), and Contrast-to-Noise Ratio (CNR).

3. Key Contributions

Explicit Behavioral Modeling: The paper introduces a framework where continuous running speed and head acceleration are treated as explicit nuisance regressors in a GLM, analogous to motion parameter correction in fMRI.
Disentanglement of Overlapping Signals: The method successfully separates motion-induced variance from genuine stimulus-evoked activity, even when motion and stimulus are temporally locked and highly correlated (as in the shock paradigm).
Superiority over Blind Denoising: The study demonstrates that behavior-informed modeling outperforms standard PCA-based denoising (PC1 removal), particularly in preserving stimulus-specific neural patterns that global component removal tends to suppress.
Validation Across Motion Regimes: The approach is validated in both low-motion (visual) and high-motion (noxious) contexts, proving its robustness across varying signal-to-confound ratios.

4. Key Results

Visual Stimulation Paradigm (Low Motion):

Motion Impact: Running speed explained a significant portion of fUSI variance ( $R^2 \approx 0.11$ ), and the spatial pattern of running correlated strongly with the first principal component (PC1).
GLM Performance: The model explicitly accounting for running and its interaction with visual stimuli (M3) produced visual response maps that were significantly more similar to the "ground truth" (stationary trials) than models ignoring motion (M2) or simple correlation methods.
PC1 vs. GLM: Removing PC1 did not improve the estimation of visual responses when running was present; in fact, explicit behavioral modeling rendered the results largely insensitive to PC1 removal.

Noxious Stimulation Paradigm (High Motion):

Challenge: Tail shocks induced intense running, making it impossible to isolate shock responses using stationary trials.
Separation of Signals:
- A model ignoring running (M6) showed shock responses that were highly correlated with running patterns (especially in the Retrosplenial Cortex and Thalamus), indicating contamination.
- The behavior-informed model (M7) successfully dissociated shock-related activity from running. The resulting shock maps showed distinct spatial patterns, with localized activation in the Primary Somatosensory Cortex (S1).
Preservation of Intensity Coding: Crucially, the behavior-informed model preserved the shock-intensity-dependent activity in S1. In contrast, removing PC1 (a common blind denoising step) attenuated or eliminated this intensity-dependent signal, effectively removing the very variance the experiment sought to measure.
Quantitative Improvement: The correlation between the estimated shock response and baseline running patterns dropped significantly (from $r \approx 0.62$ in M6 to $r \approx 0.14$ in M7), confirming superior noise separation.

5. Significance and Implications

Bridging the Gap: This work enables the use of fUSI in naturalistic behavioral paradigms (e.g., decision-making, social interaction, affective processing) that were previously inaccessible due to motion artifacts.
Methodological Shift: It advocates for a shift from "blind" statistical cleaning to hypothesis-driven, behavior-informed modeling in preclinical neuroimaging, aligning fUSI analysis practices with established fMRI standards.
Data Efficiency: Unlike censoring methods that discard large portions of data, this approach retains data while statistically accounting for motion, maximizing the utility of experimental sessions.
Future Directions: The authors suggest that as the field moves toward freely moving animals and 3D imaging, explicit behavioral modeling should become a standard component of the fUSI analysis pipeline to ensure the reliability of brain-wide hemodynamic interpretations.

In summary, the paper provides a robust, validated solution to the "motion problem" in fUSI, demonstrating that explicitly modeling behavior allows researchers to recover interpretable neural signals during complex, high-motion behaviors in awake rodents.