Summary: Scientists have long understood that faces communicate emotion, but quantifying the fine-grained link between facial muscle movement and brain state has been difficult. Researchers at Cold Spring Harbor Laboratory (CSHL) now introduce Cheese3D, an AI-driven discovery platform that uses high-speed cameras and machine learning to track mouse facial expressions with extraordinary precision. The system can non-invasively predict depth of anesthesia with accuracy comparable to EEG.
Cheese3D combines synchronized multi-angle video with computer vision to convert subtle, rapid facial motions into interpretable, quantitative data. By turning whole-face movement into a reliable readout of internal neural states, this technology promises new ways to study brain function, development, and disease.
Key Facts
- The Cheese3D rig: A calibrated array of six small, synchronized cameras films the mouse face from multiple angles to overcome the challenges posed by its conical anatomy.
- AI-driven reconstruction: Machine learning models assemble the 2D video streams into a 3D representation, extracting anatomically meaningful features and minute changes in muscle tone.
- EEG-level accuracy, non-invasive: In validation experiments, Cheese3D predicted anesthetic depth by tracking facial muscle tone with accuracy matching invasive EEG recordings, all without touching the animal.
- Development and clinical relevance: Facial movement is an early developmental milestone. Measuring face-wide dynamics offers a new approach to study social communication development and conditions such as autism.
Source: CSHL
Love, pain, joy, fear, desire — faces convey a wide emotional range. We interpret many of these signals intuitively, but turning those impressions into precise, reproducible measurements has been a long-standing challenge for neuroscience.
Mice show clearly observable facial behaviors and share conserved neural circuits that control facial movement, making them valuable models for studying how facial dynamics reflect brain states. Yet their small, conical faces have limited previous systems’ ability to capture all movements with high spatial and temporal resolution. Cheese3D was developed to fill that gap.
CSHL Assistant Professor Xun Helen Hou and her lab designed Cheese3D to capture high-speed, whole-face 3D motion including ears, eyes, whisker pad and jaw on both sides of the face. The platform uses a calibrated six-camera array and interpretable computer vision algorithms to extract features in absolute world units at sub-millimeter precision.
The idea grew out of a practical need: experienced veterinarians can often assess an animal’s condition by looking at its face, but science lacked an automated, reproducible method to quantify those signals. The team built hardware and software that lets machines perform that task, producing data that can be directly related to neural activity.
To demonstrate the system’s capabilities, the Hou lab recorded mice during everyday behaviors like eating, and performed key tests under anesthesia. In collaboration with CSHL’s Borniger lab, they compared Cheese3D’s facial readouts with invasive EEG recordings. The platform predicted anesthetic depth with comparable accuracy, while remaining non-invasive and avoiding disturbance to the animal.
“Subtle changes in facial muscle tone tell us a great deal,” Hou says. “We can predict anesthetic depth from the face in a non-invasive way.” Beyond anesthesia monitoring, the system opens paths to study disease states, developmental milestones, and how animals learn social facial movements.
Because facial movement emerges early in development — infants smile before they crawl — tools that quantify face-wide dynamics could reshape understanding of social development and help detect atypical trajectories relevant to conditions such as autism. Cheese3D provides a structured framework to pose and answer these questions experimentally.
Key Questions Answered:
A: Mice do not smile like humans, but they exhibit a rich array of face-related behaviors tied to pain, pleasure, and well-being. Veterinarians have long recognized these signals; Cheese3D quantifies them, creating a mathematical “dictionary” that links facial movement to brain state.
A: EEGs require electrodes on the scalp or implanted devices, which can be intrusive and affect behavior. Cheese3D reads brain-related states from facial motion remotely, preserving natural behavior and reducing stress from instrumentation.
A: That is an intended long-term goal. Mapping how facial muscle activity corresponds to specific neural circuits in mice may suggest non-invasive facial biomarkers for human diagnostics, anesthesia monitoring, and behavioral assessments.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- The original journal paper was reviewed in full.
- Additional context was added by staff.
About this neurodevelopment research news
Author: Samuel Diamond
Source: CSHL
Contact: Samuel Diamond – CSHL
Image: Image credited to Neuroscience News
Original Research: Closed access.
“Cheese3D enables sensitive detection and analysis of whole-face movement in mice” by Kyle Daruwalla, Irene Nozal Martin, Linghua Zhang, Diana Naglič, Andrew Frankel, Catherine Rasgaitis, Rubin Zhao, Xinyan Zhang, Zainab Ahmad, Jeremy C. Borniger & Xun Helen Hou. Nature Neuroscience
DOI: 10.1038/s41593-026-02262-8
Abstract
Cheese3D enables sensitive detection and analysis of whole-face movement in mice
Facial expressions and movements—from fleeting grimaces to rapid chewing—reflect dynamic neural and physiological processes. Mice display clear facial responses and share evolutionarily conserved facial motor circuits with other mammals, making them suitable models for linking face motion to neural state.
Existing methods have struggled to capture the full face at the spatial and temporal resolution required because of the small, conical form of the mouse face. Cheese3D addresses this by using a calibrated six-camera array to capture high-speed 3D motion across the entire face, including ears, eyes, whisker pad, and jaw on both sides.
The platform extracts interpretable dynamics of anatomically meaningful 3D facial features with sub-millimeter precision in absolute units. Proof-of-principle demonstrations include predicting anesthetic depth from changing facial patterns, inferring tooth and muscle anatomy from fast ingestion motions, measuring subtle differences evoked by brainstem stimulation, and linking neural activity to spontaneous facial movements—revealing expressive features that are measurable only in 3D, such as ear angle.
Cheese3D functions as a discovery tool that converts subtle mouse facial movements into a highly interpretable readout of otherwise hidden neural and physiological processes.