Deep Learning-Assisted Evaluation of Laryngeal Mobility in a Rat Model

This study demonstrates that a deep learning-based computer vision framework (SLEAP) can quantitatively evaluate laryngeal mobility and detect asymmetry in a rat model of recurrent laryngeal nerve injury by tracking arytenoid process displacement with high precision.

The Gist

Imagine your voice box (larynx) as a tiny, complex stage where two little doors (the vocal folds) open and close to let air in and out, and to vibrate for speech. These doors are controlled by tiny strings (nerves). If one of those strings gets cut or crushed, one door stops moving, leaving the stage lopsided and the voice broken.

This paper is about a team of scientists who wanted to figure out exactly how broken that door is in rats, but they wanted to do it without hurting the rats or killing them to check the damage.

Here is the story of their invention, explained simply:

The Problem: The "Blind" Check-Up

Usually, when scientists study nerve injuries in rats, they have to wait a few weeks, then kill the rat to look at the nerve and see if it's healing. It's like trying to fix a car engine by taking the car apart, checking the engine, putting it back together, and then taking it apart again every week. It's messy, wasteful, and you can't watch the repair happen in real-time.

Also, looking inside a rat's throat is incredibly hard. Rats are tiny, and they hate being put to sleep with gas. The old ways of checking their throats were either dangerous or just too blurry to see the tiny movements.

The Solution: A Digital "Spotlight"

The researchers came up with a clever two-part solution:

1. The Camera: Instead of a giant, scary medical scope, they used a tiny digital otoscope (the kind doctors use to look in your ears) attached to an iPhone. It's like using a high-tech flashlight to peek into the rat's throat without hurting it.
2. The Brain (Deep Learning): This is the magic part. They didn't just watch the videos; they taught a computer to watch them. They used a smart AI system called SLEAP (think of it as a super-observant robot assistant).

How the Robot Learned to Watch

Imagine you are trying to teach a toddler to spot a specific red ball in a video. You have to point at the ball in 20 different frames and say, "See? That's the ball."

The scientists did this with the rat's throat. They showed the AI computer 20 snapshots of the rat's vocal cords and pointed out the "landmarks" (the little doors). After the AI practiced on 35,000 of these snapshots, it became an expert. It could then watch the entire video and track exactly how far those little doors moved, frame by frame, down to the pixel.

The "Ruler" They Invented

The scientists needed a way to say, "Okay, the left door moved a lot, but the right door barely moved."

They created a mathematical "ruler."

• Before the injury: Both doors moved together, like a pair of synchronized swimmers. The difference between their movements was tiny.
• After the injury: The right door (the injured one) got stuck. The left door kept dancing, but the right one stood still.

The AI calculated the difference in their movements. The scientists found a magic number: 0.42.

• If the difference between the two doors was less than 0.42, the rat was symmetrical (healthy).
• If the difference was more than 0.42, the rat had a broken door (injury).

Why This Matters

Think of this like a fitness tracker for a rat's voice box. Before, scientists had to guess when a rat was getting better. Now, they can put the rat under the camera, run the AI analysis, and get a precise score on how well the nerve is healing.

This means:

• Better Science: They can test new medicines to fix nerve damage and see if they work in real-time.
• Less Suffering: They don't have to kill as many rats to get the data.
• Human Hope: Since rats are often used to test cures for humans, this better way of measuring could help us find better treatments for people who lose their voices due to nerve damage.

In short: They built a tiny camera and taught a computer to be a super-precise referee, measuring exactly how well a rat's vocal cords are dancing, so we can learn how to fix broken voices faster.

Technical Summary

1. Problem Statement

• Clinical Context: Vocal fold mobility is critical for phonation, swallowing, and respiration. Impairments often result from Recurrent Laryngeal Nerve (RLN) injuries.
• Limitations of Current Animal Models:
  • Evaluation Method: Traditional monitoring of RLN injury recovery in rats relies on euthanizing cohorts at multiple time points. This approach is imprecise, fails to capture individual variation, and prevents longitudinal tracking of the same subject.
  • Technical Challenges: Laryngoscopy in rats is difficult due to their small anatomy and the high mortality risks associated with prolonged anesthesia (specifically Ketamine/Xylazine protocols).
  • Need: There is a critical need for a non-invasive, quantitative, and repeatable method to assess subtle changes in laryngeal symmetry and mobility to evaluate functional recovery and new treatments.

2. Methodology

The study employed a combination of surgical modeling, advanced video acquisition, and deep learning-based computer vision.

• Animal Model & Surgery:

  • Subjects: Four adult male Long-Evans rats (~280 g).
  • Procedure: Animals underwent unilateral RLN crush injury at the level of the 5th tracheal ring using jeweler's forceps for 60 seconds.
  • Anesthesia: Induction with isoflurane followed by ketamine-xylazine injection to minimize respiratory depression compared to gas-only protocols.
  • Imaging: High-resolution video laryngoscopy was performed using a digital otoscope integrated with an iPhone. Recordings were taken pre-injury (baseline) and immediately post-injury.

• Deep Learning Analysis (SLEAP):

  • Framework: The open-source deep learning framework SLEAP (Social LEAP Estimates Animal Poses) was utilized for pose estimation.
  • Training: 20 randomly selected frames per video were manually labeled with predefined laryngeal landmarks. The model was trained for 35,000 iterations to generate fully labeled videos and coordinate data (.csv).
  • Landmarks Tracked:
    • Left Arytenoid (LV2) and Right Arytenoid (RV2).
    • Anatomical Midpoint (calculated between the Left Corner (LC) and Right Corner (RC) of the larynx).

• Quantitative Metrics:

  • Displacement Calculation: The distance from each arytenoid (LV2/RV2) to the anatomical midpoint was calculated frame-by-frame.
  • Normalization: Displacement values were normalized to the average displacement of the uncrushed (left) arytenoid for each video to allow cross-subject comparison.
  • Statistical Thresholding: Mean differences between the right (injured) and left (uninjured) arytenoid displacements were computed. A mean difference threshold of 0.42 was established to differentiate between symmetric movement (values < 0.42) and unilateral reduced movement (values > 0.42).

3. Key Contributions

• Quantitative Framework: Developed a novel, automated pipeline using SLEAP to track subtle laryngeal movements in rats, moving beyond qualitative visual inspection.
• Longitudinal Potential: Demonstrated a method that avoids the need for serial euthanasia, enabling the tracking of recovery trajectories in individual animals.
• Defined Threshold: Established a specific statistical threshold (0.42 mean difference) to objectively classify laryngeal asymmetry caused by RLN injury.
• Technical Adaptation: Successfully adapted digital otoscopes and deep learning pose estimation for the specific anatomical constraints of the rat larynx.

4. Results

• Injury Confirmation: Post-crush laryngoscopy confirmed grossly diminished mobility of the right arytenoid in all animals.
• Statistical Validation:
  • Pre-crush: Differences between left and right arytenoid displacements were generally non-significant, confirming baseline symmetry.
  • Post-crush: Significant differences were observed in all but one video, confirming the model's sensitivity to injury.
• Threshold Performance: The calculated mean difference of 0.42 effectively separated pre-injury (symmetric) data from post-injury (asymmetric) data.
• Visualization: The study utilized forest plots to visualize the 95% Confidence Intervals (CI) of displacement values, clearly showing the shift from symmetry (blue markers) to asymmetry (red markers) after injury.

5. Significance

• Preclinical Research: This method provides a robust tool for preclinical studies investigating nerve regeneration, bioengineered vocal fold mucosa, and pharmacological interventions for laryngeal paralysis.
• Precision Medicine: By quantifying individual variation rather than relying on group averages from sacrificed cohorts, the method offers higher statistical power and ethical efficiency (reducing animal use).
• Translational Potential: The integration of consumer-grade hardware (digital otoscopes) with advanced AI (SLEAP) suggests a scalable, cost-effective approach for assessing laryngeal function in small animal models, potentially bridging the gap between animal studies and human clinical diagnostics.