Hummingbirds use compensatory eye movements to stabilize both rotational and translational visual motion

To maintain stable vision, behaving animals make compensatory eye movements in response to image slip, a reflex known as the optokinetic response (OKR). Although OKR has been studied in several avian species, eye movements during flight are expected to be minimal. This is because vertebrates with laterally placed eyes typically show weak OKR to nasal-to-temporal motion (NT), which simulates typical forward locomotion, compared to temporal-to-nasal motion (TN), which simulates atypical backward locomotion. This OKR asymmetry is also reflected in the pretectum, wherein neurons sensitive to global visual motion also exhibit a TN bias. Hummingbirds, however, stabilize visual motion in all directions through whole-body movements and are unique among vertebrates in that they lack a pretectal bias. We therefore predicted that OKR in hummingbirds would be symmetrical. We measured OKR in restrained hummingbirds by presenting gratings drifting across a range of speeds. OKR in hummingbirds was asymmetrical, though the direction of asymmetry varied with stimulus speed. Hummingbirds moved their eyes largely independent of one another. Consistent with weak eye-to-eye coupling, hummingbirds also exhibited disjunctive OKR to visual motion simulating forward and backward translation. This unexpected oculomotor behavior, previously unexplored in birds, suggests a potential role for compensatory eye movements during flight.

Methods
Animals

Six adult male Anna’s hummingbirds (Calypte anna) were caught on the University of British Columbia campus (07 Feb 2022 – 04 March 2023). Animals were individually housed in 0.61 x 0.61 x 0.91-m cages and provided with ad libitum artificial nectar (Nektar-Plus, Nekton, Pforzheim, Germany & Nectar 9, Roudybush, Woodland, California, USA). All procedures were approved by the University of British Columbia Animal Care Committee in accordance with the guidelines set out by the Canadian Council on Animal Care. Animals were released back to the wild following the experiments described herein.

Method of restraint

To allow the release of the birds after our experiments, we developed a method for non-invasive restraint of hummingbirds. Head-restraint is required to study OKR in the absence of the vestibulo-ocular reflex. We restrained hummingbirds in a flight posture by wrapping them in an elastic “ACE Brand” bandage (3M, Maplewood, USA) and placing them on a plastic bed tilted 10° above horizontal (see Fig. 1A). To restrain the head, we fashioned a bill guide using the opening of a 30 mL Luer syringe (Henke Sass Wolf, Dudley, USA). The bill guide was oriented to hold the bill horizontally, as is typical in flight (Fig. 1B).

Birds were hand-fed a sucrose solution before each recording session. We then wrapped the bird in the elastic bandage, gently positioned the bill through the guide, and placed the animal on the plastic bed. We were careful to place the body close enough to the bill guide so that the animals were unable to retract their head. Extension of the head was prevented as the opening of the bill guide was smaller than the widest part of the bill. After each recording session, animals were returned to their home cage.

Eye tracking

An overview of the eye-tracking setup is provided in Figure 1. We used a bright pupil method to detect and track both eyes of each hummingbird [38]. Two cameras were positioned using tripods on the left- and right-side of the bird, at the height of the animal and along the optical axis of the eyes (~20 deg nasal from lateral). Bright pupil eye tracking uses an infrared light to increase the contrast between the pupil and the iris (analogous to the “red eye effect”). For illumination, we used 850 nm IR ring lights (Feyond FY76-72V4-F, Shenzhen City, China). These IR lights were attached to 50 mm Nikkor camera lenses (Kabushiki-gaisha Nikon, Tokyo, Japan). Videos were acquired using IR-sensitive cameras (Prosilicia GE680, Allied Vision, Stadtroda, Germany) operating at 120 Hz with recording performed using Streampix version 6 (NorPix Inc., Montreal, Canada). Recording occurred in the dark. The IR lights produce illumination outside of the visual range for hummingbirds [39].

Visual Stimuli

Stimuli were generated using PsychoPy (version 2023.1.3) and presented on two 144Hz LCD monitors (ASUS VG248QG, ASUSTeK Computer Inc., Taipai, Taiwan) set to sRGB color mode, resulting in gratings with a Michelson contrast of 0.99. The monitors were oriented perpendicular to the bird and positioned 30 cm away on either side (Fig. 1A). To allow a clear view of the bird for each camera, the monitors were height-adjusted so that the lower margin of the screen was just below that of the bird. Thus, visual stimuli were presented to the upper and lateral visual fields, occupying ~83 deg horizontally and ~45 deg vertically on either side of the animal.

Previous electrophysiology studies in hummingbirds used random dot fields to determine neuron direction preferences and then sinewave gratings to measure spatiotemporal tuning in the preferred direction [7,31,33]. We selected stimulus conditions to match these studies, but used sinewave gratings throughout. This allowed us to assess OKR at stimulus speeds beyond what we could produce with dots and replicate stimulus speeds across a consistent set of spatial and temporal frequencies (Table 1).

Within a recording session, we randomized the direction (NT or TN) and speed of stimuli presented on each monitor. Hummingbirds must feed frequently, limiting the maximum length of recording sessions to under 40 minutes. We therefore presented stimuli for each spatial frequency on separate days, except for spatial frequencies of 0.015625 cyc deg-1 and 4 cyc deg-1, which were presented on the same day but with a feeding break in between. Stimulus velocity was defined as the product of temporal and spatial frequency (Table 1). During binocular stimulation, stimulus speeds presented to both eyes were always matched (e.g., 1 deg s-1 TN to the right eye and 1 deg s-1 NT to the left eye).

Table 1: See manuscript

Each recording session consisted of a set of trials repeated three times and randomized within each set. Each trial began with a static pattern for 1 s, followed by a stationary sinewave grating for 1 s, followed by 10 s of motion of that same sinewave grating in either the TN or NT direction. As we did not wish to physically cover an eye for monocular trials and risk injury to the birds, two methods were used for monocular stimulation: 1) the monitor for the unstimulated eye was either turned gray after the static pattern, 2) or changed to the sinewave grating but held stationary. We analyzed data from the former method as the latter represents an unequal speed combination, though data from both methods were qualitatively similar. For binocular trials, the stimulus direction for each monitor was set independently to produce either regressive stimulation (TN & TN), progressive stimulation (NT & NT), or rotation (TN & NT) or (NT & TN). We refer to binocular NT & TN combinations as rotational stimulation; however, as our stimulus monitors did not curve around the animal, these conditions technically represent pseudo-rotation.

Analysis of oculomotor behavior

To extract the position of the pupil from each video, we used a combination of thresholding and contour segmentation in python (version 3.11.8). We trained a custom DeepLabCut [40] model (ResNet 152) to track the corneal reflection and eyelids. To track the pupil, we used data from the same DeepLabCut model to determine a region-of-interest around the eye, upscaled this portion of the video frame, and then used thresholding and contour detection in OpenCV (v. 4.9.0) for pupil detection.

Analysis of raw pupil position was performed in MATLAB R2023b (Mathworks, Natick, USA). We first removed positions if the confidence from DeepLabCut was low (< 95%). We next used the vertical distance between upper and lower eyelids to detect blinking, removing pupil position data 100 ms before and 400 ms after each detected blink to eliminate blink-artifacts. We then removed periods of pupil position data when the pupil area was changing rapidly, as asymmetrical dilation or constriction of the pupil could be erroneously interpreted as changes in position. Finally, we filled small gaps in the pupil position stream using linear interpolation to facilitate saccade detection.

To convert the filtered pupil position from pixels to degrees, we set the pupil position for each trial relative to the median position prior to motion onset. Pupil displacement was then converted to millimeters based on objects of known length in the video frame, digitized for each recording session using ImageJ 1.53k (NIH, USA). We removed data for trials in which the corneal reflection moved more than 1 mm, as this indicated that the bird rolled its head during the trial (6.7% of total trials). Finally, we converted displacement to degrees based on the assumption that the avian eye rotates around the midpoint between the pupil and the back of the eye [41]. We determined this value (2.181 ± 0.102 mm; mean ± sd) by measuring the half distance between the outer edge of the scleral ring and the medial margin of the eye for five eyes from four specimens available from previous lab projects. The eye-in-head position was then calculated as θ = tan-1 (rel. position mm / 2.181 mm). Importantly, this assumption does not impact the main conclusions of this work and resulted in a similar range of movement as determined using alternative methods [42].

We next automatically removed saccades using peak detection, analyzing the median-filtered velocity of the pupil for spikes above a speed threshold. This speed threshold was manually adjusted (between 5-15 deg s-1) for each session based on visual inspection of pupil position across trials.

Inspection of the parallel streams of data from each eye indicated that blinks and saccades often occurred simultaneously in both eyes, as in other birds [43,44]. For this reason, we considered a blink or saccade in either eye to constitute a break in OKR and computed the velocities of the eyes between each of these breaks. This allowed us to compare, for the same set of time windows, the motion occurring simultaneously in each eye. To compute OKR velocities, we fit each segment between breaks with a linear regression and extracted the slope (deg s-1), fitted values (deg), and quality of the fit (Root Mean Squared Error; RMSE). The magnitudes (deg) of rapid eye movements (blinks and saccades) were determined using these fitted values as the offset between the last and first elements in each adjacent segment.

Statistical analysis

We performed plotting and hypothesis testing in R version 4.4.0. For comparisons of OKR between conditions, we removed segments with poor fit quality (RMSE > 0.35) or with many (>10%) missing data points for either eye (together 12.7% of total fits). Significance was assessed by computing bootstrapped means and 95% confidence intervals (1000 iterations) using the R package boot [45]. Unless otherwise specified, bootstrapped values were computed by pooling fitted values from both eyes and bootstrapping across trials and individuals.