When two cardiologists trace the left ventricle on the same ultrasound loop, their contours differ. Clinical ejection-fraction readings differ by 7–13% between observers (Chao et al.). The medical imaging field explains this disagreement in two ways. One is that the image is ambiguous, through speckle, low contrast, or motion blur. The other is that raters have stable personal styles. This proposal tests a third, untested explanation. At any moment, part of the heart wall falls away from the center of the annotator's gaze. Human vision in that outer region, the periphery, handles shape and motion by different rules than central vision does, so contour errors may follow where each annotator looked. If that is true, disagreement is partly predictable from a single annotator's eye-tracking record, with no need for a second expert.
The core claim is narrow and falsifiable. We predict that the eccentricity of a boundary segment relative to the annotator's fixations while reviewing the loop explains disagreement between raters beyond what image content explains. Eccentricity here means its angular distance from where the eye was pointing. We further predict that the sign of this effect depends on whether the annotator saw the wall moving or only a still frame. The test is a causal manipulation that holds the image fixed and moves only the fixation point. Its outputs are a gaze-conditioned model of per-segment boundary uncertainty and a review tool that flags the wall segments a second expert would most likely trace differently.
Why this is not already done
Gaze has entered medical segmentation, but only as a pointer. Gaze-Assisted Medical Image Segmentation fine-tunes MedSAM, a medical version of the Segment Anything promptable segmentation model, using gaze points as prompts on static CT slices. It models gaze imprecision only as random jitter, with 80% of points placed inside the organ and 20% outside. Temporal Gaze Dynamics as Zero-Shot Prompts turns the gaze stream into a static heatmap that generates box prompts for SAM-2. The authors acknowledge that this discards the dynamics, and their result drops from 0.844 Dice with manual boxes to 0.759 with real gaze. Both papers use one reference mask and one gaze source, and both treat everything outside the fixation as unseen. Gu et al. come closest to linking gaze with inter-observer agreement. They pool fixations from 14 lay viewers into consensus mitosis labels, but on stationary pathology images and again under a foveal-only assumption.
The multi-rater modeling line treats disagreement as a property of raters and images, never of viewing. Pionono gives each rater a Gaussian in an 8-dimensional latent space whose mean captures systematic bias and whose variance captures inconsistency. The Latent Doctor Model infers a doctor's style from a set of that doctor's past labels and conditions it on image features. DiFF weights raters by how well their masks support downstream diagnosis. All three were built and tested on static images, all need many raters per image, and none can say why a particular rater was wrong on a particular segment. Our proposal keeps their goal, predicting the label distribution, and adds a measured, mechanistic input: the viewing geometry.
The perceptual grounding, including the part that cuts against us
Peripheral-vision research makes two predictions that conflict, and the experiment is designed to separate them. On one side, observers tracking several moving objects anchor their gaze on the empty centroid of the group rather than on the objects. They spend about 30% of trial time there versus about 10% on the targets, and they detect over 80% of target changes at 15–20° eccentricity without looking directly at them (Vater et al., 2016; Vater et al., 2017). The usable field of view stays wide for motion even where shape and colour judgments fail beyond 30° (Bickerdt et al.). In natural scenes, motion separates objects from background more reliably in the periphery than near the center (Friedman et al.). These findings motivate the hypothesis that a sonographer fixes on the ventricular cavity and monitors the wall peripherally. They also predict that motion rescues peripheral boundary perception.
On the other side, two findings supply directional predictions for bias. Observers widen their tolerance for misaligned contour pieces in proportion to their own sensory noise, and that noise grows with eccentricity (Zhou, Acerbi and Ma; noise-threshold correlation r = 0.88). Under peripheral uncertainty, observers also lean on a built-in assumption that things move slowly, which distorts perceived displacement (Chen, Zhang and Körding). Together these predict that peripherally viewed wall segments are traced with more scatter. They also predict that those segments are traced with an underestimated excursion between end-diastole and end-systole, which would bias ejection fraction downward.
The proposal does not assume which side wins. It measures the interaction between eccentricity and viewing condition, comparing a moving loop with a still frame.
Research Program
We build a dataset of contours traced under eye tracking. We sample 80 apical four- and two-chamber loops from CAMUS, stratified by image quality, plus 40 from EchoNet-Dynamic. Ten to twelve annotators trace end-diastolic and end-systolic left-ventricular contours on every case under a screen-mounted eye tracker. The annotators are sonographers and cardiology fellows, with prior echo experience recorded as a covariate. Each annotator first reviews the loop freely, then traces the frames. To measure intra-observer variability, 15% of cases are repeated after a delay. Every contour is resampled into 24 radial segments around the ventricle's long axis. For each segment we compute three quantities. The first is the spread of its radial position across raters, in millimetres. The second is each rater's signed deviation from the median. The third is a gaze feature: the time-weighted distribution of that segment's eccentricity relative to the rater's fixations during loop review, with dwell time at eccentricities below 2° separated out as "foveated."
The model is deliberately small. For segment $s$ of case $c$ traced by rater $r$, we model the squared deviation from the cross-rater median as
$$ \mathbb{E}[d_{s,c,r}^2] = g_\theta(x_{s,c}) + \beta_1 \bar{e}_{s,c,r} + \beta_2 \bar{e}_{s,c,r}\, v_{s,c} + u_r , $$
The terms are defined as follows. $g_\theta(x_{s,c})$ is a learned function of image content around the segment, covering local contrast, speckle statistics, and the image-only uncertainty from a multi-rater network. $\bar{e}_{s,c,r}$ is that rater's mean eccentricity for the segment during review. $v_{s,c}$ is the segment's wall speed from optical flow. $u_r$ is a per-rater offset. The scientific question is whether $\beta_1$ and $\beta_2$ are reliably nonzero once $g_\theta$ is as strong as we can make it. A parallel model on signed deviation tests the predicted inward-excursion bias. At deployment, the fitted model takes one annotator's contour and gaze record and outputs a per-segment map of where other experts would likely disagree. The tool flags the top segments for a second look.
The causal arm uses a gaze-contingent display. A fixation cross appears at controlled positions inside or beside the ventricle. If the eye leaves the cross by more than 1.5°, the loop pauses. Annotators then judge or trace the wall at targeted segments placed at 2°, 6°, 10° and 14° eccentricity, in two viewing conditions: a playing loop and the static end-systolic frame. Because the image pixels are identical across eccentricity conditions, any change in scatter or bias is attributable to viewing geometry and not to image ambiguity. This experiment answers the "it's just speckle" objection directly.
Baselines
The naive reference point assigns every segment the same disagreement level. A second reference ranks segments by a hand-built image-ambiguity score from local gradient magnitude and speckle contrast. These set the floor that any useful predictor must clear.
The strongest competitor is an image-conditioned multi-rater network. We use the conditional Latent Doctor Model (LDM) and Pionono (Schmidt et al.). These are the best existing methods for predicting how experts disagree from image content and rater identity alone, which is exactly what our gaze terms must beat. Both will be trained on the same multi-rater contours with the same rater-held-out cross-validation splits. Each gets the same hyperparameter search budget as our model and the same backbone family, a U-Net with a ResNet34 encoder as in Pionono. Their predicted per-pixel uncertainty is converted to the same 24-segment spread. The best reported numbers for these methods come from static pathology. Pionono reaches unweighted Cohen's kappa 0.758 on Gleason 2019. CLDM reaches KL divergence 0.461 there and a Spearman correlation of up to 0.728 between model entropy and pathologist entropy on ISUP grading. No reported result was found on Lacuna for multi-rater disagreement prediction on echocardiography video. We will therefore report these baselines on our own data rather than borrow a number.
Beating them means three things. First, adding gaze terms raises the held-out-rater Spearman correlation between predicted and observed segment spread by a margin whose bootstrap confidence interval excludes zero. Second, the gaze terms explain significant partial variance after $g_\theta$ absorbs the network's uncertainty. Third, segments flagged by the gaze model remove more of the cross-rater ejection-fraction disagreement on re-review than the same number of segments flagged by LDM uncertainty.
For a segmentation reference, we run MemSAM (Deng et al.), which reports 93.31% mean Dice on CAMUS-Semi. Each rater's first-frame contour is used as its prompt. This measures how much human disagreement survives propagation by a memory-based video model.
Two ablations isolate the key ingredient. The first replaces continuous eccentricity with a binary "ever foveated" indicator and with raw dwell time. The second removes the eccentricity-by-speed interaction. If the binary indicator does as well as continuous eccentricity, the finding reduces to "unexamined segments are traced badly," which is useful but not a peripheral-vision result, and we will report it that way.
Why not just…?
Why not just blame speckle and low contrast and model image ambiguity directly? This is the main threat. Deng et al. and Schilling et al. both report that ultrasound and stress-MRI image quality limit expert consistency. Our observational analysis gives the image model every advantage by putting a learned image-uncertainty network inside $g_\theta$. The gaze-contingent experiment then holds the pixels fixed. If scatter does not change with forced eccentricity on identical frames, the image-ambiguity account wins and the proposal's central claim is rejected.
Why not just collect more raters and fuse them with STAPLE or majority vote? Fusion reduces noise when raters are plentiful. Gu et al. needed 14 viewers to reach AP 0.83 against 0.85 for experts. Multi-frame cardiac contours from many experts per patient are rarely available in practice. DiFF also shows that consensus discards diagnostically useful variation, with AUC gains of up to 2.73% from keeping it. Our tool targets the single-annotator case and estimates the missing second opinion from gaze. If fusing two raters already matches what one rater plus gaze predicts, our tool has little practical value. The re-review comparison tests exactly this.
Why not just use a promptable video model and stop worrying about humans? Foundation models narrow but do not close the gap. Fine-tuned SAM still shows 5–7% ejection-fraction error (Chao et al.). MemSAM's authors report that a misplaced first prompt corrupts the memory across the whole cycle. Real gaze prompts also cost about 0.085 Dice relative to clean boxes (Shmykova and Pershin). Running MemSAM from each rater's own initial contour tells us how much human variability the model inherits. If propagated disagreement is negligible, human contouring variability matters less than we claim.
Why not just use dwell time instead of eccentricity? It is the simplest gaze feature, and it might suffice. The ablation decides. A peripheral-vision account specifically predicts graded scatter with eccentricity. It also predicts that this effect is modulated by wall speed and by whether the loop is playing, which dwell time alone cannot express.
Decisive Checks
Three results would settle the question. In the gaze-contingent experiment, the first check is whether boundary scatter rises monotonically from 2° to 14° on identical frames. The second check is whether the rise is smaller for playing loops than for still frames, as the motion-rescue findings predict, or larger, as a slowness-prior distortion of excursion would predict, with ejection fraction biased downward when the wall is peripheral. The third check is whether, in the observational data, eccentricity terms add held-out predictive power over CLDM and Pionono for where a new rater will disagree. A positive first and third result would give multi-rater modeling a measurable, collectible cause of disagreement. A null first result would be equally informative, because it would show that disagreement in echo contouring is about the image, not the viewer.
One practical risk is that a ventricle on a normal monitor spans only a few degrees. Natural eccentricities may then be too small to produce an effect. We address this by magnifying the display in the causal arm and by reporting the natural eccentricity range we actually observe, so that a null result is not overinterpreted.
References
- Foundation versus Domain-Specific Models for Left Ventricular Segmentation on Cardiac Ultrasound
- Gaze-Assisted Medical Image Segmentation
- Temporal Gaze Dynamics as Zero-Shot Prompts for Volumetric Medical Segmentation
- Supporting Mitosis Detection AI Training with Inter-Observer Eye-Gaze Consistencies
- Probabilistic Modeling of Inter- and Intra-observer Variability in Medical Image Segmentation
- The Latent Doctor Model for Modeling Inter-Observer Variability
- Calibrate the Inter-Observer Segmentation Uncertainty via Diagnosis-First Principle
- Detecting single-target changes in multiple object tracking: The case of peripheral vision
- Examining the functionality of peripheral vision
- Beyond the tracked line of sight - Evaluation of the peripheral usable field of view in a simulator setting
- Statistical regularities in natural scenes that support figure-ground segregation by neural populations
- The role of sensory uncertainty in simple contour integration
- Speed Constancy or Only Slowness: What Drives the Kappa Effect
- MemSAM: Taming Segment Anything Model for Echocardiography Video Segmentation
- Assessment of deep learning segmentation for real-time free-breathing cardiac magnetic resonance imaging at rest and under exercise stress