Eye movement dynamics are a key factor for intra-saccadic motion perception

Abstract

The objectives of this study were to investigate, using multimodal neuroimaging techniques, the involvement of an extended network of visual (V1, V2, V3, hV4 and MT/V5) and oculomotor regions (IPS and FEF) in intra-saccadic motion perception, and how the activity within these regions varies with retinal temporal frequency. It confirms the fundamental involvement of the magnocellular pathway and the mediating role of retinal temporal frequency in intra-saccadic motion perception. Perceptual efficacy is maximized within a specific temporal frequency bandwidth aligned with the tuning properties of magnocellular motion detectors. These perceptual phenomena are tightly coupled with individual oculomotor kinematics, such as saccade peak velocity and post-saccadic oscillations, demonstrating that eye movement dynamics are integral modulators of intra-saccadic visual processing.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-39420-8.

Introduction

The analysis of a visual scene by a human observer involves numerous cortical and subcortical mechanisms that remain only partially understood. While most studies focus on ocular fixation (eye stabilization), aiming to characterize its spatial position and duration during scene exploration^1,2, ocular saccades (rapid eye movements) are often neglected and interpreted merely as transitional intervals between successive fixations. However, a dual perspective places the saccade at the heart of the active visual exploration process³, during which the observer must plan both “when” and “where” to move their gaze and then execute this movement from one fixation point to another. Despite these saccadic eye movements, our perceptual experience remains stable⁴, a phenomenon called ‘saccadic omission’⁵.This raises a key question: which mechanisms account for the fact that motion of the environment is not consciously perceived during a saccade? The prevailing belief holds that information processing during saccades is actively suppressed by an extra-retinal signal⁶. However, compelling psychophysical evidence^7,8 and neuroimaging studies⁹ support the hypothesis that intra-saccadic movement perception is active and functional, and may play a critical role in both perceptual and oculomotor functions¹⁰. Recent analyses have shown that intra-saccadic smear, far from being entirely imperceptible, can be detected and modelled, suggesting a more complex intra-saccadic visual experience than previously thought, with a direct coupling between eye movement dynamics and perceptual evolution of the smear¹⁰. Idrees et al. (2020)¹¹ demonstrated that perceptual saccadic suppression arises from the coexistence of passive mechanisms, related to retinal image smear caused by rapid shifts, and active mechanisms involving extra-retinal signals. Their findings reveal that suppression starts in the retina through specific circuits that attenuate visual sensitivity independently of motor commands, whereas extra-retinal contributions modulate the duration rather than initiate suppression. Recently, Pomè et al. proposed a novel theory suggesting that habituation within sensory motor system renders it insensitive to irrelevant self-produced visual motion¹². Thus, both passive and active hypotheses coexist and together explain perceptual stability during saccades. Open questions concern the possible involvement of the earliest stages of the magnocellular visual pathway in processing extra-retinal speed signals during saccades where retinotopic coding of pursuit signals has been demonstrated¹³. Additionally, the middle temporal cortex and particularly the MT/V5 visual area, a region critically implicated in motion perception, is thought to process motion information per se during saccades, but its exact role during these rapid eye movements remains to be fully clarified.

To date, these issues have never been investigated in humans by appropriately combining multimodal neuroimaging technique: electroencephalography (EEG), functional magnetic resonance imaging (fMRI), and eye-tracking. Our study aims to elucidate the neural substrates of intra-saccadic motion perception and investigate their relationship with conscious intra-saccadic motion perception, the temporal frequency of the stimulus projected onto the retina, and individual oculomotor kinematics¹⁴. To achieve this, we replicated the pioneering experiment of Castet and Masson (2000)⁷ by simultaneously recording EEG and eye-tracking signals while participants performed horizontal rightward saccades, in the same direction as a visual stimulus optimized for the magnocellular pathway (a low spatial frequency vertical sinusoidal grating moving at high speed from left to right). The central idea was to compare conditions with and without conscious detection of intra-saccadic motion, while ensuring a horizontal rightward saccade in every trial. This may reveal pronounced activity in the human motion processing complex (MT/V5) during saccades¹⁴. More broadly, it could shed light on how modulation of the magnocellular pathway relates to conscious motion perception as a function of the temporal retinal frequency¹⁶. This study extends Nicolas et al. 2021⁹ by revealing the central role of temporal retinal frequency and by refining and extending the delineation of visual areas to better characterize how information is processed.

Results

Data corpus

The data corpus analyzed comprised two experimental conditions (see Fig. 1a for the visual stimulus and Fig. 1b for time course): STIM, where stimulus motion was aligned with the left-to-right saccade direction, and CTRL, where stimulus motion was opposite to the saccade direction. Participants performed 300 STIM trials and 180 CTRL trials in random order. For each trial, they executed a visually-guided left-to-right saccade toward a green target (Fig. 1b) at a randomly drawn eccentricity (1.5°-6.5°) and then indicated by keypress whether they had perceived stimulus motion during the saccade, labeling trials as STIMSs / CTRL-Detect (motion perceived) or STIM-NoDetect / CTRL-NoDetect (motion not perceived).

Fig. 1.

Overview of the time course of the experiment. (a) Timeline of the evolution of the spatiotemporal stimulus during a trial. Only the spatial features of the stimulus can be represented and the blue arrow indicates the direction of movement; (b) Layouts during a trial on the STIM condition.

Analyzed data

After EEG pre-processing and eye-tracking analysis, six participants were excluded: two for lacking task comprehension, two for excessively noisy EEG data, and two for too few valid trials (eye-tracking and EEG variance criteria). Table 1 displays the number of remaining trials and mean trials per participant (n = 29) for each condition. On average, 32.9% of trials were rejected. A majority of rejected trials resulted from non-compliance with six strict saccade selection criteria (see Supplementary Material (SM) A) ensuring genuine intra-saccadic motion perception. The remaining trials were rejected for poor signal quality post-preprocessing (ICA artifact rejection, etc.), standard in EEG-eye tracking studies.

Table 1.

Total number and average number (± s.e.m) per participant (n = 29), of remaining trials at the end of the EEG and eye-tracking pre-processing steps for each configuration.

Condition	STIM		CTRL
Perception	Detect	NoDetect	Detect	NoDetect
Total number of trials	3637	2209	331	3171
Mean (± s.e.m)	125.4 ± 10.0	76.2 ± 8.1	11.4 ± 3.6	109.3 ± 5.7

We focused our analyses on two key configurations for evoked potential estimation: STIM-Detect (trials where participants perceived stimulus motion during saccades) and CTRL-NoDetect (trials where they did not perceive motion). These STIM-Detect and CTRL-NoDetect configurations were balanced in trial count, ensuring comparable signal-to-noise ratios. CTRL-Detect trials were rare (6.3%) and excluded from all analyses. Model adequacy was confirmed by events-to-parameter ratios exceeding standard recommendations. STIM-Detect averaged 125 valid trials per participant (34,375 observations: 125 × 275 samples), yielding a 29:1 ratio (34,375 ÷ 1,175 regressors). CTRL-NoDetect averaged 110 trials (30,250 observations), yielding 25:1—both comfortably above the 10–15:1 usual minimum for robust GLM estimation.

Eye movements

From trials included in the analyzed data, we extracted five key oculomotor features of the unique main saccade per trial for the three configurations (STIM-Detect, STIM-NoDetect, and CTRL-NoDetect): the relative error in main saccade size, saccade latency, saccade size, saccade peak velocity, and the main sequence relationship. In addition, we also examined distributions of temporal frequency, computed from the stimulus speed independent of direction relative to saccade peak velocity (see Eq. 3, (Methods)), and saccade size for the STIM-Detect and CTRL-NoDetect configurations, as these served as predictors for EFRP estimation via spline regression (Methods).

Relative error on the main saccade size

For each trial, the relative error of the main saccade was defined as the absolute difference between the requested saccade size and the actual saccade size, divided by the requested saccade size. This measure was used to verify that participants executed the saccades as instructed toward targets at eccentricities randomly drawn from [1.5–6.5]°. The mean relative error (± s.e.m) of the main saccade was 13.62% (± 0.44%), indicating participants performed the task accurately. The two excluded participants displayed substantially higher errors (55.7% and 35.8%, respectively).

Latency of the main saccade

The latency of the main saccade was defined relative to the onset of the visual cue (timestamp , Fig. 4). Saccade latency served as an indicator of participants’ attentiveness to the cue signaling them to initiate the saccade. The mean saccade latency (± s.e.m) across all participants and configurations was 302.52 ms (± 10.79).

Fig. 4.

Right MT/V5. EFRPs over time, reconstructed at the temporal frequency  corresponding to the maximum Cohen’s d value within the cluster: (a) Cluster “Early”, range “Low”,  Hz; (b) Cluster “Late”, range “Low”,  Hz. EFRPs over temporal frequency, reconstructed at the time point t corresponding to the maximum Cohen’s d value within the cluster: (c) Cluster “Early”, range “Low”,  ms; (d) Cluster “Late”, range “Low”,  ms. The solid line represents the mean across participants, and the shaded area indicates the standard error of the mean (s.e.m) at each point, based on individual participant averages. (e) Cluster location in the time × temporal frequency domain.

Size and peak velocity of the main saccade

The mean size of the main saccade was 2.85° (± 0.04), and the average peak velocity was 226.74°/s (± 6.25). Table 2 presents the mean (± s.e.m) values of saccade size and peak velocity for each configuration.

Table 2.

Mean (± s.e.m) of saccade size (in °) and peak velocity (in °/s) based on individual means.

Condition	STIM		CTRL
Perception	Detect	NoDetect	NoDetect
Saccade size (°) (± s.e.m)	3.21 ± 0.05	2.24 ± 0.05	2.85 ± 0.04
Peak velocity (°/s) (± s.e.m)	247.03 ± 7.23	194.20 ± 7.39	227.61 ± 6.30

A repeated-measures ANOVA on main saccade size with configuration as a within-subject factor revealed a significant effect of configuration, F(2, 56) = 187.99, p < 0.0001, ηp² = 0.870. Post-hoc Bonferroni comparisons showed that saccade size was significantly larger in STIM-Detect than in both CTRL-NoDetect (p < 0.0001) and STIM-NoDetect (p < 0.0001), and also significantly larger in CTRL-NoDetect than in STIM-NoDetect (p < 0.0001).

Similarly, a repeated-measures ANOVA on peak velocity revealed a significant effect of configuration, F(2, 56) = 134.63, p < 0.0001, ηp² = 0.828. Bonferroni-corrected post-hoc tests indicated peak velocity was significantly higher in STIM-Detect compared to both STIM-NoDetect (p < 0.0001) and CTRL-NoDetect (p < 0.0001), and significantly higher in CTRL-NoDetect than in STIM-NoDetect (p < 0.0001). Larger/faster saccades in STIM-Detect may represent kinematic features enabling motion perception via optimal retinal frequency matching (Castet & Masson, 2000), compared to the smaller and slower saccades in STIM-NoDetect that appear to reduce intra-saccadic perception despite identical stimulus direction.

Population characterization based on eye kinematics

The main sequence, describing the relationship between saccade size and peak velocity^17,18, was modeled using a two-parameter exponential function () with parameters (asymptotic peak velocity) and (rate of saturation defining the shift from linear growth to saturation). These parameters were estimated for each participant based on saccade size and corresponding peak velocities. The slope at the origin, [s^-1], representing the initial linear portion for small saccades, was then computed from these parameters. Significant correlations were observed between parameters:  negatively correlated with  (r = -0.607, p = 0.0005) and positively with (r = 0.830, p < 0.0001). Figure 2a illustrates the distribution of  and across participants, revealing a continuum of main sequence profiles. Then, a K-Means clustering analysis with k = 2 identified two groups, A and B, with the corresponding average main sequences shown in Fig. 2b.

Fig. 2.

Main sequence (a) Scatter plot of the two parameters () depending on the two groups obtained after unsupervised clustering; (b) Exponential fitting of the peak velocity as a function of saccade size (line: mean, envelop: ± s.d) for the two groups of participants.

Based on this main sequence modeling distinction, comparisons of saccade duration, size, peak velocity, and frequency of post-saccadic oscillations are presented in SM B. Key findings include significantly greater saccade duration in “Group A” than in “Group B” (t(27) = 2.671, p = 0.0127), higher peak velocity (t(27) = 7.266, p < 0.0001), and increased percentage of post-saccadic oscillations (t(27) = 4.669, p < 0.0001). Conversely, saccade size was significantly smaller in “Group A” compared to “Group B” (t(27) = –2.779, p = 0.0098).

Behavioral detection rates P(STIM-Detect|STIM) confirmed monotonic decreasing functions with increasing retinal frequency (Fig. S2; SM C). The 0.5 detection rate shows a 5 Hz shift toward higher frequencies for Group B compared to Group A consistent with main sequence relationships.

Predictor’s distribution

A spline regression was used with three B-splines of order two to capture non-linear relationships between saccade size and EEG amplitude while keeping the number of parameters manageable relative to the size of datasets. B-splines were positioned according to the distribution range of the predictor for each participant (see SM D). Table 3 summarizes the minimum, median, and maximum values of the predictor distributions for the two configurations of interest, considering all participants (“All”) and the two participant groups (“A” and “B”) identified via clustering of the main sequence parameters.

Table 3.

Mean (± s.e.m) of temporal frequency (in Hz) and saccade size (in °) based on individual minima, medians and maxima, from all participants (“All”) and from both groups (“A”, “B”).

Predictor		Temporal frequency (Hz)		Saccade size (°)
Group / Configuration	Group	STIM-Detect	CTRL-NoDetect	STIM-Detect	CTRL-NoDetect
Min (mean ± s.e.m)	All	4.84 ± 0.74	5.86 ± 0.95	1.35 ± 0.05	1.18 ± 0.03
	A	2.38 ± 0.09	2.48 ± 0.15	1.33 ± 0.07	1.12 ± 0.05
	B	7.48 ± 1.19	9.48 ± 1.45	1.38 ± 0.08	1.24 ± 0.04
Median (mean ± s.e.m)	All	17.69 ± 1.24	21.00 ± 1.11	3.25 ± 0.07	2.79 ± 0.05
	A	12.84 ± 1.06	16.51 ± 0.88	3.16 ± 0.09	2.72 ± 0.07
	B	22.88 ± 1.24	25.81 ± 1.07	3.35 ± 0.10	2.88 ± 0.06
Max (mean ± s.e.m)	All	34.91 ± 1.28	39.15 ± 0.86	5.26 ± 0.13	5.00 ± 0.09
	A	30.99 ± 1.72	37.02 ± 1.35	4.96 ± 0.11	4.75 ± 0.10
	B	39.10 ± 1.11	41.45 ± 0.63	5.58 ± 0.21	5.26 ± 0.12

As expected, variation in main sequence parameters resulted in a corresponding shift in temporal frequency , where a shift from higher (“Group A”) to lower (“Group B”) values () induced a shift from lower to higher temporal frequencies. This aligned with theoretical expectations: for a given saccade size , the difference between saccade peak velocity  and stimulus velocity  was greater in “Group B” than in “Group A”, resulting in higher computed temporal frequency () for “Group B” than for “Group A” (see Eq. 3).

However, for predictor-based statistical analyses, a common range of variation had to be defined across all participants. For each predictor, a trade-off was necessary between maximizing the range of variation and including the largest possible number of participants with data in both configurations (Fig. 3).

Fig. 3.

Number of participants with data in both configurations as a function of the predictor: (a) temporal frequency F_T; (b) saccade size S_z. See text for details on data points.

For permutation tests using temporal frequency  as predictor, the range of variation encompassing data from all participants was very narrow, specifically between 19.55 Hz and 20.45 Hz (Fig. 3a). It follows that a frequency of Hz constituted a common frequency across all participants. Consistent with prior clustering, which showed temporal frequency distributions concentrated around either lower or higher frequencies, the frequency range was divided into two intervals: [6–20] Hz (range “Low”) and [20–31] Hz (range “High”). This ensured 20 participants (69%) had data in both intervals. This data-driven dichotomization maximized participant inclusion while ensuring stable permutation tests across the natural 20 Hz pivot.

Since clustering was based on main sequence parameters, it did not affect saccade size distribution. Therefore, for permutation tests with saccade size  as predictor, no such division was required, as all participants faced identical saccade sizes, specifically between 2.0° and 4.1° (Fig. 3b)—highlighting peak velocity variability as the key dichotomizing factor.

EEG data results

In the following sections, we analyze EEG data in electrode space (time × electrodes) and source space (per ROI) using non-parametric paired permutation tests (FieldTrip toolbox) with cluster-based correction for multiple comparisons (see Methods section). Eye-Fixation Related Potentials (EFRPs)—i.e., EEG responses time-locked to fixation onset (main saccade offset)—were estimated via spline-based deconvolution to address saccade size confounds. Consequently, EFRPs were modeled as a function of time (200–700 ms post-fixation onset) and predictors (temporal frequency or saccade size, depending on analysis). Cortical sources were reconstructed in visual areas along the dorsal and ventral streams, as well as in oculomotor areas.

Results in electrode space

Two-tailed non-parametric cluster-based permutation tests were applied to the eye fixation-related potential (EFRP) data at the specified temporal frequency (20 Hz) to include all participants. Analyses were conducted across time and space on all electrodes, with configuration (STIM-Detect, CTRL-NoDetect) as a within-subject factor. To separately assess the saccadic spike, complex Lambda-N1, and P300 component, three latency windows were examined. Refer to SM E for detailed configurations and results.

Three spatio-temporal clusters were identified:

1. A significant negative cluster showed a more negative amplitude for STIM-Detect vs. CTRL-NoDetect between approximately 150—250 ms, mainly in occipital regions.

2. A significant positive cluster revealed a more positive amplitude for STIM-Detect vs. CTRL-NoDetect between approximatively 300—450 ms, mainly in central region.

3. A trend positive cluster indicated a slightly more positive amplitude for STIM-Detect vs. CTRL-NoDetect between approximatively 150—180 ms, mainly in left anterior region.

No significant clusters were found at saccadic spike latencies and in occipital regions during the Lambda component latency.

Results in source space for visual areas

Latency windows for permutation tests in source space were adapted based on spatio-temporal clusters identified in electrode space. Since significant clusters appeared only during early and late latency windows, analysis in source space was restricted to these periods. Although no significant difference was observed at the Lambda component latency in electrode space, this component is strongly associated with visual perception; hence, it was included within the early latency window for source analyses. Consequently, the early latency window was defined as [50—250] ms and the late latency window as [250—450] ms, the late window starting after the early one and ending with the late cluster corresponding to the P300 component observed in electrode space.

In line with the temporal frequency distribution, ranges were divided into range “Low” from 6 to 20 Hz and range “High” from 20 to 31 Hz. Using these configurations, four one-tailed non-parametric cluster-based permutation tests were performed on EFRP data, comparing STIM-Detect and CTRL-NoDetect for each visual area in both hemispheres. Cluster-based permutation test results for each visual area are presented hereafter in the respective table, while the associated Figures are provided in SM F.

Dorsal visual areas (V1d, V2d, V3d, MT/V5)

Cluster-based permutation tests were conducted to compare STIM-Detect and CTRL-NoDetect configurations for each dorsal visual area (V1d, V2d, V3d and MT/V5) in both hemispheres, focusing on early [50—250] ms and late [250—450] ms latency windows and low [6–20] Hz and high [20–31] Hz temporal frequency ranges. Results are summarized in Table 4, with corresponding Figures (Fig. S8, S9, S10, S11) in SM F.

Table 4.

Results of cluster-based permutation tests for the right and left dorsal visual areas

Dorsal visual areas	Temporal Frequency Range	p-value	Cohen’s d Range	Cohen’s d Mean	Maximum Effect (ms, Hz)
Right V1d	Low	0.0319	0.317–0.943	0.508	~ 100 ms, ~ 10 Hz
	High	0.0345	0.317–0.736	0.448	~ 100 ms, ~ 30 Hz
Left V1d	Low	0.0172	0.316–1.069	0.511	~ 100 ms, ~ 10 Hz
Right V2d	Low	0.0212	0.316–0.914	0.509	~ 100 ms, ~ 10 Hz
Left V2d	Low	0.0300	0.316–0.763	0.486	~ 100 ms, ~ 10 Hz
Right V3d	-	-	-	-	-
Left V3d	Low	0.0012	0.317–1.016	0.499	~ 100 ms, ~ 10 Hz
Right MT/V5	Low (15–20 Hz)	0.0262	0.317–0.707	0.437	~ 100 ms, ~ 18 Hz
	Low	0.0056	0.319–0.922	0.515	~ 300 ms, ~ 10 Hz
Left MT/V5	-	-	-	-	-

In the early latency window (~ 100 ms) around the Lambda component, significant differences were observed with STIM-Detect eliciting a more positive response than CTRL-NoDetect across several dorsal visual areas. More specifically, in V1d and V2d, significant clusters were found for low temporal frequencies in both hemispheres. Additionally, in V1d, a significant effect was also found for high temporal frequencies only in the right hemisphere. For V3d, a significant effect was found for low temporal frequencies only in the left hemisphere and for MT/V5, a significant cluster was found for low temporal frequency range with a narrower frequency range only in the right hemisphere (Fig. 4a and c).

In the late latency window (~ 300 ms) around the P300 component, a significant difference was observed, with STIM-Detect eliciting a more positive response than CTRL-NoDetect, only for MT/V5 in the right hemisphere and low temporal frequency range (Fig. 4b and d).

No significant effect was observed in the right V3d and in the left MT/V5 whatever the latency window.

Ventral visual areas (V1v, V2v, V3v, hV4, PHC)

Cluster-based permutation tests were conducted to compare STIM-Detect and CTRL-NoDetect configurations for each ventral visual area (V1v, V2v, V3v, hV4 and PHC) in both hemispheres, focusing on early [50—250] ms and late [250—450] ms latency windows and low [6–20] Hz and high [20–31] Hz temporal frequency ranges. Results are summarized in Table 5, with corresponding Figures (Fig. S12, S13, S14) in SM F.

Table 5.

Results of cluster-based permutation tests for the right and left ventral visual areas. When significant differences were observed across the entire frequency range, i.e. [6–20] Hz or [20–31] Hz, only “Low” or “High” is indicated respectively. Otherwise the frequency values are specified directly in the table. Trends effects are shown in italics.

Ventral visual areas	Temporal frequency range	p-value	Cohen’s d range	Cohen’s d mean	Maximum effect (ms, Hz)
Right V1v	Low	0.0541	0.316—0.678	0.469	~ 100 ms, 10 Hz
	High	0.0345	0.323–0.860	0.550	~ 100 ms, ~ 30 Hz
Left V1v	Low	0.0180	0.321–0.973	0.474	~ 100 ms, ~ 10 Hz
		0.0189	0.322–0.808	0.473	~ 300 ms, ~ 10 Hz
	High	0.0371	0.321–0.669	0.439	~ 100 ms, ~ 30 Hz
Right V2v, V3v	-	-	-	-	-
Left V2v, V3v	-	-	-	-	-
Right hV4	-	-	-	-	-
Left hV4	Low	0.0117	0.317–0.601	0.437	~ 100 ms, ~ 10 Hz
Right, Left PHC	-	-	-	-	-

In the early latency window (~ 100 ms) around the Lambda component, significant differences were observed with STIM-Detect eliciting a more positive response than CTRL-NoDetect across several ventral visual areas. Especially for V1v and low temporal frequency range, a significant cluster was found in the left hemisphere and a trend one in the right hemisphere. For V1v and high temporal frequency range, significant clusters were observed for both hemispheres. In hV4, a significant cluster was found for low temporal frequency range in the left hemisphere.

In the late latency window (~ 300 ms) around the P300 component, a significant difference was observed, with STIM-Detect eliciting a more positive response than CTRL-NoDetect, only for V1v in the left hemisphere and low temporal frequency range.

No significant effects were observed in V2v, V3v and PHC whatever the hemisphere and the latency window. No significant effect was observed in right hV4 whatever the latency window.

Results in source space for the oculomotor areas

The early and late latency windows used for permutation tests in the oculomotor areas were the same as those for the visual areas. The saccade size variation interval was set to [2—4.1] ° consistent with the saccade size distribution (Fig. 3b). Two one-tailed non-parametric cluster-based permutation tests were conducted on the EFRP data for each oculomotor area in both hemispheres. Results are summarized in Table 6, with corresponding Figures (Fig. S15, S16) in SM F.

Table 6.

Results of cluster-based permutation tests for the oculomotor areas. Trends effects are shown in italics.

Oculomotor areas	p-value	Cohen’s d Range	Cohen’s d Mean	Maximum Effect (ms, °)
Right IPS	0.0729	0.316–0.883	0.526	~ 100 ms, 4.0°
	0.0449	0.316–0.785	0.473	~ 300 ms, 4.0°
Left IPS	0.0082	0.316–1.192	0.501	~ 100 ms, 4.0°
	0.0233	0.316–0.742	0.428	~ 300 ms, 4.0°
Right, Left FEF	-	-	-	-

In the early latency window (~ 100 ms) around the Lambda component, significant differences were observed with STIM-Detect eliciting a more positive response than CTRL-NoDetect. For the IPS and the entire saccade size interval, a significant cluster was observed in the left hemisphere and a trend one in the right hemisphere.

In the late latency window (~ 300 ms) around the P300 component, significant differences were observed, with STIM-Detect eliciting a more positive response than CTRL-NoDetect, for the IPS on the entire saccade size interval for left hemisphere and on a [3,4] ° saccade size interval for right hemisphere.

No significant effects were observed in FEF whatever the hemisphere and the latency window.

Discussion

The primary objectives of this study were to investigate: (1) the involvement of an extended network of visual and oculomotor regions in intra-saccadic motion perception, and (2) how the activity within these regions varies with retinal frequency. For this purpose, several visual and oculomotor areas—V1, V2, V3, hV4, MT/V5, FEF, and IPS—were selected based on their established roles in the network supporting intra-saccadic motion perception and eye movement control. The parahippocampal cortex (PHC) was included as a “control” region. In this study, differential activations refer to positive deviations between the STIM-Detect and CTRL-NoDetect configurations, indicating higher mean activity in the STIM-Detect configuration. This directional definition was established based on the two-tailed contrasts observed at the electrode level. The EFRP obtained in electrode space was reconstructed independently for lateralized regions in the left and right hemispheres.

Scalp EEG recordings revealed prominent N1 and P300 components

In the electrode space, differential activation was observed in the EFRPs evoked at the offset of the main saccade. The EFRPs were reconstructed using spline regression at a temporal frequency F_T of 20 Hz. This 20 Hz frequency matches both the temporal frequency at which the EFRP could be robustly reconstructed across all participants and the optimal retinal temporal frequency identified by Castet & Masson (2000)⁷for intra-saccadic motion perception. Selecting this frequency thus ensured consistent data across participants (Fig. 2a) while minimizing EFRP modulation by saccade amplitude¹⁹.

Differential activations were observed in the posterior N1 and P300 components, with a similar trend emerging in the anterior N1 component. The posterior N1 component, which was more negative in the STIM-Detect configuration between 150—250 ms, was localized over occipital regions, indicating an early modulation of cortical visual activity related to the intra-saccadic motion perception. Importantly, the “static” visual scene visible from fixation onset was identical across configurations—two dots (one red and one green) separated according to the trial-specific eccentricity—with the only difference being the presence or absence of perceived motion. The latter corresponded to a spatio-temporal pattern with a vertical spatial frequency of °/cy moving from left to right. This finding is consistent with the interpretation of the posterior N1 component (at occipito-parietal sites) as an index of early attentional and perceptual processing involved in visual feature detection²⁰. The P300 component, which was more positive in the STIM-Detect configuration from 300 to 450 ms and exhibited a central scalp distribution, is consistent with the recruitment of higher-level cognitive processes, such as pattern recognition and decision-making²¹. The absence of an effect on the Lambda component—typically associated with visual responses to retinal displacement during saccades²²—suggests that detection of intra-saccadic motion does not rely on early visuomotor signal but on more elaborated neural processes.

Source reconstruction revealed neural activation primary within the magnocellular pathway

In the source space, a temporal distinction was observed in the activity of reconstructed sources within visual areas along the magnocellular pathway: an early period encompassing the Lambda-N1 complex and a later period corresponding to the P300 component, both associated with the perception of stimulus motion during the saccade. When participants reported perceiving the motion of the stimulus, the apparent displacement of the vertical sine-wave pattern from left to right occurred around the midpoint of the saccade, near the center of the screen, corresponding to stimulation of the foveal region of the retina. Consequently, bilateral activation of early visual areas (V1, V2) was expected, together with predominant activation in the left hemisphere corresponding to the contralateral (right) visual field in high-border visual regions (V3, hV4, MT/V5) exhibiting retinotopic organization.

Dorsal areas

We first observed early dorsal differential activations within the primary visual areas V1d and V2d of both hemispheres (Table 4). This early activity likely contributes, at least in part, to the Lambda-N1 complex, reflecting the arrival of new visual input to the cortex at the onset of post-saccadic processing associated with motion detection. Consistent with previous findings, the Lambda-N1 complex originates from neural generators located in the occipital cortex^23–25. The integration of motion-related information during this early latency window suggests that primary visual areas are re-engaged after the saccade, whereas motion perception itself occurred while the eye was still in motion (prior to saccade offset), when eye velocity permitted the perception of stimulus motion. Early dorsal activations extended to V3d (Table 4), an area known to contain a higher proportion of neurons selective for global motion, speed and direction ^26,27. Subsequently, left-lateralized activation in V3d integrates visual information about both shape (vertical) and motion (horizontal, left to right), aligning with the attentional grouping effects described by Wu et al. (2023)²⁸. We interpret this early activity as contributing to the Lambda-N1 complex, which marks the arrival of new visual information in the cortex associated with motion detection. Progressing along the magnocellular pathway, our results revealed a right-lateralized MT/V5 activation (Table 4), followed by a later positive response in the same region. This finding is consistent with the literature. Although MT/V5 motion-selective neurons are not inherently lateralized^29,30, the right hemisphere generally accesses both ipsilateral and bilateral motion input more efficiently³¹. Lesion studies corroborate this asymmetry, showing that right MT/V5 damage produces broader contralateral—and often ipsilateral—motion deficits³². While inter-individual variability exists in hemispheric dominance for motion perception, a right-hemisphere bias is consistently reported^33,34. The early positive differential activity likely contributed to the Lambda-N1 complex, whereas the later response was interpreted as reflecting the P300 component, associated with decision-making regarding stimulus motion during the saccade. This interpretation aligns with previous studies identifying P300 sources in regions such as the temporo-parietal junction, in close proximity to MT/V5^35,36.

Ventral areas

Differential activation was observed exclusively in the ventral primary visual area (V1v) during both early and late latency windows (Table 5). This contrasts with dorsal areas, where differential activation persists beyond V1. These observations relate to the foundational view that V1 is a critical site where anatomically segregated magnocellular and parvocellular inputs partially converge^37,38. Recent evidence refines this view by showing that despite distinct anatomical inputs in V1’s layers, extensive intracortical connectivity enables significant functional integration of magnocellular and parvocellular signals³⁹. Beyond V1, functional segregation between dorsal and ventral streams becomes clearer. The absence of differential activation in ventral V2v and V3v likely reflects the ventral pathway’s focus on processing more static object features rather than dynamic intra-saccadic visual information. Visual areas in the parvocellular pathway, including hV4, primarily process color, complex shapes, and object identification^38,40. Our results showed early left-lateralized hV4 activation (Table 5), possibly reflecting shape processing via motion-defined borders, supported by hV4’s connections to MT/V5 and its role in encoding orientation and direction cues^37,41. Additionally, hV4 coordinates visual attention with eye movements; saccades transiently modulate hV4 receptive fields⁴², which likely facilitated motion processing during the saccade. Our moving grating stimulus likely engaged hV4 in distinguishing figure from background, consistent with motion-mediated shape perception. Regarding the PHC, our data revealed no significant differential activation, confirming that intra-saccadic motion perception does not recruit areas dedicated to complex environmental or contextual processing^43,44.

Source reconstruction revealed involvement of the IPS but not of the FEF

Regarding the activities of oculomotor areas, our analyses showed that only the IPS area exhibited differential activation (Table 6), both in early latency windows (Lambda) and late latency windows (P300). The presence of these two significant temporal clusters could reflect both early visuo-motor integration processes and decision-making stages. This observation aligns with previous studies demonstrating the involvement of the IPS on controlling spatial attention and visual selectivity⁴⁵, both of which are necessary processes for selecting and processing stimulus motion during saccades. In contrast, no significant difference was observed in the FEF area, regardless of hemisphere or time window considered. The FEF is key to the planning, preparation, and motor control of saccadic eye movements but is not specifically dedicated to the perceptual processing of visual stimuli integrated during or after the saccade^45,46. As highlighted by Andersen and Buneo (2002)⁴⁸, the posterior parietal cortex, particularly the IPS, plays a central role in sensorimotor integration and motor intention formation, in contrast to frontal regions such as the FEF dedicated to the specific motor control of saccades. These observations confirm the functional double dissociation between the IPS, a key region for perceptual and sensorimotor integration, and the FEF, primarily dedicated to saccade motor control.

Variability in saccade peak velocity and post-saccadic oscillations defines individual ocular profiles

All participants performed saccades with similar size distributions, but there was substantial variability in peak velocity across individuals, (Fig. 2) reflecting inter-individual differences in the main sequence relationship linking saccade size to peak velocity⁴⁹. As expected from oculomotor principles, larger saccades increased detection probability in STIM-Detect via prolonged retinal smear and optimal velocity matching—representing kinematic prerequisites that enable, rather than result from, intra-saccadic motion perception. To simplify interpretation, participants were divided by K-means clustering (k = 2) into Group A with higher peak velocities (thus activating lower retinal frequencies) and Group B with lower peak velocities (activating higher retinal frequencies). However, the division into ‘Group A’ and ‘Group B’ is a methodological simplification rather than an absolute categorization given the underlying continuum. Regarding post-saccadic oscillations (PSOs), they mainly reflect the relative movement of the pupil within the eye due to the viscoelastic properties of the iris, impacting estimates of saccade kinematics recorded by video eye trackers, particularly duration and in a less extend peak velocity^50,51. Hooge et al. (2015)⁵⁰ demonstrated their idiosyncratic nature, exhibiting strong within-subject consistency but considerable variability between participants. Moreover, Li et al. (2021)⁵² demonstrated a positive correlation between saccade peak velocity and the occurrence of post-saccadic oscillations. Consistent with these findings, participants with higher peak velocities exhibited more post-saccadic oscillations than participants with lower peak velocities, despite matched saccade sizes (see SM B, Fig. S1, Table S1).

Temporal frequency critically modulates differential activation linked to motion perception

As main results, we showed that differential activation in right MT/V5 was modulated by retinal frequency, with significant effects between 15 and 20 Hz during the early latency window and between 6 and 20 Hz during the late latency window (Table 4). This aligns with Castet & Masson (2000)⁷, who identified an optimal frequency range of 15—25 Hz for intra-saccadic motion perception. We observed a bell-shaped EFRP amplitude profile in STIM-Detect (Fig. 3c). When translated into retinal speed, the 6 -15 Hz range corresponds to 35.3—117.6°/s, and the 15—20 Hz range to 88.2—117.6°/s. Indeed, both intervals are compatible with the overall distribution of preferred speeds reported in MT/V5 neurons^37,53. Then, in the 15—20 Hz range, differential activation was observed dorsally, up to MT/V5, and ventrally in V1 and hV4 during the early latency window. These markers were interpreted as markers of the visual processes underlying conscious perception of the horizontal movement of the stimulus from left to right and leading to decision-making^54,55. Participants consistently reported perceiving the vertical stimulus as moving from left to right. However, in the 6—15 Hz frequency range, we observed differential dorsal activation but not in MT/V5, which plays a critical role in conscious motion perception⁵⁶. This could indicate a dissociation across the hierarchy, where V1/V2 contribute to motion perception through a combination of intrinsic lateral interactions and alternative pathways such as subcortical inputs to MT/V5 enabling them to represent apparent or local motion percepts even when MT/V5 is less active and not detected.

In addition, we observed an absence of differential activation in MT/V5 in the high retinal frequency range (20—31 Hz corresponding to 88.2°/s—182.3°/s) while participants reported motion perception. This unexpected result may stem from individual differences in oculomotor characteristics. Recent findings by Rolfs et al., (2025)¹⁴may support the hypothesis that distinct oculomotor characteristics (peak velocity and occurrence of post-saccadic oscillations) across participants would be associated with distinct cortical activation patterns across low and high retinal frequency ranges. According to these authors, participants with faster saccades (i.e., higher peak velocity and more post-saccadic oscillations) may possess a visual system with higher speed-processing capabilities than those with slower saccades. Therefore, even if participants with slower saccades (and thus exposed to higher temporal frequencies on the retina) reported perceiving intra-saccadic motion similarly to those with faster saccades, we hypothesize that the vividness of the perceived intra-saccadic motion was reduced in participants with slower saccades compared to those with faster ones. Moreover, they demonstrated that the presence of pre- and post-saccadic stationary points is critical for effective saccadic omission ^5,57,58. This suggests that post-saccadic oscillations—more pronounced for faster saccades (see SM B, Table S1)—would render saccadic omission less effective, thereby enhancing the visibility of stimulus motion during saccades. Taken together, the two oculomotor characteristics that differentiated the participants may suggest that participants with slower saccades experienced reduced visibility of stimulus motion during saccades, leading to smaller differences in activation within the visual areas involved in motion processing.

Future research should consider individual eye kinematics and temporal dynamics

Some methodological limitations of our research should be addressed in future studies. In video‑based eye tracking, post‑saccadic oscillations (PSOs) can distort estimates of saccade kinematics (particularly duration), which we detected but did not correct using biophysical models⁵¹. Although peak velocity (critical for retinal frequency estimation) is minimally affected, future studies should apply such corrections to ensure precise saccade onset/offset detection and accurate kinematic estimates. We demonstrated how individual oculomotor characteristics strongly impact the visibility of intra-saccadic percepts. To overcome this difficulty, it would be important to adapt the stimulus speed according to each participant’s eye kinematics. To personalize the spatio-temporal features of the stimulus, one could record the main sequence of each participant at the beginning of the experiment to ensure that, for the range of saccade sizes to be performed, the temporal frequency of the stimulus projected on the retina during the saccade remains confined within a predetermined temporal frequency range, for example between 15 and 20 Hz. Moreover, to consider the laterality bias observed with this protocol, it would be important to randomize the direction of the saccade and, consequently, the direction of the stimulus movement. Finally, we have discussed the spatial aspects of activations but not the propagation of activations over time following the hierarchy of visual pathways. Decoding techniques could therefore be of great interest. Note that source estimates offer only approximate spatial localization and are subject to inherent uncertainty. Activation in visual and oculomotor regions should therefore be interpreted as consistent with canonical areas rather than as definitively localized to precise cortical sites.

Conclusion

Far from being epiphenomenal, intra-saccadic signals are actively processed by the visual system to support perceptual stability and trans-saccadic integration^59,60. Pioneering work by Castet and Masson⁷ demonstrated motion perception during saccades under optimal conditions via magnocellular pathway activation⁹, while recent studies have revealed behavioral consequences such as improved object tracking and scene understanding^12,61–65.

Our study confirms the fundamental involvement of the magnocellular pathway and the mediating role of retinal temporal frequency in intra-saccadic motion perception. The perceptual efficacy is maximized within a specific temporal frequency bandwidth aligned with the tuning properties of magnocellular motion detectors. Crucially, these perceptual phenomena are tightly coupled with individual oculomotor kinematics, such as saccade peak velocity and post-saccadic oscillations, demonstrating that eye movement dynamics are integral modulators rather than mere facilitators of intra-saccadic visual processing. This strong interaction between motor and sensory processes highlights the need for future research to explicitly consider oculomotor characteristics when investigating dynamic visual perception. Refining experimental paradigms to incorporate these factors will deepen our mechanistic understanding of spatiotemporal computations underlying motion perception during saccades and contribute to a more complete model of sensorimotor integration in vision.

Methods

Participants

Thirty-five healthy adults participated in the experiment (17 women; age range: 21–38 years; mean ± s.d: 25.2 ± 4.8 years). All participants were free of any medical treatment at the time of the experiment and had no history of neurological or psychiatric disorders. They had normal or corrected-to-normal vision and none had prior experience with the experimental task. The study was approved by the French ethics committee (“Comité de Protection des Personnes Sud-Est III”, Eudra-CT 2020–100,503-36) and conducted according to the principles expressed in the Declaration of Helsinki. All participants provided written informed consent prior to the experiments and received 60 € for their participation.

Experimental protocol

The study consisted of two sessions. During the first session, an MRI scan was acquired to obtain each participant’s 3D anatomical image for EEG source reconstruction. The second session, conducted within one month, involved EEG recording using an intra-saccadic motion perception paradigm adapted from Castet and Masson (2000)⁷.

Two visual conditions were implemented, differing in the direction of the stimulus motion. In the stimulus condition (STIM), motion was aligned with the participant’s saccade, allowing potential intra-saccadic motion perception depending on eye velocity. In the control condition (CTRL), motion was opposite to the saccade, preventing motion perception while keeping all eye movement parameters identical. This design enabled direct comparison between perceptual and non-perceptual trials.

For each trial, a moving vertical grating of low spatial horizontal frequency and high horizontal speed was displayed, according to this evolution:

1

with the horizontal spatial position on the screen, the time, the condition, the average luminance, the contrast, the spatial horizontal frequency (0.17 cy/°) and the horizontal speed (360°/s). The average luminance of the grating was equal to 22 cd/m². The contrast varied from 0 to 0.17 (see below). All stimulus parameters () and their temporal evolution were based on the original experiment by Castet and Masson (2000)⁷.

Each trial comprised five sequential steps (Fig. 1). The first two phases were stabilization phases on a red fixation cross, during which two colored markers were symmetrically positioned relative to the screen center (Fig. 4b). For each trial, the inter-marker distance was randomly drawn from a uniform distribution [1.5– 6.5]°. The left marker, a red cross, served as the fixation point, and the right marker, a green solid circle, indicated the saccade target. Both markers had luminance equal to the grating’s mean luminance ( cd/m²).

From  to  (600 ms), grating contrast was null and the screen displayed uniform luminance. From  to  (400—600 ms, randomized), contrast increased from 0 to 0.17 following a raised‑cosine function, reaching its maximum at  when the red fixation cross was replaced by a red solid circle (Fig. 4b), serving as the visual cue to initiate a saccade toward the green target defined at trial onset. From  to  (1500 ms), contrast remained constant while participants fixated the target, after the saccade execution. From  to  (500 ms), contrast decayed to 0 following the same raised‑cosine profile, mirrored in time to produce a decrease. The trial ended at .

At the end of each trial, participants indicated by key press whether or not they had perceived motion of the stimulus during their left‑to‑right saccade. The next trial was initiated by another key press. Each participant completed 480 trials (300 STIM, 180 CTRL), randomly interleaved. Trials were organized into five blocks of 96 trials each (~ 8 min per block), with short pauses between blocks. Each block began with an eye‑tracker calibration.

Because the perception of intra-saccadic motion is uncommon, a familiarization phase was conducted at the beginning of the experiment. Participants performed horizontal saccades between two fixation points separated by 4°, from left to right (potential motion perception) and then from right to left (no motion perception). Saccades were executed naturally and without time constraints. During this phase, the visual stimulus was identical to that of the STIM condition with maximum contrast. The experimenter asked participants to verbally describe their percepts during saccades in both directions to ensure comprehension of the motion phenomenon relative to saccade direction. After confirming this understanding, participants completed a training block of 20 trials following the temporal sequence illustrated in Fig. 4. They were instructed to avoid blinking during trials and to rest their eyes during the inter‑trial interval before triggering the next trial with a key press. These training trials were analyzed online to provide individualized feedback based on three metrics: (1) the percentage of correct responses, (2) the relative error of the executed saccade size, and (3) the latency of the executed saccade. The first two metrics assessed task accuracy for motion perception and targeting, whereas the third was expected to fall within the typical range of 250—300 ms⁵⁷. For excessive latencies, we reminded participants of the importance of being attentive and responsive to the visual cue. Short or long latencies indicated, respectively, premature or delayed saccade initiation. The main experiment began immediately after training.

EEG data were recorded only during the main experiment. The experimental procedure was implemented in Python, using the psychopy toolbox (https://psychopy.org/about/index.html) to control the visual stimulus timing and the pylink toolbox to interface with the eye tracker.

Material

MRI acquisition

MRI data were acquired at the Grenoble MRI facility (IRMaGe) using a whole‑body 3 T scanner (Achieva, Philips Medical Systems, Best, The Netherlands) equipped with a 32‑channel SENSE head coil. The session included a high‑resolution structural T1‑weighted 3D MP‑RAGE sequence (TR/TE/TI = 8.1/13.7/678 ms; flip angle = 8°; acquisition matrix = 250 × 257 × 220; 220 sagittal slices; voxel size = 0.90 × 0.89 × 1 mm³). Participants were comfortably positioned in a supine posture, and the head was stabilized with soft foam padding to minimize motion during image acquisition.

Stimulus display

Stimuli were presented on a 21‑inch Sony CRT monitor positioned 68 cm from participants (resolution = 640 × 480 pixels; refresh rate = 160 Hz). The display subtended 27 × 21 degrees of visual angle. Luminance calibration was performed with a SpyderX Elite spectrophotometer.

Eye movements acquisition

Eye movements were recorded with an EyeLink 1000 system (SR Research) operating in pupil-corneal reflection mode. The left eye was monitored at 1000 Hz while the head was stabilized using a chin rest. A 9‑point calibration procedure was performed at the beginning of each block and repeated during the block if participants failed to maintain gaze within a 2 × 2.5° bounding box centered on the red fixation cross. Saccades and fixations were automatically detected by the EyeLink software using three criteria: a spatial displacement greater than 0.1°, a velocity exceeding 30°/s, and an acceleration above 8000°/s².

EEG acquisition

The EEG session was performed at the Grenoble EEG facility-IRMaGe. The EEG activity was continuously recorded using 64 Ag/AgCl unipolar active electrodes positioned according to the extended 10–20 system⁶⁶. The AFz and FCz electrodes were used respectively as ground and reference electrodes. In order to correct ocular artefacts, the electro-oculographic (EOG) activity was also recorded using two electrodes positioned at the eye’s outer canthi, and two respectively above and below the left eye. The ground electrode for the EOG was placed on the left shoulder. Electrodes impedances were kept below 10 kW for each electrode. The signal was amplified using a BrainAmp™ system (Brain Products, Inc.) with a sampling rate at 1000 Hz. Electrodes locations on the scalp were recorded using a CapTrack™ system (Brain Products, Inc.)

Data preprocessing

This section describes the preprocessing of eye movement and EEG data, including spline regression for evoked potential estimation, followed by behavioral data processing and neuroimaging data preparation for cortical source reconstruction. Finally, the statistical analyses are detailed.

Eye movements

Only left-eye gaze positions were recorded. For each saccade, latency (from timestamp ), horizontal amplitude, velocity profile, and horizontal peak velocity were extracted. Velocity was estimated using a nonlinear extension of the Savitzky-Golay filter, improving accuracy for abrupt time-series variations⁶⁷.

The retinal frequency (Hz) of the visual stimulus projected onto the retina during a saccade was computed as a function of stimulus speed (°/s) and the peak velocity (°/s), considering the spatial frequency (cy/°) of the stimulus and the saccade size (°). Importantly, this frequency calculation incorporates a condition-dependent factor which takes values of + 1 or -1 depending on whether the stimulus moves in the same or opposite direction as the saccade. This reflects the fact that the stimulus direction differs across the STIM and CTRL conditions. Formally:

2

with defined as in Eq. 1. Five criteria (see SM A) ensured that the main saccade between T₂ and T₃ exclusively induced intra-saccadic motion perception, excluding other saccades. Moreover, to apply consistent criteria for detecting intra-saccadic motion perception across both experimental conditions, STIM (stimulus moving with saccade) and CTRL (stimulus moving opposite), a temporal frequency was defined. This serves as a reference frequency corresponding to the retinal frequency in the STIM condition (i.e., ):

3

Behaviorally, this temporal frequency underlies intra-saccadic motion perception during the STIM condition⁷, enabling direct comparison between both our experimental conditions.

EEG

EEG data and eye-gaze positions were synchronized offline thanks to triggers sent simultaneously during the experiment on the two acquisition systems, EEG and eye tracker ⁶⁸. The preprocessing pipeline was implemented using EEGlab software⁶⁹. The EEG raw signal was first resampled at 250 Hz then band-pass filtered between 0.1 and 70 Hz and a notch filter at 50 Hz was added. Filtered EEG data were then segmented from 500 ms before to 3000 ms after . These segments were visually inspected offline and those containing muscular activities or non-physiological artifacts were rejected. The signals were also visually inspected to reject bad channels. Ocular artifacts were removed semi-automatically. After Independent Component Analysis (InfoMax algorithm⁷⁰), the components showing the highest correlation with vertical and horizontal EOG deviations were selected for correcting blink and saccade artifacts, respectively. Among these components, blink-related sources were easily identified (mainly one source, exceptionally two) and removed based on their temporal profile in source space and their spatial topography. However, the identification of components related to saccade artifacts was ambiguous for some datasets. To maintain a consistent preprocessing pipeline across all datasets, saccade artifact correction was not implemented. Finally, segments were rejected if their variance exceeded a threshold defined as the mean variance across the segments plus three standard deviations of these variances. The rejected channels were then interpolated according to a spatial spherical interpolation.

For estimation of evoked potentials, data were segmented into epochs starting 200 ms before and ending 1700 ms after , which corresponded to the onset of the visual cue instructing the participant to execute the requested saccade. Data were re-referenced to a common average. Epochs were then normalized by z-scoring each channel according to the average EEG amplitude during a baseline period between -200 and 0 ms relative to .

Estimation of evoked potentials

We compared two key configurations: STIM-Detect and CTRL-NoDetect, based on the estimation of evoked potentials. STIM-Detect included trials where participants perceived stimulus motion during saccades, while CTRL-NoDetect included trials where they did not. No EEG analysis was done for the STIM-NoDetect and CTRL-Detect configurations. Evoked potentials were estimated in the electrode space before source reconstruction on the cortical surface. The main components analyzed were the Lambda, N1, and P300, estimated as Eye Fixation Related Potentials (EFRPs), which are known to be time-locked to fixation onset (or saccade offset)²⁵. Since this study focuses on saccades, we refer to EFRPs as potentials evoked at saccade offset, knowing these two events coincide.

Estimating EFRPs by averaging poses two main challenges: (1) temporal overlap between stimulus-evoked potentials and fixation-evoked potentials, and (2) effects of eye movements on EEG amplitude^19,25,71. To address these, we used a General Linear Model (GLM) deconvolution approach^72,73. Saccades affect EFRPs, with pre-saccadic activity and Lambda components observed at occipital sites, modulated by saccade size and orientation^19,74. Our study included only left-to-right saccades, so orientation effects were not considered. Different methods have been proposed to correct for saccade size effects, including adding saccade size as a linear regressor⁷⁵ or modeling it non-linearly with spline regression^72,76. Since our experimental design included only one saccade per trial, the amount of data available for modeling non-linear effects was limited. To balance model flexibility and the risk of overfitting, we implemented a spline regression using a minimal number of basic functions. Specifically, we chose a spline regression with three B-splines of order two, which allowed us to capture non-linear relationships between saccade size and EEG amplitude while keeping the number of parameters manageable relative to the limited data. Using saccade size () as a predictor, the two configurations were compared for the same saccade sizes. As a result, residual variability in temporal frequency remained, given that intra- and inter-individual variability in the main sequence induces variability in peak velocity and therefore in temporal frequency for a fixed saccade size⁴⁹. For the same reason, the same was applied to the choice of temporal frequency as a predictor for a better correspondence of intra-saccadic perception via temporal frequency while accepting residual variability in saccade size. Consequently, the decision was made to reconstruct at a fixed temporal frequency for all analyses where control of intra-saccadic motion perception was prioritized over control of saccade size (and vice versa for the choice of reconstruction at a fixed saccade size).

Finally, for a given predictor (temporal frequency or saccade size), the i-th epoch EEG signal, , was modeled according to:

4

where is the evoked potential at , is the j-th B-spline evaluated at the value of the predictor, is the spline coefficient along time for the j-th contribution for the EFRP estimation, is the number of B-splines (), is the latency of the main saccade offset, is the potential evoked at all other saccade offsets, is the latency of the saccade offset, and is the additive noise of the ongoing activity, not correlated with the task (see SM D for the location of the B-splines). By concatenating all trials for a given configuration and a given participant, , , , and were estimated by ordinary least square regression to obtain namely , for and which was a component of non-interest in our analysis. The regression was completed for each participant and each configuration separately. Finally, the EFRP at the main saccade offset was modeled as:

5

Thus, this formulation enabled three analyses: temporal (predictor fixed), predictor-based (time fixed), and joint (time—predictor interaction). In practice, was estimated from -200 to + 900 ms (275 samples), and EFRPs from -200 to + 700 ms (225 samples each) using spline regression with three B-splines of order two. This choice constrained model complexity relative to data availability while capturing non-linear relationships. With a 250 Hz sampling rate, this yielded 1,175 regressors (275 + 4 × 225). Model adequacy will be verified in Results.

Regions of interest (ROIs)

We used Wang’s functional probabilistic atlas⁷⁷ to localize visual and oculomotor ROIs. Analyses focused on areas involved in motion perception along the magno- and the parvocellular pathways⁷⁸: V1, V2, V3 (each split into dorsal/ventral), hV4, MT/V5, and oculomotor areas FEF and IPS. One control ROI, the parahippocampal gyrus (PHC), was also included. Each ROI was defined separately for left and right hemispheres, yielding 22 ROIs. To ensure masks contained only voxels common to all participants, we applied a 50% probability threshold for V1 to hV4 to reduce overlap between adjacent areas, and 40% for MT/V5, FEF, IPS, and PHC^79,80. Masks were realigned to individual space via inverse deformation fields and projected onto cortical meshes in BrainVisa to generate surface masks for each participant.

Cortical sources reconstruction

Cortical activity evoked by the main saccade offset was reconstructed using the weighted Minimum Norm Estimation (wMNE) algorithm implemented in BrainStorm software^81,82, applied to the regression coefficients . This method was chosen for its linearity and its ability to localize deep sources, in line with the MT/V5 brain area of interest located within a deep sulcus. The cortical current densities were combined linearly to estimate EFRP-related current densities at each cortical vertex, following Eq. 6:

6

Individual cortical surfaces from BrainVisa were used. Innerskull, outerskull, and head surfaces were generated by deforming a standard template using inverse deformation fields. EEG electrodes were coregistered via fiducials. The head model was computed with the Boundary Element Method (OpenMEEG⁸³) using default conductivity values (1 S/m for head, innerskull and 0.0125 S/m for outerskull) and normal dipole orientation. Noise covariance matrices were estimated from baseline signals (-800 to 0 ms before ). The depth weighting exponent, the maximum weight limit, and the regularization parameter for noise covariance were set to 0.5, 10, and 10, respectively. Current densities were normalized by z-scoring to baseline (-200 to 100 ms pre-saccade offset), set to absolute values, and averaged within each ROI.

Behavioral data

For each valid trial (i.e., trials meeting all main saccade detection criteria and whose EEG segments passed the preprocessing pipeline), the detection of the intra-saccadic perception (participant’s response) was considered.

Statistical analysis

Concerning ocular data, statistical analyses were performed by ANOVAs using Matlab with a significance level of 0.05. Prior to the ANOVA, normality of the data was assessed using the Shapiro–Wilk test. When the assumption of sphericity was violated (indicated by a significant Mauchly’s test), degrees of freedom were adjusted using the Greenhouse–Geisser correction. Each significant effect was followed by Bonferroni post-hoc tests correcting for multiple comparisons. The p-values were rounded to the fourth decimal place. Values smaller than 0.0001 are reported as “p < 0.0001” to indicate that their exact values are not provided beyond this threshold.

Concerning EEG data, statistical analyses were performed using non-parametric paired permutation tests implemented in the FieldTrip toolbox for group-level EEG data in both electrode and cortical source space, with cluster-based correction for multiple comparisons^84,85. The null hypothesis stated that no difference existed between the two configurations (STIM-Detect vs. CTRL-NoDetect). Sample-level statistics were based on t-values, and cluster-level significance on the sum of those t-values. Cluster probabilities were estimated with a Monte Carlo method (100,000 iterations, α = 0.05). In the electrode space, three latency windows derived from ERP literature were analysed to isolate the saccadic spike, the Lambda-N1 complex, and the P300 component. Each window lasted 200 ms with a 20 ms overlap: very early [-120—80] ms, early [60—260] ms and late [240—440] ms. Three two-tailed cluster-based permutation tests were performed on EFRPs (space × time) after spline reconstruction at a temporal frequency approximating the median frequency across participants and configurations, serving as a common reference frequency for all participants and both configurations. In the source space, one-tailed cluster-based permutation tests were conducted independently for each ROI. Visual ROIs were analyzed in the temporal-frequency × time domain, and oculomotor ROIs (FEF, IPS) in the saccade-size × time domain. Latency windows were defined based on electrode-space results. For both predictors (temporal frequency or saccade size), the variation interval was chosen according to each participant’s empirical distribution to balance the widest feasible range with inclusion of the maximal number of participants, ensuring optimal statistical power.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Gaëlle Nicolas : Investigation, Methodology, Data analysis, Writing—original draft, Writing—review and editing. Emmanuelle Kristensen: Methodology, Data analysis, Writing—review and editing. Michel Dojat: Conceptualization, Supervision, Funding acquisition, Writing—original draft, Writing—review and editing. Anne Guérin-Dugue: Conceptualization, Supervision, Funding acquisition, Methodology, Data analysis, Writing—original draft, Writing—review and editing.

Funding

GN was recipient of a grant from the University Grenoble Alpes. This study was partially funded by a grant from the LabEx PERSYVAL-Lab (ANR-11-LABX-0025–01) and ANR via NeuroCog (ANR-15-Idex-02). This work was performed on the IRMaGe platform member of France Life Imaging network (grant ANR-11-INBS-0006).

Data availability

The raw and synchronized EEG and eye-tracking datasets used for this study are available in the “Recherche Data Gouv” repository, 10.57745/R0LMFV. None of the experiments was preregistered.

Code availability

The code for data analysis is available upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.