Characterizing Spatiotemporal Population Receptive Fields in Human Visual Cortex with fMRI

Validating spatiotemporal pRF models through simulation

To evaluate the robustness and validity of the spatiotemporal pRF modeling framework, we developed a simulation software with two purposes. First, to simulate experimental paradigms and test if they can be used to differentiate between predictions of different pRF models. Second, to validate the accuracy of the framework by testing if it recovers ground-truth pRF model parameters from simulated fMRI time courses with noise. By understanding how well our framework can recover parameters in simulated scenarios, we can define the scope of model interpretability given our experimental design and number of measurements.

Synthesizer: different pRF models predict distinctive time courses to identical visual stimuli

Using the stimulus design in the spatiotemporal pRF mapping experiment (Fig. 1), 300 noiseless and 300 noise-added synthetic fMRI time courses were generated for each of the pRF models: spatiotemporal pRF models (DN-ST and CST) and the conventional spatial pRF model (Dumoulin and Wandell, 2008). To compare how the temporal parameters affected fMRI time courses, synthetic fMRI time courses for each model used the same spatial parameters (xyσ ) while varying model-specific parameters for DN-ST and CST.

We found that even with the same stimuli and identical spatial parameters, distinct synthetic fMRI time courses were generated for each model. An example run (out of nine runs in total) of three noiseless synthetic time courses, one for each of the different pRF models with the same spatial parameters, is shown in Figure 4A. All the example time courses had four distinct peaks at similar times during the experiment, corresponding to the four times the bar swept across the pRF. However, the different models generated distinctive fMRI time courses in terms of their amplitudes, latencies, and widths and these differences were preserved when typical fMRI noise was applied (Fig. 4B). The different synthetic time courses confirmed that different types of pRF models can generate fMRI time courses that can be experimentally differentiated with our stimulus design.

Solver: testing the accuracy and robustness of the spatiotemporal pRF framework vis-à-vis ground-truth

We next tested whether the solver accurately recovers ground-truth pRF parameters from synthetic fMRI time courses. Specifically, using a synthetic fMRI time course with additive noise and a stimulus sequence as model inputs, the solver estimates the specific model parameters for each pRF. To evaluate the solver's accuracy, we compared the estimated parameters to the ground-truth parameters that generated the synthetic time courses. Both the synthesizer and the solver used the same pRF model.

For noiseless synthetic fMRI time courses such as the ones in Figure 4A, the solver successfully recovered the spatiotemporal pRF parameters of all models with more than 99% accuracy. This result validates the analysis code and indicates that the solver can resolve pRF parameters from the planned experimental sequence for all model types.

When a typical level of empirical fMRI noise was added to the synthetic fMRI time courses, the models are still able to successfully recover the pRF parameters. The solver accurately recovered the spatial parameters (xyσ , Fig. 4C) of all models. Significant correlations were found between the predicted and ground-truth spatial parameters (Fig. 4E, all correlations are significant p < 0.0001) for all three models. Estimated pRF center positions (xy) were highly accurate (Fig. 4C, estimated vs ground-truth along the dashed equality line). The median absolute percentage error (MAPE) for pRF centers (xy) was 4.09–7.82% across the models (Fig. 4D,E). PRF size ( σ) was also accurately recovered for all models (Fig. 4C–E), yet with a higher MAPE of 7.48–12.5%. The higher accuracy of xy than σ estimates is consistent with previous reports of lower accuracy in estimating pRF size than pRF center positions (Lage-Castellanos et al., 2020; Lerma-Usabiaga et al., 2020). Additionally, we tested whether the recovered spatiotemporal parameters interact with each other and may bias estimates of individual parameters. There were no significant relationships between estimates of CST model parameters (στn , p > 0.5) suggesting that CST parameter estimates are unbiased. However, for the DN-ST model, we observed a significant negative coupling between σDN and τ1 (F_(1,298) = 5.57, p < 0.05), and σDN and τ2 (F_(1,298) = 12.09, p < 0.001). This suggests that given the current experimental design, the solutions of the temporal parameters of the DN-ST are not independent from one another.

Although the solver was overall less accurate for estimating the temporal parameters than spatial parameters from noisy synthetic data, it successfully recovered the temporal ( τ) and compression ( n) parameters from the CST model, and the temporal parameters for the DN-ST model (τ1τ2σDNnDN) , with better estimates for the CST than the DN-ST model (Fig. 4C,D; all correlations between estimated and ground-truth parameters are significant p < 0.0001). Remarkably, the CST model was able to resolve the temporal receptive field parameter ( τ) with an MAPE < 13% (Fig. 4D,E). Despite the sluggish nature of hemodynamic responses, the stimulated result suggests that for ∼200 ms temporal receptive fields, there is only a ±26 ms margin of error. The compression parameter, n, was also reliably estimated. The estimates of the DN-ST model temporal parameters were also significantly correlated with the ground-truth temporal parameters (Fig. 4C,E, p < 0.0001), but the estimates were less accurate than the CST model, with τ2 being the least accurate parameter (Figs. 4D, 1E). This variability of τ2 estimates may be due to an insufficient number of conditions with prolonged presentation, as the τ2 parameter mainly controls the temporal decay rate of the response to prolonged stimulus durations.

To assess the impact of the experimental design choice on parameter estimation, we also tested how well parameters are recovered from a different pRF mapping experiment that uses an image presentation rate of 8 Hz (Toonotopy; Finzi et al., 2021). For the same models and pRFs, we generated simulated time courses with noise for the Toonotopy experiment (data not shown). Results from the Toonotopy experiment indicated that pRF locations were similarly well-estimated for all models (MAPE for x and y were less than 12%). However, MAPE for pRF size ( σ) estimates from the Toonotopy experiment were 48.23% for the DN-ST and 42.52% for the CST. Additionally, the estimation errors for the remaining parameters exceeded 35% for the DN-ST model and exceeded 30% for the CST model. This suggests that Spatiotemporal mapping experiment (Fig. 1) is more suitable for estimating spatiotemporal parameters than the Toonotopy experiment.

Spatiotemporal pRF models outperforms spatial only pRF model

We next assessed how well the three pRF models: Spatial, DN-ST, and CST, predicted the empirical fMRI time courses in the Spatiotemporal mapping experiment (Fig. 1) in each of our 10 participants. Note that the time courses predicted from the Spatial model depend only on the location and duration of the stimulus, whereas the predictions of the DN-ST and CST models also depend on the spatiotemporal aspects of the stimulus at each location, resulting in more complex fMRI responses.

The spatiotemporal stimulus design evoked a wide range of BOLD time courses in voxels of the visual system. Based on each voxel's fMRI time course, pRF parameters for each of the three models were estimated. Results show that fMRI responses at the voxel level were well captured by the DN-ST (example voxel, Fig. 5A, yellow) and CST (example voxel, Fig. 5A, red) models. However, the Spatial model failed to predict the fMRI time courses in multiple ways. For example, as shown in the example V1 voxel (Fig. 5A, blue), the Spatial pRF model underpredicted responses for stimulus conditions with fast temporal rates (temporal conditions 1, 2, and 4), replicating previous findings by Stigliani et al. (2017), and overpredicted responses for prolonged stimulus conditions (temporal condition 5). In contrast, the DN-ST and CST models were able to predict both the amplitude and width of these peaks that the Spatial model failed to predict.

We quantified how well the various models predicted the experimental time courses in each voxel using a threefold-cross-validation approach. On average, the Spatial, DN-ST, and CST models accounted for 25, 30, and 33% of the variance, respectively. Each model's accuracy across multiple retinotopic visual areas is plotted in Figure 5B. The DN-ST and CST models outperformed the Spatial model in predicting fMRI responses in single voxels of all tested visual areas. Indeed, the threefold-cross-validated variance explained (R²) by the DN-ST and CST models was significantly higher than the Spatial model (Fig. 5B; permutation test, V1, V2, V3, hV4, VO, LO, TO, V3AB, p < 0.01; IPS, p < 0.05). Comparing the spatiotemporal pRF models, the CST model outperformed DN-ST model across all visual regions (permutation test, V3, hV4, VO, LO, TO, V3AB, p < 0.01; IPS, p < 0.05) except V1 and V2, where performance did not differ significantly between the two spatiotemporal models.

To examine if there were systematic and spatial differences in how well models fit the data across cortex, we numerically compared the variance explained by the three pRF models in each voxel. Then, we visualized on the cortical surface which model best predicted the response of each voxel (Fig. 5C). We found that: (1) in all participants, there were almost no voxels for which the Spatial model was the best model (very small number of blue voxels in Fig. 5C), (2) for most of the voxels, the CST model was the best (except for voxels around the occipital pole for which the DN-ST model was better), and (3) consistent with the quantitative analyses in Figure 5B, the advantage of the CST over DN-ST model was more pronounced in later visual areas. To further assess the robustness of the spatiotemporal pRF models, we compared model performance across the same participants using the Toonotopy fMRI data, which had a shorter temporal duration (2-s per location) and an image presentation rate of 8 Hz. Using the Toonotopy fMRI data (data not shown), we found that both spatiotemporal pRF models significantly outperformed the Spatial model in all visual areas (permutation test, V1, V2, V3, hV4, VO, LO, TO, V3AB, p < 0.01; IPS, p < 0.05), and the CST model had significantly higher accuracy than both the DN-ST and Spatial models (permutation test, V1, V2, V3, hV4, VO, LO, TO, V3AB, p < 0.01; IPS, p < 0.05).

Together, these analyses: (1) support our hypothesis that both spatial and temporal aspects of the stimulus contribute to fMRI signals at the voxel level, (2) show that both DN-ST and CST spatiotemporal pRF models can capture spatial and temporal dynamics of neural responses, and (3) suggest that considering the nonlinear temporal dynamics, both at the stimulus and neuronal level in subsecond resolution into a pRF model is crucial for accurately predicting fMRI timeseries at the voxel level.

Spatial estimates across models

As spatiotemporal pRF models are newly developed, it is important to test if the spatial pRF parameters they estimate are topographically organized and replicate well-established retinotopic maps in human visual cortex. Thus, we next used the estimated spatial parameters to generate polar angle, eccentricity, and pRF size maps for each participant and model (Fig. 6A–C). We also quantitatively compared the estimated pRF position (polar angle), eccentricity, and size in degrees of visual angle (°) for the DN-ST and CST pRF models to those of the Spatial model (Fig. 6D–F).

• Download figure
• Open in new tab
• Download powerpoint

Figure 6.

Spatial pRF parameters estimated for the Spatial, DN-ST, and CST models are similar. A, Polar angle, B, eccentricity, and C, pRF size maps on the medial view of the right inflated cortical surface of an example participant for each of the three models. D–F, Scatter plots comparing pRFs from spatiotemporal pRF models in yellow: DN-ST and red: CST (y-axis) to estimates from the spatial pRF model (x-axis) for each of nine visual areas. Each dot indicates the parameter estimate for one voxel. Colored lines: linear regression fits. For visualization, we randomly sampled 50 voxels from each participant for each visual area. Dotted black line: reference line of identical estimates. D, Polar angle. E, Eccentricity. F, pRF size (σ) and adjusted pRF size (σ/n ) for the CST model (purple).

We found that all three models produced robust and consistent estimates of each voxel's pRF positions. From pRF center positions (xy) , we generated polar angle and eccentricity maps, which were indistinguishable across the three pRF models and had the expected topography of phase reversals and eccentricity (see example participant, Fig. 6A,B). Moreover, all models generated similar topographies of pRF size estimates with increasing sizes along a posterior–anterior axis (Fig. 6C). Although in high agreement, pRF size estimates ( σ) from spatiotemporal pRF models were smaller than the Spatial pRF size estimates in all tested visual areas (Fig. 6F). In compressive models with a static nonlinearity compression ( n), researchers (Kay et al., 2013) typically report the adjusted pRF size (σ/n) . Comparison of the adjusted pRF size (σ/n) from the CST model to the Spatial model yielded similar estimates (Fig. 6F, purple). This result suggests that the smaller pRF size estimates of the spatiotemporal pRF models are mainly due to compressive nonlinearities within the models.

Overall, the reliable estimate of spatial pRF properties across models suggests that the incorporation of additional temporal parameters into the pRF model does not reduce the statistical power to map spatial pRFs in human visual cortex.

Differences in temporal estimates cannot be explained by differences in HRFs

We next characterized the temporal parameters of spatiotemporal pRFs across voxels of the visual system. Under our experimental paradigm, the simulations show that temporal parameters of the CST model were more accurately recovered than the DN-ST model (Fig. 4C,D) and experimental data revealed that the cross-validated variance explained of fMRI responses was higher for the CST than DN-ST model for most voxels (Fig. 5B). Thus, in the following sections, we report the temporal and spatiotemporal characteristics estimated from the CST model.

We assessed the time-to-peak from the temporal parameter τ of the CST model and the temporal processing window from the FWHM of the sustained neural temporal impulse function. The neural time-to-peak ( τ) was estimated in each voxel and visual area, then the distribution of neural time-to-peak across voxels of an area averaged across participants for early visual areas (V1, V2, V3) as well as intermediate and later visual areas in the ventral (V4, VO), lateral (LO, TO), and dorsal streams (V3AB, IPS) was plotted. As evident from Figure 7A, neural time-to-peak increases from earlier to later areas but also varies within each visual area.

• Download figure
• Open in new tab
• Download powerpoint

Figure 7.

Temporal pRF estimates from the CST model. A–C, Each row shows a voxel-wise time-to-peak distribution for each visual area. The distributions were averaged across participants within each visual area. A, Time-to-peak ( τ) distribution estimated from the CST model, using an optimized HRF for each voxel. B, The distribution of optimized HRF time-to-peak. Note that unlike the CST model, the temporal resolution for the hemodynamic model is in seconds. C, Time-to-peak estimated from the CST model using a single default Vistasoft HRF for all voxels. D, The cumulative time-to-peak distribution across visual areas from the CST model with optimized voxel-wise HRF for each of the nine visual areas. The datapoints from all participants were included for each visual area.

As we estimated for each voxel its optimized HRF (Fig. 2), it is possible that this variability in the time-to-peak is of hemodynamic rather than neural origin. To examine this possibility, we conducted two additional analyses. First, we analyzed the distribution of time-to-peak of the hemodynamic functions within and across areas. We reasoned that if the temporal variability is of hemodynamic origin, then the distributions of optimized HRF time-to-peak will mirror the distributions of the estimated neural time-to-peak. However, this was not the case. The distribution of hemodynamic time-to-peak ranged from 3 to 6 s and was similar across all tested visual areas (Fig. 7B). Critically, this distribution (Fig. 7B) did not show the between-area differences in neural time-to-peak estimated by the CST model (Fig. 7A,C). Second, we repeated the calculations of the CST neural time-to-peak ( τ) using a fixed HRF for all voxels (Vistasoft default HRF). We hypothesized that if the estimated neural time-to-peak interacts with the estimated HRF parameters for each voxel, then using a fixed HRF will qualitatively change the estimates of neural time-to-peak within and across areas. However, we found that using a fixed HRF for all voxels produces quantitative but not qualitative changes in estimates of neural time-to-peak (Fig. 7C). Specifically, using a fixed HRF increased the within-area variability of neural time-to-peak estimates (e.g., compare V3 in Fig. 7C to 7A), but did not change the observation that the time-to-peak progressively increases across visual the hierarchy. These analyses give us confidence that our approach enables estimating neural temporal properties. In the next sections, we will examine in detail the temporal and spatiotemporal parameters from the CST model with the voxel-wise optimized HRF.

Hierarchical temporal latency delay across visual streams

An open question is how the time-to-peak and processing temporal window changes across the visual hierarchy and processing streams. A feedforward model of the visual system predicts that the time-to-peak will progressively increase across the visual hierarchy (Nowak and Bullier, 1997). Additionally, the ventral stream which is associated with perception of invariant properties of people or objects (such as their identity or category) may be slower and have longer times-to-peak than other streams, like the lateral stream, which is thought to process motion and dynamical aspects of the visual scene (Van Essen and Gallant, 1994; Weiner and Grill-Spector, 2013; Pitcher and Ungerleider, 2021; Wurm and Caramazza, 2022).

We found differences in time-to-peak both across the hierarchy and across streams, thus finding evidence supporting both hypotheses. Across the hierarchy, time-to-peak increased from earlier (V1, V2, V3) to later visual areas in each processing stream generating a cascade of neuronal signal propagation. We found that voxels in V1 had the earliest time-to-peak latencies (τ∼50–110ms Fig. 7A), and that time-to-peak was less than 100 ms for more than 50% of V1 voxels (Fig. 7D). V2 followed V1 (τ∼50–120ms ), and V3 followed V2 (τ∼50–130ms ). In the ventral visual stream, later time-to-peak was observed in voxels of intermediate visual area hV4 (τ∼70–170ms Fig. 7A,D, green) which were followed by responses in a late visual area, VO (τ∼80–200ms , Fig. 7A,D, dark green). Similar trends were observed in the lateral (LO to TO, pink and red in Fig. 7D) and the dorsal streams (V3AB to IPS, yellow and brown in Fig. 7D). Timing of the intermediate areas did not vary across streams with hV4 (ventral), LO (lateral), and V3AB (dorsal) having similar time-to-peaks. However, we found differences across later areas of the different streams. TO (lateral) had a later times-to-peak than VO (ventral), and IPS had the latest time-to-peak (τ∼160–230ms ) compared with other tested visual regions (Fig. 7A,D, brown).

Spatiotemporal processing across visual streams

Crucially, the CST model allows us to examine the spatiotemporal properties of pRFs across visual streams. Prior research has suggested that both spatial receptive field sizes (for review, see Wandell and Winawer, 2015) and temporal windows (Hasson et al., 2008; Honey et al., 2012; Murray et al., 2014; Baldassano et al., 2017) progressively increase across the visual hierarchy. However, before the current framework, it was not possible to estimate spatial and temporal pRF sizes in degrees and milliseconds directly at the voxel level.

Figure 8A shows the average sustained and transient spatiotemporal pRF across voxels of a visual area (see Materials and Methods, estimating spatiotemporal pRFs across voxels of a region and across cortex). Earlier visual areas had smaller pRF sizes and shorter temporal windows compared to later visual areas for both sustained and transient channels. For example, ascending the ventral stream (V1 → V2 → V3 → hV4 → VO), spatiotemporal pRF windows progressively increased both spatially and temporally (Fig. 8B). Differences between visual areas were statistically significant both for pRF size (F_(1,84) = 57.15, p < 0.001) and temporal window (F_(1,84) = 33.69, p < 0.001). In the sustained channel, V1 pRFs integrate visual information spatially across 0.41° ± 0.05° (median ± SEM) and temporally across 58.77 ± 10.71 ms, hV4 across 1.63° ± 0.13° and 88.27 ± 9.93 ms, and VO across 2.39° ± 0.18° and 98.19 ± 9.66 ms (Fig. 8A,B). Likewise, in the lateral stream, TO spatiotemporal pRFs were larger in space and time compared to LO, and in the dorsal stream, IPS spatiotemporal pRFs were larger in space and time than V3AB. In the transient channel, we find a similar progression (Fig. 8A, right).

• Download figure
• Open in new tab
• Download powerpoint

Figure 8.

Spatiotemporal pRF characteristics for nine visual areas spanning three visual streams. A, Spatiotemporal pRFs for sustained and on-transient pRFs averaged across voxels and then across participants for each visual area. The spatial location of all pRFs was zero-centered to (0,0) before averaging. The x-axis represents time (ms), and the y-axis represents a cross section of visual space (°). The red–white–blue colormap is the average of spatiotemporal pRFs, where red and blue represent the positive and negative amplitudes, respectively. B, PRF temporal window (FWHM of the sustained temporal impulse response function) versus pRF size ( σ deg). Colored dots: median values for each visual area. Marker size is related to median values of the spatiotemporal compressive exponent (marker size ∼ 1/n). That is, larger markers indicate larger compression. Error bars: ±1 SEM across 10 participants in each dimension.

To quantitatively compare the spatiotemporal pRF size and temporal window estimates, we computed the average pRF size and average temporal window of each visual area (Fig 8B). The comparison showed a positive relationship, revealing a general trend of progressively increasing spatial and temporal windows along the visual hierarchy. Specifically, spatiotemporal pRF size and temporal window progressively increased from V1 to V2 to V3 (Fig. 8B, blues). Intermediate areas of all three visual streams (hV4, LO, and V3AB) had similar spatial and temporal pRF sizes which were larger spatially, and longer temporally than V1–V3. The spatiotemporal properties of later visual areas (VO, TO, and IPS) differed across streams. Notably, spatiotemporal pRFs in VO, a ventral area, were smaller spatially and shorter temporally (Fig. 8B, dark green) than pRFs in TO, a lateral area (Fig. 8B, red), whereas IPS had pRFs that were smaller spatially, but longest temporally (Fig. 8B, brown). In addition, the amount of spatiotemporal compression increased along the visual hierarchy (Fig. 8B, marker size ∼ 1/n). The compressive exponent (n) in V1: 0.28 ± 0.03, V2: 0.25 ± 0.06, V3: 0.24 ± 0.06, hV4: 0.22 ± 0.04, VO: 0.21 ± 0.03, LO: 0.19 ± 0.02, TO: 0.18 ± 0.02, V3AB: 0.21 ± 0.02, IPS 0.14 ± 0.02, which is comparable to the increase in spatial compression levels from early to late visual areas (Kay et al., 2013). These results illustrate an interesting coupling between spatial and temporal receptive field windows that progressively increase within each of the three processing streams and diverge across streams in the later visual areas.

When comparing the contributions of the sustained and transient channels across the visual hierarchy, we found stronger transient responses than sustained responses in all visual areas except for V1 (Fig. 9), which is similar to the findings of Stigliani et al. (2017). We further examined contributions of sustained and transient channels across eccentricities in early visual cortex (data not shown). Consistent with previous findings (Horiguchi et al., 2009; Stigliani et al., 2017), we observed balanced sustained and transient contributions between 3° and 12° of eccentricity in V1, but we found higher transient responses in the central 5° than Horiguchi et al. (2009).

• Download figure
• Open in new tab
• Download powerpoint

Figure 9.

Contributions of sustained and transient responses across the visual hierarchy. Average sustained (y-axis) and transient (x-axis) response amplitudes (β weights) for each visual area estimated by the CST model. Marker size in each dimension indicates Median ±1 SEM across 10 participants. Dashed line: equal sustained and transient contributions.

Finally, we investigated the spatial topography of spatiotemporal receptive fields across visual cortex. Confirming our quantitative results in Figure 8B, qualitative examination of the maps of temporal window and pRF size reveals a topographic coupling of spatial and temporal pRF windows (Fig. 10).

• Download figure
• Open in new tab
• Download powerpoint

Figure 10.

Spatiotemporal pRF properties across eccentricities in early visual cortex. A, Average pRF temporal window and spatial window visualized on the FreeSurfer average cortical surface. Both hemispheres and two different viewpoints are shown (left: medial view, right: posterior view). Top: FWHM of average sustained temporal window estimates; Bottom: average pRF size. B, The relationship between eccentricity and temporal window in V1, V2, and V3. The eccentricities were binned into 2° of visual angle. Colored dots: individual participant. Solid line: fitted LMM regression line. C, Mean sustained spatiotemporal pRF in two different eccentricity bands (0–4° and 4–12°). Contour lines indicate 10th, 50th, and 90th percentiles.

To gain deeper insights, we examined this topography in early visual areas (Fig. 10A, left), where we observed that pRF temporal windows (Fig. 10A, top-left) increase along a posterior to anterior gradient and pRF sizes (Fig. 10A, bottom-left) also increase along the same axis. Quantitative analyses using an LMM with a fixed slope and random intercept per participant revealed a significant relationship between temporal window and eccentricity in V1, V2, V3 (Fig. 10B). Temporal windows were progressively larger in more peripheral eccentricities [V1: t₍₅₈₎ = 8.16, p < 0.001; V2: t₍₅₈₎ = 6.67, p < 0.001; V3: t₍₅₈₎ = 4.19, p < 0.001 with corresponding beta coefficients of 4.39, 3.78, and 2.83 and across areas (intercept of 77.35, 83.90, and 85.30, respectively)]. As eccentricity also increases from posterior to anterior in early visual areas, this suggests that in V1–V3, more eccentric pRFs have larger spatial and temporal windows than foveal pRFs. Figure 10C visualizes the average spatiotemporal pRFs at two eccentricity bands: 0–4° and 4–12°. Indeed, in V1–V3, both pRF size and temporal window were larger in eccentricities exceeding 4° than the central 4°. The proportional increase in both spatial and temporal dimensions with eccentricity suggests that there are computational similarities in how the visual system encodes spatial and temporal information at the center and periphery.

A similar topography showing a coupling between pRF size and temporal window is also seen in the lateral surface, which contains intermediate and higher-level regions: more posterior regions have smaller spatial and shorter temporal windows than anterior regions (Fig. 10A, right). Within later visual areas, the eccentricity-dependent relation between pRF size and temporal windows was significant in LO (t₍₅₈₎ = 2.37, p < 0.05), while the opposite effect was observed in VO (t₍₅₈₎ = −4.87, p < 0.001). No significant eccentricity-dependent effects were found in V4 (t₍₅₈₎ = −0.64, p = 0.53), TO (t₍₅₈₎ = −0.88, p = 0.38), V3AB (t₍₅₈₎ = 1.87, p = 0.07), and IPS (t₍₅₈₎ = 0.01, p = 0.99).

Together these analyses reveal three organizational principles. First, both spatial and temporal windows progressive increase across the visual hierarchy. Second, within early visual cortex, both spatial and temporal windows increase with eccentricity. Third, transient responses are dominant across all three visual processing in our experiment.