Learning Improves Peripheral Vision Via Enhanced Cortico-Cortical Communications

Dear Editor,

The adult human visual system is capable of reshaping its oculomotor control and sensory coding to adapt to impoverished visual inputs. When one’s central vision is deprived, a spared part of the peripheral retina acts as a pseudo fovea termed as preferred retinal locus (PRL). In people with normal vision, a PRL can be induced via oculomotor training with simulated central vision loss [1, 2].

Notably, the development of a PRL triggers cortical reorganization [3, 4], and is accompanied by an alleviation of the bottleneck in peripheral vision by reducing visual crowding [5, 6]. Visual crowding is believed to arise from excessive pooling in the early visual stages [7]. Converging evidence from neuroimaging studies has pointed the neural locus of crowding to V1 and V2 [8, 9], revealing a suppressive interaction among nearby items. An alternative account of crowding, however, proposes that the information from crowded items is not lost in the visual cortex. Instead, the information failed to reach the higher-level stage due to a limit in selection within or above the visual hierarchy [10, 11].

To examine how PRL training reorganizes the cortico-cortical communication of crowded visual information, we conducted functional connectivity and effective connectivity analyses on the BOLD fMRI signals recorded before and after training in normally sighted adults. During the scan, crowded letters were displayed at the PRL while subjects performed a central fixation task. Background connectivity was computed based on residual time series after removing stimulus-evoked signals.

Before and after PRL training (Fig. 1A), we measured the effect of crowding at the PRL location along the radial and the tangential axes, as the minimum distance between the target and flankers that yielded reliable target identification performance (Fig. 1B). In Chen et al. [6], a reduction of the crowding effect has been identified along the radial axis [t(8) = 2.44, P = 0.04].

Fig. 1.

Protocols and results of PRL training and fMRI experiments. A PRL training task. Subjects were asked to use the gaze marker to follow the target object (e.g., an apple) as it moved to a random location in a cluttered background. Each time a stable fixation was reached for 2 s, the target jumped to a new location. After six iterations, subjects reported whether or not the target was present among multiple objects by a key press. Central vision loss was simulated with a scotoma, which is a circular gray patch with a radius of 5°. A cross (6.5° eccentricity, 30° counterclockwise from the vertical meridian) marks the position of the PRL. Both the central scotoma and the cross were visible and gaze-contingent during training. B Crowding task in the behavioral test. Subjects were asked to report the orientation of a target “T” in the middle of two flankers in a four-alternative forced-choice (4-AFC) trial. C Stimuli configuration in the fMRI measurements. D Results of V1-based background connectivity at the PRL and the fovea in pre- and post-tests. E Sources of variance in the residuals of the current stage, with corresponding V2 BOLD signal and V1–V2 correlations at fovea and PRL locations in pre- and post-tests. Error bar/shaded area denotes ± 1 SEM across subjects. *: P < 0.05; **: P < 0.01; ***: P < 0.001; FDR corrected.

Voxels in the extrastriate cortex V2–V4, VO, and the intraparietal sulcus (IPS) that showed a stronger response to stimuli were identified as seeds for computing correlations with vertexwise V1. We compared the V1-based background connectivity before and after learning, at PRL (eccentricity range: 5°–8°) and the foveal untrained control location (eccentricity range: 1°–4°), both within a polar range of ± 30° from the stimulus center. Two-way repeated-measures ANOVAs with factors of the visual field (fovea/PRL) and training (pre/post) were conducted to examine the connectivity strength between V1 and each of the downstream areas: V2, V3, V4, VO, and IPS (Fig. 1D). An interaction between visual field and training was found in V1–V2 connectivity [F(1, 8) = 5.72, P = 0.044, BF₁₀ = 1.52]. Post-hoc t-tests revealed that the V1–V2 connectivity at PRL was significantly enhanced after learning [t(8) = 2.75, P = 0.025, BF₁₀ = 3.08]. For V1–V3 connectivity, repeated-measures ANOVA revealed a marginally significant main effect of learning [F(1, 8) = 4.17, P = 0.076, BF₁₀ = 1.02], showing an enhancement at the PRL after learning [t(8) = 4.09, P = 0.004, BF₁₀ = 14.75].

For V1-based connectivity with other areas, only a significant main effect of the visual field was found [all F(1, 8) > 5.61, P < 0.05]. As a control analysis, we examined background connectivity at the PRL area between V2 and V3, and between V3 and V4. Neither of these connections exhibited a significant increase after training [both t(8) < 1.22, P > 0.26]. These findings suggest that enhanced connectivity was not a general phenomenon across adjacent visual areas.

The question then arises as to whether the learning-dependent modulation of functional connectivity simply reflects the learning-dependent modulation of evoked responses. To address this question, we compared the BOLD signal in V2 and functional connectivity between V1 and V2 at each stage of analysis (Fig. 1E). After removing the stimulus-evoked component, the percentage of explained BOLD signal change decreased from 18.56% to 6.54% [main effect of stage: F(2, 16) = 19.41, P < 0.001]. Despite this change, the learning-dependent modulation of connectivity was stable across stages [stage × ROI × training phase interaction: F(2, 16) < 2.10, P > 0.05]. Across all stages, the interaction between visual field and training was significant or marginally significant [Stage 1: F(1, 8) = 6.52, P = 0.034, BF₁₀ = 1.82; Stage 2: F(1, 8) = 4.58, P = 0.065, BF₁₀ = 1.42; Stage 3: F(1, 8) = 5.73, P = 0.044, BF₁₀ = 1.52]. For Stage 2 and 3, post-hoc t-test showed a significant enhancement at PRL after training [Stage 1: t(8) = 1.91, P = 0.092, BF₁₀ = 1.16; Stage 2 and 3: both t(8) > 2.74, P < 0.03, BF₁₀ > 3.08]. These results suggest that learning-dependent connectivity changes were independent of stimulus-evoked responses.

Given that the learning-dependent modulation in V1–V2 and V1–V3 was observed with anecdotal and moderate evidence in ANOVA tests, we further testified the learning effect on the directional connections among these areas. Bidirectional intrinsic connections were assumed among all nodes. We examined feedforward, feedback, and recurrent models that differed in the modulatory site of learning (Fig. 2A). The optimal model with the modulatory sites in the feedforward connections showed an exceedance probability of 82.62%. In this model, the feedforward connection from V1_P to V2 [t(8) = 3.27, P = 0.013] and V3 [t(8) = 31.70, P < 0.001] was significantly enhanced after learning, while the connection from V1_F to V2 [t(8) = 3.30, P = 0.013] and V3 [t(8) = 18.20, P < 0.001] was significantly weakened.

Fig. 2.

Results of DCM analysis (A) and eccentricity-based connectivity (B). A Candidate model assumes that learning modulates the feedforward, feedback, or recurrent connections among modeled regions, respectively. A Bayesian model selection with a random effect analysis (RFX) was used to determine the optimal model. The changes in the modulatory effect at the post- relative to the pre-test were examined in the optimal model. *P < 0.05; ***P < 0.001; FDR corrected. B Top panel: V1–V2 background connectivity across eccentricities in pre- and post-tests. The yellow cluster denotes a significant difference: P < 0.05. Bottom panel: Retinotopic connectivity between V1 and downstream areas at the target and the flanker locations in pre- and post-tests. Error bar/shaded area denotes ± 1 SEM across subjects.

Finally, we characterized the retinotopic background connectivity within the visual field covering the target and its radial flankers (polar angle range: ± 30° from the center of the target) to differentiate the learning-dependent connectivity change at the target and its flanking locations. V1-based connectivity was plotted across an eccentricity range of ± 1.5°, centering at the target location with a step size of 0.1° (Fig. 2B). After training, a trend of overall enhancement was observed [main effect of training: F(1, 7) = 6.33, P = 0.04]. For each downstream area, V1-based connectivity changes at the target and its inner and outer flanker locations (± 0.5° offset from the centroid) were quantified separately. Repeated-measures ANOVA with training (Pre/Post) and location (Inner flanker/Target/Outer flanker) as two factors showed a main effect of training in V1–V2 connectivity [F(2, 8) = 12.15, P < 0.01, BF₁₀ = 5.79] and in V1–V3 connectivity [F(2, 8) = 11.95, P < 0.01, BF₁₀ = 3.47]. However, no significant interaction between training and location was found in the V1-based connectivity with other brain areas [all F(2, 16) < 1.75, P > 0.05, BF₁₀ < 0.8].

In this study, we found enhanced connectivity at the PRL between V1 and the extrastriate visual cortex. DCM analysis further revealed enhanced V1 → V2 and V1 → V3 connectivity. These findings point to an important role of the primary visual cortex in prioritizing inter-areal signal transmission as an adaptive strategy under impoverished foveal visual input. Our results complement previous findings on intra-areal changes in BOLD signals after PRL training. Along the visual pathway, crowding-related response suppression [8, 9, 12] has been found to be alleviated by PRL training [6]. Specifically, the amount of crowding-related suppression before learning and learning-induced change showed a trend that peaked in V1 [6]. Consistently, we identified an enhancement in V1-based connectivity at the PRL. This inter-areal connectivity enhancement, which accompanies the reduction of intra-areal interference at the PRL location, constitutes a critical part of plasticity in the visual hierarchy after oculomotor adjustment.

Furthermore, the learning-dependent enhancement was observed in the feedforward pathway, from V1 to the extrastriate visual cortex. These results suggest that training alleviates the bottleneck of crowding in feedforward information transmission, which provides empirical evidence in support of a reweighting theory. According to the reweighting theory, learning and other cognitive process can be implemented by adjusting the weights between visual channels in the cortical hierarchy [13–15]. The downstream visual areas may increasingly rely on channels that represent task-relevant information. Such a mechanism may help reconcile the results regarding V1 remapping following retina lesions [3, 16, 17]. Compared to intra-areal retinotopic remapping, interareal routing may help maintain a stable topographic representation while allowing for prioritizing the task-relevant information in the visual system.

In sum, the observed changes in V1-based connectivity with extrastriate visual areas may underpin the rapid and persistent adaptability of the human visual system in response to central vision loss. It should be noted that the small sample size in this study may limit the extent of statistical power to fully characterize the learning effect in cortico-cortical transmission. More comprehensive investigations are needed in future studies. Our findings shed light on understanding how adaptable neural circuits are and how to optimize clinical interventions for diseases that cause a selective loss of function.

Supplementary Information

Below is the link to the electronic supplementary material.

Data Availability

The data and scripts used to generate the results will be uploaded to the Open Science Framework (OSF) and are available from the corresponding author upon reasonable request.

Conflict of interest

The authors declare that they have no competing interests.

Ethical Approval

This study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the institutional review board of the University of Southern California.

Consent to Participate

Informed consent was obtained from all individual participants included in the study.