(C) PLOS One
This story was originally published by PLOS One and is unaltered.
. . . . . . . . . .



Human subcortical pathways automatically detect collision trajectory without attention and awareness [1]

['Fanhua Guo', 'State Key Laboratory Of Brain', 'Cognitive Science', 'Institute Of Biophysics', 'Chinese Academy Of Sciences', 'Beijing', 'University Of Chinese Academy Of Sciences', 'Jinyou Zou', 'Aier Institute Of Optometry', 'Vision Science']

Date: 2024-01

Detecting imminent collisions is essential for survival. Here, we used high-resolution fMRI at 7 Tesla to investigate the role of attention and consciousness for detecting collision trajectory in human subcortical pathways. Healthy participants can precisely discriminate collision from near-miss trajectory of an approaching object, with pupil size change reflecting collision sensitivity. Subcortical pathways from the superior colliculus (SC) to the ventromedial pulvinar (vmPul) and ventral tegmental area (VTA) exhibited collision-sensitive responses even when participants were not paying attention to the looming stimuli. For hemianopic patients with unilateral lesions of the geniculostriate pathway, the ipsilesional SC and VTA showed significant activation to collision stimuli in their scotoma. Furthermore, stronger SC responses predicted better behavioral performance in collision detection even in the absence of awareness. Therefore, human tectofugal pathways could automatically detect collision trajectories without the observers’ attention to and awareness of looming stimuli, supporting “blindsight” detection of impending visual threats.

Funding: This study was supported by Ministry of Science and Technology of China ( https://en.most.gov.cn/ ) STI2030-Major Projects (2022ZD0211900 to P.Z., 2021ZD0204200 to S.H.), National Natural Science Foundation of China ( https://www.nsfc.gov.cn/english/site_1/index.html , 31871107 and 31930053 to P.Z., 82101110 to H.Z.), Chinese Academy of Sciences ( https://english.cas.cn/ , XDB32020200, KJZD-SW-L08, YSBR-071) to S.H., and Beijing Natural Science Foundation ( https://mis.kw.beijing.gov.cn/ , 7212092) and the Capital’s Funds for Health Improvement and Research ( https://wjw.beijing.gov.cn/ , 2022-2-5041) to D.W.. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

To answer these questions, the current study used high-resolution fMRI to investigate the neural pathways involved in detecting collision trajectory in both healthy human participants and hemianopic patients. The motion trajectory of an incoming object was varied slightly to be either on a collision course or on a near-miss trajectory regarding the head of observers. For healthy participants, we first measured their visual ability to discriminate collision course from near-miss trajectories, and the associated changes in pupillary reflex (experiment 1). In a 7T fMRI experiment (experiment 2), we further investigated the neural circuitries for detecting collision trajectories and examined whether these mechanisms require top-down attention or could operate automatically without paying attention to the stimuli. In a group of hemianopic patients with unilateral lesions of the geniculostriate pathway, experiment 3 studied whether the tectofugal pathways are sensitive to collision trajectories even without awareness of stimuli presented in their blind visual field.

Moreover, it is unknown whether these collision detection mechanisms could work automatically even without the observer’s attention to and awareness of the looming stimuli, as suggested by previous behavioral studies in humans [ 13 , 14 ]. Some patients with cortical blindness can detect or even discriminate visual stimuli presented to their blind visual field, despite denial of seeing the stimuli. This phenomenon, called blindsight, attracts broad interests and has been hotly debated for almost half a century [ 19 ]. Although lots of studies have been done, the neural pathways involved and their specific functional roles in blindsight remain highly controversial. Looming-evoked avoidance behavior was observed in monkeys with V1 lesions [ 20 ], suggesting a critical role of subcortical pathways in automatic “blindsight” detection of impending visual threats. However, direct evidence for this hypothesis is still lacking.

The superior colliculus (SC) is a phylogenetically old visual nucleus lying on the roof of the mammalian brainstem. It plays important roles in visual perception and visually guided reorienting functions, such as attention, eye and head movements [ 15 ]. As a key retino-recipient region, the SC and its homologous structure in nonmammals, optic tectum, were found preferentially sensitive to looming stimuli in a number of species [ 4 , 5 , 8 ]. In rodents, specific types of neurons in the SC have been identified as the key components of several subcortical circuits to detect looming objects and trigger defensive responses, including the SC-PBGN (parabigeminal nucleus)-Amygdala [ 4 , 16 ], SC-LP (lateral posterior thalamic nucleus, homolog of primate pulvinar)-Amygdala [ 5 , 16 ] and SC-VTA (ventral tegmental area)-Amygdala [ 17 ], and LC (locus coeruleus)-SC [ 18 ] connections. Given that the subcortex is the primitive brain and that subcortical functions might be relatively conserved in mammals, the subcortical pathways found in the rodent brain might also play important roles for collision detection in the human brain. Alternatively, with the expansion of the neocortex and reduced retinal projection to the SC in higher mammals [ 15 ], it could be possible that the cerebral cortex plays a more prominent role in collision detection in the human brain. An early 3T fMRI study at a relatively low resolution suggests that the human SC, pulvinar, exhibited stronger responses to looming compared to receding stimuli [ 8 ]. However, it is unclear whether these subcortical responses were a result of cortical influence. Most importantly, the underlying neural circuitry for detecting collision trajectories remains unknown.

Detecting objects approaching on a collision course is critical for the survival of animals in the environment. Specialized neurons or brain circuits highly sensitive to looming stimuli have been identified in many species, including insects [ 1 ], fish [ 2 ], pigeons [ 3 ], mice [ 4 , 5 ], and others. In primates, both adult monkeys and human infants elicit avoidance behaviors to symmetrically looming stimuli [ 6 , 7 ], imaging studies mainly revealed looming sensitive responses in the cortical brain regions [ 8 – 12 ]. These previous studies mostly compared a large looming stimulus versus translating or receding stimuli, whose motion directions greatly deviate from the collision course. The neural mechanism for computing the time-to-contact (TTC) information has been extensively studied and was suggested to provide warning detection for large approaching objects. However, humans are not only sensitive to looming versus nonlooming stimuli but also efficient in detecting collision course from near-miss trajectories (i.e., a few centimeters away from the head at the pass point). For example, compared with a near-miss trajectory, an approaching object on a collision course with the observer could automatically capture visual attention, increase perceived object size and evoke greater pupil constrictions, even without conscious awareness of the motion direction [ 13 , 14 ]. These behavioral findings suggest early subcortical mechanisms for collision detection, through precise and sensitive measures of motion trajectories. However, little is known about the neural mechanism in the human brain for detection of collision trajectories.

To check potential “blindsight” to approaching objects on a collision course, 4 patients (P09 to P12) also performed a 2-interval forced choice (2-IFC) detection task. The stimulus was presented in one of two 330-ms intervals with a 330-ms gap in between, accompanied by a low-pitch tone and a high-pitch tone in the first and the second intervals, respectively ( Fig 6F ). Patients had to determine in which interval the stimulus was presented. The detection performance was fitted with a binomial generalized linear mixed model with a random effect of participant ( Fig 6G ). Results showed a significant main effect of trajectory (χ 2 (2) = 6.418, p = 0.04). Following tests showed that the accuracy for the object on a collision course was significantly higher than that on a receding trajectory (contrast = 10.6%, z = 2.529, p = 0.023 using Holm adjustment) and then 50% chance level (estimated marginal mean = 61.1%, z = 3.541, p < 0.001). Importantly, the detection performance showed a significant linear relationship with the response of ipsilesional SC (main effect of SC response on detection performance in a linear mixed model with participant as the random effect: F(1, 3.05) = 14.616, p = 0.031, Fig 6H ). No significant linear relationship was found between behavioral performance and the responses of vmPul or VTA. Altogether, the behavioral and fMRI results of hemianopic patients provide strong evidence in humans that the tectofugal pathways support blindsight to impending collisions.

The LGN was also investigated since previous studies suggest that it plays a critical role in blindsight [ 36 ]. However, no collision-sensitive response was found in the LGN ( S12 Fig ). In the cortical regions, V1 ( S11 Fig ) and frontoparietal areas ( Fig 6D ) showed strong activations to stimuli presented in the NVF but not in the BVF. These findings demonstrate that the tectofugal pathways can automatically detect collision trajectories even without awareness of looming stimuli and a functional geniculostriate pathway.

(a) Group-averaged activation maps in the SC to stimuli presented to the NVF (left column) and BVF (right column). Maps were thresholded at p < 0.05 uncorrected. Color bars represent percent BOLD signal change. The left and right SCs were mapped contralateral to the NVF and BVF, respectively. (b, c) Activation maps in the vmPul ( b ) and VTA ( c ). Conventions are the same as in ( a ). Only activations within the ROI were shown. Red arrows indicate the location of collision-sensitive clusters from the Hit-Miss map. (d) Stimulus activations mapped to the cortical surface (thresholded at p < 0.01 uncorrected). Left columns show contra- and ipsilateral responses to stimuli presented to the NVF. Right columns show ipsi- and contralateral responses to stimuli presented to the BVF. (e) BOLD responses in the contralateral SC, vmPul, and VTA to looming stimuli presented to the NVF (left panel) and BVF (right panel). Each dot represents data from 1 patient. + and * represent p < 0.05 before and after Holm corrections, respectively. Error bars represent the 95% confidence intervals estimated in a LME model. (f) Schematic diagram and procedure for the 2-IFC detection task. (g) The detection permanence fitted with a binomial generalized linear mixed model. Error bars represent the 95% confidence intervals. * Post hoc test p < 0.05 for comparing marginal means between conditions. *** p < 0.001 for comparing marginal means to the chance level (50%). (h) The detection performance to stimuli in the BVF showed a linear relationship with the response of ipsilesional SC (in 10% most responsive voxels, chosen separately for each stimulus type). The black solid line denotes the fixed effect (SC response) of the linear mixed model. For each participant, SC responses were normalized by dividing the mean across 3 stimulus conditions and then multiplying the mean across all 4 participants. Data underlying (e, g, h) can be found at https://osf.io/gdjwh/ . BVF, blind visual field; LME, linear mixed effect; NVF, normal visual field; Rec, receding stimulus; SC, superior colliculus; vmPul, ventromedial pulvinar; VTA, ventral tegmental area; 2-IFC, 2-interval forced choice.

Based on the findings of healthy participants, we focused our analysis on the SC, vmPul, and VTA in hemianopic patients. From the group-averaged activation maps in Fig 6A , significant clusters of activation to hit trajectories were found in the contralateral SCs regardless of whether the stimuli were presented in the NVF (cluster p = 0.023) or in the BVF (cluster p = 0.031). In the contralateral vmPul ( Fig 6B ) and VTA ( Fig 6C ), collision-sensitive clusters to stimuli presented to the BVF (voxel p < 0.05, uncorrected) were also found in similar locations as the results of heathy participants ( Fig 3 ). To increase statistical power given the small number of hemianopic patients, we performed a linear mixed effect (LME) analysis with ROI-averaged responses of visually responsive voxels significantly activated by the receding stimuli (group-level p < 0.05 uncorrected). In the LME model, trajectory (Hit/Miss), visual field (NVF/BVF), and ROI (contralateral SC/vmPul/VTA) were defined as the fixed effects and participants as the random effect. The analysis revealed a significant hit response averaged across all 3 ROIs in the BVF ( Fig 6E , z = 2.59, p = 0.038, Holm corrected across visual fields and trajectories), post hoc tests revealed significant hit responses in the SC (z = 2.054, p = 0.04 uncorrected) and VTA (z = 2.768, p = 0.017, Holm corrected across ROIs). A significant hit response was found in the contralateral SC to stimuli presented in the NVF (z = 3.468, p = 0.012, Holm corrected across all 12 conditions).

To further investigate whether the tectofugal pathways can detect collision trajectories even without awareness of visual stimuli and a functional geniculostriate pathway, we scanned a group of homonymous hemianopic patients with unilateral lesions of the geniculostriate pathway (experiment 3, n = 12). Patients lost their conscious vision of both eyes in one side of the visual field (see S1 Table for clinical characteristics and S11 Fig for visual field loss and lesioned locations). During fMRI scans, patients performed a central fixation task while stimuli with hit, near-miss, or receding trajectories were presented to their normal visual field (NVF) or blind visual field (BVF). Four patients (P09 to P12) also participated in a behavioral visibility test, in which they reported clear perception of stimuli presented to their NVF but denied seeing stimuli from the BVF. For most patients (8 of 12), there was no significant V1 activation to stimuli presented in their BVF ( S11 Fig ). Four patients showed weak uncorrected activations in their lesioned hemisphere, but no collision sensitivity.

Using a beta-series analysis [ 34 ] of the looming-evoked responses, we found significant functional connectivity between SC and vmPul and between SC and VTA in both attended and unattended conditions (all p < 0.001, Holm corrected). To identify the subcortical pathways from the SC and the information flow between cortical and subcortical areas, we performed a SEM path analysis with the beta series from these ROIs. A total of 216 candidate models were constructed with different combinations of 6 sets of alternative connections ( Fig 5C ), based on the known anatomical connections between these areas: SC-cortex [ 15 ], SC-vmPul [ 35 ], SC-VTA [ 17 ], vmPul-cortex [ 24 ], VTA-cortex [ 33 ]. Candidate models were fitted with the observed data and compared by the goodness of fit (see Methods for model comparisons). The best-fitted models for both attended (fit index: χ 2 = 18.481, df = 2, CFI = 0.975, GFI = 0.997, AGFI = 0.978, PGFI = 0.133, RMSEA = 0.057, RMR = 0.019) unattended (fit index: χ 2 = 21.370, df = 2, CFI = 0.968, GFI = 0.997, AGFI = 0.975, PGFI = 0.133, RMSEA = 0.062, RMR = 0.022) conditions revealed highly significant SC-vmPul and SC-VTA connections (all p < 0.001). Importantly, the model shows no cortical influence to the SC in the unattended condition. These findings support a pivotal role of the SC-vmPul and SC-VTA pathways in imminent collision detection even without attention. The collision sensitivity found in the SC was unlikely a result of cortical influence.

(a) Correlation matrix of collision sensitivity. Each grid in the matrix shows the Pearson’s correlation coefficient between the collision sensitivities of 2 ROIs, in the attended (left) and unattended (right) conditions. + indicates uncorrected p < 0.05 and * for p < 0.05 after FWE correction by permutation test. Color bars indicate the size of correlation coefficient. (b) Correlations between the collision sensitivity in the SC, and those in subcortical (vmPul/VTA) and cortical (VC/AttNet) areas. Each dot represents 1 participant. (c) Candidate SEM models for the effective connectivity of beta series between ROIs. Gray solid single-headed arrows represent fixed one-way connection. Red solid double-headed arrows indicate connections with 2 alternative directions. Dotted double-headed arrows indicate that the connection may follow one of 2 alternative directions or may not exist. All combinations of alternative connections yield 216 candidate models. (d, e) Best fitted models in the attended (d) and unattended (e) conditions. Arrows with solid lines indicate significant connections. There was no insignificant connection. Data underlying (b) can be found at https://osf.io/gdjwh/ . AttNet, frontoparietal attention network; FWE, family-wise error; SC, superior colliculus; VC, visual cortex; vmPul, ventromedial pulvinar; VTA, ventral tegmental area.

In the attended condition ( Fig 5A , left panel), we found significant correlations of collision-sensitive responses between the SC and its downstream subcortical target vmPul (r = 0.556, uncorrected p = 0.014), between VC and VTA (r = 0.478, uncorrected p = 0.033), and between VC and AttNet (r = 0.651, p = 0.021, after family-wise error (FWE) correction of the correlation matrix). In the unattended condition ( Fig 5A , right panel, and 5b), there were significant correlations between the SC and both downstream subcortical targets (SC-vmPul: r = 0.593, uncorrected p = 0.007; SC-VTA: r = 0.564, uncorrected p = 0.012), but not between cortical and subcortical regions (all uncorrected p > 0.1).

To further investigate the subcortical circuits involved in collision detection and the potential influence from cortical brain areas, we calculated across-participant correlations between collision-sensitive responses in the subcortical nuclei (SC, vmPul, and VTA) and cortical areas including the visual cortex (VC) and frontoparietal attention network (AttNet), followed by a path analysis with structural equation modeling (SEM) on the beta series of looming-evoked responses to infer their effective connectivity. We selected these ROIs because significant collision-sensitive responses were found in these regions. The vlPul was not included in this analysis because the SC’s projection to this area is weak and controversial [ 25 , 26 ], and the collision-sensitive responses in the SC and vlPul showed no significant correlation (uncorrected p > 0.4 in both attended and unattended conditions).

(a) BOLD responses in the visual cortex. *, **, or *** denote p < 0.05, p < 0.01, or p < 0.001 after FDR correction, and + for p < 0.05 (uncorrected). (b) Collision-sensitive activations in the high order visual cortex and the frontoparietal areas. The color bar indicates the percent signal change of Hit-Miss. Maps were thresholded at uncorrected p < 0.01. No significant clusters can be found after FWE correction. Data underlying (a) can be found at https://osf.io/gdjwh/ . FEF, frontal eye field; FWE, family-wise error; IFG, inferior frontal gyrus; INS, insular; LH, left hemisphere; MTG, middle temporal gyrus; RH, right hemisphere; TO, temporal occipital; TPJ, temporal parietal junction.

We further investigated whether collision sensitivity can also be found in the cortical brain regions. ROI-averaged retinotopic responses in a visual cortical area were determined by a leave-one-run-out (LORO) cross-validation approach at the individual level (see Methods for details). As shown by Fig 4A , significant collision sensitivity can be found in the early visual cortex in both attended and unattended conditions (attended condition: false discovery rate (FDR) p = 0.030 in V3, p = 0.030 in V3b, and p = 0.030 in V4 after BH-FDR correction; uncorrected p = 0.027 in V1; unattended condition: uncorrected p = 0.044 in V4, uncorrected p = 0.031 in TO2). Fig 4B shows the group-averaged collision-sensitive (Hit-Miss) activations on the cortical surface. A few clusters with collision sensitivity (p < 0.01, uncorrected) can be found from bilateral middle temporal gyrus (MTG) in the attended condition and from bilateral temporal parietal junctions (TPJ), and the frontal eye field (FEF), inferior frontal gyrus (IFG), and insular (INS) in the right hemisphere in the unattended condition. However, no significant cluster can be found after family-wise error correction. These results revealed collision-sensitive responses in the visual cortex and possibly in the frontoparietal areas of attention network.

The VTA receives direct input from the SC [ 17 , 31 , 32 ] and projects to many brain areas including the amygdala [ 17 ] and frontal lobe [ 33 ]. In the attended condition ( Fig 3C , left), the VTA showed significant activations to looming stimuli but without collision sensitivity. In the unattended condition, there was a significant collision-sensitive cluster (cluster p = 0.004). The averaged responses of the collision-sensitive cluster were plotted in the right panel of Fig 3D . For the ROI-averaged response of the whole VTA ( Fig 3D , left panel), we found a marginally significant effect of collision sensitivity in the unattended condition (t(19) = 1.970, Cohen’s d = 0.441, p = 0.064). LOSO analysis further revealed a significant collision sensitivity in the unattended condition (t(19) = 3.234, Cohen’s d = 0.723, p = 0.004).

The ventromedial pulvinar (vmPul) receives strong inputs from the SC and possibly sparse input from the retina and reciprocally connects with visual areas in the dorsal visual stream [ 26 – 28 ]. The dorsal part of vmPul may also connect with the amygdala and frontoparietal cortex [ 29 , 30 ]. In the group-averaged activation maps in the unattended condition ( Fig 3A ), there was a small cluster of collision sensitivity ( Fig 3A , bottom right). The ROI-averaged responses of vmPul ( Fig 3B , right panel) showed a significant interaction between trajectory and attention (F(1,19) = 7.06, p = 0.016, = 0.271). Post hoc t tests revealed a significant collision sensitivity in the unattended condition (t(19) = 3.613, p = 0.002, Cohen’s d = 0.808), but not in the attended condition (p = 0.321). LOSO cross-participants validation did not find significant effect of collision sensitivity, likely due to a relatively large interindividual variability of collision-sensitive response in the vmPul. No significant collision-sensitive response was found in other subnuclei of pulvinar.

As shown by Fig 3A , looming stimuli activated both lateral and medial portions of the ventral pulvinar. The ventrolateral pulvinar (vlPul) mainly connects with the early visual cortex but also with the parietal cortex [ 24 ] and may also receive sparse input from the SC [ 25 , 26 ]. As shown by Fig 3A , a collision-sensitive cluster was found in the vlPul in the attended condition (cluster p = 0.091). In Fig 3B (left panel), the ROI-averaged responses of vlPul to stimuli presented to the contralateral visual field showed significant interaction of attention and trajectory (F(1,19) = 4.62, p = 0.045, = 0.196). Post hoc paired t tests revealed a significant collision sensitivity in the attended condition (t(19) = 2.750, p = 0.013, Cohen’s d = 0.615), but not in the unattended condition (p = 0.347). The LOSO cross-validation analysis revealed a significant collision sensitivity in the vlPul in the attended condition (t(19) = 2.646, p = 0.016, Cohen’s d = 0.592).

(a, c) Activation maps to looming stimuli in coronal views of the pulvinar (a) and VTA (c) . Maps were thresholded at p < 0.05 uncorrected. Dotted lines mark the anatomical boundaries of the vlPul, vmPul, and VTA. Activation maps of the whole pulvinar were smoothed with a 2.8-mm FWHM Gaussian filter for display purpose (1.4 mm FWHM for other nuclei). (b) ROI-averaged responses of the vlPul and vmPul. (d) Left panel shows the ROI-averaged responses of the whole VTA. Right panel shows the averaged responses of the collision-sensitive cluster in (c) . * or *** above a long line indicate ANOVA p < 0.05 or p < 0.001. *, **, or *** above a short line for t test p < 0.05 or p < 0.001. Other conventions as in Fig 2 . Data underlying (b, d) can be found at https://osf.io/gdjwh/ . FWHM, full-width half-maximum; vlPul, ventrolateral pulvinar; vmPul, ventromedial pulvinar; VTA, ventral tegmental area.

We further investigated collision sensitivity in other subcortical regions, including the pulvinar, VTA, PBGN, amygdala, and LC, which form looming sensitive circuits with the SC as demonstrated by previous rodent studies [ 4 , 5 , 16 – 18 ]. The lateral geniculate nucleus (LGN) was also included as a potential control area. S8 Fig shows the anatomical masks for these subcortical regions in MNI space. For each ROI, we performed similar analyses as for the SC to check its activation map and ROI-averaged responses. Collision-sensitive responses were found in the pulvinar and VTA ( Fig 3 ), but not in other subcortical nuclei ( S9 Fig ).

To evaluate whether looming stimuli lead to any measurable head movement, we investigated head movements after the onset of the looming stimuli using the motion parameters estimated with EPI volumes. Results showed negligible head motion to looming stimuli (less than 0.01 mm on average). We further confirmed that the observed collision sensitivity in the SC was unlikely influenced by sparsely and randomly presented fixation changes or due to retinotopic difference in luminance changes and optical flows between hit and near-miss trajectories ( S2 and S3 Figs, S1 Text ).

We further divided the SC into a rostral (foveal) part and a caudal (extrafoveal) part (enclosed by blue dotted lines in Fig 2B ) based on the significant contralateral-ipsilateral activations (green voxels in the upper middle panel of Fig 2B ; see also S7 Fig ). In the foveal SC ( Fig 2E ), there was a significant interaction between trajectory and attention (F(1,19) = 8.06, p = 0.010, = 0.298). Post hoc t tests showed significant collision sensitivity in the foveal SC in the unattended condition (t(19) = 3.015, p = 0.007, Cohen’s d = 0.674), but not in the attended condition (p = 0.600). To further validate the collision sensitivity found in the foveal SC in the unattended condition, we performed a leave-one-subject-out (LOSO) cross-validation analysis. For each participant, the ROI to calculate the collision-sensitive response was defined by voxels with significant group-level collision sensitivity from the remaining participants. The LOSO results revealed a significant collision sensitivity in the foveal SC in the unattended condition (t(19) = 2.761, p = 0.012, Cohen’s d = 0.617). In the extrafoveal SC ( Fig 2F ), there was a significant main effect of trajectory (F(1,19) = 6.75, p = 0.018, = 0.262), but no interaction between attention and trajectory (p = 0.943). This finding suggests collision sensitivity in the extrafoveal SC independent with the attentional state of observers.

The ROI analysis focused on the responses to stimuli from the contralateral visual field. We first investigated the responses of the whole SC ( Fig 2C ). Two-way repeated measures (rm) ANOVA revealed a significant main effect of trajectory (Hit/Miss, F(1,19) = 8.67, p = 0.008, = 0.313) and attention (Attended/Unattended, F(1,19) = 49.27, p < 0.001, = 0.722). Post hoc paired t tests showed a significant collision sensitivity in the unattended condition (t(19) = 2.363, p = 0.029, Cohen’s d = 0.528), but not in the attended condition (p = 0.247). We then investigated SC responses to looming stimuli in the upper and lower visual fields in the unattended condition ( Fig 2D ). Although collision sensitivity in the upper and lower visual fields showed no significant difference (trajectory by visual field interaction: p = 0.124), the SC responses to looming stimuli on a near-miss trajectory were significantly stronger in the upper visual field compared to those in the lower visual field (t(19) = 2.798, p = 0.011, Cohen’s d = 0.626), suggesting higher looming sensitivity in the SC to approaching objects from the upper visual field. This finding is consistent with the behavioral results of slightly better looming-trajectory discrimination and stronger pupillary reflex in the upper visual field (Figs 1 , S5 , and S6 ).

Previous studies suggest that the SC, a phylogenetically old midbrain nucleus, might play important roles for collision detection in the human brain [ 8 ]. Thus, in the current study, we carefully investigated the response of the SC to looming stimuli with direct-hit or near-miss trajectories. From the group-averaged activation maps ( Fig 2C ), bilateral SCs showed robust responses to the looming stimuli in both attended and unattended conditions. This confirmed that the human SC was indeed highly sensitive to looming stimuli. We further investigated the collision sensitivity in each voxel defined as the response difference between the direct-hit and near-miss trajectories ( Fig 2C , “Hit-Miss”). In the attended condition, no significant cluster of voxels with collision sensitivity was found in the SC. However, significant or marginally significant clusters can be found in the unattended condition (permutation test with small volume correction: cluster p = 0.053 for the contralateral SC, and p < 0.001 for the ipsilateral SC. The cluster defining threshold is voxel p = 0.05). These collision-sensitive clusters located more rostral (or anterior) ( Fig 2B , foveal SC, “Hit-Miss,” unattended), corresponding to the foveal part of the SC [ 21 , 22 ]. More caudal (or posterior) part of the SC showed strong responses to both looming stimuli presented in the contralateral visual field ( Fig 2C , extrafoveal SC, “Hit” and “Miss”), but no significant collision-sensitive clusters. Based on the normalized depth map of the SC ( S5 Fig in our previous study [ 23 ]), we divided the SC into 3 compartments with equal thickness, roughly correspond to the superficial (depth from 0 to 1/3), intermediate (1/3 to 2/3), and deep layers (2/3 to 1). The center-of-mass location of the contralateral (depth = 0.2) and ipsilateral (depth = 0.36) clusters locate in the superficial and intermediate layers, respectively.

(a) Looming objects with hit and near-miss trajectories were presented in 2D images with an MRI-compatible projector for 330 ms, with 7.67~11.67 s of intertrial intervals. The stimulus was presented in one of the 4 quadrants of the VF in each trial. The right panel shows the size and position of the stimulus at the last frame in each quadrant. (b) SC activation maps. Red and green squares on the coronal and sagittal slices indicate the locations of the SC. From the zoomed-in sagittal view, blue dashed lines outline the ROIs of the rostral/anterior and caudal/posterior SCs representing the foveal and extrafoveal visual fields, respectively. The ROIs was determined by the retinotopic activations (contra-ipsi; S7 Fig ) shown here as the green overlay. Activation maps were thresholded at p < 0.05 (uncorrected). Color bars indicate percent signal change. Left and right SCs were mapped with responses to the contralateral and ipsilateral stimuli, respectively. Activation maps to stimuli from the left visual field were horizontally flipped and averaged with those to stimuli from the right VF. C, I, D, and V in the compass abbreviate contralateral, ipsilateral, dorsal, and ventral, respectively. Dotted lines indicate an approximate boundary between the superficial and deeper layers of the SC. (c and d) ROI-averaged BOLD responses of the whole SC to looming stimuli from the contralateral visual field. (e and f) The responses in the foveal ( c ) and extrafoveal ( f ) parts of the SC. Red/blue horizontal bars, black vertical bars, and red/blue circles denote mean, SE, and individual data, respectively. * above a long black line indicates ANOVA p < 0.05 for the interaction between attention and trajectory. * and ** above a short black line indicate paired t test p < 0.05 and 0.01, respectively. * In between the red and blue lines in (f) indicates p < 0.05 for the main effect of trajectory. The original fMRI resolution is 1.5-mm isotropic; here, the functional maps were up-sampled at 0.6-mm isotropic to match the resolution of MNI template (see Methods for details). Data underlying (c, d, e, f) can be found at https://osf.io/gdjwh/ . SC, superior colliculus; SE, standard error; VF, visual field.

To investigate potential subcortical pathways for the detection of collision trajectory in the human brain, we performed a 7T fMRI experiment (1.5-mm isotropic resolution) in a group of healthy human participants in experiment 2 (n = 20). In a long event-related design ( Fig 2A ), participants were asked to discriminate whether an incoming object would hit or miss their head in an attended condition or to detect occasionally and randomly presented fixation-color changes in an unattended condition. Participants’ performance was 97.2% ± 3.6% (mean ± STD) in the attended condition and 95.6% ± 4.7% in the unattended condition, indicating that they followed the task instructions very well.

The discrimination sensitivity, calculated based on the slope of the psychometric curve, was slightly but significantly higher for objects approaching from the upper visual field than from the lower visual field (permutation test p < 0.001; Fig 1C ), indicating that observers could more precisely discriminate hit from near-miss trajectories for objects coming from the upper visual field. Pupil size was smaller around 1,100 ms after stimulus onset in the hit than miss conditions ( Fig 1D–1F ), consistent with previous findings [ 13 ]. Similar results were observed when testing with a much brighter background or in an irrelevant task condition ( S5 and S6 Figs). Fixational eye movements showed no significant difference between the hit and miss events ( S4B Fig ). These behavioral results show that observers and their pupillary responses can precisely discriminate collision from near-miss trajectories of an incoming object.

(a) Visual stimuli depicting an incoming ball from one of the 4 quadrants of the VF were presented with a 3D LCD monitor in the behavioral experiment. The trajectory of the looming object varied slightly to either hit (hit nasion, hit eye) or miss (near miss, far miss) the head of observers. (b) The percentage of hit responses at different impact points was fitted with a normal CDF. Black circles indicate the extrapolated impact points of different trajectories. (c) The discrimination sensitivity, calculated as in which σ is the standard deviation of the fitted normal CDF, is significantly higher in the upper VF than in the lower VF. Error bars indicate SE. **p < 0.01. (d) The change of pupil size with respect to baseline (−200~0 ms) was normalized by standard deviations and then plotted for hit (purple) and miss (dark gray) trajectories, respectively (data from the upper and lower VFs were combined here). Dark gray bars indicate the time points when the pupil size was significantly smaller (permutation test p < 0.05 using cluster-size based adjustment) in the hit condition. The vertical gray bar indicates the time interval of visual stimulus presentation. (e, f) The averaged change of pupil size from 1,000 to1,364 ms after stimulus onset was plotted for different looming stimuli from the upper (e) and lower (f) VFs. *Bonferroni corrected p < 0.05, + uncorrected p < 0.05. Shaded areas in (b, d) and vertical lines in (c, e, f) indicate SEM. Data underlying (c, e, f) can be found at https://osf.io/gdjwh/ . CDF, cumulative distribution function; SEM, standard error of the mean; VF, visual field.

In a behavioral experiment (experiment 1, n = 15), we measured the ability of healthy human observers to discriminate hit from near-miss trajectories. Participants responded whether an incoming object would hit or miss their head (Figs 1A and S1A ). Pupil size and eye movements were also recorded. In Fig 1B , the percentage of hit responses was plotted as a function of the extrapolated impact point of the approaching objects. The psychometric functions were further fitted with cumulative normal distributions. The steep slope of the psychometric function around the edge of the head (about 6 cm from nasion) indicates that participants could precisely discriminate the incoming objects on a collision course from those with near-miss trajectories, consistent with the prediction from previous studies [ 13 , 14 ]. S4A Fig shows the results for all individuals.

Discussion

Detecting imminent collision is crucial for our survival. In experiment 1, we found that human observers can precisely discriminate whether an approaching object was on a collision course or a near-miss trajectory with their head. Collision events also induced significant changes in pupillary reflex. In experiment 2, high-resolution 7T fMRI revealed collision-sensitive responses in several subcortical nuclei, including the SC, ventral pulvinar and VTA. Correlation and path analyses further demonstrated collision sensitivities in the SC-vmPul, and SC-VTA pathways without attention and cortical influence. In experiment 3, for hemianopic patients with unilateral lesions of the geniculostriate pathway, the ipsilesional SC showed collision sensitivity to stimuli presented to their BVF. Finally, stronger response in the SC was associated with better detection performance of the collision event. These findings clearly demonstrate a critical role of the human tectofugal pathways in automatic detection of collision trajectories without attention and awareness, supporting “blindsight” to threating visual information.

In the optic tectum and the downstream nucleus rotundus (homologues of the SC and pulvinar in mammals) of pigeons, different types of neurons encode several optical variables of a looming stimulus, including the time-to-collision, absolute rate of expansion, and object size [3]. In the mouse SC, Shang and colleagues identified parvalbumin-positive (PV+) excitatory projection neurons in the superficial layers encoding the optical parameters of a looming stimulus in their receptive fields [4,16]. The response onset depended on the stimulus size and moving velocity, and the response peaked at the time of collision. Here, we show that the human SC was highly sensitive to slight trajectory differences of looming stimuli with similar times to contact, rates of expansion, and object sizes, suggesting that the primate SC contains neurons sharply tuned to looming trajectories for accurate collision detection. Interestingly, although the looming stimuli were presented in the parafovea, the strongest collision-sensitive response was observed in the anterior part of the SCs corresponding to the central visual field (Fig 2C). One possibility is that SC neurons are sensitive to the would-be point of collision, making predictions about the impact point in the immediate future. Consistent with this explanation, predictive remapping activity has been observed for SC neurons before saccades [37]. Collision-sensitive activations in the foveal SCs cannot be explained by retinotopic difference in optical flow or luminance, because the hit stimulus shows no overall retinotopic bias in the central visual field and the collision-sensitive clusters is clearly located outside the retinotopic region (S2B Fig). It was also unlikely due to fixational eye movements, because observers can maintain stable fixations that showed no significant difference in fixational eye movements between the hit and miss conditions (S4B). The lack of collision sensitivity in the attended condition was likely due to the saturation of SC responses.

Our results consistently show that the behavioral performance and pupillary reflex can better discriminate looming trajectories from the upper visual field (Figs 1, S5, and S6). The SC response was also stronger to looming (near-miss) stimuli in the upper visual field (Fig 2D). This upper visual field advantage of looming sensitivity suggests that the phylogenetically conserved tectal pathways are also ecologically adaptive, as threatening looming objects such as diving predators or falling stones appear more often in the upper than in the lower visual field due to gravity. A recent study showed that the primate SC, like that of rodents, also overrepresents the upper visual field, with sharper, stronger, and faster visual representations than in the lower visual field [22]. Together, these findings support Previc’s ecological perspective of the primate visual system, suggesting that the primate SC may be optimized and specialized for detection of transient and biological salient information from extrapersonal space in the upper visual field [38]. This upper visual field advantage forms an interesting contrast with previous findings that sustained visual attention resolves and tracks objects better in the lower visual field [39].

The results of experiment 2 clearly show that compared with the looming-evoked responses during a central fixation task, paying attention to and judging the trajectory of looming stimuli strongly enhanced the responses in both superficial and deeper layers of the SC (Fig 2B). However, in experiment 3, the SC’s responses to collision trajectories were comparable when the stimulus was presented to the NVF or to the BVF of hemianopic patients (Fig 6A). Thus, looming-sensitive responses in the SC were strongly modulated by top-down attention but may not be sensitive to the awareness of visual stimuli. These findings demonstrate the critical role of the SC in both top-down attention and subconscious processing of threatening visual information.

Our data clearly demonstrate that the SC-vmPul pathway in the human brain automatically detects collision trajectories without attention to the looming stimuli (Fig 5). Results from the hemianopic patients further suggest collision sensitivity in the ipsilesional SC and vmPul in the absence of visual awareness (Fig 6). These findings are consistent with recent rodent studies that the SC-LP (or pulvinar) pathway processes looming information in an anesthetized state [16] and triggers defensive freezing behavior [5,16]. In Wei and colleagues’ study, optogenetic mapping and electrophysiology also revealed a disynaptic circuit from the SC through LP to the lateral amygdala, which directly mediated the innate fear-related defensive response. In our study, although looming objects on a collision course changed pupillary reflex, participants did not report fear to these stimuli, and no significant response was found in the amygdala (S10 Fig). One possible explanation is that human observers quickly adapted to repetitive presentations of the virtual looming objects, which may not be effective to trigger fear response or evoke strong activation in the amygdala. In contrast, collision-sensitive responses can be reliably observed in the SC and vmPul. We speculate that the role of the SC-vmPul pathway might be automatically processing collision-related visual information, such as the looming trajectory or time-to-collision, which could be used by the downstream brain areas (e.g., the dorsal stream) to guide quick and subconscious actions to avoid the impending threats [40].

Collision-sensitive activity was also found in the VTA without attention (Fig 3C and 3D) and awareness (Fig 6C and 6E), showing correlation with and effective connectivity from the SC (Fig 5B and 5E). These findings are consistent with the rodent studies that VTA neurons responded in short latency to biologically salient or aversive stimuli [31] and that VTA neurons receiving direct input from the SC mediated defensive flight behaviors to large overhead looming stimuli [17]. Populated mainly with dopaminergic (DA) neurons, the VTA plays important roles in reward, motivation, and attention [41]. The role of SC-VTA pathway might be modulating the level of arousal [42] or changing the attentional state [43] of the observer to mediate rapid defensive responses [44]. Although rodent studies also suggest the PBGN and LC in processing looming information and mediating looming-evoked defensive behaviors [4,18], we did not find significant collision-sensitive response from these areas in the human brain. Given their small sizes and deep locations, the negative finding from these small subcortical nuclei could be due to the low SNR and partial volume effect of BOLD signals.

The tectopulvinar pathway was originally proposed to support “blindsight” [19]. A critical role of the SC in visuomotor functions has been repeatedly confirmed by studies of visually guided saccades in V1-lesioned monkeys [45,46]. FMRI studies of blindsight patients also indicated visually evoked activation from the SC [47,48]. However, given the small sample sizes and the low-resolution fMRI approaches in these studies, the involvement of the SC in human blindsight still lacks conclusive evidence. Using high-resolution fMRI, here we showed clear evidence from 12 hemianopic patients that visually evoked response in the ipsilesional SC was highly sensitive to looming objects on a collision course (Fig 6A and 6E), which also predicted the above-chance detection performance of impending collisions (Fig 6G and 6H). These findings provide strong evidence for the critical role of the SC in detecting impending visual threats in human blindsight.

Compared with the SC, the role of pulvinar in blindsight is more controversial. Human fMRI studies suggest that the tectopulvinar-amygdala pathway may underlie blindsight of emotionally and socially salient face stimuli [49,50]. A recent study in V1-lesioned monkeys also revealed a critical role of the tectopulvinar pathway in visually guided saccades [51]. However, there is accumulating evidence that the geniculo-extrastriate pathway plays a major role in blindsight in both Monkey [36] and Human [52]. These discrepancies might depend on the scope of lesion, time of lesion, or distinct roles of the pulvinar and LGN pathways in blindsight [53,54]. Our results may suggest collision sensitivity in the ipsilesional ventromedial pulvinar to looming stimuli presented to the BVF (Fig 6B). Collision-sensitive response was also found in the VTA (Fig 6C and 6E). While visually evoked responses can be found in the LGNs of both healthy participants (S10 Fig) and hemianopic patients (S12 Fig), there was no significant activation of collision sensitivity. These findings support distinct roles of the subcortical pathways in blindsight: The tectofugal pathways are specialized to detect impending visual threats, while the geniculo-extrastriate pathway processes basic visual information (e.g., contrast and motion).

Human fMRI studies reported cortical representations of the TTC and possibly trajectories of approaching objects [55]. While the cortical responses may indicate a complex and relatively slow neural calculation for 3D motion perception, the subcortical pathway found in our study suggests an efficient and rapid computational strategy for collision trajectory detection. In support of this, studies have demonstrated that simpler equations for calculating looming trajectories could better predict human performance, especially the systematic errors made by observers. For example, Duke and Rushton [56] suggested that the perceived trajectory is based on the ratio of lateral angular speed to the sum of looming and changing disparity signals. Our findings revealed the neural substrate for an efficient and rapid neural computation of collision trajectory in the human brain, which may inspire the optimization of computer vision algorithms for collision detection [57].

[END]
---
[1] Url: https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3002375

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/

You are viewing proxied material from magical.fish. The copyright of proxied material belongs to its original authors. Any comments or complaints in relation to proxied material should be directed to the original authors of the content concerned. Please see the disclaimer for more details.