Three ways round a missing cortex

About nine in ten fibres leaving the eye go to the lateral geniculate nucleus, or LGN, in the thalamus, and from there almost all of its output goes to V1. That is the main road of human vision. When V1 is destroyed, the main road ends in a wall. But the eye also sends fibres elsewhere, and the LGN itself sends a small number of fibres past V1 to higher visual areas. Any visual signal that guides Meera’s hand must use one of these minor roads.

Three candidates have been studied most, and the first runs from the retina to the superior colliculus, a layered structure on the roof of the midbrain that orients the eyes and head. From the colliculus, signals pass to the pulvinar, a large thalamic nucleus, and on to cortical areas such as the motion area MT and the parietal cortex. The second runs from the LGN directly to extrastriate areas, especially MT, which is called hMT+ in humans. The third links the damaged hemisphere to the intact one, so that the healthy side’s visual areas help out.

What happens when each route is switched off

Correlation cannot settle which route matters, because all three may be active at once. The decisive experiments switch a route off and see what fails. These have been done in macaque monkeys with surgical V1 lesions, which recover the ability to make eye movements to blind-field targets over several weeks. In a 2021 review, Tadashi Isa and Masatoshi Yoshida summarise that recovery and what their laboratory found about its pathways.

The first decisive result implicated the LGN. In a 2010 Nature study, Michael Schmid, David Leopold and colleagues scanned monkeys with chronic V1 lesions. High-contrast stimuli in the blind field produced fMRI activation in V2, V3, V4, MT, FST and the lateral intraparietal area, and the animals correctly located the stimuli. Then the team temporarily inactivated the LGN on the lesioned side. Their abstract reports that “fMRI responses and behavioural detection were abolished”. The title of the paper is its conclusion: blindsight depends on the lateral geniculate nucleus.

The second decisive result implicated the colliculus. In a 2011 study, Rikako Kato, Isa and colleagues injected muscimol, a drug that silences neurons, into the superior colliculus of V1-lesioned monkeys. After injection on the lesioned side, the monkeys failed to make visually guided eye movements to targets in the part of the field represented at the injection site. Spontaneous eye movements towards that region were not abolished. The deficit was therefore in vision, not in moving the eyes, and the authors concluded that the retino-tectal pathway “plays an essential role in residual vision after V1 lesion”.

The third result traced the colliculus pathway one step further. In a 2019 Nature Communications paper, Masaharu Kinoshita, Isa and colleagues silenced the ventrolateral pulvinar and found impaired visually guided eye movements to the blind field. They then used a two-virus method to block only the fibres running from the colliculus to that part of the pulvinar. That selective block also impaired the eye movements. Their title says the circuit reveals “the critical role of pulvinar and superior colliculus”.

Both routes, and why they can both be necessary

These results look contradictory. If silencing the LGN abolishes blind-field detection, how can silencing the colliculus pathway also abolish blind-field eye movements? Kinoshita and colleagues address this directly. They suggest that the difference may depend on how much the LGN has degenerated after the V1 lesion, since many LGN neurons die once their target is gone, and on the task used. A route can be necessary in one animal and one task and less important in another.

There is also a simpler reading, in which two routes can each be necessary because they do different jobs in the same chain. The LGN route may carry the visual signal to cortex, while the colliculus may be needed to turn that signal into an eye movement, or the reverse. An experiment that silences one part of a chain stops the whole chain, whichever link carries the visual information. Necessity shows that a route is in the chain, not that it is where the seeing-without-seeing happens.

Patients, tracts and the human evidence

Monkeys are not patients, and the human evidence comes from imaging, not inactivation. The most direct study is by Sara Ajina, Franco Pestilli, Ariel Rokem, Christopher Kennard and Holly Bridge, published in eLife in 2015. They studied 17 patients with V1 damage acquired in adulthood and 9 controls, dividing the patients by psychophysics into blindsight-positive and blindsight-negative groups. They then traced white-matter pathways with diffusion imaging.

The result was clean, and in their words “all blindsight positive individuals showed intact geniculo-hMT+ pathways”. In blindsight-negative patients, by contrast, “this pathway was significantly impaired or not measurable”. Pathways from the colliculus to hMT+, and between hMT+ in the two hemispheres, “were not consistently present in blindsight positive cases”. In this sample, the LGN route marked who had blindsight, and the colliculus route did not.

A companion study by the same group in Brain in 2015 scanned seven patients viewing moving dots. Without V1, the response of hMT+ to motion coherence resembled that of early visual cortex, as if hMT+ had become the first cortical stop for vision in the damaged hemisphere. That fits Riddoch’s soldiers, who reported movement in their blind fields.

The colliculus is not out of the human picture. In a 2006 Brain study, Sandra Leh, Heidi Johansen-Berg and Alain Ptito tested four people who had had a whole cerebral hemisphere removed. Two showed type I blindsight on an indirect test, and those two had reconstructed fibre tracts from the colliculus that the other two lacked. With no cortex at all on one side, the LGN-to-cortex route is gone, so whatever remains must run through the colliculus.

A test the colliculus fails

One clever test uses colour. The superior colliculus receives little or no input from S-cones, the short-wavelength cones that respond to violet and blue. A stimulus defined only by S-cone contrast should therefore be invisible to the colliculus. In a 2010 study, Marco Tamietto, Beatrice de Gelder and colleagues used this with a patient who had lost V1 on one side.

A grey stimulus in his blind field speeded his responses to a stimulus he could see, an effect called implicit bilateral summation, and activated the colliculus and occipito-temporal areas. When the unseen stimulus was purple, drawing mainly on S-cones, “any evidence of implicit visuomotor integration disappeared”. Activation in the colliculus also dropped. For that patient and that task, the colliculus was the route from unseen light to a motor response.

The S-cone test matters for Dr Rao because it can be run in a clinic. If Meera’s blind-field responses survive grey stimuli and vanish with purple ones, the colliculus route is carrying them. If they survive both, the LGN route or another one is involved. It is not a perfect dissection, because S-cone signals travel through the LGN’s koniocellular layers, and in monkeys Lawrence Sincich and colleagues traced a direct projection from those layers to MT that skips V1. Still, it splits the hypotheses at almost no cost.

The emerging picture is that the question’s premise is half right. The colliculus and pulvinar route is necessary for some blindsight behaviours, especially fast orienting, and the geniculate route is necessary for others, especially detection and motion. Which route matters in a given patient depends on what survived the injury, what the task asks for, and probably how much the brain has rewired since. That still leaves the main puzzle untouched, because whatever the route, the signal reaches the hand and not experience, and nothing so far says why.

Three models, one patient’s scores

Parts three and four left three accounts standing, each able to explain why Meera might guess well and deny seeing. Words alone make all three sound equally good, so I built a small simulation to see where they come apart. The instrument is a toy, and it shows only what follows from its assumptions, and its value is in making those assumptions explicit and finding the tests that separate them.

What the model contains

Each simulated trial presents one stimulus to one simulated patient. Model A is the two-channel account. An action channel carries a strong signal, with d′ of 1.5, and feeds forced choice and reaching. Seventy per cent of its input comes from the colliculus route and thirty per cent from the geniculate route. After the stimulus ends, its trace decays with a time constant of half a second. A separate awareness channel carries a weak signal, with d′ of 0.3, and feeds yes-no detection and confidence.

Model B is Phillips’s degraded-vision account. A single conscious channel with d′ of 1.0 feeds every response, and the patient sets a cautious yes-no criterion of 1.8. Because the signal is conscious, it is held in memory, which adds noise slowly over time. Model C is Ko and Lau’s metacognitive account. The single channel is the same as in B, but the criterion jitters from trial to trial, with a standard deviation of 1.2, and confidence is read from the signal through extra noise.

Each model is run on four measures, with 4,000 trials per measure. The first is forced-choice d′ for left versus right, and the second is yes-no d′ for present versus absent. The third is type-2 sensitivity, the area under the curve relating confidence to correctness, where 0.5 means confidence says nothing about accuracy. The fourth is reaching d′, measured both immediately and after a two-second delay. A fifth measure repeats forced choice with S-cone stimuli, which the colliculus share of the signal cannot see.

# model A: two channels
a = mu_a * ((0 if scone else f_sc) + (1 - f_sc))   # action strength
choice = argmax(noise + a * target)                 # forced choice
conf   = |weak awareness evidence, mu_w = 0.3|      # confidence
yes    = noise + mu_w * present > crit              # yes-no
reach(d) uses a * exp(-d / tau) and, after a delay, the weak trace

# models B and C: one channel, mu = 1.0
yes  = noise + mu * present > crit + normal(0, jitter)
conf = |evidence| + normal(0, meta)
reach(d) adds memory noise 0.35 per second of delay

The output

I ran five conditions. The first three are models A, B and C. The fourth is a negative control: the sighted half of the field with a near-threshold stimulus and an ordinary criterion. The fifth is a counterexample, described below, and each condition used its own random seed, starting at 39.

seed 39, 4000 trials per measure; d-prime values; type-2 = AUROC of confidence for correctness
condition                                   dFC    dYN  ratio  type2  reach0  reach2  S-cone
two channels (A)                           1.56   0.36   4.32  0.486    1.57    0.33    0.49
degraded conscious (B)                     1.02   1.02   1.00  0.722    1.00    0.82    1.01
metacognitive failure (C)                  1.00   0.70   1.43  0.627    1.02    0.83    0.98
sighted field, near threshold (control)    1.03   0.98   1.05  0.709    0.95    0.80    0.99
C with extreme jitter (counterexample)     1.50   0.34   4.38  0.535    1.55    1.19    1.50

The three models leave three different fingerprints. Model A shows a large gap between forced choice and yes-no, a ratio of 4.3, and type-2 sensitivity at chance, 0.49. Its reaching collapses after the delay, from 1.57 to 0.33, and its S-cone performance falls by two-thirds. Model B shows no gap at all, a ratio of 1.0, with confidence that tracks accuracy and reaching that survives the delay. Its S-cone performance is untouched, and model C sits in between, with a modest gap of 1.4 and confidence that tracks accuracy somewhat.

The first real test, Azzopardi and Cowey’s bias-free comparison, already rules against model B for their patient, because B predicts equal d′ in both tasks. Pure caution cannot produce a sensitivity gap, because d′ corrects for caution. Phillips’s account survives only if the conscious signal itself differs between tasks, for instance if forced choice draws on cues that the yes-no task does not. That is a coherent position, but it is no longer the simple one.

A counterexample that copies model A

The fifth condition is the one that keeps the instrument honest. I took model C and pushed it to extremes: criterion jitter of 4.5, confidence noise of 6 and a stronger signal of 1.5. On the standard tests it now looks just like model A. Its ratio is 4.38 against A’s 4.32, and its type-2 sensitivity is 0.54 against A’s 0.49, close to chance in both.

Where the argument fails

The instrument also finds the boundary of my own claim. A sensitivity gap is the classic evidence for a separate action channel. The question is how large the gap must be before a jittering criterion can no longer explain it. I swept the criterion jitter in model C from zero to four, keeping everything else fixed.

sweep (seed 390): criterion jitter in model C (mu 1.0, crit 1.8, meta 1.5)
  jitter 0.0  dFC 1.05  dYN 0.95  ratio 1.11  type2 0.619
  jitter 0.5  dFC 1.05  dYN 0.88  ratio 1.20  type2 0.619
  jitter 1.0  dFC 1.05  dYN 0.72  ratio 1.47  type2 0.619
  jitter 1.5  dFC 1.05  dYN 0.59  ratio 1.78  type2 0.619
  jitter 2.5  dFC 1.05  dYN 0.39  ratio 2.68  type2 0.619
  jitter 4.0  dFC 1.05  dYN 0.25  ratio 4.19  type2 0.619

The result is a warning, because a jittering criterion can produce a ratio of 2.7 when its standard deviation is 2.5, and 4.2 when it reaches 4. Those jitters are large, more than twice the signal itself, but nothing in the yes-no data rules them out. A sensitivity gap alone does not prove two channels, because a sufficiently unstable criterion can manufacture any gap. Ko and Lau’s point survives in the model, and the burden falls on tests that a criterion cannot touch.

The delay sweep shows what one such test looks like. I varied the delay before reaching from zero to three seconds in models A and B.

sweep (seed 391): delay before reach onset, model A vs model B
  delay 0.00 s  reach d' A 1.43  B 0.96
  delay 0.25 s  reach d' A 0.89  B 0.96
  delay 0.50 s  reach d' A 0.58  B 0.94
  delay 1.00 s  reach d' A 0.31  B 0.92
  delay 2.00 s  reach d' A 0.25  B 0.77
  delay 3.00 s  reach d' A 0.25  B 0.66

The curves cross between a quarter and half a second. Before that, model A’s strong action signal beats B’s weaker conscious one, and after it, B’s memory wins. Model A levels off at about 0.25 because the weak awareness trace remains. A real patient tested at only two seconds, as in Striemer’s study, shows only the right-hand end of this curve. Testing at several delays would show its shape, and the shape is what distinguishes a decaying trace from a noisy memory.

The last sweep varies the colliculus share of model A’s action signal, which sets how much the S-cone test can reveal.

sweep (seed 393): share of action input from the colliculus, model A, S-cone forced choice
  SC share 0.0  dFC 1.48  S-cone dFC 1.52
  SC share 0.3  dFC 1.48  S-cone dFC 1.05
  SC share 0.5  dFC 1.48  S-cone dFC 0.77
  SC share 0.7  dFC 1.48  S-cone dFC 0.47
  SC share 0.9  dFC 1.48  S-cone dFC 0.16
  SC share 1.0  dFC 1.48  S-cone dFC 0.01

S-cone performance falls in step with the colliculus share. That makes the colour test a rough dial for the route, not a switch. A patient whose S-cone performance drops by a third would have roughly a third of her action signal arriving through the colliculus, if the model’s assumptions hold. Tamietto’s patient, whose effect vanished entirely, would sit near the right-hand end.

For Meera, the instrument turns Dr Rao’s list of tests into a decision rule. A large forced-choice to yes-no ratio is necessary for calling her case blindsight but not sufficient. If her reaching collapses with a delay and her responses weaken with purple stimuli, the two-channel account fits her best. If neither happens, she may be a model C patient, or a model B patient with faint vision, and a trial that trains “residual vision” might suit her for different reasons.