Visual Cortex Prosthesis
A cortical visual prosthesis is the write-in side of BCI — letting the blind "see" via cortical stimulation. It sits at the three-way intersection of medicine, neuroscience, and AI. Fernández 2021 Science Advances is the modern milestone for this direction.
1. Visual Cortex Organization (Recap)
- V1: retinotopy — each location on the cortex corresponds to a region in the visual field
- Central vision (fovea) occupies a disproportionate share of V1
- Stimulating V1 evokes "phosphenes" — the perception of a spot of light at a specific location
This is the anatomical basis of visual cortical prostheses.
2. Three Visual Prosthesis Paths
| Path | Location | Resolution | Limits |
|---|---|---|---|
| Retinal prosthesis | Retina | 60 pixels (Argus II) | Retina must still be healthy |
| Optic nerve | Optic nerve | ~40 pixels | Not applicable after optic-nerve degeneration |
| V1 cortex | Occipital V1 | 96 channels (Fernández) | Bypasses all downstream — broadly applicable |
V1 is the "last chance" for blindness BCI — after retinal degeneration or optic-nerve loss, the cortical path is the only option.
3. History: Dobelle to Fernández
Dobelle 1968 (first generation)
William Dobelle made the first attempt in 1968: - 64 electrodes on the cortical surface - Evoked phosphenes - Patients could "see" simple shapes (letters, crosses) - But extremely low resolution and poor stability
2000s stagnation
The Dobelle system never went commercial. Research shifted toward retinal prostheses (Argus II approved in 2013).
Fernández 2021 (second generation)
Fernández et al. (2021, Sci Adv) delivered the modern breakthrough.
4. Fernández 2021 Science Advances
Subject
- Bernardeta Gómez (blind for 16 years, with prior residual visual concepts)
- 96-channel Utah Array implanted in V1
Experiment
- Stimulation threshold: 70 μA sufficed to evoke phosphenes
- Spatial resolution tests: adjacent electrodes evoke phosphenes at different spatial positions
- Identification task: user could recognize letters (E, O, H)
- Vision + camera: wear a camera + run computer vision → image features → electrode stimulation patterns
Key findings
- V1 cortical stimulation evokes stable, repeatable visual percepts
- Low threshold (70 μA, far below retinal prosthesis requirements)
- First cortical prosthesis to achieve "letter reading"
Significance
- Proves the V1 cortical visual prosthesis path is viable
- 96 channels is not enough for a clear image, but the proof-of-principle is achieved
- Opens the path to the next generation with 1000+ channels
5. Phosphene Engineering
Controllability
By tuning current amplitude, frequency, and pulse width, phosphene properties can be modulated:
| Parameter | Effect |
|---|---|
| Current amplitude | Brightness |
| Frequency | Flicker |
| Pulse width | Size |
| Electrode selection | Spatial position |
Multi-electrode stimulation
Simultaneously stimulating multiple electrodes produces a mosaic pattern — akin to a low-resolution display.
But there is nonlinear cross-interaction: co-stimulating adjacent electrodes does not equal the superposition of two single phosphenes.
Phosphene shapes
Empirically, phosphenes are usually dots, but also include: - Line segments - Color patches - Even simple shapes
6. Differentiable Phosphene Simulation
Latest 2024 progress: differentiable phosphene simulation.
Problem
Directly optimizing "electrode pattern → user percept" requires extensive patient feedback — slow and subjective.
Solution
- Build a differentiable phosphene generation model: input electrode pattern, output predicted phosphene image
- Train the generative model with user feedback
- Given a target image, use gradients to optimize the electrode pattern
Advantages
- End-to-end optimizable: target image → best electrode pattern
- No need to poll the user each time
- Supports dynamic vision
Representatives
- Beauchamp 2020 and related work at the Moran Eye Institute
- Duret 2024 Differentiable phosphene simulation
7. Next-Generation High-Channel-Count Prostheses
Moran Eye / BLV (Baylor-Illinois-NIH-VA)
- 1000+ channel silicon-based
- Multiple regions (V1 + V2 + LGN)
- Human trials expected 2026+
Second Sight (restructured after bankruptcy)
- Orion project (V1 prosthesis)
- 60+ channels
- Preclinical
Chinese visual-prosthesis projects
- Related research at SJTU, Tsinghua Neuracle
- Human trials expected 2025–2027
8. Integration with Computer Vision
A visual prosthesis doesn't work in isolation — it requires camera + computer vision processing:
Camera image
↓
Object recognition + salient region extraction
↓
Low-resolution phosphene mapping
↓
Electrode stimulation pattern
↓
User percept
AI optimization
- Saliency awareness: stimulate only key objects (YOLO + attention)
- Edge enhancement: sharpen boundaries
- Motion emphasis: moving objects brighter
- Personalization: optimize per user
BrainGate also needs AI for vision processing — making 96 "pixels" as useful as possible.
9. Brain-to-Video Decoding vs Visual Prosthesis
These two paths are dual:
- Brain-to-video decoding: brain → visual content (read)
- Visual prosthesis: visual content → brain (write)
They share: - Visual-cortex anatomy - CLIP / visual representations - Generative-model priors
Future fused systems might include: - Blind user wears a camera → AI processes → V1 stimulation (write) - Simultaneously record V1 response → decode "what the user sees" (read) - Closed-loop optimization
This is the direction of bidirectional visual BCI.
10. Ethics and Patient Experience
Visual prosthesis ≠ restoring vision
Users do not see "normal images", but rather low-resolution phosphene mosaics — they have to learn to interpret them.
Adaptation process
The brain needs months to integrate phosphenes into meaningful vision — neural plasticity is key.
Risks
- Electrode infection
- Long-term stability
- Psychological adjustment (vision vs expectation)
Informed consent
Complex expectation management — patients must understand "it won't look like normal seeing."
11. Logical Chain
- V1 retinotopy is the anatomical basis of cortical visual prostheses.
- Fernández 2021 used a 96-channel Utah to let a 16-year-blind patient see letters.
- Phosphenes are the fundamental visual unit of cortical stimulation; multiple electrodes form low-resolution images.
- Differentiable phosphene simulation lets electrode patterns be optimized by AI.
- Next-generation 1000+ channel prostheses are the goal — approaching practical vision.
- AI + camera + visual prosthesis is the complete system: CV processing + cortical write-in.
- Read-write fusion is the future direction of visual BCI.
References
- Fernández et al. (2021). Visual percepts evoked with an intracortical 96-channel microelectrode array in a blind patient. Science Advances. https://www.science.org/doi/10.1126/sciadv.abf8986
- Dobelle et al. (1974). Phosphenes produced by electrical stimulation of human occipital cortex, and their application to the development of a prosthesis for the blind. J Physiol.
- Beauchamp et al. (2020). Dynamic stimulation of visual cortex produces form vision in sighted and blind humans. Cell. https://www.cell.com/cell/fulltext/S0092-8674(20)30437-2
- Chen et al. (2020). Shape perception via a high-channel-count neuroprosthesis in monkey visual cortex. Science.
- Duret et al. (2024). End-to-end optimization of phosphene patterns via differentiable simulation. bioRxiv.