Zebrafish brain projection: a step-by-step gallery project

The dots on your screen aren't the room yet

You open ImageJ, you see 100,000 neurons glowing in their anatomical neighborhoods, and you think: this should be in a gallery. Then you blow the image up to poster size, paste it on a wall, and a visitor walks past in four seconds. The biology is extraordinary; the installation is forgettable. That's the gap we're closing today.

We're going to walk through how to take that head-of-a-pin brain — 100,000 neurons packed into less than half a cubic millimeter — and turn it into a room-sized, physically printed, sonified, interactive installation that makes a non-STEM visitor actually stop and stay. The key isn't the projector or the printer. It's the framework under the data that lets you map cleanly to a new scale. That framework is the Z-Brain atlas, and if you don't already have it as your project's north star, let's fix that right now.

From Micro-Anatomy to Gallery Scale: The Z-Brain Framework

Here's the thing about a larval zebrafish brain at 6 dpf: it's so small and so reproducible that we can treat it like a standard part. The whole organ is under 0.5 cubic millimeters, but every animal has roughly the same 100,000 neurons in roughly the same places. That reproducibility is what makes the leap from wet lab to gallery possible. You're not trying to abstract a unique brain — you're trying to expose a canonical one.

The Z-Brain atlas, first published by the Engert lab and now maintained as a community resource, was built from 899 individual brain scans of nacre/mitfa mutants at 6 dpf. Those 899 brains were computationally averaged into a single reference brain, and the result was partitioned into 294 anatomically defined regions — forebrain, optic tectum, cerebellum, the works. When you register your own dataset to Z-Brain, you are not just saying "this is the tectum." You're saying "this is layer 4 of the tectum, at coordinate (x, y, z) in the standard reference space," and that same coordinate can be the same coordinate your installation software uses to place a projection, fire a MIDI note, or light an LED.

This is why we keep coming back to Z-Brain. It's the scaffold that connects your raw tiff stack to your gallery's coordinate system. Skip it, and you'll be hand-annotating every neuron's "neighborhood" in your own private ontology, and six months later you'll realize your collaborator's lab can't reproduce your mapping. Use it, and your installation and your preprint are describing the same brain.

Treat the brain like a standard part. The reproducibility is what lets you ship a room-sized thing.

Precision Alignment: Mapping Neural Data to 3D Physical Models

Once you have Z-Brain as your coordinate target, the next move is registration. For most labs the workflow is ANTs SyN (Advanced Normalization Tools – Symmetric Normalization), the open-source diffeomorphic registration algorithm that handles the soft, slightly variable larval brain with eerie grace. The published alignment precision against Z-Brain is about 8 micrometers — roughly one cell diameter. That's your floor. You cannot, with this pipeline, claim to be mapping subcellular features; you can only claim to be mapping cells and clusters of cells. Be honest about that in your wall text, because visitors will trust you more when you tell them what the system can't do.

So now you have a registered volume: every active neuron in your experiment tagged with a 3D coordinate inside the Z-Brain reference. The next move is to pick your physical medium. There are essentially three choices, and they each have a different feel in the room:

• 3D-printed translucent resin — the whole brain, printed at 50× or 100× scale, sitting on a plinth. The visitor walks around it. Projection mapping lands on the resin's surface and follows the dorsal-ventral axis as the visitor moves. This is the route Cartography of Touch took, and it works because the object itself is the artifact.
• Etched glass or acrylic slabs — serial optical sections printed at 200×–400× scale and stacked. Light projected through each slab at a different angle creates volumetric depth. Cheaper than resin, more fragile, beautiful in a darkened room.
• Suspended volumetric projection — no physical model at all, just a haze of water vapor or a thin foil screen, with a tightly aligned projector painting the brain in mid-air. Looks magical, costs a fortune, and a single sneeze ruins your alignment.

Whichever you pick, the math is the same: take Z-Brain coordinate (x, y, z), apply your affine transform to the installation's coordinate system, and that's where the pixel goes. For a 100×-scale resin print, one cell in the larva becomes roughly 0.8 mm in the physical model — visible from across the room, but still feeling intimate.

Translating Neural Activity into Auditory and Visual Experiences

Here's where the lab geek in you gets to play. You have hundreds of thousands of spatial coordinates, each tagged with a time of activity from your calcium imaging run. You can't show all of them. You have to downsample, and you have to choose your mapping function carefully.

A common approach, and one that works surprisingly well sonically, is to take a downsampled subset of your active neurons — say, 23,743 neurons thinned to 8,000 representative points — and assign each neuron a pitch within a four-octave diatonic scale based on its anterior-posterior coordinate. The result, when you trigger a note every time a neuron fires, is a melodic line that traces the brain's active regions over time. The tectum (visual processing) sings in a different register than the hindbrain (motor). Visitors hear the brain thinking.

Pair this with projection mapping, and you have a coupled audio-visual layer: every note corresponds to a flash of light on the printed or projected brain. The Hong Lab and others have built these in environments like Max/MSP or TouchDesigner, with a small Python service reading the registered coordinates and emitting OSC messages to the visual server. The point is to keep the two layers synchronized to a single clock; even 50 ms of drift between the flash and the note will break the spell.

This is also where the sci-art canon is your friend. Pieces like Neuro Memento Mori — a rotating projected brain that decays as the audio slows — show that the mapping doesn't have to be literal. You can map age to color, mutation to timbre, or developmental trajectory to melodic contour. The Z-Brain scaffold gives you the anatomy; the mapping function is where your artistic voice lives. teamLab's interactive projection work, which kicked off in 2013 with pieces like Sketch Aquarium, showed the gallery world that biology-scale responsiveness was commercially viable at scale; your lab can do the same trick on a much smaller budget, with much more rigorous data underneath.

The clean signal is the magic. Choose your mapping function before you choose your software.

Managing Latency and Artifacts in Live Interactive Displays

Now the lecture hall part of the lab meeting. If your installation is static — a pre-rendered loop, a printed model — you can stop reading here. If your installation responds to a visitor (motion sensor, EEG cap, joystick, breath, gaze), you have a real-time systems problem, and this is where most student-built installations die.

The latency budget in a live neuroscience art piece has at least four layers, and each one can spike independently:

1. Sensor sampling — your Kinect, camera, or biosignal device is sampling at some fixed rate. At 30 Hz, you're already at 33 ms of jitter.

2. Data buffering — you need to smooth the sensor input, usually with a rolling window of 100–250 ms. Otherwise the room is too twitchy.

3. Feature extraction — pulling out the meaningful signal (e.g., "visitor is moving their right hand") takes another 20–80 ms in optimized code.

4. Transport — sending the result to your visual and audio engines via MIDI or OSC adds 5–15 ms, more on congested networks.

Stack these and you're somewhere between 150 and 400 ms. That's perceptible. Visitors will feel the lag as "the room is drunk." Your job is to either hammer the latency down to under 100 ms (expensive, requires native code and fast hardware) or to design the interaction so the lag is invisible (e.g., the visitor's input modulates a slow-evolving soundscape rather than triggering an immediate flash).

The second disaster is artifacts. A loose cable, a sunbeam through a window, a visitor's black shirt fooling your depth sensor — any of these will inject garbage into your signal and corrupt the projections. The fix is an artifact gate: a hard-coded set of rules that drops any input not matching physical plausibility (depth outside expected range, velocity above biological limits, brightness jumps inconsistent with the lighting model). Log every dropped frame. Review the log after opening night. You'll find at least three bugs you didn't know existed.

The Physiological Impact of Physical Sci-Art Installations

Here's the part that makes this work matter beyond the aesthetic. A 2023 study published in a peer-reviewed science communication journal tracked STEM and non-STEM visitors through a neuroscience art exhibit and found something striking: a single visit closed the baseline interest gap between the two groups, and two weeks later, knowledge retention was equivalent to reading scientific abstracts. The art did the outreach that the abstracts never could.

The mechanism is physiological. A separate 2024 study found that viewing original physical artworks in a gallery produced a 22% reduction in salivary cortisol and measurable decreases in inflammatory markers IL-6 and TNF-alpha. The same study found that viewing digital reproductions of the same artworks on screens produced only an 8% cortisol reduction. The body knows the difference between a thing and a picture of a thing. Your installation is a thing.

This is why the printed 3D brain, the projected resin model, the darkened room with synchronized audio — all of it matters. You're not just decorating a science result. You're building a regulatory environment for the visitor's nervous system, and in that environment, the brain you're showing them becomes a piece of their own experience. The head-of-a-pin brain becomes a room-sized event in the listener's body, and the science lands.

What to do tomorrow

If you've been nodding along and now have a stack of two-photon files on your bench, here's the path I'd take. First, download the Z-Brain atlas from the public repository and register your brightest dataset against it using ANTs SyN. Don't fight the 8 μm floor — describe it accurately in your wall text. Second, do a one-week prototype: print a 50× translucent brain on a standard resin printer, set up a short-throw projector, and map one channel of your registered activity to it. No audio yet. Just light. Get the geometric alignment perfect before you add anything else. Third, add one audio layer — sonify a single anatomical region to a single instrument — and synchronize it to the projection. Run it on a friend. If the lag feels magical, ship it. If it feels drunk, return to the latency budget.

That's the whole pipeline, and every piece of it is in your hands. The Z-Brain scaffold, the registration math, the 3D print, the projection mapping, the OSC routing, the sonification — none of it requires a museum budget or a curation team. It requires a wet lab that knows its anatomy, a dry lab that knows its coordinates, and the patience to ship a small thing before you ship a big one. Let's get your brain out of the microscope and into the room.