QWE AI Academy By the end of this guide you’ll have a working understanding of how Aleph Neuro’s released ultrasound imaging of the brain pipeline turns sound waves bouncing off blood cells into a 3D vascular map of a living human head – and the practical steps to pull their open-source code, feed it sample data, and read the output without falling into the traps the press releases skipped.
Aleph Neuro went public on June 25, 2026. The claim: the highest-resolution 3D images of the brain taken from outside the skull – 100 times greater volumetric resolution than comparable CT, through an intact adult skull. The kicker for builders: they open-sourced the entire pipeline along with the dataset. (BusinessWire/Yahoo Finance announcement)
What you’re actually reproducing
Not a generic brain scan. The first 3D ultrasound localization microscopy in a human brain through an intact skull, at 100× volumetric resolution vs comparable CT. The mechanism is microbubbles – tiny gas pockets that act like individual reflective dots you can track.
Short version: inject FDA-approved sulfur hexafluoride microbubbles (lipid-shell, the same approved contrast agent used in cardiac imaging) into the bloodstream. Fire ultrasound through the skull. Each bubble throws a sharp echo. Localize the center of that echo at sub-wavelength precision using a sub-pixel Gaussian fit. Track those centers across frames. Stack them into a 3D vessel map.
Turns out the bubbles do something ordinary ultrasound can’t – they beat the diffraction limit. Normally, two objects closer than about one wavelength apart collapse into a single blob on the image. A single bubble still blurs into a wavelength-wide spot, but fitting a curve to that spot pins its center far below that blur radius. That’s the mechanism. (Aleph technical blog)
Setting up the open-source pipeline
The repo link is in the Aleph blog post. Three requirements before running anything: the raw RF dataset they published, a Python environment with the standard scientific stack, and a block of uninterrupted time for bubble tracking (it’s slow; don’t kill it).
- Clone and install. Follow the repo README. As of the June 2026 launch, dependency versions aren’t fully pinned – pin yours before you start.
- Pull the dataset. Tens of gigabytes of raw radiofrequency data, not processed images. Budget storage and download time.
- Beamform. Unfocused plane waves insonify the whole field. Tilted emissions get re-focused (beamforming) and coherently compounded – this is what buys the high frame rate and depth needed for long sequences.
- Filter clutter. Compound Doppler signals are processed to strip strong tissue returns, then squared and time-integrated. Output: power Doppler frames where pixel brightness ∝ cerebral blood volume.
- Detect and localize bubbles. Sub-pixel Gaussian fits over each bright spot per frame.
- Track and accumulate. Bubble centers stacked into a 3D volume across the full acquisition.
The tracking step is where most first-time runs stall. Don’t kill it.
Reading the output without fooling yourself
A volume rendering comes out looking like an MRA. Don’t read it like one. Each voxel is an accumulation count, not a snapshot of blood flow at one instant:
# Mental model of the output
volume[x, y, z] = sum over time of bubble_localizations_at(x, y, z)
# Bright voxel != more blood right now
# Bright voxel == more bubbles passed through here during the 4-min acquisition
Sharp vessel trees appear because every voxel along a real vessel gets hit by many bubbles. Noise voxels get hit once or zero times and fall below threshold.
Before tweaking the bubble detector: check the raw localization density map. Blotchy in one hemisphere? That’s skull phase aberration, not a code bug. The phase aberration correction in the pipeline is doing real work – disable it “to see what’s happening” and you’ve made the problem invisible, not gone.
Three traps the launch coverage didn’t mention
The 4-minute infusion window is non-negotiable. The bubbles – SF6 pockets in lipid shells, continuously infused – have a fixed circulation window. Interrupt the scan halfway and you can’t just resume. Bubble concentration in circulation has shifted; localization density across the volume becomes uneven in ways that won’t crash the pipeline but will silently distort the output. Plan the full acquisition sequence before pressing record. (Aleph technical blog)
The catch: motion ruins everything, and quietly. A 2025 fUS denoising study puts it plainly – sensitivity to motion artifacts, cardiac pulsatility, anesthesia, and respiration limits the capacity to detect milder cerebral alterations. (PMC12406050) A subject who breathes deeply or shifts mid-scan won’t crash your pipeline. You’ll get a softer image and never know why. Standard countermeasure for functional work: canonical correlation analysis preprocessing before any connectivity analysis.
The contrast-enhanced result is a milestone, not an endpoint. Aleph says it directly: the contrast-enhanced images are a step on the path, not the destination. The destination is contrast-free neurovascular imaging. (Aleph technical blog) If you’re building on this pipeline expecting a stable final API – don’t. Expect breaking changes as they move toward that goal.
What “performance” actually means here
Numbers people throw around in comparison threads, grounded:
| Modality | Spatial | Temporal | Practical reach |
|---|---|---|---|
| fMRI (BOLD) | ~1-2 mm | seconds | Whole brain, room-sized scanner |
| Standard fUS | ~100 µm | ~100 ms | Whole brain in rodents; adult skull is a hard obstacle |
| EEG | cm-scale | ms | Wearable, spatially blurry |
| Aleph ULM (June 2026) | Sub-wavelength via bubble localization | 4-min integrated | Through intact adult human skull, contrast-enhanced |
The fUS row comes from a 2025 drug-screening fUSI study: ~100 µm spatial, ~100 ms temporal, sensitive to flow under 1 mm/s. (PMC12423641) That’s the baseline before ULM. Adult skulls strip resolution hard – significant acoustic impedance contrast between skull bone and soft tissue is the core physics problem ultrasound has always had to work around.
What does contrast-free imaging actually enable, if Aleph ships it? No IV line, no infusion window, potentially wearable. But that’s still ahead. For now the open question is whether the localization accuracy holds without the microbubble signal enhancement – and nobody’s published that answer yet.
When NOT to use this
Research pipeline. Treat it as one.
- Clinical decisions. Established clinical fUS contexts – as of a 2020 Deep-fUS preprint – are intraoperative tumor removal and neonatal imaging through the anterior fontanel window. (bioRxiv preprint) Hobbyist reconstructions aren’t in that category.
- Functional connectivity studies right now. The pipeline images vasculature with contrast agent. Neural activity readout needs the contrast-free version, which hasn’t shipped.
- If IV contrast is off the table. SF6 microbubbles are FDA-approved but still an injection. Required for this method as it stands.
One thing worth sitting with: a pipeline you can clone changes who gets to iterate. The gap between EEG’s spatial blur and the invasiveness of implanted electrodes has been stuck for years. Ultrasound brain imaging is one of the few non-invasive candidates in that space. Whether Aleph’s open-source drop accelerates that or turns into a research footnote depends on what the community actually builds on top of it.
FAQ
Is this the same as the fUS used in neuroscience labs?
Same family, different scale. Lab fUS typically images rodents through a thinned skull. Aleph pushes through a full adult human skull using ULM with microbubbles to recover the resolution the skull destroys. The foundational math – plane-wave compounding, power Doppler – is the same; the skull correction and localization layer on top are what’s new. The original fUS method is described in Macé et al. (2011, Nature Methods).
Can I run this on a Butterfly iQ3 I bought as a clinician?
Almost certainly not. Aleph used Butterfly’s Embedded licensing program – a separate co-development track, not the commercial iQ3. Same chip family, but raw RF access and the custom acquisition modes needed for ULM aren’t exposed on the consumer device. Use the open-source pipeline as a data-processing reference, not a recipe to scan anyone.
How long until this becomes a wearable?
No timeline is public.
Next move: open the Aleph technical post, grab the repo, and run the beamforming step on the sample data. Once a clean Doppler frame comes out, the rest of the pipeline is bookkeeping.