Mapping the Mind: Advancements in Brain Ultrasound Imaging

A few years ago, a groundbreaking study captured our imagination by demonstrating that it is possible to reconstruct the images a person is viewing simply by decoding their neural activity.

This glimpse into a potential "telepathic" future—exemplified by projects like MindEye—relies on fMRI data. However, there is a massive practical hurdle: you cannot wear an MRI machine on your head.

The Hardware Bottleneck

The primary obstacle to viable brain-computer interfaces (BCIs) is the hardware. Currently, we are forced to choose between two flawed extremes:

~~Invasive Surgery:~~ Drilling into the skull to implant electrodes.
~~Low Resolution:~~ Using EEG to capture "blurry" signals from outside the skull.

We are developing a third way: hardware that provides MRI-level precision without the need for surgery.

The Neurovascular Link

Our approach leverages the biological relationship between neurons and blood flow. When neurons activate, the vascular system responds by delivering more blood to those specific areas. By sending ultrasound waves through the skull, we can detect the scattering caused by red blood cells, allowing us to map blood volume and flow across the brain.

Defining the Ideal Mind Interface

For a BCI to be truly general-purpose, it must satisfy two critical criteria:

Broad Field of View: It must monitor a significant portion of the brain.
High Spatial Resolution: It must see fine details.

The following table illustrates why current methods fall short compared to neurovascular ultrasound:

Modality	Field of View	Resolution	Limitation
Electrodes	Very Low	Very High	Captures $\approx 0.001\%$ of brain activity
EEG / MEG	Very High	Very Low	Fundamental physics of field propagation
NV Ultrasound	Very High	Very High	Requires advanced signal processing

Because of the underlying physics, neurovascular ultrasound can record roughly a million independent pixels throughout the brain, each smaller than $1\text{mm}$ .

A Major Milestone: First Light

We have achieved what we believe is the most detailed vascular image of a living human brain ever captured through an intact skull.

This 3D image, created via ultrasound localization microscopy, reveals pial arteries, arterioles, and larger vessels. Volumetrically, this resolution is 100 times greater than that of a standard CT scan.

Because this technology has immense potential for diagnosing conditions such as:

Stroke
Alzheimer’s Disease
Traumatic Brain Injury (TBI)

...we are open-sourcing the entire dataset and processing pipeline.

The Science of Super-Resolution

To achieve this, we must overcome the diffraction limit. Normally, ultrasound cannot distinguish between two objects closer than approximately one wavelength ( $\lambda$ ).

$\text{Resolution Limit} \approx \frac{\lambda}{2}$

The Microbubble Strategy

We use FDA-approved contrast agents: pockets of sulfur hexafluoride encased in lipid shells.

By injecting these bubbles sparsely, their "blobs" do not overlap. We can then pinpoint the center of each bubble with extreme precision—far below the diffraction limit. Over a 4-minute window, we track these bubbles as they flow, creating a high-fidelity map of the microvasculature.

The Path to Contrast-Free Imaging

While microbubbles provide a proof-of-concept for the detail achievable through the skull, our ultimate goal is contrast-free neurovascular imaging.

The Challenges & Opportunities:

Hardware: Ultrasound tech has evolved from $\$ 100,000$ carts to smartphone-sized devices (e.g., Butterfly).
Signal: Red blood cells scatter sound much more weakly than microbubbles.
Data: Current pipelines are inefficient.

Standard Pipeline: Raw Data $\rightarrow$ Hand-engineered Features $\rightarrow$ 0.1% Data Retention

We believe that end-to-end machine learning, trained on massive datasets, can recover the signals that current manual pipelines discard. To facilitate this, we are currently assembling the largest neurovascular ultrasound dataset in existence.

Note: The super-resolution techniques described above currently apply only to the contrast-enhanced version of our imaging process.