Synchrotron radiation-based tomography of an entire mouse brain with sub-micron voxels: augmenting interactive brain atlases with terabyte data

This study presents a publicly available, 3.3-teravoxel synchrotron radiation-based X-ray microtomography dataset of an entire mouse brain with 0.65 μm isotropic resolution, which has been non-rigidly registered to the Allen Mouse Brain Common Coordinate Framework and made accessible via browser-based viewers to significantly enhance existing interactive brain atlases.

The Gist

Imagine you have a tiny, incredibly complex city inside a mouse's head. This city is the brain, and it's packed with billions of tiny streets (neurons), buildings (cells), and power lines (synapses). For a long time, scientists could only take blurry, low-resolution photos of this city, or they could zoom in so close that they could see the bricks on a single building, but they lost the map of the whole city.

This paper is about a team of scientists who finally managed to take a crystal-clear, 3D photo of an entire mouse brain at a resolution so high you could see individual cells, all while keeping the whole brain in the frame.

Here is how they did it, explained with some everyday analogies:

1. The "Puzzle Piece" Problem

Think of the mouse brain as a giant jigsaw puzzle. The scientists wanted to see every single piece, but their camera (a super-powerful X-ray machine at a giant particle accelerator called a Synchrotron) was like a tiny magnifying glass. If they used it normally, they could only see one small square of the brain at a time.

To solve this, they didn't just take one photo. They took 64 different photos (8 rows by 8 columns) of the brain, slightly overlapping each other like tiles on a bathroom floor. Then, they used a computer to stitch these tiles together into one giant, seamless image. It's like taking 64 photos of a football field with a phone camera and stitching them together to make one massive, ultra-high-definition photo of the whole field.

2. The "Terabyte" Mountain

The result of this stitching was a digital file so huge it would crash a normal computer. It's 3.3 Teravoxels (a voxel is a 3D pixel). To put that in perspective:

• If a standard photo is a single sheet of paper, this dataset is a library of books stretching from the Earth to the Moon.
• It weighs in at 3.3 Terabytes of data.

Because the file is so massive, you can't just email it or put it on a USB stick. The team had to invent a special way to "zip" and organize this mountain of data so that anyone with a web browser could explore it without needing a supercomputer.

3. The "Google Maps" for Brains

Having a giant, high-res photo is great, but it's useless if you don't know where you are. Imagine having a perfect photo of a city but no street names.

The scientists took their new X-ray brain and registered it (matched it up) with the "Allen Mouse Brain Atlas." Think of the Atlas as the official Google Maps for mouse brains. It has all the street names (brain regions) and landmarks.

• The Challenge: The X-ray brain and the Atlas map looked slightly different (like comparing a photo of a city in winter vs. summer).
• The Solution: They used a smart algorithm to gently stretch and warp their X-ray brain until it perfectly overlaid the Atlas map. Now, if you click on "Hippocampus" in the Atlas, the computer instantly shows you the high-res X-ray view of that specific area.

4. The "Interactive Museum"

Finally, they didn't just lock this data away in a lab. They uploaded it to the internet (via a platform called Ebrains) and made it viewable through special web tools called Neuroglancer and siibra-explorer.

Think of this as opening a virtual museum where:

• You can fly through the brain in 3D.
• You can zoom out to see the whole brain, then zoom in until you can see the nucleus of a single neuron (like zooming from a satellite view of a country down to a specific person's face).
• You can share a link with a friend, and they will land exactly where you are looking, ready to explore together.

Why Does This Matter?

Before this, scientists had to choose between seeing the "big picture" or seeing the "tiny details." This study bridges that gap. It allows researchers to study how the tiny wiring of a single cell connects to the massive structure of the whole brain.

It's like finally having a map where you can see the entire country, but also zoom in to read the license plate on a car in a specific town. This opens the door for new discoveries in how the brain works, how diseases affect it, and how we might treat them in the future.

In short: They built a massive, ultra-detailed 3D map of a mouse brain, matched it to the official guidebook, and put it online so anyone with a browser can explore it.

Technical Summary

Here is a detailed technical summary of the paper "Synchrotron radiation-based tomography of an entire mouse brain with sub-micron voxels: augmenting interactive brain atlases with terabyte data."

1. Problem Statement

Understanding the relationship between brain structure and function requires imaging across multiple length scales, from the whole organ (~1 cm) down to individual cells (1–10 µm) and synapses (<100 nm). While X-ray micro- and nanotomography offer non-destructive, high-resolution 3D imaging, a significant gap exists:

• Field-of-View (FOV) Limitation: Standard X-ray microtomography detectors typically capture volumes of ~10 mm³, whereas a full mouse brain is ~500 mm³.
• Data Scale: Imaging a full brain at sub-micron resolution generates terabyte-scale datasets (e.g., 3.3 teravoxels), creating challenges in storage, processing, memory-limited registration, and accessibility for neuroscientists with limited image processing expertise.
• Integration: Existing brain atlases (like the Allen Mouse Brain CCFv3) lack the isotropic sub-micron resolution provided by X-ray tomography, and merging these modalities is computationally difficult due to memory constraints and multi-modal registration issues.

2. Methodology

A. Data Acquisition (Extended-Field X-ray Microtomography)

• Source: Synchrotron radiation at the Anatomix beamline (Synchrotron Soleil, France) using a polychromatic beam (~38 keV mean energy).
• Sample: A C57BL/6JRj mouse brain, perfused with PFA, dissected, and dehydrated in ethanol to enhance X-ray contrast.
• Scanning Strategy: To overcome detector limitations (2048 × 2048 pixels), the authors employed an 8 × 8 offset acquisition scheme:
  • The rotation axis was displaced laterally and vertically relative to the detector.
  • 8 height steps were performed; at each height, 4 rotations (360°) were recorded with offset axes to ensure ~200-pixel overlap between tiles.
  • Total projections: 4,495 (after stitching logic).
  • Resolution: 0.65 µm isotropic voxel size.
  • Scan Time: ~8 hours total.

B. Image Reconstruction and Processing

• Stitching: Projections from the 8 × 8 grid were stitched to form extended-field projections (14,982 × 14,982 pixels). Translational offsets were refined using normalized cross-correlation on overlapping regions.
• Artifact Correction:
  • Phase Retrieval: Propagation-based phase retrieval applied with $\delta/\beta = 140$.
  • Ring Artifact Correction: High-pass filtering of the mean projection to isolate and subtract inhomogeneities.
• Reconstruction: The gridrec algorithm (via tomopy) was used on blocks of 32 sinograms.
• Output: A reconstructed volume of $15,000^3$ voxels (approx. 3.3 teravoxels), representing ~911 mm³ of raw FOV.

C. Registration to Allen Mouse Brain CCFv3

• Challenge: Registering a 3.3-teravoxel volume to the Allen CCFv3 atlas (2-photon microscopy template) exceeds standard RAM limits.
• Solution: A distributed hierarchical registration framework:
  1. Downsampling: The microtomography (µCT) volume was downsampled 32× (to 20.8 µm voxels) to fit in memory.
  2. Coarse Registration: Affine and non-rigid registration performed on the low-res volume using elastix. The cost function combined:
     • Negative Mutual Information (for multi-modal intensity).
     • Mean Squared Difference (for edge alignment).
     • Kappa Statistics (for segmentation overlap).
     • Landmark Constraints: 70 manually selected anatomical landmark pairs.
  3. Local Transformation: The resulting transformation field was used to define local regions of interest (ROIs) in the full-resolution volume.
  4. Block-wise Warping: Full-resolution blocks were transformed independently using the low-res transformation parameters, circumventing memory bottlenecks.
• Result: A non-rigidly registered volume aligned with the Allen CCFv3 coordinate system.

D. Data Dissemination and Visualization

• Format Conversion: The registered volume was converted into sharded, pre-computed Neuroglancer format.
  • Data was chunked ($64 \times 64 \times 64$ voxels) and compressed (gzip).
  • Six resolution levels were generated (0.65 µm to 20.8 µm).
  • Sharding reduced the file count from ~20 million to ~5,000 files, optimizing I/O performance.
• Access: Data was uploaded to the EBRAINS repository.
• Viewers: Accessible via siibra-explorer (for semantic navigation using the Allen ontology) and Neuroglancer (for high-performance, multi-resolution 3D visualization). URLs encode specific view coordinates for easy sharing.

3. Key Results

• Dataset Scale: The final dataset covers 3.3 teravoxels (6.6 TB raw, 5.4 TB compressed).
• Resolution: Achieved isotropic 0.65 µm resolution across the entire brain, visualizing individual neurons, fiber tracts, and nuclei (e.g., hippocampus, striatum, cerebellum) in 3D.
• Registration Accuracy:
  • Manual pre-alignment error: 2.49 mm.
  • Final non-rigid registration error (with landmarks): 0.08 mm (mean landmark error).
  • The combination of intensity-based metrics and landmark constraints proved most robust.
• Volumetric Analysis: The registration revealed a mean volumetric strain of -48% (shrinkage) when mapping the ethanol-dehydrated sample to the CCFv3 template. This aligns with previous findings of ~39% shrinkage from formalin to ethanol, validating the physical deformation model.
• Visualization: The hierarchical format allows users to navigate from macroscopic views to sub-cellular details seamlessly. Specific anatomical regions (e.g., parabrachial nucleus) can be queried directly via the atlas ontology.

4. Key Contributions

1. Full-Brain Sub-Micron Imaging: Demonstrated the first successful acquisition and reconstruction of an entire mouse brain at 0.65 µm isotropic resolution using synchrotron radiation, bridging the gap between whole-organ and cellular imaging.
2. Scalable Registration Pipeline: Developed a novel distributed, hierarchical registration method that enables the alignment of terabyte-scale images to standard atlases without requiring supercomputing memory resources.
3. Open Data Infrastructure: Established a workflow for converting massive virtual histology datasets into the Neuroglancer format, making them accessible to the broader neuroscience community via standard web browsers without specialized software installation.
4. Enhanced Atlas Utility: Augmented the Allen Mouse Brain CCFv3 with true 3D isotropic resolution, allowing for the verification of atlas boundaries and the exploration of structures in any virtual plane.

5. Significance

This work represents a paradigm shift in neuroimaging data sharing. By solving the "big data" problem in structural neuroscience, it allows domain experts (neuroscientists) to interact with terabyte-scale datasets using familiar web tools.

• Scientific Impact: It provides a reference dataset for correlating structure and function at the cellular level across the whole brain, potentially aiding in the mapping of connectomes and the study of neurodegenerative diseases.
• Methodological Impact: The pipeline (acquisition $\to$ stitching $\to$ hierarchical registration $\to$ web-based visualization) serves as a blueprint for future large-scale organ imaging projects.
• Collaboration: The ability to share specific 3D coordinates via URL facilitates precise collaboration and reproducibility, allowing researchers to overlay their own data (e.g., from other modalities) onto this high-resolution reference.

The dataset is publicly available via EBRAINS (DOI: 10.25493/759K-JVU), and the source code for reconstruction and transformation is available on GitHub.