Â鶹Éç

The brain structure spans multiple scales, from nanoscale synapses to macroscale brain circuits, whose organization underpins brain function. Understanding how brain works and is disrupted in disease requires 3D analysis of its microarchitecture across ALL scales. This has only recently become possible via Synchrotron X-ray Nano-Holotomography (XNH) imaging. XNH captures nanometer-resolution 3D images over millimeter volumes in ~2 hours. 

For the first time, we visualize the entire brain hierarchy in a single XNH scan—a feat beyond MRI or traditional microscopy. Yet, each XNH scan is terabytes of data, growing exponentially with every advance in synchrotron technology. Segmentation—critical for any downstream analysis—is not feasible manually (takes years). Today's AI tools collapse at this scale; their memory demands scale exponentially with data size, requiring multi-GPU supercomputers even to attempt(each GPU >300,000DKK). Moreover, these models operate at a single scale, missing cross-scale relationships.I take a different direction: instead of building ever-larger models, I propose a memory-efficient, multi-scale AI designed from the ground up to segment terabytes of 3D brain scans in minutes using standard GPUs. 

The model will mimic expert analysis —zooming in for fine details and out to capture brain organization—via a novel dual-branch design. To make it memory-efficient yet 3D, I propose a novel 2D-to-2.5D training scheme, enabling the model to reach 3D-level performance with at least ~500 times less memory. I'll further compress the model using new pruning methods. Data for model training/testing is already available via the collaborator. I'll share all pipeline/data openly, providing the first accessible tool to gain novel insights from the most advanced brain scans we have ever had. By cutting the analysis time from years to minutes, it'll accelerate discoveries in neuroscience, paving the way for more advanced diagnostic and treatments.