NextBrain: A 3D AI-powered Brain Atlas

Let’s say you come across two neuroimaging studies, both reporting an effect in the amygdala. While one study links it to fear conditioning, the other links it to predicting uncertainty. Although this might seem comparable, the amygdala isn’t a single, uniform structure - it contains subregions with slightly different functions. So if those two studies are using inconsistent boundaries, the same label may actually refer to different sub-regions of the amygdala, and the findings could be interpreted quite differently.

Brain atlases exist to prevent this confusion, providing a consistent manner to define and compare brain regions across individuals and datasets. In this blog, we will go through the types of atlases widely used today - what they do well, where they fall short, and why a new AI-powered atlas called NextBrain is particularly exciting.

What is a brain atlas?

A brain atlas is a reference map that labels different parts of the brain, much like Google Maps for navigation. You can drop a pin at any location, and someone else can find the same spot, as long as you are using the same template. Similarly, when researchers report their findings using a common atlas, they can be more confident of referring to the same anatomical region (and its boundaries). This shared frame of reference makes findings generalisable across studies.

Brain atlases that already exist

The most widely used neuroimaging coordinate system is the MNI/ICBM system, which aligns individual brains onto a shared 3D grid derived from population averages. Reporting results in MNI standard coordinates gradually became the field's common language, allowing findings from labs and scanners worldwide to be placed on the same spatial map. But not every research question requires the same template, and over time, different atlases have evolved, each optimised for varying scientific questions.

Many popular atlases are macro-anatomical, delineating borders along large-scale landmarks that structural MRI reliably captures, such as the folds and grooves of the cortical surface (also known as gyri and sulci). A classic example of this is the Desikan–Killiany atlas, which subdivides the cortex into 34 gyral-based fixed regions per hemisphere, making it convenient for automated region-of-interest (ROI) analyses. In contrast, the Harvard–Oxford atlas takes a probabilistic approach; rather than drawing a hard boundary around a region, it encodes how consistently a given point in space (known as a voxel, which is essentially a 3D pixel) belongs to that region across many individuals. This is particularly useful when anatomy varies, better reflecting the inter-individual variability present in real life. 

"A brain atlas is a reference map that labels different parts of the brain, much like Google Maps for navigation."

However, not all atlases are structurally focused - for example, the Allen Brain Atlas asks "what is this region made of, molecularly?". By mapping gene expression across brain regions, it serves as a powerful tool for linking neuroanatomy to the biological underpinnings of cognition and disease. More recently, the Human Connectome Project (HCP) produced a multimodal parcellation of the cerebral cortex by integrating complementary MRI-derived signals, function, connectivity, and topography across 210 healthy young adults. Rather than relying on a single contrast, this approach defines anatomical borders by combining biological evidence from structural MRI, myelin-sensitive contrasts, and functional MRI (resting-state, task, and topographic mapping)..

Despite this progress, nearly all of these atlases are constrained by what MRI can see. Standard MRI operates at roughly 1mm³ resolution, which is excellent for capturing gross anatomy, but too coarse to resolve the fine-grained subregions that lie beneath. Structures like the amygdala contain multiple functionally distinct nuclei, yet can be reduced to a single label under these parameters. This may cause biologically meaningful distinctions to be averaged away before downstream analysis.

How does NextBrain bridge this gap?

NextBrain tackles this issue by starting from the tissue level. To create the atlas, the researchers took post-mortem human brain tissue and produced approximately 10,000 ultra-thin serial sections, each examined under a microscope. These sections were then digitally reconstructed into a detailed 3D volume and used to produce precise delineations of 333 brain regions of interest. Crucially, this includes the smaller subregions that standard MRI cannot resolve reliably. 

To reconstruct the 3D volume, NextBrain uses a contrastive learning-based alignment method, which is an AI technique that matches corresponding features across very different image types, allowing accurate registration of histology to MRI at scale. Once built, the NextBrain atlas becomes a practical tool for the wider research community. The authors pair it with a Bayesian segmentation tool that uses the atlas as prior information to automatically parcellate any new MRI scan into those same 333 regions. This means microscopic structures can now be applied to existing MRI datasets at scale.

"NextBrain uses a contrastive learning-based alignment method, which is an AI technique that matches corresponding features across very different image types (...)"

In Google Maps terms, it’s the difference between pinning “the city centre” and pinning a specific street address: two studies can finally point to the same location in a way that’s anatomically meaningful.

Future directions

NextBrain is already a substantial leap forward, setting up clear next steps. Expanding the dataset to include more donors, particularly across different age groups, capturing the full range of human neuroanatomical variation (across age, sex, or population), moving closer to a more representative reference.

The clinical relevance is also visible in the paper’s proof of concept: NextBrain improved Alzheimer’s disease classification accuracy from 86.9% to 90.3% compared with other atlases. This suggests that more anatomically specific segmentation might yield more informative features, making disease-related signals easier to detect and localise.

Notably, NextBrain is open source and extensible. The authors have publicly released the raw and aligned tissue data, the atlas itself, the segmentation tool, and an online visualisation tool. Researchers can download the data, add new labels, and rebuild a customised atlas for their specific region of interest. This means the resource will grow in value as the community contributes to it.

References

Casamitjana, A., Mancini, M., Robinson, E. et al. A probabilistic histological atlas of the human brain for MRI segmentation. Nature 648, 678–685 (2025). https://doi.org/10.1038/s41586-025-09708-2 

Dadar, M., Manera, A. L., Fonov, V. S., Ducharme, S., & Collins, D. L. (2021). MNI-FTD templates: Unbiased average templates of frontotemporal dementia variants. Scientific Data, 8(1), 222. https://doi.org/10.1038/s41597-021-01007-5 

Desikan, R. S., Ségonne, F., Fischl, B., Quinn, B. T., Dickerson, B. C., Blacker, D., Buckner, R. L., Dale, A. M., Maguire, R. P., Hyman, B. T., Albert, M. S., & Killiany, R. J. (2006). An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest. NeuroImage, 31(3), 968–980. https://doi.org/10.1016/j.neuroimage.2006.01.021 

Glasser, M. F., Coalson, T. S., Robinson, E. C., Hacker, C. D., Harwell, J., Yacoub, E., Ugurbil, K., Andersson, J., Beckmann, C. F., Jenkinson, M., Smith, S. M., & Van Essen, D. C. (2016). A multi-modal parcellation of human cerebral cortex. Nature, 536(7615), 171–178. https://doi.org/10.1038/nature18933 

Hawrylycz, M. J., Lein, E. S., Guillozet-Bongaarts, A. L., Shen, E. H., Ng, L., Miller, J. A., van de Lagemaat, L. N., Smith, K. A., Ebbert, A., Riley, Z. L., Abajian, C., Beckmann, C. F., Bernard, A., Bertagnolli, D., Boe, A. F., Cartagena, P. M., Chakravarty, M. M., Chapin, M., Chong, J., … Jones, A. R. (2012). An anatomically comprehensive atlas of the adult human brain transcriptome. Nature, 489(7416), 391–399. https://doi.org/10.1038/nature11405 

Iglesias, J. E., Billot, B., Balbastre, Y., Tabari, A., Conklin, J., González, R. G., Alexander, D. C., Golland, P., Edlow, B. L., Fischl, B., & Alzheimer’s Disease Neuroimaging Initiative. (2021). Joint super-resolution and synthesis of 1 mm isotropic MP-RAGE volumes from clinical MRI exams with scans of different orientation, resolution and contrast. NeuroImage, 237, 118206. https://doi.org/10.1016/j.neuroimage.2021.118206 

Lancaster, J. L., Tordesillas-Gutiérrez, D., Martinez, M., Salinas, F., Evans, A., Zilles, K., Mazziotta, J. C., & Fox, P. T. (2007). Bias between MNI and Talairach coordinates analyzed using the ICBM-152 brain template. Human Brain Mapping, 28(11), 1194–1205. https://doi.org/10.1002/hbm.20345 

Schaefer, A., Kong, R., Gordon, E. M., Laumann, T. O., Zuo, X.-N., Holmes, A. J., Eickhoff, S. B., & Yeo, B. T. T. (2018). Local-global parcellation of the human cerebral cortex from intrinsic functional connectivity MRI. Cerebral Cortex, 28(9), 3095–3114. https://doi.org/10.1093/cercor/bhx179 

---

This article was written by Anwesha Chakraborty and edited by Julia Dabrowska, with graphics produced by Suzana Sultan. If you enjoyed this article, be the first to be notified about new posts by signing up to become a WiNUK member (top right of this page)! Interested in writing for WiNUK yourself? Contact us through the blog page and the editors will be in touch.