TL;DR
Researchers have developed Brain/MINDS 3D Digital Marmoset Brain Atlas Version 2.0 (BMA2.0), a comprehensive, population-based atlas that overcomes limitations of single-subject atlases. This new atlas integrates multimodal imaging data (myelin, Nissl, ex vivo and in vivo MRI) from numerous marmosets, providing a more realistic and robust framework for studying marmoset brain structure and function.
What Problem Does This Paper Address?
Creating accurate brain atlases is crucial for understanding brain structure, function, and disease. However, previous atlases of the common marmoset, a widely used primate model, were often based on a single individual. This approach is problematic because it can introduce biases due to individual anatomical variations, potentially misrepresenting the true diversity of the marmoset brain. Furthermore, integrating data from different imaging modalities (like histology and MRI) can be challenging due to differences in resolution, contrast, and spatial distortion. The authors aimed to address these limitations by creating a population-based atlas that incorporates multiple imaging modalities and provides a standardized space for data integration and analysis.
Methods
The creation of BMA2.0 involved a multi-stage process leveraging advanced computational techniques:
- Population-Based Delineation: Instead of relying on a single brain, BMA2.0 utilized manually delineated cortical regions from 10 individual marmoset brains. These delineations were then used to create a generalized cortical parcellation.
- Multimodal Integration: The atlas integrates various data types:
- Histology: Myelin and Nissl-stained brain sections were used for detailed anatomical parcellation.
- MRI: Ex vivo MRI (from 91 individuals) and in vivo MRI (from 446 individuals) were used to create average templates.
- Advanced Registration Techniques: To align these diverse datasets, the researchers employed sophisticated image registration pipelines, including:
- AI-based Registration: Leveraging Artificial Intelligence (AI) algorithms, such as CycleGAN and Pix2Pix, to translate between different image modalities (e.g., Nissl to synthetic myelin) and generate segmentation masks. This improved the accuracy of aligning histological data with MRI.
- Streamline-Based Label Assignment: A novel method using Laplace's equation to assign labels to voxels based on streamline flow from white matter to pial surfaces, ensuring smooth transitions and respecting cortical structure.
- Template Construction: Average templates for myelin, Nissl, and in vivo MRI were created by registering and averaging data from multiple individuals. Symmetrization techniques were used to create unbiased population average templates.
- Surface Generation: Cortical flat maps and pial, mid-thickness, and white matter surfaces were generated for enhanced visualization and analysis.
Datasets
The development of BMA2.0 utilized several key datasets:
- Histological Data: Myelin and Nissl-stained sections from 10 individual common marmoset brains were used for manual delineation.
- Ex Vivo MRI Data:
- eNA91 Dataset: 91 individuals' ex vivo T2WI MRI scans were used to construct the population average ex vivo T2WI MRI template.
- BMA2017/BMA2019 Datasets: Previous Brain/MINDS atlas data was used for reference and transformation.
- In Vivo MRI Data:
- NA216 Dataset: Data from 446 individuals' in vivo T2WI MRI scans were used to create the in vivo T2WI population average template.
- NA216 DWI Data: Diffusion-weighted imaging data from 126 individuals was used for population-based tractography and connectome construction.
The paper introduces BMA2.0 as a new, comprehensive atlas that builds upon and integrates with previous Brain/MINDS atlases (BMA2017, BMA2019) and other widely used marmoset brain atlases.
Key Insights and Ideas
- Population-Based Atlases are Crucial: The study strongly advocates for population-based atlases over single-subject atlases to account for natural biological variation and provide a more generalizable representation of brain structure.
- AI Revolutionizes Image Registration: The successful application of AI models like CycleGAN and Pix2Pix for cross-modal image translation and segmentation significantly improved the accuracy and efficiency of integrating histological and MRI data.
- Multimodal Data Enhances Anatomical Precision: Combining myelin (fiber architecture) and Nissl (neuronal cell bodies) staining, along with MRI, provides complementary information that allows for more robust and accurate anatomical delineation.
- Streamline-Based Parcellation Captures Cortical Complexity: The novel streamline-based approach for cortical parcellation respects the complex, non-linear organization of the cortex, leading to more accurate regional assignments.
- BMA2.0 Enables Advanced Morphometric Studies: The atlas not only provides anatomical labels but also serves as a foundation for quantitative morphometric analyses, such as regional variations in cortical layer thickness, revealing systematic gradients related to functional hierarchies.
Key Results
- Comprehensive Parcellation: BMA2.0 includes a total of 323 regions per hemisphere, comprising 117 cortical regions, 156 subcortical regions, and 45 cerebellar lobule regions.
- Multimodal Templates: The atlas package includes co-registered average templates for myelin, Nissl, ex vivo MRI (91 individuals), and in vivo MRI (446 individuals).
- Functional Validation: Comparisons of multivariate time-series similarity between BMA2.0 ROIs and non-anatomical K-means clustering ROIs showed that BMA2.0 significantly better captured functional activity, demonstrating the functional relevance of its anatomical segmentation.
- Connectomics Validation: A population-based connectome derived from the BMA2.0 in vivo template showed a substantial correlation (Pearson's r = 0.46) with an axonal tracer-based connectivity atlas, validating the atlas's utility for network-level analyses.
- Supragranular Thickness Gradients: Analysis revealed systematic dorsal-ventral and rostral-caudal gradients in supragranular layer thickness, correlating with presumed functional hierarchies across different cortical systems.
Limitations and Future Work
- Subcortical Region Refinement: While BMA2.0 refines subcortical regions, some were combined (e.g., inferior olive subregions) due to difficulties in precise delineation with the available average template contrasts. Future work could involve further manual refinement of these areas.
- Interoperability with Other Atlases: The authors have provided transformations to several other marmoset brain atlases, but further efforts to ensure seamless integration with an even wider range of resources would be beneficial.
- Probabilistic Laminar Maps: The authors suggest that BMA2.0 can facilitate future efforts to build probabilistic laminar maps across individuals and species, analogous to developments in rodent and human brain mapping.
- Further Functional Validation: While functional validation was performed, continued exploration of how BMA2.0 can be used to integrate and analyze diverse functional datasets (e.g., fMRI, electrophysiology) is a promising avenue.
Summary updated