High-resolution mri guided whole mouse brain neuronal cell type atlas using deep learning

Abstract

Cell types represent groups of cells with shared anatomical and functional properties. Traditional brain cell type atlases rely on single-cell sequencing, which provides molecular detail but lacks whole-brain, isotropic resolution. Diffusion magnetic resonance imaging (dMRI) offers a complementary approach for probing cytoarchitecture and myeloarchitecture, with quantitative metrics increasingly used as biomarkers of brain development and neurodegenerative disorders. However, the capacity of dMRI to predict cell types remains unclear. Here, we develop a multimodal framework by integrating high-resolution dMRI and three-dimensional light-sheet microscopy of adult mouse brains through registration to the Allen Mouse Brain Common Coordinate Framework. We investigate correlations between dMRI and spatial transcriptomics-derived cell types and generate a whole-brain neuronal cell type atlas at 10 µm isotropic resolution using deep learning. Together, these results establish an efficient, high-resolution strategy for brain neuron atlas generation and underscore the potential of advanced imaging techniques to illuminate cellular mechanisms of the brain.

Diffusion MRI predicts whole-brain neuronal cell-type distributions, providing a noninvasive framework for mapping cellular organization and linking macroscopic imaging signals to underlying molecular architecture.

Introduction

Cell types are the basic structural and functional units of organisms. In the mammalian brain, cell types typically have a high degree of diversity and specialization. Studies^1,2 have shown that cell types are highly diverse in the midbrain, hindbrain, hypothalamus, and cortex. Yao et al.³ found that cell types have a high degree of transcriptomic identity and spatial specificity. More importantly, brain cell types also have a hierarchical feature in terms of taxonomies, including (from the highest to the lowest): (1) neuronal and non-neuronal (glial) classes, (2) cell classes by neurotransmitter type and major brain regions, and (3) cell subclasses within each cell class^3,4. Characterization of brain cell types is important, as it will facilitate our understanding of the development, evolution, and pathology of the brain⁵. One of the most widely used approaches is single-cell RNA-sequencing (scRNA-seq), which can profile the expression of thousands of genes per cell. By combining scRNA-seq and spatial transcriptomics, comprehensive cellular and molecular brain atlases can reveal cell types and their epigenomic information in their native tissue context. Brain Initiative Cell Census Network (BICCN) is developing several mouse brain atlases, including transcriptomic and spatial atlas^1,3,6, epigenetic landscape, and gene regulatory elements^7,8. Although many of the aforementioned atlases have been integrated into the Allen Mouse Brain Common Coordinate Framework (CCFv3)⁹, limited efforts have been made to enable predictive inference of one modality from the other¹⁰. As cell atlas research is fundamental for guiding future therapeutic innovations for brain disorders¹¹, cross-modal prediction becomes increasingly essential for validating the consistency and alignment of cell type identities across modalities.

Novel imaging techniques, such as magnetic resonance imaging (MRI), are one potential alternative to scRNA-seq for characterizing brain cellular profile. Among various MRI techniques, diffusion MRI (dMRI) provides a unique approach to measure the diffusion of water molecules in tissue and has been demonstrated to reveal tissue microstructure and cytoarchitecture¹². For example, the apparent diffusion coefficient (ADC) derived from diffusion-weighted imaging (DWI) is highly correlated with cellular size changes during acute brain ischemia¹³. By using different statistical and biophysical models, dMRI can detect specific cellular changes. Diffusion tensor imaging (DTI) applies the diffusion tensor model to extract the anisotropy and diffusivity metrics. The degree of diffusion anisotropy during brain development reflects the migration and organization of glial cells and neurons¹⁴ and the myelination process¹⁵. Diffusion kurtosis imaging (DKI) quantifies the non-Gaussian distribution of water diffusion and has a high sensitivity for pathological changes in neuronal tissues¹⁶. Neurite orientation dispersion and density imaging (NODDI) employs a biophysical model to extract the neurite density index (NDI) and orientation dispersion index (ODI) of tissue. NODDI has better specificity for certain tissue properties, such as its superior sensitivity to tissue-specific white-matter tract damage in Alzheimer’s Disease¹⁷. Soma and neurite density imaging (SANDI) is more specific to cellular and neurite densities, as the tissue component model explicitly includes the soma of any brain cell type, enabling the estimation of apparent soma density and size¹⁸. SANDI parameters such as Fsoma, representing the signal fraction associated with cell bodies, and Fneurite, representing the signal fraction associated with axonal and dendritic processes, offer complementary insights into cellular composition and microstructural organization, and are therefore well suited for characterizing spatial variations in cell density and neurite architecture across the brain. Current studies have investigated the association of dMRI parameters with cell type-specific spatial transcriptomics. De Santis et al.¹⁹ found that mean diffusivity (MD) is sensitive to microglia density. Patel et al²⁰. found that fractional anisotropy (FA) is associated with CA1 pyramidal cells. Shen et al.²¹ found that FA is correlated with myelination and oligodendrocytes, and the diffusivity indices are correlated with neurons and their connecting compounds. However, whether dMRI can predict brain cell types is still unknown^22,23. Investigating the predictive performance of dMRI in brain cell types will enhance its applications in neuroscience and deepen our understanding of the brain across different scales.

Light-sheet microscopy (LSM) enables rapid, high-resolution volumetric imaging of whole brains at near-cellular resolution and is widely used to provide detailed anatomical references^24–26. In parallel, deep learning has emerged as a powerful approach for modeling high-dimensional neuroimaging data and capturing nonlinear relationships between imaging signals and underlying biological structure^27,28. In diffusion MRI, where signals encode complex tissue microstructure, deep learning provides a promising framework for linking imaging-derived features to cellular organization^29–31. Imaging-based atlas construction has leveraged MRI and LSM to generate standardized anatomical references and high-resolution cellular maps^32–34, with recent efforts integrating multimodal imaging to bridge spatial scales^35,36. Here, we extend this paradigm by combining diffusion MRI, LSM, and deep learning to predict neuronal cell types across the whole brain.

In this study, we integrated in-house acquired high-resolution diffusion MRI with publicly available high-resolution 3D LSM³⁷ in the CCFv3 space and developed a deep learning-based whole-brain neuronal cell type prediction model to map diffusion-derived features to cell type labels defined by the Allen Brain Cell (ABC) Atlas³. This study integrates multimodal high-resolution imaging data into the CCFv3, a mouse brain common coordinate framework, enabling spatially resolved measures of water diffusivity and microstructure to be directly aligned with other modalities and reference datasets. This study also demonstrates the usage of dMRI in predicting different taxonomy levels of neuronal cell types in the adult mouse brain. The resulting complete whole-brain neuronal cell type atlas may serve as a detailed visualization of brain structures and functions, enabling insights into the organization of neural circuits, developmental processes, and the underlying mechanisms of neurological diseases.

Results

Multimodal dMRI-ABC Atlas-LSM integration and prediction framework at the CCFv3 space

We developed a multimodal dMRI-ABC Atlas-LSM integration and prediction framework at the CCFv3 space (Fig. 1): Firstly, high-resolution (45 µm isotropic) ex vivo dMRI data were acquired and modeled using DTI, DKI, NODDI, and SANDI, and were registered to the CCFv3 space to enable cross-modal alignment. Secondly, we trained a supervised multi-diffusion-MRI fusion network (MFNet, Fig. 1A) incorporating a three-level taxonomical classification scheme and a co-localization-based brain division stratification strategy (Fig. 1C), leveraging the hierarchical and region-specific structure of the cell type dataset³ (Fig. 1B). Training and validation were performed using individual ABC Atlas neurons as ground truth, with cell type labels integrated with spatially aligned diffusion MRI features. Model performance was evaluated using leave-p-groups-out cross-validation. Thirdly, we examined the correlation between dMRI parameters and spatial transcriptomics from the ABC Atlas (Fig. 1D). Finally, neurons were identified from 3D NeuN-stained (staining for neuronal nuclei) LSM by segmenting NeuN-positive voxels (Supplementary Fig. 1), and then the trained model was applied to NeuN-positive voxels in LSM to perform neuronal cell type classification, enabling the generation of a whole-brain neuronal cell type atlas in CCFv3 space at 10 µm isotropic resolution (Fig. 1E).

Fig. 1. Overview of the multimodal dMRI-ABC Atlas-LSM integration and prediction framework at the CCFv3 space.

A Neuronal cell type classification model by MFNet. All dMRI datasets were included in each of the DTI, DKI, NODDI, and SANDI models. B The hierarchical tree of ABC Atlas with level I: Neuron Type, level II: Neighborhoods, and level III: Classes. C Overview of the model training and validating process at CCFv3-2D. D Correlation between a representative dMRI metric (Fsoma, blue box) and ST (Slc17a7, green box), both registered in CCFv3 space (CCFv3 label boundaries overlaid) E Overview of the generation of volumetric whole-brain neuronal cell type atlas at CCFv3 space. The cell type atlas is color-coded by neuron type, cell neighborhood, and cell class, shown on example orthographic views. Abbreviations: ABC Allen Brain Cell, CB cerebellum, CCFv3 Allen Mouse Brain Common Coordinate Framework, CTX cerebral cortex, DKI diffusion kurtosis imaging, dMRI diffusion magnetic resonance imaging, DTI diffusion tensor imaging, HB hindbrain, LSM light-sheet microscopy, MB midbrain, MFNet multi-diffusion-MRI fusion network, MRI magnetic resonance imaging, MY medulla, NODDI neurite orientation dispersion and density imaging, PAL pallidum, SANDI soma and neurite density imaging, ST spatial transcriptomics.

A multi-modality brain imaging data (Supplementary Table 1), including dMRI maps with fourteen different image contrasts (Fig. 2B, C) and 3D LSM images (Fig. 2A), has been integrated into CCFv3 space. The alignment accuracy that was measured by the Dice similarity coefficient of each pair of spatially matched coronal slice (n = 1320) between dMRI and CCFv3 is 0.96 ± 0.06, and between 3D LSM and CCFv3 is 0.92 ± 0.09. The robust alignment enabled accurate downstream analyses and yielded a multimodal atlas in a common coordinate framework.

Fig. 2. CCFv3 space integrated dMRI and 3D LSM.

A NeuN image overlaid with an ABA label. Scale bar = 2 mm. B B₀ image overlaid with ABA label. C dMRI maps, including metrics from DTI (in red box), NODDI (in yellow box), DKI (in green box), and SANDI (in blue box) models. See Supplementary Fig. 2 for the dMRI maps in native space. Abbreviations: ABA Allen Brain Atlas, DKI diffusion kurtosis imaging, DTI diffusion tensor imaging, LSM light-sheet microscopy, MRI magnetic resonance imaging, NODDI neurite orientation dispersion and density imaging, SANDI soma and neurite density imaging.

Whole-brain neuronal cell type prediction evaluation

Using the common coordinate framework (CCF)-integrated dMRI, we predicted the spatial distribution of neuronal cell types across the whole mouse brain. MFNet outperformed baseline deep learning models, a multi-layer perceptron (MLP) (Supplementary Table 2). Division-specific training further improved performance compared to a unified model (Supplementary Fig. 3). Within MFNet, the multi-scale convolutional neural network (CNN) provided the primary predictive signal, the Transformer supplied complementary contextual information, and cross-modal fusion was essential for achieving these gains, particularly in cell type classification. Training and validation loss curves, together with performance metrics, confirmed that MFNet is both effective and generalizable (Supplementary Fig. 4).

Figure 3 shows the prediction results in the hippocampal formation (HPF), isocortex, and the whole brain. In HPF, the weighted average prediction accuracy is 88.8 ± 6.3% for the excitatory (E) neuron type (Fig. 3A, left), 94.8 ± 0.01% for the inhibitory (I) neuron type (Fig. 3B, left). In the isocortex, the weighted average prediction accuracy is 87.7 ± 0.9% for the excitatory neuron type (Fig. 3A, right), 94.8 ± 0.2% for the inhibitory neuron type (Fig. 3B, right). Across the whole brain, MFNet predicted cell neighborhoods (Fig. 3C) better than cell classes (Fig. 3D); the weighted average accuracy is 97.8 ± 2.9% or 98.6 ± 2.0% for E or I neighborhood predictions (Figs. 3C) and 88.6 ± 3.9% or 82.6 ± 10.9% for E or I class predictions (Fig. 3D). Overall, there is no significant difference between the performance of MFNet in excitatory and in inhibitory neuron types (Fig. 3E).

Fig. 3. Neuronal cell type prediction validation.

A, B Neuron type prediction results on excitatory (A) and inhibitory (B) neuron types in the whole brain, HPF (zoomed left), and isocortex (zoomed right) regions. C, D Cell neighborhoods (C) and classes (D) prediction results for excitatory and inhibitory neurons in the whole brain from each testing slice. E Confusion matrices for excitatory neuron type prediction, inhibitory neuron type prediction, excitatory cell class prediction, and inhibitory cell class prediction. Bar graphs on the right and bottom of the confusion matrices indicate the corresponding cell neighborhoods. Dashed lines indicate true positive values. F Bar graph comparing the weighted average prediction accuracy and F1 score in MFNet-dMRI (in blue bars) and MFNet-CCF structure (in red bars). Statistical tests used paired t-tests. Error bars represent the standard deviation. n = 5 biologically independent testing slice samples. Statistical significance is indicated by FDR-corrected p-values shown between the compared groups. The results shown in this figure were obtained using MRI data from the high b-value protocol. See results using the high angular and spatial resolution protocols in Supplementary Figs. 6 and 7. Abbreviations: ABC Allen Brain Cell, CCF common coordinate framework, dMRI diffusion magnetic resonance imaging, E excitatory neuron, HPF hippocampal formation, I inhibitory neuron, MFNet multi-diffusion-MRI fusion network.

To investigate the importance of dMRI over anatomical locations in predicting neuronal cell types, we compared the performance of MFNet using dMRI or CCF structures (i.e., baseline feature, as the ABC Atlas cell types are defined based on their location in the CCF broad region). From MFNet-CCF structure (Fig. 3F, red bars) to MFNet-dMRI (Fig. 3F, blue bars), the performance has increased substantially in neuron type predictions (+15.1% or +12.8% on E or I accuracy and +22.4% or +34.7% on E or I F1 score) and excitatory cell class predictions (+17.6% on accuracy). The most significant increases occurred in the ventricular systems (VS), pallidum (PAL), fiber tracts, and unassigned regions. In contrast, they performed similarly in E/I neighborhood and inhibitory cell class predictions. We also compared the performance of MFNet with and without the cell coordinates and found that excluding the coordinates caused 12.4% or 24.3% decreasing on E or I class prediction only (Supplementary Fig. 5). Furthermore, we evaluated the robustness of MFNet for image acquisition setting dependency by training with high angular and spatial resolution protocols. MFNet showed a consistently high degree of prediction accuracy (Supplementary Figs. 6 and 7).

Diffusion MRI distinguishes neuronal cell type through co-localization-based division-stratification

Diffusion MRI contained sufficient microstructural information to classify neuronal cell types using a co-localization-based division stratification framework. At the whole-brain level, U-MAPs exhibited partially overlapping clusters across neuronal cell types (Fig. 4A). In contrast, U-MAPs computed within individual brain divisions showed markedly reduced overlap and more distinct spatial clustering, particularly for neighborhood- and class-level classifications (Fig. 4B).

Fig. 4. Co-localization-based division stratification enhances diffusion MRI discrimination of neuronal cell types.

A U-MAP visualizations of diffusion MRI-derived feature representations for neurons across the whole brain, color-coded by division–excitatory subtype, division–inhibitory subtype, neuron type, cell neighborhood, and cell class. B U-MAP visualizations of diffusion MRI-derived feature representations for neurons within excitatory and inhibitory subtype divisions, color-coded by neuron type, cell neighborhood, and cell class. C Violin plots showing distributions of the SANDI-derived metrics Fsoma (left) and Fneurite (right) across neuron types, cell neighborhoods, and cell classes.

These findings indicated that stratifying brain regions based on excitatory–inhibitory neuronal co-localization patterns (Supplementary Fig. 8) enhanced the separability of neuronal populations and suggested that regional microstructural heterogeneity captured by diffusion MRI is critical for resolving neighborhood- and class-specific classification. Furthermore, feature importance analysis consistently identified SANDI-derived metrics, particularly Fsoma and Fneurite, as the most informative predictors (Supplementary Fig. 9). Both metrics exhibited clear and systematic differences across neuronal types, neighborhoods, and classes (Fig. 4C), further supporting their sensitivity to underlying cellular composition and organization.

Diffusion MRI correlates with spatial transcriptomic signatures of oligodendrocyte, axonal, and neuronal marker genes across the whole brain

Diffusion MRI exhibited strong associations with spatial transcriptomic profiles of oligodendrocyte/myelin, axonal organization, and neuronal marker genes across the whole brain. As shown in Fig. 5A, among the 500 MERFISH-sequenced genes, oligodendrocyte-associated genes (e.g., Klk6, Cdkn1a, Col23a1, Car4) demonstrated robust positive correlations with DKI-derived metrics (AK, KFA, MKT, RTK), as well as FA, NDI, and Fneurite, and negative correlations with MD, RD, ODI, Fextra, and Fsoma. Additional correlations with axonal and vascular-related genes (e.g., Cntnap3, Kcnh8, Gfra1) further support that diffusion MRI captures integrated signatures of cellular composition and tissue microenvironment.

Fig. 5. Correlation between dMRI metrics and spatial transcriptomics.

A, B Heatmaps of PCCs between dMRI metrics and the spatial transcriptomics of the top 30 significant genes from the 500-MERFISH (A) and neuron type marker genes (B). Statistical tests used Pearson’s correlation analysis. n = 686 biologically independent ROI samples. Statistical significance is indicated by FDR-corrected SA-corrected p-values. * Corrected p-values < 0.05, ** corrected p-values < 0.01. C Heatmap comparing the normalized expression energy of neuron type marker genes and the normalized dMRI values distributed in CCF regions. The row represents one CCF label region. Bar graphs on the left show the broad CCF regions. D Linear regression plots of four independent variable-dependent variable pairs: Slc17a7 (x-axis)-Fsoma (y-axis), Aldh1a1 (x-axis)-Fsoma (y-axis), Slc17a7 (x-axis)-Fneurite (y-axis), Aldh1a1 (x-axis)-Fneurite (y-axis). Statistical tests used Pearson’s correlation analysis. n = 686 biologically independent ROI samples. Statistical significance is indicated by FDR-corrected SA-corrected p-values shown in each plot. The PCC r-values and 95% confidence intervals (in grey shadows) were also shown in each plot. E Representative slices of the spatial transcriptomics of neuron type marker gene Slc17a7 (row 1) and Aldh1a1 (row 2), as well as Fsoma (row 3) and Fneurite (row 4) maps. Black and blue arrows indicate the isocortex and HPF, respectively. Abbreviations: DKI diffusion kurtosis imaging, DTI diffusion tensor imaging, NODDI neurite orientation dispersion and density imaging, PCC Pearson’s correlation coefficient, SANDI soma and neurite density imaging.

Neuronal marker genes, including excitatory (Slc17a7, Chat) and inhibitory (Aldh1a1, Gad2, Htr3a) populations, were strongly associated with SANDI-derived metrics, particularly Fsoma and Fneurite (Fig. 5B). Consistently, major neuronal markers (Slc17a6, Slc17a7, Slc32a1, Gad1, Gad2, Aldh1a1) exhibited spatial patterns closely aligned with dMRI-derived maps (Fig. 5C), with Slc17a7/Aldh1a1–Fsoma/Fneurite relationships among the most significant (Fig. 5D). Regionally, in the isocortex and hippocampal formation (Fig. 5E), high Slc17a7 expression and lower Aldh1a1 levels corresponded to elevated Fsoma and reduced Fneurite. Compared with CCF-based structural correlations, diffusion MRI identified a greater number of significant gene associations, with notably stronger effects for inhibitory neuron markers (Supplementary Fig. 10).

MFNet Atlas covers the whole brain volume and is more complete compared to the ABC Atlas

Using MFNet, we generated a whole-brain mouse neuronal cell type atlas, termed the MFNet Atlas (Fig. 6), in CCFv3 space at 10 µm isotropic resolution, with labels across all three taxonomic levels. Compared to the ABC Atlas, the MFNet Atlas exhibits comparable anatomical and spatial detail (Fig. 6A), with average similarity scores across 35 spatially matched slices of 0.99 ± 0.01 or 0.98 ± 0.02 for E or I neuron type, 0.98 ± 0.02 or 0.98 ± 0.02 for E or I neighborhood, and 0.99 ± 0.01 or 0.93 ± 0.05 for E or I class; observed discrepancies (arrows in Fig. 6A) primarily reflect differences in neuron segmentation (Supplementary Fig. 11A).

Fig. 6. MFNet Atlas, a complete, whole mouse brain neuronal cell type atlas at CCFv3 space with 10 μm isotropic resolution.

A Comparisons between one representative slice from the MFNet Atlas and the corresponding ABC Atlas. The two atlases are color-coded by neuron type (row 1), cell neighborhoods (row 2), and classes (row 3), with excitatory, inhibitory, and all neurons shown separately. Arrows indicate the major mismatched areas. B Comparisons between three representative slices of cell class from the MFNet Atlas and the corresponding ABC Atlas that were damaged. Arrows in the enlarged regions indicate the damaged areas. See comparisons on more slices in Supplementary Fig. 11. C Comparisons between three representative slices of inhibitory neuron types from MFNet Atlas and corresponding interpolated ABC Atlas with reference Gad1 and Gad2 ISH images (Allen Mouse Brain Atlas, mouse.brain-map.org). Arrows indicate the spatially matched areas, including cerebellar cortex (orange arrows), isocortex (dark blue and purple arrows), and hippocampus (green and light blue arrows). D Representative slices of cell type from axial and sagittal views from the MFNet Atlas and the corresponding ABC Atlas. Abbreviations: ABC Allen Brain Cell, MFNet multi-diffusion-MRI fusion network.

The MFNet Atlas also offers several key advantages over the ABC Atlas. First, 34% of slices in the ABC Atlas were damaged, resulting in missing cell type information (arrows in Fig. 6B and Supplementary Fig. 11B), whereas MFNet recovers these regions to provide a complete atlas. Second, MFNet achieves higher accuracy than the interpolated ABC Atlas; for example, inhibitory neuron distributions in regions highlighted by arrows (Fig. 6C) show greater concordance with Gad1 and Gad2 (two common inhibitory neuron marker genes) in situ hybridization (ISH) ground truth. Third, MFNet provides whole-brain coverage, including 1320 coronal, 1140 sagittal, and 800 axial views, compared to 53 coronal views in the ABC Atlas. MFNet generates a continuous 3D volume, whereas the ABC Atlas is limited to discrete 2D slices (Fig. 6D, Supplementary Movie 1). In summary, the MFNet Atlas delivers more complete cell type information than the ABC Atlas, establishing a high-resolution resource for investigating the brain at the whole-volume level.

Discussion

Understanding the structure and function of the brain remains one of the most challenging topics in neuroscience. Significant efforts have been devoted to refining high-resolution imaging techniques for more detailed anatomical insights^38,39, while simultaneously other efforts have focused on creating cellular taxonomies to elucidate underlying mechanisms^40,41. Integrating imaging techniques with taxonomy findings will significantly enhance the understanding of the brain at different scales⁴². In this study, we integrated MRI images and LSM into the Allen Mouse Brain Common Coordinate Framework, providing an atlas with both structural and functional information for the whole mouse brain. Compared to the previous studies^43,44 that were solely on the basic DTI model, this integrated atlas incorporates both DTI and advanced diffusion models, enabling a more precise characterization of tissue microstructure. Furthermore, it provides the neuroimaging community with a high-resolution multi-parameter MRI template in the CCFv3 space with detailed cell types across the whole brain.

The predictive power of dMRI for cell types arises from its sensitivity to anatomical and cellular microstructural organization. dMRI detects microstructural changes, including fiber orientations in white matter⁴⁵, cortical soma size and density, and fine cortical and olfactory bulb structures¹². High-resolution dMRI also reveals neuroinflammation⁴⁶, neurodegeneration¹⁶, glial activation⁴⁷, and cellular-level tissue features such as cell morphology, membrane integrity, and local cell density^48,49. dMRI differentiates intracellular and extracellular space as intracellular molecules generally diffuse much more slowly than extracellular molecules⁵⁰, leading to increased signal intensity in regions of high cell density. Beyond microstructure, diffusion metrics such as NDI and ODI correlate with gene expression profiles, linking MRI signals to molecular pathways^51,52. Spatial imaging-transcriptomics further confirms that regional diffusion markers reflect underlying gene expression patterns, highlighting dMRI as a non-invasive probe of cellular and genetic features²¹. Our previous study also demonstrated that dMRI metrics are associated with the genes involved in brain development⁵³.

MRI outperforms CCF-based anatomical labels in distinguishing neuronal cell types. While CCF regions exhibit regional specificity and contain distinct neuronal populations, many cell types—including glutamatergic and GABAergic neurons—colocalize within the same regions, limiting the discriminatory power of anatomical labels alone. Differentiating these colocalized cell types, therefore, requires higher-order structural and functional information, which MRI provides at high resolution. In contrast, high-resolution 3D LSM primarily captures spatial distributions of neurons^37,54. To leverage the strengths of both modalities, we assigned complementary roles to LSM and MRI. NeuN-stained LSM was used to segment NeuN-positive voxels that defined the spatial locations of neurons, while diffusion MRI provided co-registered microstructural features after both datasets were aligned in the CCFv3 space. The trained MFNet then mapped these spatially matched dMRI features onto the segmented LSM voxels to predict neuronal cell types. Integrating LSM-based spatial segmentation with MRI-based cell type prediction enabled the generation of a high-resolution whole-brain neuronal cell type atlas.

Notably, SANDI parameters—particularly Fsoma and Fneurite—ranked as the top two features among all dMRI metrics for predicting cell types. This is consistent with the design of the SANDI model, a biophysical framework specifically developed to estimate cellular and neurite densities¹⁸, rendering it highly sensitive to cellular and molecular differences between neuronal types. More specifically, Fsoma describes the fraction of soma that produces neurotransmitters, while Fneurite would signal the presence of a prominent apical dendrite on cortical pyramidal neurons¹⁸. Given that neuronal cell types are defined by their neurotransmitter types in the ABC Atlas, the high predictive power of Fsoma and Fneurite can also be attributed to their strong correlation with the spatial transcriptomics of neurotransmitter marker genes. Slc17a7 encodes Vesicular Glutamate Transporter 1 (VGLUT1), a key protein for Glutamate neurotransmission^55,56. Aldh1a1 is crucial both for GABA metabolism and for enabling dopaminergic neurons to co-release GABA^57,58. As shown in Fig. 5, both Fsoma and Fneurite correlate strongly with the spatial expression of these marker genes, supporting their role in distinguishing neuronal cell types.

Beyond neuronal marker genes, dMRI captures spatial transcriptomic signatures of oligodendrocyte and axonal genes across the whole brain, providing molecular interpretability of microstructural imaging. Strong associations with oligodendrocyte-related genes indicate sensitivity to myelin organization and tissue complexity, while robust correlations with axonal genes suggest that dMRI reflects integrated features of cellular composition and tissue microenvironment. Notably, a substantially larger number of genes were significantly correlated with dMRI than with CCF structure, and dMRI exhibited stronger associations with inhibitory neuronal markers. Although dMRI–gene correlations are partly constrained by shared anatomical organization, dMRI explains considerably more transcriptomic variation beyond structural effects alone.

Leveraging these insights, we generated a 3D, whole-brain MFNet Atlas with 10 µm isotropic resolution in CCFv3 space. Existing mouse brain cell type atlases offer high in-plane resolution but limited axial resolution (200 µm)^3,6, resulting in incomplete coverage across many brain regions. Leveraging the high isotropic resolution of diffusion MRI, we generated a whole-brain, MRI-guided neuronal cell type atlas with 10 µm axial resolution, fully consistent with CCFv3⁵⁹. The MFNet Atlas also provides micrometer-scale in-plane resolution (10 µm), comparable to the average size of a mouse brain cell⁶⁰. A key advantage of MFNet is its ability to accurately annotate cell types in regions where the ABC Atlas is incomplete, including damaged or intervening slices. Beyond completeness, the MFNet Atlas may be applied to study brain development, evolution, function, and pathology^3,61,62, facilitating detailed investigation of molecular and anatomical architecture, characterization of morphological and physiological properties, and precise spatial mapping of distinct neuronal populations.

Diffusion MRI-based cell type prediction has broad research and clinical utility. Traditionally, characterizing cell types requires labor-intensive approaches, including transcriptomic profiling, molecular identification, developmental analyses, and constructing conceptual frameworks for cell type definition⁴. In contrast, deep learning–based prediction models can infer cell types without extensive domain expertise, and high-resolution MRI enables cellular-level microstructural information to be captured. Given that diffusion MRI is already widely applied in studies of brain disorders^16,63, integrating cell type prediction into existing imaging workflows could provide additional cellular-level insights, enhancing both clinical and translational research. The MFNet framework can also be used to quantify the excitatory/inhibitory (E/I) ratio, which is critical for maintaining proper brain function and network stability⁶⁴. Dysregulation of this balance has been implicated in neurodevelopmental disorders⁶⁵ and is also an early hallmark of AD progression in mouse models⁶⁶. Therefore, using the MFNet prediction framework may assist early detection and treatment assessment of neurological disease.

Several limitations remain. First, the number of mice was limited due to lengthy image acquisition times. Second, the prediction model currently cannot be applied to non-neurons, which are widely distributed, highly colocalized, and less region-specific, leading to similar MRI values that are difficult to differentiate. Moreover, because the native dMRI resolution (45 μm isotropic) is substantially coarser than MERFISH resolution (167 nm)⁶⁷, multiple adjacent MERFISH cells may share the same registered dMRI values, limiting classification performance. The model, therefore, has reduced performance for cell types with limited data points or substantial spatial overlap with other cell populations. Third, the accuracy of MFNet partly depends on prior knowledge of the cellular composition characteristic of the genotype. For example, the model infers cell type proportions from reference labels, which may be less reliable in slices that do not match the ABC Atlas data or that originate from different genotypes. In addition, the brain-division stratification step relies on predefined degrees of cell type co-localization, information that may be unavailable or inaccurate for some genotypes. Fourth, although diffusion time allows water to travel in the micrometer range, making this technique sensitive to cellular-level organization rather than just macroscopic tissue morphology, the imaging voxel size remains coarser than the 10 μm isotropic CCFv3 atlas space, such that partial volume effects and interpolation artifacts may persist in the generated atlas. Further increasing spatial resolution could enhance the accuracy of the predicted cell type atlas. Last, neuron density in the MFNet Atlas is higher than in the ABC Atlas, primarily because MFNet utilizes neuronal segmentation directly from 3D NeuN images, whereas the ABC Atlas has excluded low-quality cells to improve sequencing accuracy³. Future improvements could include enhancing the spatial resolution of dMRI, refining models for non-neuronal cell types, applying unsupervised or transfer learning approaches to blank slices, employing spatially independent models, and integrating sparse cell distributions into the MFNet Atlas.

In this study, we integrated high-resolution diffusion MRI and three-dimensional light-sheet microscopy within the Allen Mouse Brain Common Coordinate Framework to generate a comprehensive neuronal cell type atlas. A deep learning-based network accurately predicted neuronal cell types across the entire mouse brain, demonstrating that dMRI parameters capture key cellular properties and strongly correlate with the spatial distribution of neurons. We produced a 3D whole-brain neuron atlas with high in-plane and through-plane resolution. Collectively, these results highlight the power of advanced imaging techniques for mapping brain structure and function and provide a high-resolution resource for understanding the brain at cellular and molecular scales.

Methods

Animal study

Six adult C57BL/6 J (B6, wild type) male mice (Jackson Laboratory, Bar Harbor, ME) were selected for MR imaging, with three protocols: high b-value protocol (n = 1), high angular resolution protocol (n = 1), and high spatial resolution protocol (n = 4). All animal experiments were approved by the Duke University Institutional Animal Care and Use Committee (Approval code A226-17-09). We have complied with all relevant ethical regulations for animal use. Mice were euthanized with Nembutal and perfusion-fixed with a 1:10 mixture of ProHance-buffered (concentration of 10%, Bracco Diagnostics, Princeton, NJ) formalin. The specimens were immersed in buffered formalin for 24 h and then transferred to a 1:200 mixture of ProHance-saline solution (concentration of 0.5%) to shorten the T1 (to approximately 100 ms) and reduce scan time. Ex vivo MRI was chosen over an in vivo study to achieve higher anatomic resolution and better tissue contrast, which improves the detection of microstructural changes^68,69 and to maintain consistency with ex vivo spatial transcriptomics data.

MRI acquisition

MRI data were acquired using a 9.4 T Oxford MRI scanner (Oxford Instruments, Abingdon, UK) with the maximum gradient amplitude of 2000 mT/m along each axis. Diffusion MRI measurements were performed with a modified three-dimensional diffusion-weighted spin-echo pulse sequence incorporating k-space undersampling. The readout dimension was fully sampled, whereas undersampling was applied to both phase-encoding dimensions using a sparsity-based acquisition strategy described previously³⁸. An acceleration factor of 4 was employed. Diffusion gradient directions were generated according to the scheme developed by Koay et al.⁷⁰ to ensure uniformity within each shell and across all shells.

The dMRI undersampled datasets were acquired with three protocols (Supplementary Table 3): (1) high b value protocol (n = 1): 16 shells b value = 1000, 2000, 3000, 4000, 5000, 6000, 7000, 8000, 9000, 10,000, 11,000, 12,000, 13,000, 14,000, 15,000, 16,000 s/mm² at 45 μm isotropic resolution. (2) high angular resolution protocol (n = 1): 8 shells b value = 1000, 2000, 3000, 4000, 5000, 6000, 7000, 8000 s/mm² at 50 μm isotropic resolution. (3) high spatial resolution protocol (n = 4): 4 shells b value = 1000, 4000, 6000, 8000 s/mm² at 32 μm isotropic resolution. Animals acquired under the same protocol were averaged (for protocol 3 only), and each resulting averaged dMRI dataset was included individually in the MFNet model training and validation. Because the SANDI model performance depends on high b values¹⁸, only the high b value protocol dataset was used for correlation with spatial transcriptomics and for generating the whole-brain cell type atlas in CCFv3 space.

The reconstruction procedure for undersampled k-space data is by minimizing^38,71:

$$f\left(x\right)={\parallel Fx-y\parallel }_2^2+{\lambda }_1{\parallel \Psi x\parallel }_1+{\lambda }_{2}TV\left(x\right)$$ 1

Where $x$ denotes the reconstructed image, $y$ represents the acquired k-space measurements, F is the Fourier transform operator, $\Psi$ is the sparse-transform basis, ${\lambda }_1$ and ${\lambda }_2$ are regularization parameters, and $TV$ denotes the total variation regularization term. For the present study, ${\lambda }_1$ was set to 0.006 and ${\lambda }_2$ was set to 0.0012 to provide robust reconstruction performance with the Daubechies wavelet transform⁷¹. With the acceleration factor of 4, the multi-shell diffusion MRI dataset was acquired in about 48 h, whereas an equivalent fully sampled acquisition would have required 192 h. Image reconstruction was performed on a Dell Cluster. Initial Fourier transformation along the readout direction generated multiple two-dimensional image slices from each acquisition; the reconstruction workload was distributed across the computing cluster to enable parallel processing.

MRI processing

For ex vivo mouse brain diffusion MRI, raw diffusion-weighted images were converted from NIFTI to MRtrix3 (RRID:SCR_024123) format while preserving the diffusion gradient table. Gradient orientations were verified and corrected. To improve signal quality at high spatial resolution, thermal noise was reduced using MP-PCA denoising, followed by Gibbs ringing suppression. A mean b0 image was extracted and corrected for low-frequency intensity inhomogeneity using N4 bias field correction implemented in Advanced Normalization Tools (ANTs, v2.0 RRID:SCR_004757). Eddy-current distortions and localized signal outliers affecting individual slices within the reconstructed 3D diffusion volumes were corrected using FSL through MRtrix3, with outlier replacement and linear slice-to-volume modeling enabled. Residual intensity nonuniformity in the full diffusion dataset was further corrected using ANTs-based bias correction with a brain mask. Gradient consistency was rechecked after preprocessing. The fully corrected diffusion dataset was then exported to NIFTI format for subsequent DTI, DKI, NODDI, and SANDI model fitting.

The DTI reconstruction method was performed for reconstruction using DSI Studio (“Chen” version, RRID: SCR_009557)⁷². From data with b values of 1000–4000 s/mm², the diffusion tensor was calculated. The scalar indices include fractional anisotropy (FA), mean diffusivity (MD), axial diffusivity (AD), and radial diffusivity (RD).

Diffusion Imaging in Python (Dipy, version 1.10.0, RRID:SCR_000029) was used to compute DKI metrics from data with b values of 1000–8000 s/mm². Because accurate kurtosis estimation is particularly sensitive to noise at high b-values, Self-Supervised Denoising via Statistical Independence (Patch2Self)⁷³ was used before fitting to the DKI model as a model-free denoising step that preserves diffusion signal characteristics. The DKI model was fitted using constrained linear least squares (CLS) with fit_method = “CLS” and cvxpy_solver = “MOSEK.” CLS fitting was selected to enforce physically plausible parameter bounds, particularly non-negative kurtosis estimates, thereby reducing biologically implausible values and spatial artifacts that can arise in low-SNR or high-b-value regimes. This constrained framework improves the robustness and interpretability of DKI-derived measures. The resulting parameters included axial kurtosis (AK), kurtosis fractional anisotropy (KFA), mean kurtosis tensor (MKT), and radial tensor kurtosis (RTK).

The NODDI toolbox for Matlab (v1.05, RRID:SCR_006826) was used to fit the diffusion MRI data with b values of 1000–8000 s/mm²⁷⁴. For the ex vivo NODDI model, a relatively small dot-compartment (the diffusion is restricted in all directions) was included for the fitting⁷⁵. Axial diffusivity (Din = 9 × 10⁻⁴ mm²/s) and isotropic diffusivity (Diso = 3.0 × 10⁻³ mm²/s) were fixed model parameters. The parameters derived from the NODDI model were the orientation dispersion index (ODI) and neurite density index (NDI).

SANDI Matlab Toolbox (latest release)¹⁸ was used to calculate SANDI metrics from data with all b values. Given that SANDI relies on multi-shell compartment fitting and benefits from explicit noise estimation, MP-PCA denoising (dwidenoise, MRtrix3) was performed prior to model fitting to generate both denoised data and an estimated noise map. The SANDI metrics recorded for further analysis were soma fraction (Fsoma), neurite fraction (Fneurite), radius of soma (Rsoma), and ratio of neurite (Rneurite).

The diffusion MRI images were registered to an MD template in CCFv3 space using ANTs⁷⁶. Briefly, each subject MD image was first aligned to the template using affine registration to correct for global differences in position, orientation, and scale, followed by nonlinear registration using the SyN algorithm to account for local anatomical variation. The estimated transformations were then applied to all diffusion MRI metrics, yielding spatially normalized maps in CCFv3 space for voxel-wise analyses.

Allen Mouse Brain Atlas processing

Allen Mouse Brain Atlas (AMBA)’s serial two-photon tomography (STP, a population average of 1,675 young adult mouse brains created from tissue autofluorescence)⁵⁹ image templates in Allen Mouse Brain Common Coordinate Framework version 3 (2020 version, RRID:SCR_020999) with 10 μm isotropic resolution (CCFv3-3D) and resampled CCF mapped coordinates with 10 μm in-plane and 200 μm through-plane resolution (CCFv3-2D), and cell type labels in CCFv3-2D were downloaded from the ABC Atlas (https://portal.brain-map.org/atlases-and-data/bkp/abc-atlas) from BICCN (RRID:SCR_015820) and converted to NIFTI format⁵⁹.

To preserve consistent voxel dimensions between the CCFv3-3D (x × y × z = 1140 × 800 × 1320) and CCFv3-2D (x × y × z = 1100 × 1100 × 76) spaces, we first resampled the CCFv3-2D STP image along the z-axis from 76 to 1320 slices using linear interpolation, generating an intermediate reference volume with matched spatial resolution across all three axes. The CCFv3-3D STP image was then aligned to this upsampled reference using ANTs’ affine registration, yielding the transformation chain CCFv3-3D → affine transform → upsampled CCFv3-2D space (Supplementary Fig. 12). This transformation was subsequently applied to enable consistent spatial mapping across modalities.

Histology image processing

The 3D NeuN staining (staining for neuronal nuclei, Fox3/NeuN, mouse monoclonal, Cat# MCA-1B7, RRID:AB_2572267, resolution = 1.8 × 1.8 × 4 μm) images (n = 3) for adult male B6 mice were downloaded from the Duke Mouse Brain Atlas (DMBA, https://civmimagespace.civm.duhs.duke.edu/)³⁷ and converted to NIFTI format. The Glutamate Decarboxylase 1 (Gad1, inhibitory GABAergic neuron marker gene) and Glutamate Decarboxylase 2 (Gad2, inhibitory GABAergic neuron marker gene) in situ hybridization (ISH) staining images for adult male B6 mice were downloaded from Allen Mouse Brain Atlas (https://mouse.brain-map.org/. Experiment # 479 and # 79591669, respectively.

The LSM NeuN data were not used for deep learning model training or validation. Instead, NeuN-positive voxels identified from LSM were used solely for neuronal segmentation during the generation of the whole-brain neuronal cell type atlas in CCFv3 space. The ISH data were used as ground truth for comparison between the generated atlas and interpolation-based results.

Neuron cells were segmented from 3D NeuN images at their native space. Labkit⁷⁷ automatic segmentation was used for each image slice (Fig. 1A and Supplementary Fig. 1). The segmentation result images were saved as.tiff format and were pre-processed by firstly averaging across animals and secondly binarizing using a locally adaptive threshold.

The 3D NeuN images were registered to CCFv3-3D space using ANTs’ affine and syn image registration. The registration was refined manually, guided by anatomical landmarks such as hippocampal subfields and cortical layers. The transformation matrices were applied to their corresponding pre-processed segmentation result images.

Mouse whole-brain cell type prediction model

The mouse whole-brain transcriptomic cell type atlas consisted of whole adult mouse brain MERFISH spatial transcriptomics and taxonomy of cell type dataset, which were exported from Allen Brain Cell Atlas (ABC Atlas, https://portal.brain-map.org/atlases-and-data/bkp/abc-atlas). The MERFISH spatial transcriptomics dataset included the coordinate information of each cell centroid from the whole mouse brain. Such coordinates were provided in CCFv3-2D space. The taxonomy dataset integrated several whole-brain single-cell RNA-sequencing (scRNA-seq) datasets and contains cell type clusters in a hierarchical structure. Four hierarchical levels were selected in this study (from top to bottom): (1) neurons versus non-neurons; (2) three neuron types—excitatory (E; glutamatergic (Glut) neurons), inhibitory (I; GABAergic (GABA), GABA–glycinergic (GABA-Glyc), and glutamatergic–GABAergic (Glut-GABA) neurons), and other neurotransmitter types (including cholinergic (Chol), dopaminergic (Dopa), histaminergic (Hist), noradrenergic (Nora), and serotonergic (Sero) neurons); (3) 7 cell neighborhoods; and (4) 34 cell classes (Fig. 1C). All levels followed the hierarchical organization defined in the ABC Atlas. Non-neuronal cell types and other neuron types were deleted from the data as they are out of the scope of the current topic.

A total of 53 MERFISH coronal section slices (MERFISH-C57BL6J-638850.05–MERFISH-C57BL6J-638850.67 Reconstructed Coordinates) were used, covering the whole anterior-to-posterior brain volume. We selected one section slice from each fifth of the brain volume as a testing section slice (MERFISH-C57BL6J-638850.14, MERFISH-C57BL6J-638850.26, MERFISH-C57BL6J-638850.36, MERFISH-C57BL6J-638850.51, and MERFISH-C57BL6J-638850.58), with the remaining section slices used as training data. This will ensure that the testing data is representative of the entire dataset and independent of the training data.

Heterogeneous colocalization levels were observed in the spatial distribution of E/I neuron types across particular brain regions³. In alignment with this observation, we developed a division-specific neuron type prediction model (Fig. 1D) to separately train the whole brain into three brain divisions for each E/I subtype according to the degree of colocalization (highly/moderately/lightly colocalized by the other subtype, corresponding to Division 1–3). Each neuron from the whole brain was assigned to one division based on its CCFv3 labels³. The testing dataset was separated in the same way. At the end, the divisions were combined into a whole brain for cell neighborhood and class classification.

For model development and statistical analysis, including MFNet training, testing, and correlation with spatial transcriptomics, dMRI data were analyzed in the CCFv3-2D space. We observed that the image resolution of MERFISH imaging (167 nm⁶⁷) is significantly higher than that of the CCFv3-2D image template. To avoid overlapping data, we adjusted the xy-plane image resolution of all images acquired in this study by upsampling each 2D plane by a factor of 10. We then removed the overlapping data (~1.0% of the population) that retained the same xyz coordinates after image upsampling.

Each training/testing dataset contained a table sized n × 21, where the nth row corresponded to the nth cell centroid of the ABC Atlas cells within the given division; column #1–14 corresponded to the values of 14 spatially matched dMRI metrics; column #15–21 corresponded to the z, x, y coordinates, cell class, cell neighborhood, neuron type, and brain division, respectively. In training by CCF structures, columns #1–14 were substituted by the grayscale values of the AMBA’s STP image. Coordinates were included as predicting variables as they offered spatial information that is directly correlated to brain anatomy⁵⁹. Each predicting variable was rescaled to [0, 1] for data normalization. The cells with empty cell type labels, or outside of the brain, or mismatched during image registration, were deleted.

We developed MFNet, a multi-diffusion-MRI fusion network for cell type prediction that integrates convolutional and transformer-based components (Fig. 1D). The 14 dMRI-derived metrics were incorporated as complementary representations of tissue microstructure rather than independent predictors, and their joint inclusion enhances model expressiveness. The architecture consists of three modules—multi-scale convolutional feature extraction (Fig. 1D, in green box), enhanced transformer encoding (Fig. 1D, in orange box), and cross-modal feature fusion (Fig. 1D, in red box)—followed by a classifier (Fig. 1D, in blue box). In detail, the convolutional module uses parallel 1D convolutions (kernels 3, 5, 7) to capture multi-scale features. Outputs are concatenated, refined with depthwise-separable and pointwise convolutions⁷⁸, and modulated by a channel-attention mechanism to enhance discriminative features⁷⁹. A residual connection stabilizes training; The transformer module captures long-range dependencies⁸⁰ by embedding inputs via convolutional projection with positional encodings⁸¹, followed by a multi-layer transformer encoder. Global self-attention and a parallel local-attention branch extract complementary contextual features⁸², which are combined via adaptive average pooling; cross-modal fusion aligns CNN and transformer⁸³ features in a shared latent space using cross-attention, with a gating network adjusting their relative contributions⁸⁴. The fused representation is normalized and fed through fully connected layers with dropout to generate final cell type predictions.

The model was trained using a two-stage, three-level classification framework. In the first stage, for each predefined brain division under the excitatory or inhibitory (E/I) subtype scheme, the model classified whether a given cell belonged to the excitatory or inhibitory subtype. To address division-specific class imbalance, we defined loss-weight sets

$$W_j=\left\{w_i,\mid ,i=1,2\right\},j\in \left\{1,2,3\right\}$$ 2

where $w_i$ represents the inverse prevalence of subtype $i$ within division $j$. Model optimization used a weighted cross-entropy loss,

$$\mathcal{L}\left(y,\widehat{y}\right)=-\sum _{i=1}^2{w}_i\cdot y_i\cdot \log \left({\widehat{y}}_i\right)$$ 3

with $y_i\in \left\{\mathrm{0,\; 1}\right\}$ denoting the one-hot ground-truth label and ${\widehat{y}}_i$ the predicted probability for subtype $i$. This formulation increases the influence of underrepresented sfthe predicted probability ubtypes and compensates for imbalance across highly, moderately, and weakly co-localized divisions. Division-specific probability thresholds were estimated using held-out ground-truth data and subsequently applied during inference. To preserve anatomical consistency, predicted probabilities were further calibrated using a Z-slice spatial prior that adjusted scores according to each slice’s baseline E/I ratio prior to thresholding.

In the second stage, for all cells across the whole brain previously assigned an E/I subtype, the model simultaneously predicted each cell’s neighborhood and class. The training process was conducted over 100 epochs for each division using an Adam optimizer with an initial learning rate of 1 × 10⁻⁴. To ensure reproducibility, a fixed random seed (42) was established across Python (v3.11), NumPy, and PyTorch (v2.6). All experiments were executed on an Intel I9-14900K CPU paired with an NVIDIA RTX 4090 GPU equipped with 24 GB of VRAM.

To validate the proposed model, we evaluated its performance by leave-p-groups-out validation using five independent slices as testing data. Classification performance was evaluated separately for the three classification levels. The classification metrics included weighted average accuracy and weighted average F1-score. Model selection relied solely on validation performance. We monitored training loss and validation accuracy throughout training. We used mild explicit regularization, including dropout with p = 0.3 in the classifier and L2 weight decay of 1 × 10⁻⁵, together with Batch Normalization. To measure the contribution of each module to the overall model, we ablated each component (Transformer, CNN, and fusion block, sequentially) on both neighborhood and class classification for excitatory and inhibitory neurons, and then the prediction accuracies after each ablation were recorded and compared. Meanwhile, we also compared the prediction accuracy on MFNet with MLP.

We performed a 2D U-MAP projection to visualize the spatial clustering of neuron types, cell neighborhoods, and classes. To avoid memory issues, we down-sampled the complete dataset to a maximum of 100 cells per cell type class. We then used all predicting variables as input to create UMAP projections (for neighborhood and class: n_neighbors = 6, min_dist = 0.5; for neuron type: n_neighbors = 6, min_dist = 0.2). The U-MAP projections were color-coded according to the corresponding cell types.

We ranked the predictive features based on their importance scores using the Analysis of Variance (ANOVA) filter. Features are ranked based on their F-statistics, enabling the selection of the most impactful ones. Feature ranking was done separately on neuron type, cell neighborhood, and cell class classifications.

MRI correlation tests with spatial transcriptomics

To examine the correlation between MRI and spatial transcriptomics, we performed a linear correlation test on the fourteen CCF-registered dMRI metrics and the spatial transcriptomic data of (i) neurotransmitter (NT) type marker genes for excitatory (i.e., Glutamatergic) and inhibitory (i.e.,GABAergic and Glycinergic) neurons; (ii) the 500-gene MERFISH spatial transcriptomic data. The NT type marker genes and the MERFISH spatial transcriptomics were obtained from Yao et al.³. Regions of interest (ROIs; n = 686), each corresponding to a distinct anatomical region in the CCFv3, were analyzed by averaging the values of all voxels within the region. The resulting regional mean value was treated as a single data point in the correlation analysis (Fig. 1E). Pearson’s correlation analysis was performed to assess linear associations between MRI and spatial transcriptomics. We recorded all the pairwise Pearson’s correlation coefficients (PCC) as observed PCCs.

Spatial autocorrelation (SA) in dMRI maps was corrected using BrainSMASH⁸⁵ to generate spatially autocorrelated surrogate maps (n = 1000 per dMRI map) whose spatial dependence matched that of the dMRI data (Supplementary Fig. 13). For each gene, we calculated the PCCs between its expression and these surrogate maps as the surrogate PCCs, forming a null distribution. For each gene-dMRI pair, the percentage of surrogate PCCs exceeding the observed PCC was calculated as the SA-corrected p-value. Multiple-comparison adjustment by Benjamini-Hochberg for false-discovery-rate (FDR) correction was then applied on the SA-corrected p-value, and the FDR-corrected SA-corrected p-value<0.05 was considered statistically significant.

To assess whether dMRI–gene correlations were stronger than those based on anatomical structure (CCF), we applied Steiger’s test for comparing dependent correlations. For each gene, Pearson correlations were computed between gene expression and (i) dMRI metrics and (ii) CCF structural measures across matched spatial locations. Because both correlations share the same gene expression variable, their dependency was accounted for by incorporating the correlation between dMRI and CCF maps. Two-sided p-values were obtained and corrected for multiple comparisons using the Benjamini–Hochberg FDR (q < 0.05). Genes with significant positive test statistics were considered to show stronger associations with dMRI than with CCF structure.

Complete whole-brain neuronal cell type atlas at CCFv3 space generation, validation, and evaluation

Using the MFNet, we generated a complete whole-brain neuronal cell type atlas (MFNet Atlas, Fig. 1F) in CCFv3 space. First, the CCFv3 registration transformation was subsequently applied to dMRI and LSM datasets in CCFv3-3D space to obtain co-registered dMRI maps and LSMs in the common reference space. Second, we applied the pre-processed and CCFv3-3D-registered neuron cell segmentation result images from 3D NeuN images to spatially matched dMRI maps to filter the neuron voxels. Next, we predicted firstly the E/I neuron type and then the cell neighborhoods and classes on these voxels. The MFNet-generated 3D whole-brain neuron atlas, including neuron type, cell neighborhood, and class labels, was saved as NIFTI files and made available through Zenodo (https://doi.org/10.5281/zenodo.19324441⁸⁶; version v1.0). The same color scheme as the ABC Atlas was used. Lastly, to evaluate the MFNet Atlas, we compared the similarity between the MFNet Atlas and the ABC Atlas by using histogram correlation metrics between pairwise spatially matched coronal slices.

To examine the advantages of the MFNet Atlas over the ABC Atlas, we first interpolated the interval z-slices in between the available ABC Atlas z-slices by using k-nearest-neighbor interpolation across z. For each target slice, we locate the k closest ABC slices in z, weight them by inverse distance, and propagate x–y coordinates and labels via distance-weighted majority vote; ties follow the empirical class prior. At the boundaries, we clamp to the nearest observed slice. This yields a dense 1320-slice atlas without altering in-plane geometry or applying smoothing. Then we compared the inhibitory neuron type maps of MFNet Alas with the interpolated ABC Atlas, with Gad1 and Gad2 ISH images as the ground truth.

Statistics and reproducibility

All statistical analyses were performed using GraphPad Prism version 10.6.1 (GraphPad Software, Inc. RRID:SCR_002798), Python (v3.11, RRID:SCR_008394), and MATLAB R2024b (MathWorks, RRID:SCR_001622). Unless otherwise stated, all statistical tests were two-sided. Multiple-comparison correction was performed using the Benjamini–Hochberg false-discovery-rate (FDR) procedure, with FDR-corrected p-values < 0.05 considered statistically significant.

For MFNet model development and evaluation, diffusion MRI datasets acquired using three imaging protocols were analyzed independently: a high b-value protocol (n = 1), a high angular-resolution protocol (n = 1), and a high spatial-resolution protocol (n = 4). MERFISH sections were used as image samples for model training and testing. For the high b-value and high angular-resolution datasets, training and testing were performed using different sections from the same animal; therefore, testing samples represented held-out technical samples rather than independent biological replicates. For the high spatial-resolution dataset, four biologically independent animals were included. Biological replicates were defined as individual animals, whereas image sections derived from the same animal were treated as technical samples and not as independent biological replicates. For paired comparisons used to evaluate MFNet performance across diffusion MRI protocols and to compare analyses performed with and without CCF structure stratification, two-sided paired t-tests were applied. Given the small sample sizes and the exploratory nature of these analyses, formal tests of normality were not performed. Results were interpreted cautiously, and effect sizes are reported alongside p-values where appropriate.

For region-based correlation analyses, diffusion MRI measurements from a single mouse brain (n = 1) were analyzed. The CCFv3-region defined 686 ROIs represented spatial observations within a single biological specimen and were not treated as independent biological replicates. Assumptions of linearity and the absence of extreme outliers were assessed by visual inspection of scatter plots. Pearson correlation was used to quantify associations between diffusion MRI-derived measurements and reference transcriptomic variables because the analyzed variables were continuous and the large number of spatial observations reduced sensitivity to moderate departures from normality.

For atlas construction and multimodal integration, one mouse brain was used for diffusion MRI acquisition (n = 1). LSM and ST data were obtained from publicly available datasets and were not generated from the same animal. Therefore, the sample size for whole-brain neuronal cell type atlas generation was one animal, whereas LSM and ST sample sizes correspond to those reported in the original source studies. No biological replication was performed for atlas construction.

No formal statistical methods were used to predetermine sample size. Sample sizes were selected based on the availability of high-resolution diffusion MRI, LSM, and MERFISH datasets. All available datasets acquired under each imaging protocol that passed image preprocessing and registration quality control were included. Approximately 1.0% of image voxels whose coordinates were identical after MRI image upsampling were excluded. No additional animals, sections, or ROIs were excluded from model development or statistical analyses. For atlas construction and multimodal integration, only the dataset acquired using the high b-value diffusion MRI protocol was included because the principal SANDI-derived metrics used for atlas generation are sensitive to high b-value diffusion encoding¹⁸. No additional exclusion criteria were applied.

Ethical approval

All animal studies have been approved by the appropriate ethics committee: Duke University Institutional Animal Care and Use Committee. Approval code A226-17-09.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

X.H. and R.H. conceived and designed the study. X.H., R.H., and Z.L. performed data analysis and model development. J.C. and M.J. contributed to data preprocessing and quality control. H.S. contributed to methodology development. Y.Z. assisted with statistical analysis and visualization. M.L. provided domain expertise on deep learning. L.E.W. and G.A.J. contributed imaging resources and technical guidance. N.W. supervised the project and secured funding. X.H. drafted the paper, and all authors reviewed and approved the final paper.

Peer review

Peer review information

Communications Biology thanks the anonymous reviewers for their contribution to the peer review of this work. Primary Handling Editors: Marta Vallejo and Jasmine Pan. A peer review file is available.

Funding

N.W. discloses support for the research of this work from National Institutes of Health grant [R01NS125020]. G.A.J. discloses support for publication of this work from National Institutes of Health grants [P41EB015897] and [S10OD010683]. X.H., R.H., Z.L., J.C., M.J., H.S., Y.Z., M.L., and L.E.W. declare no relevant funding. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the paper.

Data availability

The MFNet-derived whole-brain neuronal cell type atlas is publicly available at Zenodo (https://doi.org/10.5281/zenodo.19324441⁸⁶; version v1.0) under a Creative Commons Attribution 4.0 (CC BY 4.0) license. The dataset is provided in NIFTI format and includes voxel-wise labels of neuronal type, cell neighborhood, and cell class in CCFv3 space. The atlas can be accessed and visualized using standard neuroimaging software (e.g., ITK-SNAP, FSL, MRtrix3) or Python-based tools (e.g., NiBabel, Nilearn). Code and documentation for accessing and using the atlas are available at https://github.com/selinahxy/mricelltype. Raw and processed dMRI and LSM data are available at our data portal (https://utsw.app.box.com/s/mhjwdck148p3y3vofd22j7pd0h5jdhby). Availabilities about publicly available datasets can be found in Supplementary Table 1. The source data of Figs. 3–5 are available in Supplementary Data 1–5.

Code availability

The code used to generate the MFNet model, perform diffusion MRI preprocessing, and produce the neuronal cell type atlas is publicly available at https://github.com/selinahxy/mricelltype (https://doi.org/10.5281/zenodo.20736482⁸⁷; release v2.0). The repository includes scripts for preprocessing, model training and validation, and atlas generation, along with documentation for reproducing the main results. This study used the following software and packages: Matlab R2024b, Python v3.11, DSI Studio “Chen” version, NODDI Matlab Toolbox v1.05, Dipy 1.10.0, SANDI Matlab Toolbox Latest Release, ANTs v2.0, PyTorch v2.6, UMAP v0.5.7, MRtrix3, GraphPad Prism version 10.6.1.

Competing interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the paper preparation process

During the preparation of this work the authors used ChatGPT (GPT-5.1, RRID:SCR_023775) in order to improve clarity and grammar. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s42003-026-10608-y.