Atlas-Free Brain Network Transformer

• Atlas-Free Brain Network Transformer is a framework that generates individualized brain parcellations from resting-state fMRI to overcome limitations of fixed atlases.
• It computes standardized ROI-to-voxel connectivity features that are processed through a transformer architecture to produce comparable subject-level embeddings.
• Experimental evaluations show that it achieves superior accuracy (e.g., 89.2% for sex classification) and robustness to spatial misalignments compared to atlas-based methods.

Searching arXiv for the cited paper and closely related atlas-free brain network work. Searching arXiv for the paper by title and related atlas-free fMRI models. Atlas-Free Brain Network Transformer (atlas-free BNT) is a brain network analysis framework that leverages individualized brain parcellations derived directly from subject-specific resting-state fMRI data rather than standardized anatomical or connectivity-driven brain atlases (Huang et al., 30 Sep 2025). It was proposed to address spatial misalignment across individuals, functional heterogeneity within predefined regions, and atlas-selection biases, and it computes ROI-to-voxel connectivity features in a standardized voxel-based feature space that are subsequently processed using the BNT architecture to produce comparable subject-level embeddings (Huang et al., 30 Sep 2025). In the reported evaluations, atlas-free BNT consistently outperforms state-of-the-art atlas-based methods, including elastic net, BrainGNN, Graphormer, and the original BNT, on sex classification and brain-connectome age prediction (Huang et al., 30 Sep 2025).

1. Motivation and problem setting

Current atlas-based approaches to brain network analysis rely heavily on standardized anatomical or connectivity-driven brain atlases. In the formulation motivating atlas-free BNT, these fixed atlases often introduce significant limitations, such as spatial misalignment across individuals, functional heterogeneity within predefined regions, and atlas-selection biases, collectively undermining the reliability and interpretability of the derived brain networks (Huang et al., 30 Sep 2025).

The central methodological difficulty is that atlas-based region definitions enforce the same parcel geometry across subjects even when subject-specific functional organization differs. Atlas-free BNT addresses this by generating individualized parcellations for each subject directly from resting-state fMRI data and then restoring cross-subject comparability in a standardized voxel-based feature space. This design separates subject-specific parcel discovery from group-level representation.

Related atlas-free work in functional neuroimaging frames the same problem in similar terms. AFBR-KAN and ABFR-KAN describe atlas-based functional connectivity analysis as vulnerable to selection bias and reduced subject specificity, and they likewise replace atlas-defined ROIs with subject-adaptive patch-based representations (Ward et al., 4 Apr 2025, Ward et al., 1 Jan 2026). In structural connectivity, continuous atlas-free analysis replaces discrete ROI adjacency matrices with smooth random functions over a product manifold, again removing the need for a pre-defined atlas (Consagra et al., 2023). This suggests that atlas-free BNT belongs to a broader shift from fixed parcellation toward individualized or continuous connectome representations.

2. Individualized parcellation from resting-state fMRI

The individualized parcellation stage is based on unsupervised clustering of subject-specific resting-state fMRI data. Two strategies are described: agglomerative clustering and spectral clustering (Huang et al., 30 Sep 2025).

In agglomerative clustering, the procedure begins with each voxel as its own cluster and iteratively merges spatially adjacent voxel clusters with the highest pairwise functional connectivity, measured by Pearson correlation,

$$\rho_{ij} = \frac{\mathrm{cov}(x_i, x_j)}{\sqrt{\mathrm{Var}(x_i)\,\mathrm{Var}(x_j)}},$$

where $x_i, x_j \in \mathbb{R}^L$ are BOLD timeseries for voxels $i,j$. The spatial constraint permits only neighboring voxels or clusters to merge, which brings complexity to $\mathcal{O}(N \log N)$. Merging stops when either a specified number of clusters is reached or the maximum correlation between any two clusters drops below threshold $\nu$.

In spectral clustering, a Pearson correlation similarity matrix $S$ is constructed among voxels, and the normalized graph Laplacian is formed as

$L_{\text{norm}} = I - D^{-1/2} S D^{-1/2},$

where $D$ is the degree matrix. The top $k$ eigenvectors are computed to embed voxels, followed by $k$-means clustering. For tractability, the similarity matrix is sparsified so that $x_i, x_j \in \mathbb{R}^L$0 if $x_i, x_j \in \mathbb{R}^L$1.

These procedures produce functionally homogeneous ROIs tailored for each subject. The reported interpretation is that this directly addresses spatial misalignment and within-ROI heterogeneity. The comparison between agglomerative clustering and spectral clustering is empirical rather than purely theoretical: both are used within the same downstream transformer pipeline, but agglomerative clustering yields the best overall results in the reported benchmarks.

3. Standardized ROI-to-voxel representation

A central issue in any individualized parcellation framework is that parcellations differ across subjects, so direct ROI-to-ROI features are not comparable. Atlas-free BNT resolves this by computing ROI-to-voxel connectivity in a canonical voxel space, usually MNI, thereby standardizing subject data without reverting to a fixed atlas (Huang et al., 30 Sep 2025).

For each ROI $x_i, x_j \in \mathbb{R}^L$2, the average BOLD signal yields mean timeseries $x_i, x_j \in \mathbb{R}^L$3. The functional connectivity profile of ROI $x_i, x_j \in \mathbb{R}^L$4 is then defined as

$x_i, x_j \in \mathbb{R}^L$5

where $x_i, x_j \in \mathbb{R}^L$6 are timeseries of each voxel and $x_i, x_j \in \mathbb{R}^L$7 is the number of voxels in the standardized space. The resulting feature matrix has size $x_i, x_j \in \mathbb{R}^L$8 per subject, where $x_i, x_j \in \mathbb{R}^L$9 is the number of ROIs.

These features are mapped into a 3D voxel grid as a multi-channel brain map

$i,j$0

where each channel corresponds to connectivity with a voxel. Because this representation is high-dimensional, dimensionality reduction by PCA is applied to $i,j$1, producing low-dimensional ROI features:

$i,j$2

A common misconception is that an atlas-free method necessarily dispenses with all standardized spatial reference systems. Atlas-free BNT does not do so. Its atlas-free property lies in individualized parcellation rather than in abandoning common space registration. The canonical voxel space remains necessary because, as stated explicitly, direct ROI-to-ROI features are not comparable across subjects when the parcellations differ.

4. Transformer architecture and processing pipeline

The downstream model is a BNT architecture applied to the standardized ROI-to-voxel representation. Input preparation begins by blocking the multi-channel brain map $i,j$3 into overlapping $i,j$4 voxels, or patches, with stride $i,j$5, resulting in $i,j$6 blocks or nodes (Huang et al., 30 Sep 2025).

For block $i,j$7, voxel features are sum-pooled across all channels, yielding an input node feature $i,j$8. All block features are concatenated to form the transformer input,

$i,j$9

Multi-Head Self-Attention is then applied:

$\mathcal{O}(N \log N)$0

$\mathcal{O}(N \log N)$1

where at each layer $\mathcal{O}(N \log N)$2, $\mathcal{O}(N \log N)$3 indexes heads, and $\mathcal{O}(N \log N)$4 are learned parameters.

After $\mathcal{O}(N \log N)$5 such layers, the global embedding is constructed via an orthonormal clustering readout function, producing the subject-level embedding $\mathcal{O}(N \log N)$6. This embedding is used as input to a multi-layer perceptron for downstream classification or regression.

Three design choices are identified as key innovations. First, the pipeline is atlas-free and individualized to each subject. Second, ROI-to-voxel connectivity in a standardized voxel feature space makes subject data comparable. Third, blockwise input with overlapping 3D blocks increases local robustness to residual misregistration. The last point is especially important because it clarifies that the method does not assume perfect spatial normalization; instead, it attempts to mitigate imperfect alignment during feature aggregation.

5. Experimental evaluation and reported findings

The reported experiments evaluate sex classification on the ABCD dataset with 6,738 subjects and brain connectome age prediction on the EHBS dataset with 2,255 sessions (Huang et al., 30 Sep 2025). The sex classification metrics are accuracy, sensitivity, specificity, and AUROC; the age prediction metric is mean absolute error in years. Baselines include elastic net, BrainGNN, Graphormer, and the original BNT, each using several atlases: AAL, Craddock-400, Shen-368, and HCP-360.

The best atlas-free result is obtained by atlas-free BNT with agglomerative clustering: 89.2% accuracy, 95.9 AUROC, and 4.03 mean absolute error. Atlas-free BNT with spectral clustering reports 88.8% accuracy, 95.7 AUROC, and 4.06 mean absolute error. The best atlas-based BNT reports 87.9% accuracy, 94.6 AUROC, and 4.21 mean absolute error. The best Graphormer result is 86.4% accuracy, 93.9 AUROC, and 4.71 mean absolute error; the best elastic net result is 86.8% accuracy, 93.9 AUROC, and 4.86 mean absolute error; and the best BrainGNN result is 77.3% accuracy, 84.6 AUROC, and 4.63 mean absolute error.

The interpretation attached to these results is that atlas-free BNT outperforms all atlas-based methods on both tasks, regardless of parcellation approach, with agglomerative clustering yielding the best results, likely due to maximizing ROI homogeneity. Additional qualitative and quantitative findings are also reported. Intra-ROI BOLD correlation is higher with individualized parcellations than with atlas-based methods, which is presented as evidence of improved functional homogeneity. Blockwise pooling is said to ensure that minor spatial misalignments do not disrupt feature extraction, supporting robustness to misregistration. Grad-CAM analyses indicate that the model identifies interpretable, task-relevant networks for both sex and age prediction tasks.

The broader implications stated for these findings are precision, robustness, and generalizability. Individualized parcellation avoids anatomical misalignment and enhances ROI homogeneity; projecting individualized connectivity into a common voxel-based space enables group-level analyses and across-study harmonization; and the reported predictive gains suggest potential to improve neuroimaging biomarkers and clinical diagnostic tools for personalized precision medicine.

6. Relation to adjacent atlas-free and atlas-based transformer models

Atlas-free BNT is part of a larger methodological landscape, but its design is distinct from other atlas-free transformer approaches. AFBR-KAN and ABFR-KAN are also transformer-based and atlas-free, yet they construct functional connectivity from randomized anchor patch selection and iterative, multi-scale patch sampling rather than from individualized parcellations and ROI-to-voxel features (Ward et al., 4 Apr 2025, Ward et al., 1 Jan 2026). In those models, MLP components are replaced by Kolmogorov-Arnold Network blocks, and the primary application is autism spectrum disorder diagnosis rather than sex classification or brain-connectome age prediction.

Omni-fMRI takes a different route again: it operates directly on voxel-level signals, uses a dynamic patching mechanism to reduce computational cost while preserving informative spatial structure, and scales self-supervised pretraining to 49,497 fMRI sessions across nine datasets (Wang et al., 30 Jan 2026). By contrast, atlas-free BNT still constructs an explicit connectivity representation before transformer processing. A plausible implication is that atlas-free BNT occupies a middle position between patch-token voxel models and conventional atlas-based connectome transformers.

The contrast with atlas-based transformer models is sharper. BrainNPT represents each brain functional network as a graph whose nodes correspond to atlas-defined ROIs, explicitly notes that the model as presented is not atlas-free, and uses a learnable <cls> token as the classification embedding vector (Hu et al., 2023). Atlas-free BNT differs at the level of representation before readout: it does not assume a fixed ROI correspondence supplied by an atlas, and it does not rely on atlas ordering as the basis for token consistency.

More broadly, continuous atlas-free analysis of structural connectivity models connectomes as smooth random functions

$\mathcal{O}(N \log N)$7

with reduced-rank approximation

$\mathcal{O}(N \log N)$8

and reports strong reproducibility and statistical power without discrete atlas parcellation (Consagra et al., 2023). This suggests that atlas-free BNT can be read as one member of a broader atlas-free program in neuroimaging: individualized, continuous, or voxel-level representations are replacing fixed parcel systems, while different architectures impose comparability through standardized spaces, learned embeddings, or reduced-rank function bases.