Atlas-Free Brain Representation Learning

• Atlas-Free Brain Representation Learning is a computational framework that extracts high-resolution brain features directly from raw data, bypassing predefined anatomical atlases.
• It leverages techniques like contrastive learning, clustering, and adaptive token-based architectures to achieve invariant, subject-specific patterns across diverse neuroimaging modalities.
• Empirical studies demonstrate enhanced prediction accuracy, spatial fidelity, and computational efficiency compared to atlas-based methods, supporting advanced neuroscience applications.

Atlas-free brain representation learning encompasses a set of computational methodologies that learn high-dimensional, information-rich encodings of the brain’s structure, function, or multimodal activity without reliance on a predefined anatomical or functional atlas. By circumventing the use of fixed parcellations or atlas-based region-of-interest (ROI) definitions, these methods preserve native spatial/frequency detail, adapt to inter-individual variability, and enable representation discovery aligned with intrinsic biological patterns. Atlas-free approaches have fundamentally transformed analysis pipelines in structural MRI, diffusion MRI, fMRI, EEG/MEG, and histological studies, and continue to underpin advances in foundational brain modeling, cross-modal alignment, and neurobiomarker development.

1. Fundamental Principles of Atlas-Free Representation Learning

Atlas-free brain representation learning eliminates all dependency on region-based segmentation or template registration. Instead, the models learn directly from minimally processed raw signals—voxels, surface vertices, frequency bands, or image patches—through unsupervised, self-supervised, or supervised training objectives. Core principles include:

• Native data granularity: Raw input (e.g., voxel timeseries, cortical surface points, histological image patches) is used directly, avoiding any aggregation or averaging within predefined regions (Wang et al., 30 Jan 2026, Consagra et al., 2023, Schiffer et al., 2020).
• Learned parcellation or clustering: Any spatial grouping emerges from clustering over features or connectivity, or as a by-product of representation learning, not as a prior input (Mohan et al., 2020, Huang et al., 30 Sep 2025).
• Invariant and robust features: Data augmentations or architectures are used to enforce robustness to anatomical deformation, contrast/appearance changes, subject variability, and acquisition artifacts, enhancing generalization across datasets and populations (Liu et al., 2023, Jiang et al., 2023).
• Parameter sharing and cross-subject alignment: Models often share weights or mapping functions across all voxels/patches/sensors and all subjects, requiring no atlas-dependent label harmonization or per-subject retraining (Li et al., 13 Jul 2025, Gong et al., 2023).
• Contrastive and latent space objectives: Embeddings are supervised or regularized via similarity in feature or target space, not via explicit region boundaries (Nguyen et al., 2024, Wang et al., 26 Dec 2025).

This approach stands in strict opposition to region-based pipelines, where features are extracted from or aggregated within standardized regions, and all downstream modeling is constrained by the selection and granularity of the atlas.

2. Core Architectures and Methodological Innovations

Atlas-free representation learning employs diverse architectures tailored to the spatial and temporal characteristics of each data type. Representative household models include:

• Foundation models for 3D/4D MRI/fMRI: Vision Transformers (ViT), 3D convolutional autoencoders, masked language image models (MAE), and hierarchical patch-based encoders learning end-to-end from full volumes, maintaining patch-level or voxel-level features throughout (Wang et al., 30 Jan 2026, Wang et al., 26 Dec 2025, Jaiswal et al., 2018, Mohan et al., 2020).
• Spectrally-informed encoders: Frequency-domain networks employing DFT or power-spectrum compression, exploiting the global structure and denoising capabilities of the frequency representation while mapping directly from raw voxels (Gong et al., 2023).
• Contrastive learning and mutual information objectives: Encoders trained with cycle-consistency, triplet, or contrastive loss, optimizing agreement across modalities (fMRI, images, text, EEG, MEG), and across augmentations, to learn semantically meaningful, modality- and region-agnostic representations (Nguyen et al., 2024, Schiffer et al., 2020, Li et al., 13 Jul 2025).
• Unsupervised and generative frameworks: U-Net-based unsupervised partitioners ["AutoAtlas"], patch-level CAEs, and fully continuous latent function modeling on cortical product manifolds, removing arbitrary boundaries and facilitating continuous, high-resolution mapping (Mohan et al., 2020, Consagra et al., 2023).
• Token selection, saliency ranking, and adaptive masking: Transformers with Mutual-Attention Token Selection or adaptive patching at multiple spatial/temporal resolutions to maximize efficiency and biological relevance, as in SLIM-Brain and MCIAT (Wang et al., 26 Dec 2025, Jiang et al., 2023).

Several methods integrate top-down (task/semantic information) and bottom-up (data-driven structure) signals, with architectures supporting multi-modal integration (EEG/MEG/fMRI), time-frequency attention, and subject-specific adaptive weighting (Li et al., 13 Jul 2025, Nguyen et al., 2024).

3. Theoretical and Mathematical Frameworks

Technical rigor in atlas-free representation learning is achieved through custom loss functions, latent space regularization, and algorithmic advances in clustering and decomposition, including:

• Contrastive loss for cross-modal and patchwise alignment: Formulations involving the negative log-likelihood ratio of positive (matched) and negative (mismatched) pairs, e.g.,

$$\mathrm{Contras}(X,Y)\;=\;-\frac1N\sum_{i=1}^N\Bigl[\log\frac{\exp(x_i\!\cdot\!y_i/\sigma)}{\sum_j\exp(x_i\!\cdot\!y_j/\sigma)}+\log\frac{\exp(y_i\!\cdot\!x_i/\sigma)}{\sum_j\exp(y_i\!\cdot\!x_j/\sigma)}\Bigr]$$

as in BRACTIVE for text–visual–fMRI triplet alignment (Nguyen et al., 2024).

• Atlas-free partitioning via clustering/spectral decomposition: Multi-stage clustering (e.g., spatially-constrained agglomerative clustering, spectral clustering with graph Laplacians) to define individualized ROIs, and principal component analysis (PCA) or eigenfunction decomposition for high-dimensional connectivity tensors (Huang et al., 30 Sep 2025, Consagra et al., 2023).
• Continuous function space modeling: Representing the structural connectome $U:M\times M\to \mathbb R_+$ as a continuous function on a cortical manifold, parameterized by a low-rank expansion in smooth basis functions $\{\varphi_k\}$ (Consagra et al., 2023).
• Hierarchical reconstruction and regularization losses: Compose reconstruction, smoothness/neighborhood, and anti-collapse/volume-preservation losses for unsupervised embedding extraction and interpretable partitioning (Mohan et al., 2020, Jaiswal et al., 2018).
• Multi-objective and adversarial training: Simultaneous optimization over denoising, reconstruction, age-prediction, and adversarial discriminators, to achieve semantically disentangled representations aligned with global and fine-grained properties (Jiang et al., 2023, Liu et al., 2023).

Each methodology is formulated to maximize per-sample spatial specificity and representation informativeness, with hyperparameters tuned to balance reconstruction fidelity, clustering granularity, subject alignment, and downstream predictive accuracy.

4. Empirical Performance and Benchmarking

Atlas-free models have shown consistent performance benefits over atlas-based analogs across core neuroimaging benchmarks:

• Improved subject- and group-level prediction: In functional and structural MRI, atlas-free models outperform established atlas-based tools (e.g., FreeSurfer features, Graphormer, BNT with fixed atlases) on predictive tasks such as sex classification, connectome age regression, diagnosis (ADHD, MCI, AD), and fingerprinting (Huang et al., 30 Sep 2025, Wang et al., 26 Dec 2025, Mohan et al., 2020, Jaiswal et al., 2018).
• Superhuman spatial and functional fidelity: ROI localization (e.g., BRACTIVE mean dice ≈69.4% vs GradCAM ≈58.0% (Nguyen et al., 2024)), image/fMRI retrieval (Lite-Mind NSD top-1 ≈94.6% vs. MindEye ≈93.4%, with 98% fewer parameters (Gong et al., 2023)), and high-fidelity brain segmentation/generation (Brain-ID T1w PSNR ≈33.8 dB, SSIM ≈0.993 (Liu et al., 2023)).
• Cross-subject and cross-modal generalization: Models trained from pooled subject data (BrainFLORA, Omni-fMRI) can align, retrieve, and reconstruct concept representations across EEG, MEG, and fMRI, without ROI harmonization or atlas registration (Li et al., 13 Jul 2025, Wang et al., 30 Jan 2026).
• Compute and data efficiency: Hierarchical, patch-based, and saliency-ranked pipelines (SLIM-Brain, Omni-fMRI) achieve 10× speedup and ~70% GPU memory reduction compared to naïve dense approaches, while pushing SOTA across >7 external benchmarks (Wang et al., 26 Dec 2025, Wang et al., 30 Jan 2026).

Empirical evaluations also highlight the robust clustering of learned features into anatomically meaningful groups, enhanced flexibility for novel concept localization, and preservation of fine-scale structural information often lost through ROI averaging or parcellation (Schiffer et al., 2020, Nguyen et al., 2024).

5. Applications and Interpretability in Neuroscience and Machine Intelligence

Atlas-free brain representations support applications beyond conventional brain mapping:

• Unbiased ROI and subnetwork discovery: Subject-specific, concept-driven ROI masks are inferred directly from model activations and similarity maps (e.g., BRACTIVE can localize arbitrary object categories) (Nguyen et al., 2024).
• Continuous connectome analysis: Group differences or phenotypic correlates are mapped by projecting statistical contrasts or permutation-based feature importances back onto the native voxel/fiber space, removing ROI boundary bias (Consagra et al., 2023, Mohan et al., 2020).
• Cross-modal decoding and brain–machine interface (BMI): Joint latent spaces facilitate multimodal (EEG/MEG/fMRI) BCIs and cognitive prediction pipelines without ROI engineering (Li et al., 13 Jul 2025).
• Augmented vision and computer vision transfer: Human-guided, atlas-free pretraining improves downstream visual task performance in generic detectors and segmenters (e.g., BRACTIVE-pretrained ViTs outperform ImageNet/CLIP on COCO/ADE20K) (Nguyen et al., 2024).

Atlas-free models often afford improved interpretability via explicit mapping of feature saliency, learned partition importance, or concept clustering, and have been empirically linked with behavioral and cognitive domains—typically via correlation of embedding dimensions with neuropsychological test scores or anatomical measurements (Jiang et al., 2023, Mohan et al., 2020, Schiffer et al., 2020).

6. Limitations, Challenges, and Future Directions

Despite clear advances, atlas-free representation learning faces distinct challenges:

• Scalability to ultra-high resolution and real-time data: Efficient routing of tokens, saliency-based pruning, or hybrid learned partitioning is still an open area, particularly with increasing data resolutions or multimodal extensions (Wang et al., 26 Dec 2025, Wang et al., 30 Jan 2026).
• Robustness to artifacts and out-of-distribution variation: Frequency-based smoothing, augmentation with synthetic deformations, and explicit artifact simulation improve, but do not guarantee, full robustness across all acquisition conditions and device configurations (Liu et al., 2023, Gong et al., 2023).
• Interpretability and biological plausibility: While unsupervised clustering aligns with anatomical/functional patterns, absolute classification and segmentation accuracy can be limited by ground-truth granularity and intersubject anatomical variability (Schiffer et al., 2020, Mohan et al., 2020).
• Hybrid frameworks and domain adaptation: Integration of atlas-free and weakly-atlas-informed cues, dynamic scale adaptation, and subject-specific fine-tuning via token selection or mixture-of-experts remain active research topics with the goal of enhancing interpretability, efficiency, and transferability (Li et al., 13 Jul 2025, Jiang et al., 2023).

A plausible implication is that further success in this domain will depend on advances in scalable, hierarchical architectures; more sophisticated data-driven or biologically motivated partitioning; and expanded cross-modal joint training frameworks.

7. Representative Methods and Comparative Summary

Selected atlas-free model classes, their key innovations, and benchmark results:

Model	Key Feature	Notable Results/Applications
BRACTIVE (Nguyen et al., 2024)	ViT triplet, text-driven ROI	SOTA concept-localization, boosts vision model transfer
Lite-Mind (Gong et al., 2023)	DFT-based encoder, tiny FreMLP	94.6% NSD image-retrieval, 98.7% param. reduction vs. MindEye
BrainFLORA (Li et al., 13 Jul 2025)	Multimodal (EEG/MEG/fMRI) MoE	Concept alignment, 28.3% fMRI retrieval, cross-modal transfer
AutoAtlas (Mohan et al., 2020)	Joint U-Net partitioning & AE	Outperforms FreeSurfer in behavioral prediction
Omni-fMRI (Wang et al., 30 Jan 2026)	Dynamic patch-ViT (MAE), fast	+10 pt accuracy over atlas-based, 4,300 vs. 14k tokens
SLIM-Brain (Wang et al., 26 Dec 2025)	Temporal extractor + Hiera-JEPA	SOTA on 7 tasks, 30% GPU memory, 4k pretrain sessions
Brain-ID (Liu et al., 2023)	3D U-Net, domain-agnostic augm.	SOTA on MRI/CT segmentation, synthesis, robust low-res perf.
MCIAT (Jiang et al., 2023)	Multi-task ViT, mutual-attn tokens	Best ADHD-200/OASIS accuracy, interpretability
Contrastive histology (Schiffer et al., 2020)	Patchwise contrastive learning	50.1% top-1, tight anatomical clustering in microstructure
Continuous connectome (Consagra et al., 2023)	Manifold, functional basis L²(M×M)	100% test–retest match, .2 corr. gain in trait prediction
Atlas-free BNT (Huang et al., 30 Sep 2025)	Individual parcellation + transformer	89.2% ABCD sex classif., 4.03 MAE age pred. (best)
CAE-family (Jaiswal et al., 2018)	Staged/joint/3D autoencoders	0.81–0.86 ADNI AUROC (equals FS), <1s/scan feature extraction

These results collectively demonstrate the breadth and empirical superiority of atlas-free brain representation learning pipelines across modalities, granularity levels, and downstream tasks. The field is rapidly evolving, with foundational studies now providing reproducible, open benchmarks for future development and comparison.