BrainFIBRE: A Foundation Model via Information Decomposition for Brain Microstructure

Abstract: Diffusion MRI probes brain microstructure with particular sensitivity to early cerebrovascular and neurodegenerative changes. Neurite Orientation Dispersion and Density Imaging (NODDI) decomposes the diffusion signal into three biophysically interpretable maps: neurite density index (NDI), orientation dispersion index (ODI), and free water fraction (FWF), capturing neurite packing, fiber coherence, and extracellular fluid. These 3D maps offer a rich substrate for transferable microstructural representations, yet integrating them is challenging: standard representation learning struggles to disentangle the unique information in each map from their shared and synergistic interactions. We present BrainFIBRE, the first foundation model for brain microstructure, pretrained on NODDI-derived maps from 55,592 UK Biobank participants. We propose Self-supervised Partial Information Decomposition (SPID), which extends PID-guided multimodal learning to the self-supervised regime for the first time. A novel Counterfactual Candidate Construction (CCC) paradigm perturbs inter-modality alignment through modality dropping and swapping, providing the contrastive signal for a Mixture-of-Experts architecture to disentangle unique, synergistic, and redundant information without any downstream label. On both Caucasian and Asian cohorts, BrainFIBRE achieves state-of-the-art performance across diverse tasks predicting age, sex, cerebrovascular and neurodegenerative markers, and cognition, while yielding neurobiologically interpretable representations that reveal task- and cohort-specific interaction patterns. BrainFIBRE establishes a versatile foundation for neuroimaging analysis at the microstructural level.

• The paper introduces BrainFIBRE, a model that uniquely decomposes multimodal NODDI maps using a self-supervised MoE framework to isolate unique, redundant, and synergistic information.
• It employs Self-supervised Partial Information Decomposition (SPID) with counterfactual candidate construction, enhancing interpretability and achieving state-of-the-art performance on clinical and demographic predictions.
• Ablation studies validate that full integration of NODDI modalities is crucial for capturing neurobiologically meaningful microstructural variations across diverse populations.

BrainFIBRE: Information Decomposition Foundation Modeling for Brain Microstructure

Introduction

The "BrainFIBRE" model (2607.00573) proposes a new paradigm for foundation modeling in neuroimaging, focusing specifically on the domain of brain tissue microstructure. While previous self-supervised brain foundation models have been restricted to macroscopic MRI modalities, this work leverages Neurite Orientation Dispersion and Density Imaging (NODDI), which yields three biophysically grounded parametric maps: Neurite Density Index (NDI), Orientation Dispersion Index (ODI), and Free Water Fraction (FWF). These maps partition the diffusion MRI signal into meaningful compartments corresponding to key biophysical properties and have demonstrated superior specificity for neurodegenerative, vascular, and aging-related processes compared to conventional diffusion metrics.

The principal challenge addressed by BrainFIBRE lies in the integration of these three highly informative, yet statistically intertwined, spatial modalities without collapsing or entangling their interpretative power. Standard fusion methods used in multimodal learning are inadequate for this problem, as they tend to obscure the compartmental specificity that NODDI was designed to isolate.

Figure 1: Overview of NODDI-derived microstructural maps (the input to BrainFIBRE): NDI, ODI, and FWF provide compartment-specific sensitivity to brain tissue changes.

Model Architecture and Self-Supervised Partial Information Decomposition

BrainFIBRE is architected around a Mixture-of-Experts (MoE) framework and introduces Self-supervised Partial Information Decomposition (SPID), an extension of classical Partial Information Decomposition (PID) to the self-supervised setting. The aim is to decompose the information provided by the three NODDI modalities into unique (modality-specific), redundant (shared), and synergistic (jointly emergent) information components.

Three modality-specific 3D Vision Transformers serve as encoders, projecting NDI, ODI, and FWF volumes to compact embeddings. These are then processed by five interaction experts corresponding to the PID atoms: uniqueness for each modality, redundancy, and synergy. An adaptive re-weighting module dynamically fuses these experts into a global subject-level representation. Importantly, the routing to different experts is task- and data-adaptive, allowing the model to learn the interaction patterns most predictive for downstream targets.

SPID is operationalized in a label-free contrastive pretraining regime via Counterfactual Candidate Construction (CCC): the model generates challenging perturbations of the NODDI triplet, such as modality-dropping (blanking one map) or cross-subject modality swapping, thereby producing contrastive views that break specific compartmental correspondences. Expert-specific training targets are defined using formal information-theoretic rules, ensuring that each expert isolates its intended information component even in the absence of downstream supervision.

Figure 2: Schematic of BrainFIBRE, illustrating the SPID-driven multimodal encoder, five interaction experts, adaptive fusion, and counterfactual candidate construction for self-supervised learning.

Empirical Results and Ablation Analyses

Extensive experiments validate the efficacy and interpretability of BrainFIBRE. Pretrained on over 55,000 UK Biobank participants, the model is evaluated across several downstream tasks: demographic prediction (age, sex), clinical endpoints (hippocampal atrophy, white matter hyperintensity (WMH) volume), and cognitive measures, in both internally held-out UKB subjects and two external datasets (HCP-Aging and SINGER, the latter representing an Asian and high-cardiometabolic-risk population).

Compared to unimodal models, classical self-supervised fusion, and other MoE-based multimodal baselines, BrainFIBRE achieves consistently superior results. For example, on UKB age prediction, it attains a mean absolute error of 3.95 years (substantially outperforming single-modality and baseline fusion approaches) and achieves the highest accuracy and correlation for demographic and cognitive predictions across all cohorts.

Ablation studies demonstrate that the performance improvements are contingent on both the inclusion of all three NODDI maps and the explicit modelling of unique, synergistic, and redundant interaction components. Removing any expert, or omitting the dedicated interaction losses, yields notable reductions in predictive performance and representation disentanglement.

Figure 3: Distribution of learned expert weights by the adaptive re-weighter across multiple tasks, highlighting task- and population-specific shifts in the contribution of unique, redundant, and synergistic information.

Figure 4: Ablation analysis showing the importance of full expert fusion and multimodal integration—removing any expert or modality diminishes downstream task performance.

Interpretability: Expert Specialization and Spatial Patterns

One of BrainFIBRE's central claims is that information decomposition not only boosts representational power but also yields neurobiological interpretability. Analyses of the adaptive expert weights reveal that certain tasks (e.g., age prediction) predominantly rely on synergetic fusion, while others (e.g., WMH prediction) are driven largely by the FWF uniqueness channel, reflecting the established neuropathological role of free water in cerebrovascular disease.

Spatial attention patterns extracted from each expert further corroborate the biological plausibility: uniqueness experts highlight canonical cortical and tract regions consistent with their map's specificity (e.g., medial temporal decline for NDI with aging), while synergy and redundancy experts localize to multimodal hubs (e.g., cingulum bundle) associated with integrative network aging and disease vulnerability.

Figure 5: Spatial attention maps by expert for key tasks, indicating distinctive neuroanatomical correlates of unique, redundant, and synergistic microstructural information.

Theoretical Foundations and Empirical Validation

The self-supervised PID framework employed by BrainFIBRE is mathematically substantiated: the model's expert-conditional discrimination losses are shown to possess a unique, PID-grounded optimum, such that each expert learns only its target information atom when presented with sufficient variability in counterfactual perturbations. Posthoc analyses using CCA confirm that the learned representations align as expected: uniqueness experts are maximally correlated with their specific modality embeddings, redundancy aligns broadly, and synergy exhibits minimal correlation with any single input, supporting the model's theoretical motivation.

Implications and Future Directions

On the practical front, BrainFIBRE establishes a scalable, modality-aware foundation for brain microstructure analysis, supporting a variety of clinical and cognitive endpoints across ethnically diverse datasets. Its information decomposition mechanism ensures that future lines of research can probe for population- or disease-specific shifts in microstructural information content with formal interpretability, rather than relying on opaque global representations.

Theoretically, BrainFIBRE's demonstration that explicit information decomposition—grounded in formal PID and implemented via MoE within a self-supervised regime—can be instantiated at scale without supervision is significant for the field of multimodal representation learning. It suggests that further extensions to higher-order PID architectures, fine-grained interaction modeling, and applications to other domains with compartmentalized data (multi-omics, multi-modal sensor data) are now tractable.

Figure 6: Empirical validation of representation disentanglement via CCA, confirming the alignment of each expert with its intended PID information atom.

Figure 7: White matter tract attention rollout comparing BrainFIBRE and late-fusion baselines, demonstrating BrainFIBRE's tract-specific specialization for disease-relevant pathways.

Conclusion

BrainFIBRE introduces a theoretically principled, information-decomposing foundation model for brain tissue microstructure, demonstrating robust state-of-the-art performance and interpretability across diverse downstream tasks and cohorts. By fusing advances in information theory, contrastive self-supervision, and multimodal MoE architectures, BrainFIBRE provides a new template for biophysically interpretable, scalable neuroimaging representation learning with immediate clinical and research utility. Key challenges for future work include extending PID decompositions to capture even more granular interaction structures, leveraging transfer to additional populations and disorders, and integrating dynamic or longitudinal microstructural states.