TL;DR
This paper proposes a novel method using a Stochastic Block Model (SBM) to find "canonical processing units" in the brain by jointly analyzing structural and functional connectivity data. The approach reveals a shared underlying organization despite modality-specific biases, leading to improved representations of brain networks.
What Problem Does This Paper Address?
Understanding how the brain's structure (how regions are physically connected) relates to its function (how regions communicate dynamically) is a fundamental challenge. While both diffusion MRI (dMRI) and functional MRI (fMRI) provide insights into these networks, their connectivity profiles often don't directly align. This mismatch makes it difficult to understand the relationship between structure and function and to identify consistent "processing units" across the brain. The authors aim to bridge this gap by looking for shared modular organization rather than direct connection-level correspondence.
Methods
The core of this work is the application of a Stochastic Block Model (SBM). SBMs are generative models that can identify community structures (clusters or "parcels") within networks. The key innovation here is using a multi-layer SBM. This model assumes a single underlying parcellation (grouping of brain regions) exists across both structural and functional connectomes. However, it allows each modality to have its own independent connectivity structure between these shared parcels. This means the model can capture concurrent functional and structural units while acknowledging that the way these units connect might differ between structure and function. The authors then evaluate the quality of these parcellations by using them to predict "hold-out" networks, essentially testing how well a learned parcellation from one set of data can predict unseen data from the same or different modalities. They also employ a "blocked permutation procedure" to create a null hypothesis, breaking the anatomical correspondence between modalities to test if the observed agreement is genuine.
Datasets
The study utilizes data from the Human Connectome Project (HCP), specifically the resting-state fMRI and diffusion MRI data from 250 healthy subjects. The data was preprocessed according to standard HCP pipelines. The authors also compare their results to the HCP_MMP1.0 atlas, a well-established multimodal parcellation of the human cerebral cortex.
Key Insights and Ideas
- Shared "Processing Units" Exist: Despite significant differences in how structural and functional connectivity are measured and represented, the brain appears to have underlying, shared organizational units.
- Modality-Specific Biases are Real: Functional connectivity tends to be more bilateral and exhibits a wider range of parcel sizes, likely due to how fMRI captures widespread co-activation. Structural connectivity, on the other hand, is more unilateral and shows more uniform parcel sizes, potentially limited by tractography methods.
- Joint Modeling Bridges the Gap: By modeling structural and functional connectomes together, the SBM can find a "consensus" parcellation that is less biased by individual modality limitations and better represents the shared underlying organization.
- Functional Data Benefits from Structure: The noisier functional connectivity estimates show improved prediction accuracy when integrated with structural information, suggesting structure acts as a regularizer.
- Permutation Breaks the Consensus: When the anatomical correspondence between modalities is deliberately broken (using the blocked permutation), predictive performance drops significantly, confirming that the identified shared units are not random.
Key Results
- Improved Predictive Performance: The joint multimodal parcellations consistently outperformed modality-specific parcellations in predicting hold-out networks, particularly for functional connectivity. For example, predicting function using the joint model achieved an Area Under Curve (AUC) of 0.9957, compared to 0.9927 for structure-only and 0.9806 for permuted structure.
- Structural Parcellations Predict Function Well: Structural parcellations were surprisingly good predictors of functional connectivity, often performing comparably to or even better than functional parcellations predicting structure.
- Functional Parcellations Struggle with Structure: Functional parcellations were poor predictors of structural connectivity, to the point where permuted structural data sometimes predicted better.
- High Agreement with HCP_MMP1.0: The data-driven multimodal parcellations showed high Normalized Mutual Information (NMI) scores with the HCP_MMP1.0 atlas, indicating a strong agreement in the identified brain regions. The NMI between the joint parcellation and structural parcellations was comparable to the NMI within structural parcellations themselves (around 0.85), and similarly for functional (around 0.7).
Limitations and Future Work
- Inter-Subject Alignment: The reliance on surface morphometry for alignment might introduce biases.
- Arbitrary Thresholding: The networks were thresholded at 1% density, and future work could explore the impact of different thresholding strategies.
- Resolution: The study used a fixed number of parcels (360) to match the HCP_MMP1.0 atlas. However, structural and functional data might benefit from different resolutions in different brain regions. Future work could explore using more parcels in the joint model to capture finer-grained resolutions where supported by the data.
- Permutation Procedure: The blocked permutation procedure is an approximation and future work could develop more advanced methods that better preserve spatial homogeneity, shape, and size.
- Model Assumptions: The SBM assumes independent connectivity profiles between modalities, which is a simplification. Future work could explore more advanced nonlinear modeling approaches (e.g., deep learning) to capture more complex relationships.
- Ground Truth: The lack of a definitive "ground truth" for optimal brain units makes evaluation qualitative. Future work could explore testing the framework on a third modality to see if the consensus representation better characterizes it.
- Functional Connectivity Metrics: The study primarily used zero-lag Pearson correlation. Future work could investigate other functional connectivity metrics (e.g., mutual information, wavelet correlation) and their impact.
- dMRI Noise: The inherent noise and indirect nature of dMRI data can affect structural connectivity inference. Future work could benefit from improved dMRI acquisition and processing techniques.
Summary updated