TL;DR
This paper introduces a novel approach to modeling brain functional interactions by allowing different brain network subnetworks (subdomains) to have varying model complexity. By using Bayesian optimization (TPE) to tune the depth of neural network branches for each subdomain, they achieve improved schizophrenia classification and reveal that brain regions implicated in the disorder require more complex models to extract discriminatory information.
What Problem Does This Paper Address?
Traditional machine learning approaches to neuroimaging data often treat all features or brain regions equally, assuming a uniform complexity in how they encode information. This is problematic because different brain networks likely interact and encode information with varying degrees of complexity. For instance, some interactions might be straightforward, while others involve intricate, non-linear relationships. This assumption of homogeneity can limit both predictive accuracy and our ability to understand which brain interactions are most important for a given condition. The paper specifically addresses this by investigating schizophrenia, a disorder known to involve widespread brain network disruptions, to see if a more flexible modeling approach can yield better insights and predictions.
Methods
The core of the methodology involves a multi-branched Multilayer Perceptron (MLP) architecture. Instead of a single, uniform network, this MLP has separate branches for different "subdomain interactions" (SDIs), which represent functional connectivity between pairs of brain networks. The key innovation is that the depth (number of hidden layers) of each branch can be independently optimized.
To navigate the vast search space of possible branch depths, the authors employ the Tree Parzen Estimator (TPE), a Bayesian optimization algorithm. TPE efficiently explores hyperparameter spaces by building probabilistic models of the objective function (in this case, classification accuracy) and using them to guide the search towards promising configurations. The process involves:
- scICA (spatially-constrained Independent Component Analysis): Used to identify functional brain networks (components) from fMRI data.
- Functional Connectivity Calculation: Static correlations between the time courses of these components are computed to create a functional connectivity matrix.
- Subdomain Interaction (SDI) Definition: These connectivity features are grouped into 28 SDIs, representing intra-network (within a network) and inter-network (between networks) connections.
- Multi-branched MLP Architecture: Each SDI serves as input to a dedicated branch of the MLP. The number of fully connected hidden layers in each branch is the hyperparameter to be optimized.
- TPE for Hyperparameter Optimization: TPE is used to find the optimal number of hidden layers (0, 1, or 2) for each of the 28 SDI branches, aiming to maximize classification accuracy.
- Model Analysis: After optimization, the learned architectures are analyzed to understand which SDIs require more complex models (deeper branches) and how this relates to their importance in classification.
Datasets
The study utilizes two independent datasets:
- fBIRN (Function Biomedical Informatics Research Network) dataset: A large dataset with both healthy controls (HC) and individuals with schizophrenia (SZ).
- COBRE (Center of Biomedical Research Excellence) dataset: Another dataset also containing HC and SZ subjects.
Both datasets consist of resting-state fMRI data. The authors ensured that HC and SZ subjects were matched by age and gender to minimize confounding factors.
Key Insights and Ideas
- Brain Network Interactions Vary in Complexity: The central hypothesis, supported by the results, is that different functional brain network interactions (SDIs) encode information with varying levels of complexity. This means a "one-size-fits-all" model complexity for all brain interactions is suboptimal.
- Adaptive Architectures Improve Performance: Allowing model complexity to vary per subdomain interaction leads to better classification performance compared to models with uniform complexity across all branches. This highlights the benefit of flexible, adaptive deep learning architectures for neuroimaging.
- Schizophrenia-Related Networks Need Deeper Models: Brain regions and their interactions known to be affected in schizophrenia (e.g., cognitive control, visual, and sensorimotor networks) consistently require more complex (deeper) models to extract discriminatory information. This suggests that the underlying neural mechanisms in these affected networks are more intricate.
- Intra-network vs. Inter-network Complexity: Interestingly, intra-network connections (within a single brain network) generally required shallower models, even though these connections are also implicated in schizophrenia. This suggests that while intra-network disruptions are important, the way information is encoded might be less complex than in certain inter-network interactions.
- Cross-Dataset Consistency: The patterns of required model depth for specific SDIs were remarkably consistent across the two independent datasets, lending strong support to the findings and suggesting generalizable principles about brain network complexity.
Key Results
- Improved Classification Accuracy: The TPE-optimized multi-branched MLP achieved a mean prediction accuracy of 0.81 for fBIRN and 0.78 for COBRE on held-out test data. This was significantly higher (p < 0.05) than baseline methods, including uniformly branched MLPs (UNIF0, UNIF1) and a non-branched MLP (MLP1).
- Consistent Feature Importance: The importance of SDIs learned by the TPE-optimized architecture (in the fusion layer) was highly correlated (0.9 for fBIRN, 0.81 for COBRE) with the importance of the same SDIs in the input layer (functional connectivity features). This indicates that the learned representations are meaningful and retain the predictive power of the original features.
- Identified Complex Subdomains: Connections involving the cognitive control (CC) to visual (VIS) and sensorimotor (SM) subdomains consistently required deeper models, suggesting more complex information encoding. Auditory (AUD) to VIS subdomain connections also showed this trend.
- Consistent Depth Patterns: For 19 out of 28 SDIs, the required model depth was identical across both the fBIRN and COBRE datasets, a statistically significant finding (p = 2.09 × 10⁻⁴) indicating a common underlying complexity structure.
Limitations and Future Work
Limitations:
- The study focused on a specific number of hidden layers (0, 1, or 2) for each branch. While this kept the search space manageable, exploring a wider range of depths or other architectural variations could yield further improvements.
- The analysis primarily focused on functional connectivity. Integrating structural connectivity or other modalities could provide a more comprehensive understanding.
Future Work:
- Extend to Other Disorders: Apply this adaptive modeling approach to other brain disorders to identify domain-specific complexity patterns.
- Explore More Optimization Methods: Investigate other hyperparameter optimization techniques beyond TPE.
- Increase Model Flexibility: Introduce more types of flexibility in the frameworks, such as automatically determining model structure/depth and incorporating additional measures of model complexity.
- Multimodal Integration: Extend these frameworks to exploit complex interrelationships from subdomains in a multimodal setting (e.g., combining fMRI with structural MRI or EEG).
Summary updated