TL;DR
This paper introduces a novel method for creating unified brain atlases by aggregating individual brain connectivity maps. Their "consensus clustering" approach outperforms existing anatomical atlases in accurately representing brain connectivity, predicting biological sex, and achieving better inter-hemispheric symmetry.
What Problem Does This Paper Address?
Understanding the human brain's intricate network of connections (the connectome) is crucial for neuroscience. However, analyzing these complex networks often requires simplifying them into distinct brain regions, known as parcellation. Existing methods for creating these brain atlases have limitations: population-based atlases can obscure individual differences, while individual-level parcellations lack a common framework for comparison. This paper addresses the challenge of creating a unified, yet individually-informed, brain atlas that accurately reflects the brain's connectivity.
Methods
The authors propose a two-stage approach:
- Hierarchical Clustering of Individual Connectomes: They start with a "continuous connectivity" representation of individual brain networks. This representation is then subjected to a hierarchical clustering process (specifically, the Louvain algorithm) to create multiple levels of parcellations for each individual brain. This results in a three-level hierarchical partition for each subject.
- Consensus Clustering: To create a unified atlas, they aggregate the individual parcellations. They employ two methods for this "consensus clustering":
- Cluster-based Similarity Partitioning Algorithm (CSPA): This method defines similarity between brain regions based on how often they are grouped together across individual parcellations.
- Greedy Approach (Hard Ensemble - HE): This method iteratively combines individual partitions to find a "pseudo-average" partition that minimizes the sum of distances to all individual partitions. The distance metric used is the average squared distance between membership functions.
The quality of the resulting parcellations is assessed using Kullback-Leibler (KL) and Jensen-Shannon (JS) divergences against the original dense connectome, inter-hemispheric symmetry, and performance in a binary sex classification task.
Datasets
The study utilizes data from 400 subjects from the Human Connectome Project (HCP) S900 release. This dataset is used for both training (deriving individual connectomes and parcellations) and evaluation (testing the consensus parcellations). No new benchmarks are introduced; rather, the performance is compared against the standard Desikan-Killiany anatomical atlas.
Key Insights and Ideas
- Data-driven atlases can capture biological relevance: By basing the parcellation directly on connectivity data, the resulting atlas shows improved performance in predicting biological sex, suggesting it captures functionally relevant distinctions.
- Consensus clustering bridges individual and group levels: The method effectively balances the need for a unified atlas with the preservation of individual connectivity patterns, offering a more nuanced representation than purely population-based atlases.
- Hierarchical structure is informative: The hierarchical clustering approach reveals multiple levels of brain organization, and the consensus method can leverage these different levels for optimal atlas generation.
- Symmetry emerges naturally: Remarkably, the parcellation method, which is purely driven by graph connectivity and has no explicit knowledge of brain topology, produces highly symmetrical parcellations between the left and right hemispheres.
Key Results
The greedy consensus clustering approach, particularly at a hierarchical depth of 3 (referred to as HE PIII), consistently outperformed other methods, including the standard Desikan-Killiany atlas, across several metrics:
- Connectivity Representation: Achieved lower KL and JS divergences, indicating a better approximation of the dense connectome.
- Biological Sex Classification: Demonstrated higher ROC AUC scores (0.86 vs. 0.81 for DKT), suggesting better capture of sex-related connectivity differences.
- Inter-hemispheric Symmetry: Showed higher symmetry scores (0.66 vs. 0.64 for DKT).
- Ensemble Goodness: The consensus parcellations also showed high similarity to individual partitions, indicating good agreement.
- Reproducibility: When splitting the dataset into two halves, the resulting consensus parcellations remained highly similar to each other and to the full consensus parcellation.
Limitations and Future Work
- Computational Cost: While not explicitly detailed as a limitation, consensus clustering can be computationally intensive, especially with large datasets and complex hierarchical structures.
- Choice of Distance Metric: The performance of consensus clustering can be sensitive to the chosen distance metric between partitions. The paper uses average squared distance, and exploring other metrics could be a future direction.
- Interpretation of Hierarchical Levels: While the paper shows that different hierarchical levels have varying performance, further investigation into the specific functional or anatomical interpretations of these levels could be beneficial.
- Generalizability to Other Modalities: The current approach is focused on structural connectivity. Extending this framework to functional connectivity or multi-modal data could be a promising avenue for future research.
Summary updated