Related Experiment Video
Updated: Jan 7, 2026
Multi-Tracer Studies of Brain Oxygen and Glucose Metabolism Using a Time-of-Flight Positron Emission Tomography-Computed Tomography Scanner
Published on: June 7, 2024
Which fMRI clustering gives good brain parcellations?
Bertrand Thirion1, Gaël Varoquaux1, Elvis Dohmatob1
1Parietal Project-Team, Institut National de Recherche en Informatique et Automatique Palaiseau, France ; Commissariat à l'énergie Atomique et Aux Énergies Alternatives, DSV, Neurospin, I2 BM Gif-sur-Yvette, France.
This study compares brain parcellation clustering techniques, finding Ward's method superior for accuracy and reproducibility in neuroimaging. The optimal model involves a trade-off between accuracy and stability.
Area of Science:
- Neuroimaging
- Computational Neuroscience
- Brain Mapping
Background:
- Neuroimaging analysis requires dividing the brain into homogeneous regions (parcels).
- Predefined atlases lack individual subject adaptability.
- Clustering techniques offer data-driven parcellation for improved signal homogeneity.
Purpose of the Study:
- To determine the most appropriate clustering technique for brain parcellation.
- To optimize clustering models based on accuracy and reproducibility.
- To evaluate clustering algorithms on simulated and functional Magnetic Resonance Imaging (fMRI) data.
Main Methods:
- Comparison of Ward, spectral, and k-means clustering algorithms.
- Assessment using goodness of fit (accuracy) and cross-validation reproducibility criteria.
- Application to simulated data and two task-based fMRI datasets.
Main Results:
- Ward's clustering generally outperformed spectral and k-means in both accuracy and reproducibility.
- The criteria for accuracy and reproducibility diverged, favoring more conservative solutions for reproducibility.
- A trade-off between parcellation accuracy and stability is necessary for practical decision-making.
Conclusions:
- Ward's clustering is a robust method for neuroimaging parcellation.
- Optimizing parcellation requires balancing accuracy with reproducibility.
- The choice of clustering model depends on the desired balance between fit and stability.
