[Paper Review] Small-sample Brain Mapping: Sparse Recovery on Spatially Correlated Designs with Randomization and Clustering

Gaël Varoquaux, Alexandre Gramfort+1 more|arXiv (Cornell University)|Jun 27, 2012

Sparse and Compressive Sensing Techniques27 references50 citations

TL;DR

This paper proposes a sparse recovery framework for small-sample functional MRI brain mapping by leveraging spatial clustering and randomization to improve variable selection under strong spatial correlation. By transforming original voxels into clustered features and using bootstrap resampling, the method enhances support recovery accuracy, demonstrating superior performance on simulated and real fMRI datasets compared to standard sparse regression.

ABSTRACT

Functional neuroimaging can measure the brain?s response to an external stimulus. It is used to perform brain mapping: identifying from these observations the brain regions involved. This problem can be cast into a linear supervised learning task where the neuroimaging data are used as predictors for the stimulus. Brain mapping is then seen as a support recovery problem. On functional MRI (fMRI) data, this problem is particularly challenging as i) the number of samples is small due to limited acquisition time and ii) the variables are strongly correlated. We propose to overcome these difficulties using sparse regression models over new variables obtained by clustering of the original variables. The use of randomization techniques, e.g. bootstrap samples, and clustering of the variables improves the recovery properties of sparse methods. We demonstrate the benefit of our approach on an extensive simulation study as well as two fMRI datasets.

Motivation & Objective

• Address the challenge of brain mapping in fMRI with limited samples and highly correlated spatial variables.
• Overcome poor recovery performance of standard sparse regression in high-dimensional, correlated designs.
• Improve support recovery in small-sample settings by exploiting spatial structure in neuroimaging data.
• Develop a robust framework combining clustering, randomization, and sparse regression for enhanced detection of active brain regions.
• Validate the method on both synthetic data and real fMRI datasets to demonstrate practical utility.

Proposed method

• Apply spatial clustering to group highly correlated voxels into super-voxels, reducing dimensionality and correlation structure.
• Transform the original design matrix into a new set of clustered variables to improve sparsity recovery.
• Use bootstrap resampling to generate multiple randomized datasets, enhancing stability and robustness of feature selection.
• Apply sparse regression (e.g., Lasso) on the clustered and randomized designs to identify relevant brain regions.
• Aggregate results across bootstrap samples to improve estimation reliability and reduce false positives.
• Integrate clustering and randomization to jointly improve the recovery of true active brain regions under small-sample constraints.

Experimental results

Research questions

• RQ1 Can clustering spatially correlated voxels improve sparse recovery in small-sample fMRI data?
• RQ2 Does randomization via bootstrap resampling enhance the stability and accuracy of brain mapping with limited samples?
• RQ3 How does the proposed method compare to standard sparse regression in terms of true positive and false positive detection rates?
• RQ4 To what extent does spatial clustering reduce the impact of high correlation in neuroimaging predictors?
• RQ5 Can the combination of clustering and randomization lead to more reliable identification of active brain regions in real fMRI datasets?

Key findings

• The proposed method significantly improves true positive detection rates in simulated fMRI data with high spatial correlation and small sample sizes.
• Clustering reduces the effective dimensionality and correlation structure, leading to more stable sparse recovery compared to unclustered designs.
• Randomization through bootstrap resampling enhances the robustness of feature selection, reducing variability in identified active regions.
• On real fMRI datasets, the method outperforms standard Lasso and other sparse regression baselines in identifying known brain activation patterns.
• The combination of clustering and randomization leads to lower false positive rates and improved reproducibility across multiple runs.
• The method demonstrates consistent performance gains across both synthetic and empirical fMRI data, confirming its utility in practical brain mapping applications.