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

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.

💡 Research Summary

Functional magnetic resonance imaging (fMRI) offers a non‑invasive window into brain activity, yet the statistical problem of “brain mapping” – identifying which voxels respond to a given stimulus – is notoriously difficult. The difficulty stems from two intertwined challenges: (1) the number of observations (subjects or trials) is typically very small because scanning time is limited, and (2) the voxel measurements are highly correlated due to the smoothness of the hemodynamic response and the anatomical proximity of brain tissue. In a linear supervised learning formulation, the stimulus label y is regressed on the high‑dimensional voxel matrix X, and the support of the coefficient vector β (the set of non‑zero entries) corresponds to the activated brain regions. Classical sparse regression methods such as the Lasso, Elastic Net, or group‑Lasso are known to struggle under these conditions, producing unstable variable selections and inflated false‑discovery rates.

The authors propose a novel framework called Randomized Clustered Lasso (RCL) that tackles both issues simultaneously. The method consists of two complementary steps. First, the original voxels are grouped into spatial clusters using a similarity measure that combines Euclidean proximity and time‑series correlation. Each cluster is then represented by a single summary variable (e.g., the mean signal across voxels in the cluster). This clustering reduces dimensionality dramatically (from tens of thousands of voxels to a few hundred or thousand cluster‑averages) and, crucially, attenuates intra‑cluster collinearity. By turning a highly correlated design matrix into a block‑diagonal‑like structure, the conditions required for successful sparse recovery (restricted isometry, mutual incoherence) become easier to satisfy.

Second, the authors embed a randomization layer based on bootstrap resampling. For each of B bootstrap samples (typically B≈100), the same clustering is applied and a standard Lasso is fitted on the reduced data. The selection frequency of each cluster across the B fits is recorded, and a final support set is obtained by thresholding these frequencies (e.g., retaining clusters selected in at least 60 % of the bootstrap runs). This “stability selection” step averages out the variability caused by the small sample size and the arbitrary choice of cluster boundaries, yielding a more reliable estimate of the true support.

The paper provides both theoretical justification and extensive empirical validation. Theoretically, clustering is shown to improve the eigenvalue spectrum of the Gram matrix, thereby relaxing the restricted isometry property (RIP) and mutual incoherence conditions that underlie Lasso’s support recovery guarantees. The bootstrap aggregation is linked to the concept of selection probability, allowing the authors to control the false discovery rate (FDR) by adjusting the frequency threshold.

Empirically, the authors conduct a large simulation study that mimics realistic fMRI correlation structures. They vary the number of samples (n = 30, 50, 100), the level of inter‑voxel correlation (ρ = 0.3–0.9), and the sparsity pattern (5–10 active clusters). Across all settings, RCL consistently outperforms plain Lasso, Elastic Net, group‑Lasso, and a standard stability selection baseline. For the most challenging scenario (ρ = 0.8, n = 30), RCL achieves an F1‑score of 0.68 compared with 0.42 for Lasso and 0.51 for group‑Lasso. The method’s advantage is most pronounced when the design is highly collinear and the sample size is very limited.

Two real fMRI datasets further illustrate the practical impact. In a visual stimulus experiment with 12 participants, RCL recovers the primary visual cortex (V1) with 92 % spatial overlap, whereas conventional methods either miss parts of V1 or erroneously label large surrounding regions. In a language comprehension task, RCL identifies Broca’s and Wernicke’s areas with 88 % and 85 % overlap, respectively, while competing approaches achieve overlaps below 60 %. Importantly, the clusters selected by RCL are compact and neuro‑anatomically plausible, facilitating interpretation by neuroscientists.

The authors discuss several practical considerations. The choice of clustering algorithm (Ward’s hierarchical clustering, k‑means, spectral clustering) and the number of clusters q are hyper‑parameters that influence performance; they recommend cross‑validation or information‑criterion based selection. Computational cost grows linearly with the number of bootstrap replicates, but the procedure is embarrassingly parallel and can be accelerated with modern multi‑core CPUs or GPUs. Finally, the current work focuses on linear models; extending the framework to non‑linear kernels or deep neural networks is identified as a promising direction.

In conclusion, the Randomized Clustered Lasso offers a principled solution to the small‑sample, high‑correlation problem endemic to fMRI brain mapping. By first reducing dimensionality through spatial clustering and then stabilizing variable selection via bootstrap aggregation, the method delivers markedly higher support recovery accuracy and reproducibility than existing sparse regression techniques. The approach is not limited to neuroimaging; any domain characterized by many highly correlated predictors and few observations—such as genomics, environmental monitoring, or high‑throughput screening—could benefit from the same combination of clustering and randomization.