Abstract
Continuous improvement in medical imaging techniques allows the acquisition of higher-resolution images. When these are used in a predictive setting, a greater number of explanatory variables are potentially related to the dependent variable (the response). Meanwhile, the number of acquisitions per experiment remains limited. In such high dimension/small sample size setting, it is desirable to find the explanatory variables that are truly related to the response while controlling the rate of false discoveries. To achieve this goal, novel multivariate inference procedures, such as knockoff inference, have been proposed recently. However, they require the feature covariance to be well-defined, which is impossible in high-dimensional settings. In this paper, we propose a new algorithm, called Ensemble of Clustered Knockoffs, that allows to select explanatory variables while controlling the false discovery rate (FDR), up to a prescribed spatial tolerance. The core idea is that knockoff-based inference can be applied on groups (clusters) of voxels, which drastically reduces the problem's dimension; an ensembling step then removes the dependence on a fixed clustering and stabilizes the results. We benchmark this algorithm and other FDR-controlling methods on brain imaging datasets and observe empirical gains in sensitivity, while the false discovery rate is controlled at the nominal level.
Introduction
In neuroimaging, researchers often face a "curse of dimensionality." A single fMRI scan contains hundreds of thousands of voxels, but a study might only include a few dozen or hundred participants. When trying to determine which voxels are actually "predictive" of a behavior or disease, standard statistical tools often break down. A major challenge is controlling the False Discovery Rate (FDR)—the proportion of "discovered" voxels that are actually just noise. While Knockoff Inference provides elegant FDR control in multivariate settings, it requires estimating a covariance matrix that is impossible to compute when features far outnumber samples ($p \gg n$). Furthermore, it relies on randomly generated "fake" variables (knockoffs), leading to unstable results.
Methods
The authors introduce ECKO (Ensemble of Clustered Knockoffs), which solves these issues through a three-step pipeline involving dimension reduction, knockoff generation, and a double-layer ensembling approach.
- Dimension Reduction via Clustering: Instead of individual voxels, ECKO groups them into spatially constrained clusters. This reduces the problem to a size where knockoff inference is mathematically tractable.
- Clustered Knockoffs (CKO): The knockoff procedure is applied to these cluster averages. This involves creating "synthetic" versions of the clusters that mimic the correlation structure of the original data but are independent of the response variable.
- Double Ensembling: To fix the instability of random knockoffs and arbitrary cluster boundaries, ECKO uses two layers of aggregation. Internal ensembling runs multiple knockoff generations for each clustering, while external ensembling repeats the process across many different random clusterings to average scores at the voxel level.
Additionally, the framework introduces $\delta$-FDR, a "spatial tolerance" version of FDR. Since the method works with clusters, it controls the probability that a discovery is more than a distance $\delta$ away from a true signal.
Key Results and Insights
The researchers evaluated ECKO using synthetic 3D data, the Haxby Dataset (faces vs. houses), and the Oasis Dataset (age prediction from brain density).
- Superior Stability: In synthetic tests, a single run of Clustered Knockoffs (CKO) produced noisy results. ECKO consistently recovered ground truth regions with high precision through its aggregation layers.
- Robust FDR Control: Unlike simpler methods that failed to control the false discovery rate as the Signal-to-Noise Ratio (SNR) increased, ECKO remained reliable, keeping the FDR below the target 10% threshold.
- Real-World Sensitivity: On the Oasis dataset, ECKO was the only method to detect significant predictive regions at a strict 10% FDR level. Other state-of-the-art methods like ECDL and APT returned empty results, suggesting ECKO has higher statistical power in difficult, high-dimensional scenarios.
Conclusion
ECKO demonstrates that FDR control can be resolution-invariant: controlling the FDR of clusters effectively controls the FDR of the constituent voxels, provided a small spatial tolerance is accepted. By ensembling results across different random clusterings, the algorithm moves away from the arbitrariness of single solutions, leading to reproducible and sensitive brain maps. While computationally intensive, the approach is highly parallelizable and offers a robust framework for multivariate inference in high-dimensional medical imaging.
Summary updated