vix.ing · top · new · best · stats · spec

Electromagnetic neural source imaging under sparsity constraints with\n SURE-based hyperparameter tuning

2021/10/28 by Pierre‐Antoine Bannier, Bannier, Pierre-Antoine, Quentin Bertrand +5
Engineering · Computer Science · Medicine · #Sparse and Compressive Sensing Techniques #Distributed Sensor Networks and Detection Algorithms #Advanced MRI Techniques and Applications

paper · pdf · doi:10.48550/arxiv.2112.12178

Abstract

Estimators based on non-convex sparsity-promoting penalties were shown to\nyield state-of-the-art solutions to the magneto-/electroencephalography (M/EEG)\nbrain source localization problem. In this paper we tackle the model selection\nproblem of these estimators: we propose to use a proxy of the Stein's Unbiased\nRisk Estimator (SURE) to automatically select their regularization parameters.\nThe effectiveness of the method is demonstrated on realistic simulations and\n30 subjects from the Cam-CAN dataset. To our knowledge, this is the first\ntime that sparsity promoting estimators are automatically calibrated at such a\nscale. Results show that the proposed SURE approach outperforms\ncross-validation strategies and state-of-the-art Bayesian statistics methods\nboth computationally and statistically.\n

Related