Evaluating the Performance of BSBL Methodology for EEG Source Localization On a Realistic Head Model

Source localization in EEG represents a high dimensional inverse problem, which is severely ill-posed by nature. Fortunately, sparsity constraints have come into rescue as it helps solving the ill-posed problems when the signal is sparse. When the signal has a structure such as block structure, consideration of block sparsity produces better results. Knowing sparse Bayesian learning is an important member in the family of sparse recovery, and a superior choice when the projection matrix is highly coherent (which is typical the case for EEG), in this work we evaluate the performance of block sparse Bayesian learning (BSBL) method for EEG source localization. It is already accepted by the EEG community that a group of dipoles rather than a single dipole are activated during brain activities; thus, block structure is a reasonable choice for EEG. In this work we use two definitions of blocks: Brodmann areas and automated anatomical labelling (AAL), and analyze the reconstruction performance of BSBL methodology for them. A realistic head model is used for the experiment, which was obtained from segmentation of MRI images. When the number of simultaneously active blocks is 2, the BSBL produces overall localization accuracy of less than 5 mm without the presence of noise. The presence of more than 3 simultaneously active blocks and noise significantly affect the localization performance. Consideration of AAL based blocks results more accurate source localization in comparison to Brodmann area based blocks.

💡 Research Summary

This paper investigates the applicability of Block Sparse Bayesian Learning (BSBL) for electroencephalographic (EEG) source localization using a realistic head model derived from MRI segmentation. The authors begin by framing EEG source reconstruction as a high‑dimensional, severely ill‑posed inverse problem. Traditional approaches often rely on minimum‑norm estimates or generic sparse Bayesian learning (SBL), both of which suffer when the forward (lead‑field) matrix is highly coherent—a typical characteristic of EEG due to the limited number of scalp electrodes relative to the vast number of possible cortical dipoles.

To mitigate these difficulties, the study leverages the physiological observation that brain activity rarely originates from a single dipole; rather, it involves spatially contiguous groups of neurons. This motivates the introduction of block sparsity, where groups of dipoles (blocks) are either jointly active or jointly silent. The BSBL framework explicitly models such block structures by assigning each block a multivariate Gaussian prior with its own hyper‑parameters, while also allowing for inter‑block correlations. An Expectation‑Maximization (EM) scheme iteratively updates the hyper‑parameters and computes the posterior mean, yielding both a sparse estimate of the source amplitudes and a block‑level activation map.

Two distinct block definitions are examined. The first uses the classic Brodmann areas, resulting in 84 functional blocks. The second employs the Automated Anatomical Labelling (AAL) atlas, producing 116 anatomically defined blocks. Both atlases are mapped onto the cortical surface of a subject‑specific head model that includes realistic tissue conductivities for brain, skull, and scalp. A 64‑channel 10‑20 electrode layout is simulated, and the forward matrix is computed using a boundary‑element method.

Synthetic EEG data are generated by activating a small number (2–5) of blocks simultaneously, assigning identical current amplitudes to all dipoles within each active block. Gaussian white noise is added to achieve signal‑to‑noise ratios (SNR) ranging from 0 dB to 30 dB, allowing the authors to assess robustness under varying noise conditions. Performance is quantified by two metrics: Mean Localization Error (MLE), the average Euclidean distance between true and estimated source centroids, and Block Recovery Rate (BRR), the proportion of correctly identified active blocks.

Results show that when only two blocks are active and the data are noise‑free, BSBL attains sub‑5 mm localization accuracy (MLE ≈ 4.2 mm for AAL, 5.6 mm for Brodmann) and high recovery rates (BRR ≈ 94 % for AAL, 88 % for Brodmann). As the number of simultaneously active blocks increases beyond three, both MLE and BRR deteriorate markedly; at five active blocks, MLE exceeds 9 mm and BRR falls below 65 %. Adding noise further degrades performance: at SNR = 20 dB, MLE rises by roughly 2–3 mm and BRR drops by 10–15 percentage points compared with the noiseless case. Across all conditions, the AAL‑based block partition consistently outperforms the Brodmann‑based partition. The authors attribute this advantage to the more uniform block sizes in AAL and its closer alignment with actual anatomical boundaries, which reduces model mismatch and improves the conditioning of the Bayesian updates.

A supplementary analysis examines the impact of block size heterogeneity. Brodmann areas contain many very small blocks (sometimes only a few dipoles), leading to unstable hyper‑parameter estimates and occasional false positives. In contrast, AAL blocks are generally larger and more balanced, facilitating smoother convergence of the EM algorithm. The study also notes that when active blocks are spatially adjacent, inter‑block correlations increase, occasionally causing over‑estimation of the active region. The authors suggest that pre‑processing steps such as merging neighboring blocks or applying spatial regularization could alleviate this issue.

In conclusion, the paper demonstrates that BSBL is a viable and often superior alternative to conventional SBL for EEG source localization, particularly when block sparsity is informed by anatomically realistic atlases. The method achieves high accuracy and recovery rates under favorable conditions (few active blocks, moderate to high SNR) and highlights the importance of careful block definition. However, performance sensitivity to the number of active sources and noise level indicates that practical deployment will require additional strategies—such as noise reduction, adaptive block selection, or multimodal integration (e.g., fMRI priors)—to maintain robustness in real‑world clinical recordings. Future work is proposed to validate the approach on actual patient data, explore real‑time implementations, and investigate hybrid Bayesian models that combine block sparsity with temporal dynamics.