Brain Source Reconstruction Solution Quality Assessment with Spatial Graph Frequency Features

Meng Jiao, Feng Liu, Onur Asan, Roshanak Nilchiani, Xinglong Ju, Jing Xiang

Abstract

Different EEG/MEG source imaging (ESI) algorithms can render different reconstructions, so as to the same algorithm with different hyperparameters. Moreover, we found the locations of active sources also have an impact on the performance of ESI algorithms. For the real EEG/MEG source reconstruction, as the ground true activation is unknown, it is hard to validate which algorithm performs better. In this paper, we proposed to use statistical features from source space to predict whether the reconstruction is a satisfactory solution. The training data and testing data are from solutions from different algorithms based on synthetic EEG data where ground truth activations are available. The good and bad solutions are determined by Area Under Curve (AUC) and localization error (LE). We extract spatial and general statistical features from solutions, then we used machine learning models to classify good vs. bad solutions, and showed the feasibility of judging the quality of solution without knowing ground truth, which can serve as a feedback for further hyperparameter tuning.
BibTeX
@inproceedings{jiao2022brain,
  author = {Meng Jiao and Feng Liu and Onur Asan and Roshanak Nilchiani and Xinglong Ju and Jing Xiang},
  title = {{Brain Source Reconstruction Solution Quality Assessment with Spatial Graph Frequency Features}},
  booktitle = {BI},
  year = {2022},
  doi = {10.1007/978-3-031-15037-1\_15},
}

Loading related content...

Figure