TY - GEN
T1 - Brain Source Reconstruction Solution Quality Assessment with Spatial Graph Frequency Features
AU - Jiao, Meng
AU - Liu, Feng
AU - Asan, Onur
AU - Nilchiani, Roshanak
AU - Ju, Xinglong
AU - Xiang, Jing
N1 - Publisher Copyright:
© 2022, Springer Nature Switzerland AG.
PY - 2022
Y1 - 2022
N2 - 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.
AB - 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.
KW - Brain source imaging
KW - EEG/MEG
KW - Graph fourier transform
KW - Solution quality
UR - http://www.scopus.com/inward/record.url?scp=85136999863&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85136999863&partnerID=8YFLogxK
U2 - 10.1007/978-3-031-15037-1_15
DO - 10.1007/978-3-031-15037-1_15
M3 - Conference contribution
AN - SCOPUS:85136999863
SN - 9783031150364
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 175
EP - 183
BT - Brain Informatics - 15th International Conference, BI 2022, Proceedings
A2 - Mahmud, Mufti
A2 - He, Jing
A2 - Vassanelli, Stefano
A2 - van Zundert, André
A2 - Zhong, Ning
T2 - 15th International Conference on Brain Informatics, BI 2022
Y2 - 15 July 2022 through 17 July 2022
ER -