Page 146 - MEGIN Book Of Abstracts - 2023

data underwent subsequent time-domain averaging spikes. Neurophysiologists visually identify spikes from
and source localization of left and right primary motor the MEG waveforms and estimate the equivalent cur-
(M1) and somatosensory (S1) cortices was conducted rent dipoles (ECD). However, presently, these analyses
using a single equivalent dipole model. Successful are manually performed by neurophysiologists and
mapping was determined based on dipole goodness are time-consuming. Another problem is that spike
of fit metrics ∼ 95%, as well as an accurate and conceiv- identification from MEG waveforms largely depends
able spatial correspondence to precentral and postcen- on neurophysiologists' skills and experiences. These
tral gyri for M1 and S1, respectively. Our results suggest problems cause poor cost-effectiveness in clinical MEG
that mapping M1 in epilepsy and tumor patients was examination. To overcome these problems, we fully
on average 94.5% successful, when patients only com- automated spike identification and ECD estimation us-
pleted motor mapping protocols. In contrast, mapping ing a deep learning approach fully automated AI-based
S1 was successful 45-100% of the time in these patient MEG interictal epileptiform discharge identification
groups when they only completed somatosensory and ECD estimation (FAMED). We applied a semantic
mapping paradigms. Importantly, Z-tests for inde- segmentation method, which is an image processing
pendent proportions revealed that the percentage of technique, to identify the appropriate times between
successful S1 mappings significantly increased to ∼ 94% spike onset and peak and to select appropriate sensors
in epilepsy patients who completed both motor/so- for ECD estimation. FAMED was trained and evaluated
matosensory mapping protocols during MEG. Together, using clinical MEG data acquired from 375 patients.
these data suggest that ordering more comprehensive FAMED training was performed in two stages: in the
mapping procedures (e.g., both motor and somatosen- first stage, a classification network was learned, and
sory protocols for a collective sensorimotor network) in the second stage, a segmentation network that
may substantially increase the accuracy of presurgical extended the classification network was learned. The
functional mapping by providing more extensive data classification network had a mean AUC of 0.9868 (10-
from which to base interpretations. Moreover, clinicians fold patient-wise cross-validation); the sensitivity and
and magnetoencephalographers should be considerate specificity were 0.7952 and 0.9971, respectively. The
of the major contributors to mapping failures (i.e., low median distance between the ECDs estimated by the
SNR, excessive motion and magnetic artifacts) in order neurophysiologists and those using FAMED was 0.63
to further increase the percentage of cases achieving cm. Thus, the performance of FAMED is comparable to
successful mapping of eloquent cortices. that of neurophysiologists, and it can contribute to the
efficiency and consistency of MEG ECD analysis.
Keywords: Epilepsy, Magnetoencephalography, Postcen-
tral gyrus, Precentral gyrus, Presurgical mapping, Tumor IEEE transactions on medical imaging (2022), Vol. 41, No.
10 (35536808) (2 citations)
NeuroImage. Clinical (2022), Vol. 35 (35597033) (1 citation)

Classification of EEG Signal-Based Encephalon
Fully-Automated Spike Detection and Dipole Magnetic Signs for Identification of Epilepsy-Based
Analysis of Epileptic MEG Using Deep Learning Neurological Disorder (2022)
(2022)
Kaur, Arshpreet; Gupta, Suneet; Kathiravan, M;
Hirano, Ryoji; Emura, Takuto; Nakata, Otoichi; Nasrullah, Syed; Paul, Chayan; Rahin, Saima Ahmed
Nakashima, Toshiharu; Asai, Miyako; Kagitani-Shimono,
Kuriko; Kishima, Haruhiko; Hirata, Masayuki GNA University, Village Hargobindgarh, Phagwara, Punjab,
India; Department of CSE, School of Engineering and Tech-
ABSTRACT Magnetoencephalography (MEG) is a useful nology, Mody University, Lakshmangarh, Rajasthan, 332311,
tool for clinically evaluating the localization of interictal India; Department of Computer Science and Engineering,

 ontents  Index 125
C