Paper Abstract and Keywords |
Presentation |
2020-03-03 09:00
[Poster Presentation]
EpiNet: Convolutional Neural Network for Epileptic Seizure Localization from Interictal Intracranial EEG Kosuke Mori, Kosuke Fukumori, Toshihisa Tanaka (TUAT), Yasushi Iimura, Takumi Mitsuhashi, Hidenori Sugano (Juntendo Univ.) EA2019-157 SIP2019-159 SP2019-106 |
Abstract |
(in Japanese) |
(See Japanese page) |
(in English) |
The electroencephalogram (EEG) recording is necessary for epileptic diagnosis. In particular, the intracranial EEG (iEEG) data is essential to localize the seizure onset zone (SOZ). However, it is time-consuming and heavy load for specialists to interpret the iEEG recorded for long time. Therefore, an automatic technology to localize the SOZ is required. Based on VGG, the well-known convolutional neural network for image recognition, we propose a model for analyzing one-dimensional signals and try to localize the SOZ. For this model, the input is a signal obtained by dividing the iEEG in a short time, and the output is whether the recording area of the iEEG is the SOZ or not. Class-Balanced Focal Loss, which is reported to be effective for imbalanced data, was used as the loss function for training. We conducted time-series split cross-validation for the iEEG from cortical dysplasia epilepsy patients. As a result, the proposed model achieved generally higher AUC, F-measure, sensitivity, and specificity than the conventional method using feature extraction and SVM. In consequence, it is possible to outperform the conventional method by an appropriate neural network and loss function. |
Keyword |
(in Japanese) |
(See Japanese page) |
(in English) |
Epilepsy / Seizure onset zone / Electroencephalogram / Supervised learning / Convolutional Neural Network / / / |
Reference Info. |
IEICE Tech. Rep., vol. 119, no. 440, SIP2019-159, pp. 325-330, March 2020. |
Paper # |
SIP2019-159 |
Date of Issue |
2020-02-24 (EA, SIP, SP) |
ISSN |
Print edition: ISSN 0913-5685 Online edition: ISSN 2432-6380 |
Copyright and reproduction |
All rights are reserved and no part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. Notwithstanding, instructors are permitted to photocopy isolated articles for noncommercial classroom use without fee. (License No.: 10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
Download PDF |
EA2019-157 SIP2019-159 SP2019-106 |
|