Robust Epilepsy Classification via Integrating Plural Neural Network Weak Classifiers

View Statistics Email Alert RSS Feed

Information

Details

Project title

Code/計畫編號

NSC99-2221-E019-040

Translated Name/計畫中文名

整合複數個神經網路弱分類器之癲癇腦波判讀

Project Coordinator/計畫主持人

Jung-Hua Wang

Funding Organization/主管機關

National Science and Technology Council

Department/Unit

Department of Electrical Engineering

Website

https://www.grb.gov.tw/search/planDetail?id=2118286

Year

2010

Start date/計畫起

01-08-2010

Expected Completion/計畫迄

31-07-2011

Bugetid/研究經費

587千元

ResearchField/研究領域

資訊科學--軟體

腦波信號對於特定疾病具有獨特的特徵表現性，將特徵表現歸納出對應人體之疾病係眾多研究人員努力的目標。然而特定病徵如癲癇，其需處理的腦波資料量相當龐大且屬於非穩態(non-stationary)、非線性，故研發一套精確且客觀的癲癇腦波輔助診斷系統是兼具學術及應用價值。承續前期計畫已完成針對數個spatio-temporal神經網路模型(TDNN, SRNN, ACNN等)適用性篩選及HHT拆解多通道腦波信號，基於該研究成果與經驗累積，本期計畫擬進一步提出一整合複數個弱分類器之癲癇判讀方法，每一個弱分類器係指一具有不同延遲窗口之癲癇認知神經網路(ERNN，修改自TDNN)，其特點有三 (1)不同於SRNN以補零擴張輸入資料，吾人擬藉鏡射擴張以達到對神經網路的充分訓練(2)由於窗口的大小可以控制輸入資料的解析度(resolution)，具有等效量化的效果，藉此克服龐大的腦波資料量(3)藉由腦波訊號在希爾伯特頻譜能量上的表現選取數個延遲窗口大小，據以決定ERNN的數量。最後，使用Adaboost學習法將數個延遲窗口大小不同的ERNN結合成為一強建性癲癇腦波分類器(Robust epilepsy waves classifier)，由於係採用整合複數個弱分類器的架構，並非僅有固定之單一窗口大小，可有效克服腦波的非穩態、非線性。期能輔助醫師作客觀的診斷，提昇醫療品質。特別感謝財團法人長庚紀念醫院基隆院區-神經內科主任彭宗義醫師定期提供專業意見與病患腦波資料，藉此吾人得以將神經網路技術與醫療應用做跨領域的結合。Abstract-Brain disorders often present specific and unique EEG wave patterns, how to utilize this unique characteristic to assist medical diagnosis is an important research subject in medical field, and developing an objective and accurate epilepsy diagnosis system is both academically and practically potential. However, the key difficulty lies in that EEG wave associated with disorders such as epilepsy contains enormous (not to mention multiple channels), non-stationary and non-linear data. This project proposes a robust epilepsy classifying method constructed by integrating plural weak classifiers called ERNN (Epilepsy Recognition Neural Network), each with different delay window size. ERNN is based on modified TDNN in corporation with HHT (Hilbert-Huang Transformation) features. ERNN characterizes in threefold: (1) Unlike SRNN, ERNN can be sufficiently trained by a mirroring process (2) The problem of enormous brain data is dealt with changing resolution of input data through varying the delay window size (3) The energy distribution of Hilbert spectrum of an input EEG wave is used to determine the number of delay window, so is the number of ERNN.Finally, a robust epilepsy classifier is constructed by training plural ERNN with Adaboost training method, and it is expected that problems of non-stationary and non-linear inherent with epilepsy brain waves can be effectively overcome with various ERNNs, each with different window size and quantized input data.

Keyword(s)

癲癇
分類器
希爾伯特-黃轉換
時間延遲神經網路
AdaBoost 演算法
Epilepsy
classifier
Hilbert-Huang Transform (HHT)
Time delay neural network
Adaptive Boosting

DSpace CRIS

Robust Epilepsy Classification via Integrating Plural Neural Network Weak Classifiers

Details

Description