本研究的目的在於開發利用卷積神經網路模型分析心電圖完整波形，並利用完整波形來進行正常波形的判斷，輔助醫師加速判斷特徵波形的時間。研究方法中首先直接採用MIT-BIH資料庫中心電圖原始訊號，將原始訊號以圖片的型式存儲，並把資料庫中所有正常波形、VT病徵波形與非標準正常波形經過整理和收集，放入卷積神經網路學習特徵，最後建立判斷為正常、VT病徵與非標準正常波形的模型，未來只需要將病人心電圖以圖片的方式輸入已訓練好的卷積神經網路模型作分析，即可判斷是否為正常、VT病徵或非標準正常波形。利用harris角點檢測法擷取特徵波形，並將其標籤，使用AlexNet與GoogLeNet兩種神經網路架構進行卷積神經網路模型建立，探討訓練樣本數對模型訓練的設定及比較其結果，發現GoogLeNet模型比AlexNet模型於本研究所應用之心電圖圖片，具有比較好的準確率。

The purpose of this research is to develop convolutional neural network models to analyze the complete waveform of electrocardiogram.Use the complete waveform to determine the normal waveform, and to help doctors speed up the time to determine the characteristic waveforms. On the research method, the original signal of the central electrogram of the MIT-BIH database is directly used, and the original signal is stored in the form of a photo picture. All normal waveforms, VT symptom waveforms and non-standard normal waveforms are sorted and collected in the database.This study analyze neural networks to learn features, and finally build a model that can judge waveform classification as normal, VT symptoms and non-standard normal waveforms. In the future, We only need to calibrate the patient's ECG and transform ECG to picture format that trained convolutional neural network model to determine whether it is normal, VT symptoms or non-standard normal waveforms. By using harris corner detection method, one can obtain characteristic waveforms, label them, use two neural network architectures (AlexNet and GoogLeNet ) to build models, and discuss the number of training samples for model training settings. In comparison with AlexNet and GoogLeNet structures, I found that the GoogLeNet model is better than the AlexNet model. The GoogLeNet model accuracy is better than the AlexNet model.

摘　要 i
ABSTRACT ii
致　謝 iv
目　錄 v
表目錄 viii
圖目錄 x
第一章 緒論 1
1.1 研究背景 1
1.2 研究動機與目的 2
1.3 相關文獻探討 3
1.4 論文架構 8
第二章 實驗原理介紹 9
2.1 心臟概述 9
2.1.1 心臟相關生理訊號 10
2.2 心電圖 10
2.2.1 心電圖導程 11
2.2.2 心電圖波形 14
2.3 Harris角點檢測法 16
2.4 局部最小值 20
2.5 圖像角點閾值選取 21
2.6 卷積神經網路 22
2.6.1 卷積層(Convolutional Layer) 23
2.6.2 激活函數層(Activation Layer) 24
2.6.3 池化層(Pooling Layer) 25
2.6.4 完全連接層(Fully Connected Layer) 26
2.6.5 損失函式層(Loss Function Layer) 27
2.7 K-摺交叉驗證(K-fold Cross-Validation) 28
2.8 混淆矩陣(Confusion Matrix) 29
2.9 欠擬合(Underfitting)與過擬合(Overfitting) 31
第三章 實驗架構與研究方法 32
3.1 實驗架構 32
3.2 分析軟體 33
3.3 ECG資料庫 33
3.4 資料庫讀取 37
3.5 角點特徵標記 40
3.6 PQRST特徵波形標記 41
3.6.1 R波、Q波與S波標記 41
3.6.2 P波、P波起始點標記 44
3.6.3 T波標記 45
3.7 特徵波形擷取 46
3.7.1 正常波形 46
3.7.2 VT病徵波形 47
3.7.3 非標準正常波形 48
3.8 卷積神經網路模型建立 50
第四章 實驗結果與分析 52
4.1 訓練樣本 52
4.2 交叉驗證設定 58
4.3 卷積神經網路模型訓練結果 59
4.3.1 AlexNet 59
4.3.2 GoogLeNet 63
4.4 心電圖實際應用辨識 66
第五章 結論與未來展望 73
5.1 心電圖特徵波形擷取 73
5.2 訓練樣本數與模型比較結果 74
5.3 未來展望 75
參考文獻 76

[1]世界衛生組織-心血管疾病, 2020/06/06取自：
https://www.who.int/en/newsroom/factsheets/detail/cardiovascular-diseases-(cvds)
[2]衛生福利部，107年國人死因統計結果，2020/6/6取自:
https://www.mohw.gov.tw/cp-16-48057-1.html
[3]Chris Harris and Mike Stephens,"A Combined Corner and Edge Detector," Proceedings of the 4th Alvey Vision Conference,1988, pp.147-151.
[4]S.S. Mehta and N.S. Lingayat,"SVM-based algorithm for recognition of QRS complexes in electrocardiogram",ScienceDirect,IRBM Volume 29,Issue 5,2008,pp.310-317.
[5]白宗學，以心電圖辨識為應用之脈衝類神經網路波形分類架構，國立臺灣大學碩士，2013。
[6]Mohamed Hammad,Shanzhuo Zhang and Kuanquan Wang,"A novel two-dimensional ECG feature extraction and classification algorithm based on convolution neural network for human authentication", ScienceDirect,Future Generation Computer Systems,Volume 101,2019,pp.180-196.
[7]Shuang Wang,Shugang Zhang,Zhen Li,Lei Huang and Zhiqiang Wei, "Automatic digital ECG signal extraction and normal QRS recognition from real scene ECG images ",Computer Methods and Programs in Biomedicine,Volum 187,2020.
[8]常見的心臟病症狀，2020/07/15取自:
https://blog.xuite.net/tracychchen/twblog1/124444065%E5%B8%B8%E8%A6%8B%E7%9A%84%E5%BF%83%E8%87%9F%E7%97%85%E7%97%87%E7%8B%80
[9]心臟，2020/07/15取自: http://www.hkpe.net/hkdsepe/human_body/heart.htm
[10]林晏廷，基於心電圖之心臟疾病診斷與專家系統研究，國立臺北科技大學碩士，2019。
[11]張敏軒，12導程影像模式心電圖系統改進與疾病資料庫建立，國立陽明大學碩士，2003。
[12]ECG Learning Center，2020/07/15取自: https://ecg.utah.edu/lesson/1
[13]心電圖，2020/07/15取自: https://smallcollation.blogspot.com/2013/07/electrocardiogram-ekgecg.html#gsc.tab=0
[14]Agateller (Anthony Atkielski),Schematic diagram of normal sinus rhythm for a human heart as seen on ECG (with English labels),2007。
[15]邱琪雅，基於小波轉換之單一導程心電圖重構12導程心電圖分類，國立中央大學碩士，2018。
[16]Hans P. Moravec, " Obstacle Avoidance and Navigation in the Real World by a Seeing Robot Rover,"Computer Science Department Stanford University Ph.D.,1980.
[17]Matlab，突出度，2020/06/15，取自: https://www.mathworks.com/help/signal/ug/prominence.html
[18]Gonzalez Woods著、繆紹綱譯，數位影像處理，臺北:台灣培生教育出版股份有限公司，2009。
[19]劉平，圖像分割閾值選取技術綜述，中科院成都計算所，2020/06/25，取自: http://blog.csdn.net/ywywcy/article/details/1704566
[20]"Convolutional Neural Networks (LeNet)–DeepLearning 0.1 documentation".DeepLearning 0.1.,LISA Lab.,2013.
[21]Habibi Aghdam,Hamed,Jahani Heravi,Elnaz, "Guide to convolutional neural networks : a practical application to traffic-sign detection and classification ,"Switzerland: Springer International Publishing,2017.
[22]徹底理解數字圖像處理中的卷積-以Sobel算子為例，2020/07/16取自:
https://read01.com/zDkAMP.htmlhttps://read01.com/zDkAMP.html#.Xztc0-gzaUk
[23]Xavier Glorot, Antoine Bordes,Y. Bengio, "Deep Sparse Rectifier Neural Networks ," Journal of Machine Learning Research 15,2010.
[24]卷積神經網路的運作原理，2020/07/17取自:
https://brohrer.mcknote.com/zh- Hant/how_machine_learning_works/how_convolutional_neural_networks_work.html
[25]M. Jordan,J. Kleinberg,B. Scholkopf, "Pattern Recognition and Machine Learning ," Springer-Verlag New York,2006.
[26]交叉驗證，2020/06/15，取自: https://zh.wikipedia.org/wiki/%E4%BA%A4%E5%8F%89%E9%A9%97%E8%AD%89
[27]Powers,David M. W.,"Evaluation: From Precision, Recall and F-Measure to ROC, Informedness, Markedness & Correlation,"Journal of Machine Learning Technologies,Volume 2,Issue 1,2011,pp.37-63.
[28]使用 TensorFlow 了解 overfitting 與 underfitting，2020/07/17取自:
https://medium.com/%E6%89%8B%E5%AF%AB%E7%AD%86%E8%A8%98/%E4%BD%BF%E7%94%A8-tensorflow-%E4%BA%86%E8%A7%A3%E9%81%8E%E6%93%AC%E5%90%88%E8%88%87%E6%AC%A0%E6%93%AC%E5%90%88-a75a26cc87e0
[29]William J.Palm III、張智星、翁展翔，MATLAB在工程上的應用，臺北:美商麥格羅希爾國際股份有限公司臺灣分公司，2013。
[30]PhysioNet，2019/10/19取自:
https://physionet.org/content/mitdb/1.0.0/
[31]Alex Krizhevsky,Ilya Sutskever and Geoffrey E. Hinton, "ImageNet Classification with Deep Convolutional Neural Networks." Advances in Neural Information Processing Systems,2012.
[32]Christian Szegedy,Wei Liu,Yangqing Jia,Pierre Sermanet,Scott Reed,Dragomir Anguelov,Dumitru Erhan,Vincent Vanhoucke,Andrew Rabinovich," Going deeper with convolutions ," Google Inc.,IEEE Conference on Computer Vision and Pattern Recognition (CVPR),2015.