成功大學博碩士論文系統

簡易檢索 / 詳目顯示

回結果列表

研究生：	陳可維 Chen, Ke-Wei
論文名稱：	機器學習與深度神經網路用於心臟電位影像重建 Solving Inverse Electrocardiographic Mapping Using Machine learning and Deep Learning Frameworks
指導教授：	林哲偉 Lin, Che-Wei
學位類別：	碩士 Master
系所名稱：	工學院 - 生物醫學工程學系 Department of BioMedical Engineering
論文出版年：	2021
畢業學年度：	109
語文別：	英文
論文頁數：	36
中文關鍵詞：	心臟電位影像重建、反向問題、機器學習、深度學習
外文關鍵詞：	Electrocardiographic imaging, Inverse Problem, Machine learning, Deep Learning
相關次數：	點閱：58 下載：0
分享至:	分享至facebook 分享至twitter

查詢本校圖書館目錄查詢臺灣博碩士論文知識加值系統勘誤回報

心臟電位影像重建(Electrocardiographic imaging)是基於體表所測得之心電圖，重建心臟表面或內部的電位分佈的過程。此問題亦被稱為反向問題(Inverse Problem)。由於此問題在本質上無法得出唯一解，以及不容易針對體表與心臟間的電傳導性建模，目前針對此問題所發展出的方法，其準確度約0.7(所重建的心電圖與實際心電圖之間相關系數之中位數)。本研究嘗試使用神經網路解決此問題，以增加模型的準確度。
本研究使用之兩組資料，一是來自於豬體表所量測之多個心電圖訊號，二是同時量測之豬心臟表面的心電圖。第一部分的研究，訓練與測試模型的資料皆來自同一之豬。所使用之神經網路有兩種。一是由數個全連接層(Fully Connected Layer FCN)所組成之神經網路，二是由長短期記憶(Long Short-Term Memory LSTM)神經網路所組成。第二部分的研究，我們將來源於不同的五隻豬之資料進行轉換，加以統合。接著我們使用卷積神經網絡(Convolutional Neural Network CNN)來建立模型。本研究使用留一驗證（leave-one-out cross-validation）作為模型準確度的驗證方法。
若比較重建的心臟表面心電圖與實際心電圖，在第一部分研究中，使用全連接層FCN之部分，整體的相關系數之中位數以及前四分之一與後四分之一的數值為0.90 [0.68–0.96]。使用全連接層LSTM之部分則為0.82 [0.54–0.93]。在第二部分研究，使用CNN，整體的相關系數之中位數以及前四分之一與後四分之一的數值則為0.74 [0.22–0.89]。
若比較重建的激發時間圖(activation map)，在第一部分研究中，使用全連接層FCN之部分，整體的相關系數之中位數以及前四分之一與後四分之一的數值為0.86 [0.61–0.93]。使用全連接層LSTM之部分則為0.52 [0.05–0.80]。在第二部分研究，使用CNN，整體的相關系數之中位數以及前四分之一與後四分之一的數值則為0.82 [0.67–0.93]。
若比較激發點位置誤差(localization error)，在第一部分研究中，使用全連接層FCN之部分，整體的距離之中位數以及前四分之一與後四分之一的數值為10.4 [3.6–22.6] mm。使用全連接層LSTM之部分則為18.5 [6.4–41.5] mm。在第二部分研究，使用CNN，整體的相關系數之中位數以及前四分之一與後四分之一的數值則為9.3 [3.4–17.0] mm。
本研究顯示，針對心臟電位影像重建，我們可以使用相對少量的資料解決。我們所達到的最佳結果為0.74(所重建的心電圖與實際心電圖之間相關系數之中位數)。此外，本研究也顯示並不需要精確的座標資訊來重建心電圖。針對模型的準確性而言，在不同的豬隻與不同的量測結果之歧異度仍然很大。這可能與資料量相對較小有關，這可以在第二部分研究中看到。其整體準確度較佳，可能與其整合了所有的資料，其資料量較第一部分多有關。臨床應用部分，此研究顯示可使用非侵入性的體表量測心電圖，來找尋心臟的電刺激源點。此資訊可用於心室早期收縮病人之治療。

Electrocardiographic imaging reconstructs the heart surface as an electrogram using the potentials recorded from the body surface. This problem is called the inverse problem. Due to the ill-posed nature and the difficulty in modeling the conductive property of the body, currently, the overall accuracy for a reconstructed electrogram is only 0.7(median correlation coefficient for activation time map). This study tries to improve the model’s accuracy using a neural network.
Electrocardiograms are simultaneously recorded from pigs’ hearts and their body surfaces. For part I of the study, we trained and tested the model with the same pig. The neural network is composed of Fully Connected Neuro network (FCN) and Long Short-term Memory (LSTM) neural network. For part II of the study, we align the data from five different pigs by transforming the torso potential data into 2D data and transforming the epicardial potential data with a registration method. A Convolutional Neural Network is used to construct the model. We evaluated the method using leave-one-out cross-validation.
For the reconstructed electrogram in part I, the overall median of correlation efficient with the first to third quantiles are 0.90 [0.68–0.96] and 0.82 [0.54–0.93] for FCN and LSTM, respectively. In part II, the overall median of the correlation efficient with the first to third quantiles is 0.74 [0.22–0.89].
For the reconstructed activation map in part I, the overall medians of the correlation efficient with the first to third quantiles are 0.86 [0.61–0.93] and 0.52 [0.05–0.80] for FCN and LSTM, respectively. In part II, the overall median of the correlation efficient with the first to third quantiles is 0.82 [0.67–0.93].
For the localization error of the predicted pacing site in part I, the overall medians of the correlation efficient with the first to third quantiles are 10.4 [3.6–22.6] mm and 18.5 [6.4–41.5] mm for FCN and LSTM, respectively. In part II, the overall median of the correlation efficient with first to third quantiles is 9.3 [3.4–17.0] mm.
In conclusion, a neural network can be used to solve the inverse problem of ECGi with relatively small datasets. Our best result shows overall median of the correlation efficient to be 0.82. Our study also shows that a rough geometrical information of torso and heart may be enough to reconstruct the epicardial gram. Performance of the model is inconsistent between different recording and pig. This may be due to relatively small dataset and may improve with larger dataset. As shown in part II study, it has better result when model is trained with more data. In clinical setting, this study shows the potential to identify source of pacing site with non-invasive electro-cardiogram recorded from the surface, which can be applied to evaluation for patient with premature ventricular contractions.

摘 要 ....................................................................................................................................... i
Abstract ................................................................................................................................ iii
Acknowledgements ............................................................................................................... v
Table of Contents ................................................................................................................. vi
List of Tables ...................................................................................................................... viii
List of Figures .................................................................................................................... viii
List of Abbreviations ............................................................................................................ ix
Chapter 1 Introduction and Background ........................................................................... 1
1. Inverse Electrocardiographic Mapping ......................................................................... 1
2. The importance of inverse electrocardiographic mapping ........................................... 2
3. Traditional Methods ........................................................................................................ 2
4. Problems faced in current methods ................................................................................ 3
5. Neural network for prediction ........................................................................................ 3
Chapter 2 Material and Methods ........................................................................................ 4
1 Overall design of the study .............................................................................................. 4
2 Data collection .................................................................................................................. 5
3 Final data used in the study ............................................................................................. 7
4 Part I study: without Considering the Geometry .......................................................... 7
4.1 Model selection ............................................................................................................. 7
5 Part II study: Add Geometrical Information ................................................................ 8
5.1 Torso node registration ................................................................................................ 8
5.2 Transforming 1D data into 2D .................................................................................... 9
5.3 Epicardial surface node registration ........................................................................ 10
5.4 Transforming 1D data into 1D data with the same geometrical sequence ............ 12
5.5 Model selection ........................................................................................................... 13
6 Model evaluation ............................................................................................................ 14
6.1 Leave-one-out cross-validation. ................................................................................ 14
6.2 Evaluation metric: potential prediction ................................................................... 15
6.3 Activation time reconstruction and pacing site localization ................................... 15
6.4 Evaluation metric: activation time ........................................................................... 16
6.5 Evaluation metric: localization error ....................................................................... 16
Chapter 3 Results ............................................................................................................... 17
1 Potential visualization .................................................................................................... 17
2 Median Correlation Coefficient .................................................................................... 18
3 Activation Time Correlation ......................................................................................... 19
4 Localization error ........................................................................................................... 21
Chapter 4 Discussion, Conclusion and Future Works .................................................... 23
1 Interpretation of the results .......................................................................................... 23
2 Comparison with Previous Reported Accuracy .......................................................... 23
3 How important is the geometrical information? ......................................................... 24
4 Potential clinical application ......................................................................................... 25
5 Limitations ...................................................................................................................... 25
6 Conclusions ..................................................................................................................... 25
References ........................................................................................................................... 26
                                    

[1] B. T. Howard S. Oster, Robert L. Lux, Philip R. Ershler, Yoram Rudy, "Noninvasive electrocardiographic imaging: reconstruction of epicardial potentials, electrograms, and isochrones and localization of single and multiple electrocardiac events.," Circulation, vol. 96, no. 3, pp. 1012–1024, 1997.
[2] D. H. B. Matthijs Cluitmans1, Rob MacLeod, Olaf Dössel, María S. Guillem,Peter M. van Dam, Jana Svehlikova, Bin He, John Sapp, Linwei Wang, Laura Bear, "Validation and Opportunities of Electrocardiographic Imaging: From Technical Achievements to Clinical Applications," Front. Physiol, vol. 9, p. 1305, 2018.
[3] P. S. C. Laura Bear, Olivier Bernus, Igor Efimov, Rémi Dubois, "Introduction to Noninvasive Cardiac Mapping," Cardiac Electrophysiology Clinics, vol. 7, no. P1-16, 2015.
[4] R. Y. Wang Y, "Application of the method of fundamental solutions to potential-based inverse electrocardiography," Ann Biomed Eng, vol. 34, pp. 1272-1288, 2006, doi: 10.1007/s10439-006-9131-7.
[5] L. R. Bear et al., "How Accurate Is Inverse Electrocardiographic Mapping? A Systematic In Vivo Evaluation," Circ Arrhythm Electrophysiol, vol. 11, no. 5, p. e006108, May 2018, doi: 10.1161/CIRCEP.117.006108.
[6] S. F. Duchateau J, Pambrun T, Derval N, Chamorro-Servent J, Denis A, Ploux S, Hocini M, Jaïs P, Bernus O, et al., "Performance and limitations of noninvasive cardiac activation mapping," Heart Rhythm, vol. 16, pp. 435–442, 2019.
[7] L. S. Joe Horvath, Tommy Peng, Avinash Malik, Mark Trew, Laura Bear, "Deep learning neural nets for detecting heart activity," arXiv:1901.09831 [physics.med-ph], 2019.
[8] G. W. Aras K, Tate J, Burton B, Brooks D, Coll-Font J, Doessel O, Schulze W, Potyagaylo D, Wang L, van Dam P, MacLeod R. , "Experimental data and geometric analysis repository-EDGAR," J Electrocardiol, no. 48, pp. 975–981, 2015.
[9] C. L. Bear LR, LeGrice IJ, Sands GB, Lever NA, Paterson DJ, Smaill and BH., "Forward problem of electrocardiography: is it solved?," Circ Arrhythm Electrophysiol., vol. 8, pp. 677–684, 2015.
[10] P. M. Duchateau J, Dubois R, "Spatially coherent activation maps for electrocardiographic imaging," IEEE Trans Biomed Eng, vol. 64, no. 5, pp. 1149 - 1156, 2017.
[11] "Noninvasive electrocardiogram imaging of substrate and intramural ventricular tachycardia in infarcted hearts," J Am Coll Cardiol, vol. 38, pp. 2071–2078, 2001.
[12] J. P. Ghanem RN, Ramanathan C, Ryu K, Markowitz A, Rudy Y, "Noninvasive electrocardiographic imaging (ECGI): comparison to intraoperative mapping in patients.," Heart Rhythm, vol. 2, no. 4, pp. 339-354, 2005.
[13] D. F. Sapp JL, Clements JC, Horácek BM, "Inverse solution mapping of epicardial potentials: quantitative comparison with epicardial contact mapping," Circ Arrhythm Electrophysiol, vol. 5, pp. 1001–1009, 2012.
[14] O. M. Graham AJ, Zacur E, Dhillon G, Daw H, Srinivasan NT, Lane JD, Cambridge A, Garcia J, O’Reilly NJ, Whittaker-Axon S, Taggart P, Lowe M, Finlay M, Earley MJ, Chow A, Sporton S, Dhinoja M, Schilling RJ, Hunter RJ, Lambiase PD, "Simultaneous comparison of electrocardiographic imaging and epicardial contact mapping in structural heart disease," Circ Arrhythm Electrophysiol, vol. 12, no. 11, 2019.
[15] P. P. B. P. Matthijs J.M.Cluitmans MD, Joël M.H.Karel PhD, MarcoDas MD,PhD, Bas L.J.H.Kietselaer MD, PhD Monique M.J. Jonga Frits W.Prinzen PhD, Ralf L.M.PeetersPhD. Ronald L.Westra PhD, Paul G.A.Volders MD, PhD, "In Vivo Validation of Electrocardiographic Imaging," JACC Clin Electrophysiol, vol. 3, no. 3, pp. 232-242, 2017.
[16] M. E. R. Stephan Hohmann, Hiroki Konishi, Anna Borenstein, Songyun Wang, Atsushi Suzuki, Gregory J. Michalak, Kristi H. Monahan, Kay D. Parker, L. Katie Newman, Douglas L. Packer, "Spatial Accuracy of a Clinically Established Noninvasive Electrocardiographic Imaging System for the Detection of Focal Activation in an Intact Porcine Model," Circulation: Arrhythmia and Electrophysiology, vol. 12, no. 11, 2019.

校內：2024-10-31公開
校外：2024-10-31公開

簡易檢索 / 詳目顯示

相關論文