探討知識蒸餾方法於心肌梗塞心電圖訊號之應用

Exploring the application of knowledge distillation method in myocardial infarction ECG signal

賴漢樺(HAN-HUA LAI)

指導教授 ： 葛宗融

中原大學/工學院/生物醫學工程學系/碩士(2022年)

https://doi.org/10.6840/cycu202201330

摘要

背景與動機：心電圖是一種非侵入式且價格低廉的心肌梗塞診斷工具。現今各項電腦輔助診斷系統普遍使用卷積神經網路(Convolutional neural network, CNN)，透過自動辨識心電圖來診斷心肌梗塞，得以早期診斷及預防。然而，現今診斷心肌梗塞模型為了追求高檢測性能，模型架構龐大和高耗能成為一大隱憂，因此，本研究以六分類心電圖心肌梗塞訊號作為深度學習分類資料集，並導入知識蒸餾方法縮小模型大小和減少耗能。
材料與方法：訊號預處理使用濾除基線飄移、反鋸齒濾波、下採樣、隨機切片採樣和數據增強。深度學習分類模型使用基於CNN深度學習網路的ML-ResNet和VGG-6模型進行訓練及分類；同時，透過知識蒸餾方法提升小模型準確率，並驗證5種不同蒸餾模式的有效性。模型測試是以準確率、精確率、召回率、F1-score和本研究所提出的成長率作為評估指標，並及以無母數檢定來驗證不同知識蒸餾方法之差異性，最後利用Qt designer工具及PyQt5套件建立心肌梗塞檢測系統，並驗證其系統預測性能。
結果與討論：使用知識蒸餾方法在二和六心肌梗塞分類之K折交叉平均驗證下，剩餘誤差知識蒸餾在準確率和成長率都有最好的性能表現，二分類準確率達86.69%，成長率為5.2%，六分類準確率達42.25%，成長率為9.76%。
結論：本研究使用知識蒸餾方法提升於心肌梗塞檢測，並於結果中有顯著提升差異。在心肌梗塞檢測系統中也有顯著的心肌梗塞檢測標記。透過知識蒸餾方法，能維持高準確率及模型減量，未來有望搭載於穿戴式或移動式裝置中，以及建立即時心肌梗塞檢測的健康APP，實際導入臨床應用。

關鍵字

知識蒸餾 ； 心電圖 ； 心肌梗塞 ； 卷積神經網路

並列摘要

Background and motivation: The electrocardiogram is a non-invasive and inexpensive diagnostic tool for myocardial infarction (MI). Nowadays, various computer-aided diagnosis systems generally use a convolutional neural network (CNN) to diagnose MI through automatic identification of electrocardiogram, to enable early diagnosis and prevention. However, to pursue high detection performance, the current diagnostic MI model has a huge model structure and high energy consumption. Therefore, this study uses the six-category ECG myocardial infarction signal as the deep learning classification data set, and introduces the knowledge distillation (KD) method to reduce the model. size and reduce energy consumption.
Materials and methods: Signal preprocessing used baseline drift removal, antialiasing, downsampling, random slice sampling, and augmentation. The deep learning classification model uses ML-ResNet and VGG-6 model based on the CNN deep learning network for training and classification. At the same time, the KD method is used to improve the accuracy of the small model and verify the effectiveness of 5 different distillation modes. Model testing is based on accuracy, precision, recall, F1-score, and the growth rate proposed in this study as evaluation indicators, and the no-matrix test to verify the differences between different knowledge distillation methods, and finally use the Qt designer tool and the PyQt5 suite to build a myocardial infarction detection system and verify its predictive performance.
Results and discussion: Using the KD method, Residual KD has the best performance in both accuracy and growth rate under the K-fold cross-average validation of two-class and six-class MI classification. The accuracy of the two-class classification is 86.69%, and the growth rate is 5.2%, the six-category accuracy rate reached 42.25%, and the growth rate was 9.76%.
Conclusion: This study used KD to improve MI detection, and there were significant differences in the results. There are also significant myocardial infarction detection markers in the myocardial infarction detection system. Through the knowledge distillation method, high accuracy and model reduction can be maintained. In the future, it is expected to be installed in wearable or mobile devices, and establish a health APP for real-time myocardial infarction detection, which will be practically introduced into clinical applications.

並列關鍵字

knowledge distillation ； electrocardiogram ； myocardial infarction ； convolutional neural network

參考文獻

[1] Benjamin, "Heart Disease and Stroke Statistics-2019 Update: A Report From the American Heart Association (vol 139, pg e56, 2019)," Circulation, vol. 141, no. 2, pp. E33-E33, Jan 2020, doi: 10.1161/cir.0000000000000746.

[2] C. Xu, L. Xu, Z. Gao, S. Zhao, H. Zhang, Y. Zhang, X. Du, S. Zhao, D. Ghista, and H. Liu, "Direct delineation of myocardial infarction without contrast agents using a joint motion feature learning architecture," Medical image analysis, vol. 50, pp. 82-94, Feb 2018.

[3] P. Kligfield, L. S. Gettes, J. J. Bailey, R. Childers, B. J. Deal, E. W. Hancock, G. Van Herpen, J. A. Kors, P. Macfarlane, and D. M. Mirvis, "Recommendations for the standardization and interpretation of the electrocardiogram: part I: the electrocardiogram and its technology a scientific statement from the American Heart Association Electrocardiography and Arrhythmias Committee, Council on Clinical Cardiology; the American College of Cardiology Foundation; and the Heart Rhythm Society endorsed by the International Society for Computerized Electrocardiology," Journal of the American College of Cardiology, vol. 49, no. 10, pp. 1109-1127, Mar 2007.

[4] K. Thygesen, J. S. Alpert, A. S. Jaffe, B. R. Chaitman, J. J. Bax, D. A. Morrow, H. D. White, and E. G. o. b. o. t. J. E. S. o. C. A. C. o. C. A. H. A. W. H. F. T. F. f. t. U. D. o. M. Infarction, "Fourth universal definition of myocardial infarction (2018)," Journal of the American College of Cardiology, vol. 72, no. 18, pp. 2231-2264, Oct 2018.

[5] Z. Chen, E. N. Brown, and R. Barbieri, "Characterizing nonlinear heartbeat dynamics within a point process framework," IEEE Transactions on Biomedical Engineering, vol. 57, no. 6, pp. 1335-1347, June 2010.