類神經網路在心導管心室攝影與單光子斷層掃描之分析與臨床應用

Analysis and Clinical Applications of Backpropagation Neural Network in Catheterization Ventriculography and GSPECT

蕭聿謙(Yu-Chien Shiau)

指導教授 ： 朱樹勳 ； 郭德盛

國立臺灣大學/電機資訊學院/電機工程學研究所/博士(2004年)

https://doi.org/10.6342/NTU.2004.01381

摘要

動態心臟影像的運動弁鄐尷R對於心臟病的診斷與病情的評估是非常重要的。由於影像處理技術的進步，目前已有多種方法可提供動態影像運動分析的弁遄A其中，應用類神經網路於醫學影像處理為近年來的新發展領域，值得深入研究。
本論文提出新的方法，應用逆傳導神經網路分析多種動態心臟影像，包括：心導管心室攝影 (catheterization ventriculography)、核醫心肌灌注門控單光子斷層掃描 (myocardial perfusion gated single photon emission computed tomography, GSPECT)、核醫血池門控單光子斷層掃描 (blood-pool GSPECT)、與核醫心室攝影 (radionuclide ventriculography)等，以提供動態心臟影像的運動弁鄐尷R。針對各項動態心臟影像，設計合適的假體模擬影像以訓練逆傳導神經網路，當誤差函數呈現收斂，神經網路訓練完成之後，該神經網路即可用於動態心臟影像之運動偵測與分析，心臟各部分之運動方向，以運動向量場與都卜勒顏色法顯示。並由專科醫師判讀原始動態心臟影像、運動向量場、與都卜勒顏色法顯示影像，以「接收者操作特性 (receiver operating characteristic)曲線分析」比較統計差異。
統計結果顯示，以運動向量場、都卜勒顏色法影像評估心臟運動弁遄A與原始動態心臟影像之影像判讀無顯著差異。以逆傳導神經網路分析所得之運動向量場與都卜勒顏色法影像，較不受雜訊干擾，可呈現客觀的判讀資訊，提供醫師作為心臟病的診斷、病情評估、與治療追蹤的參考工具。
逆傳導神經網路用於分析多種動態心臟影像，所得到的運動向量場與都卜勒顏色法影像是有用的，並且逆傳導神經網路具有潛力可供應用於其他心臟動態影像分析，以及其他醫學影像分析，提供豐富的資訊與臨床應用。

關鍵字

心臟影像 ； 類神經網路 ； 運動分析

並列摘要

Cardiac motion analysis is very important for the clinical diagnosis and disease evaluation of heart disease. There have been several image processing techniques for motion detection and motion analysis. Among those techniques, artificial neural network has been an important new approach in recent years.
We propose a new approach and a novel algorithm, which uses a three-layer backpropagation neural network for cardiac motion analysis of catheterization ventriculography, myocardial perfusion gated single photon emission computed tomography (GSPECT), blood-pool GSPECT, and radionuclide ventriculography. The backpropagation neural network is trained by phantom images or sample images, which simulate different kinds of dynamic cardiac images. After the training is completed, the neural network can perform motion detection on patients’ dynamic cardiac images. The results of motion detection are displayed in the formats of motion vector fields and Doppler color display, which are superimposed on the original cardiac images. The physician specialists interpret cardiac wall motion on patients’ original images, motion vector fields, and Doppler color display. Receiver operating characteristic (ROC) curve analysis is used to test the statistical differences among the accuracy of the image interpretations.
There is no significant statistical differences among the areas under ROC curves for the image interpretations made on original dynamic cardiac images, motion vector fields, and Doppler color displays.
Cardiac wall motion detection and analysis by backpropagation neural network is useful for dynamic cardiac images. It is highly suggestive that the algorithm can be also useful for cardiac cine MRI, cardiac multidetector CT, and cardiac ultrasonography.

並列關鍵字

neural network ； heart image ； motion analysis

參考文獻

[1] R.D. White and R.M. Setser, “Integrated approach to evaluating coronary artery disease and ischemic heart disease”, American Journal of Cardiology, vol. 90 (10C), 2002, pp. 49L-55L.

[4] E. Atsumi, M. Takagi, and K. Yokosawa, “Local motion detection by hierarchical neural network”, Systems and Computers in Japan, vol. 25, No. 11, 1994, pp. 24-35.

[7] J.P. Ioannidis, T.A. Trikalinos, P.G. Danias, “Electro-cardiogram-gated single-photon emission computed tomography versus cardiac magnetic resonance imaging for the assessment of left ventricular volumes and ejection fraction: a meta-analysis,” Journal of the American College of Cardiology, vol. 39, pp. 2059-2068, 2002.

[10] C.E. Metz, “ROC methodology in radiologic imaging,” Invest Radiology 21:720-733, 1986.

[11] C.E. Metz, “Some practical issues of experimental design and data analysis in radiological ROC studies,” Invest Radiology 24:234-245, 1989.