Calculating the Dual Energy Decomposition by Neural Network

• Main Authors: ML Chiang, 江孟良
• Other Authors: KS Chuang
• Format: Others
• Language: zh-TW
• Published: 1999
• Online Access: http://ndltd.ncl.edu.tw/handle/93248001736687892900

碩士 === 國立清華大學 === 原子科學系 === 87 === Dual energy technique has been applied in chest radiography successfully. It enhances the image of desired tissues, and suppresses the others. Traditionally, the dual energy decomposition by Newton-Raphson method is severely affected by the quantum mottle. The neural network with its high error tolerance can be used to approximate any function from the noisy data. In this experiment, two types of networks: (1) a multilayer perceptron with back-propagation algorithm and (2) a linear network are employed to process the nonlinear relation between attenuation ratio and thickness of medium. The recomposition ability at various SNR (signal to noise ratio) situation of these two methods are compared with traditional Newton method. In this study, we applied the three method of decomposition on the calibration data set to generate the parameters, these parameters were then employed on a test data set. The mean squared errors of the three methods are compared. In calibration experiment, 7 set different thickness for each of aluminum and Plexiglas to combine step phantom of 49 set thickness. In test experiment, 10 set different thickness for equal materials to combine step phantom of 100 set thickness. The spectrum of x-ray for high energy (140kVp) and low energy (80kVp) are generated by computer simulation. The photon number of x-ray spectrum has a Poisson distribution to simulate the actual situation. Two calibration experiments is performed in this research, the first is training in optimized state, and the second is training in real state. The results of the first calibration experiment show that no large deviation between three methods. The second experiment indicates that the variation of neural network with SNR is small. Besides, the mean squared error for inputs with fraction is more stable than the other inputs.