usually 13 which cannot describe the scanned object exactly In order to enlarge the visual field instead of considering the neighborhood of pixel only we adopt deep learning technique to solve the multimaterial decomposition ID: 777624
Download The PPT/PDF document "Introduction Traditional post-process ma..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
IntroductionTraditional post-process material decomposition algorithms are based on pixel-local usually [1-3], which cannot describe the scanned object exactly. In order to enlarge the visual field instead of considering the neighborhood of pixel only, we adopt deep learning technique to solve the multi-material decomposition problem [4]. We build a convolutional neural network(CNN) [5] and simulate some reconstruction images of spectral CT to generate the training data to train the network. Compared to the results of solving linear equations, the CNN method turns out to work much better in the test samples. As the conclusion, we think CNN is useful in the multi-material decomposition, but there still remains many researches to work, such as the source of training data, the balance between the priori knowledge and measurement.
Fig. 3 shows the result of simulation in 100 phantoms. Ordinate is MSE of the 100 decomposition result. The blue stars are the MSE of solving equations, while the red are CNN method’s. The abscissa is the harmonic average of the distances from each test phantoms to all the training phantoms.In this work, CNN shows its effectiveness to solve multi-material decomposition problem, reducing the MSE of the result by 1~2 orders comparing to direct decomposition. One of the decomposition result is shown in Fig. 4.
Application
of Deep Learning in Multi-Material Decomposition of Spectral CTZhengyang Chen, Liang LiDepartment of Engineering Physics, Tsinghua University, China.
References
[1] Y. Long, et al., IEEE T Med Imag, 33, 1614, (2014).[2] M. Patino, et al., Radiographics, 36, 1087,( 2016).[3] P. R. Mendonca, et al., IEEE T Med Imag, 33, 99, (2014).[4] G. Wang, et al., IEEE Access, 4, 8914, (2016).[5] K. Simonyan, et. al., Comput Sci, (2014).
MethodThe flow chart of the CNN method is in Fig. 1. According to the work [3], the volume fraction of basis material in the spatial position should meet:which is exactly the result of a CNN used in the application of image classification should meet. We generate some phantoms, and use them to retrain the CNN to solve the material decomposition task.
Figure 3
.
R
esult
of simulation in 100 phantoms.
Figure 1. Flow chart of the CNN method.
Figure
2. The phantoms used to train CNN.
Figure 4. One of the simulation results. The first line is the generated phantom, the second line is the results of CNN, and the last line is the results of solving equations.
Contact info
Liang Li, Associate Professor, Tsinghua University, Email: lliang@Tsinghua.edu.cn
Acknowledgement
Supported by
NSFC 61571256