Presentation file for "Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition" presented at the Workshop on Deep Learning for Biomedical Image Reconstruction of IEEE Internetional Symposium on Biomedical Imaging, ISBI 2020.
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
1. Workshop: Deep Learning for Biomedical Image Reconstruction
Multi-energy Bone Subtraction in Chest Radiography
by Eigenvalue Decomposition
Boah Kim, Junyoung Kim, Wontaek seo, Choul Woo Shin, Jong Chul Ye
2. Contents
ISBI2020 Workshop Presentation 2
Boah Kim
1. Introduction
2. Theory
3. Experiment
4. Conclusion
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
3. Introduction
Bone subtraction on chest radiography
Chest radiography (CR)
ISBI2020 Workshop Presentation 3
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
• Fundamental imaging method to diagnose disease
(ex. lung cancer, arterial stenosis, pleural abnormalities)
• Often overlap subtle lesions with normal structure
→ Hard to be detected due to low sensitivity of the lesions
Standard CR images
Kuhlman, Janet E., et al. Radiographics, 2006
Bone subtraction
• Enhance the visibility of soft tissues
• Detect abnormal nodules with high accuracy
4. Introduction
Dual-energy subtraction
ISBI2020 Workshop Presentation 4
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
MacMahon, Heber, et al. Journal of thoracic imaging, 2008
Standard radiograph Dual energy soft tissue Dual energy bone
Employ differences of attenuation coefficients of body tissues
• Principle: Bone contained calcium absorbs more photons at lower energy than soft tissues.
• Be used to generate bone-subtracted soft tissue images
• Superior sensitivity for the detection of calcification within a pulmonary nodule
5. Introduction
Dual-energy subtraction
ISBI2020 Workshop Presentation 5
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
MacMahon, Heber, et al. Journal of thoracic imaging, 2008
Standard radiograph Dual energy soft tissue Dual energy bone
Challenge
• Computationally expensive and sensitive to compute weight of bone subtraction
- Estimate thickness of bone & soft tissue on images
- Use of nonlinear polynomial approximation algorithms
6. Introduction
Dual-energy subtraction
ISBI2020 Workshop Presentation 6
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
Challenge
• Computationally expensive and sensitive to compute weight of bone subtraction
- Estimate thickness of bone & soft tissue on images
- Use of nonlinear polynomial approximation algorithms
Multi-energy subtraction
Based on eigenvalue decomposition (EVD)
Can be used for application of deep-learning-based
multi-energy image synthesis
Multi-energy chest radiography
7. Contents
ISBI2020 Workshop Presentation 7
Boah Kim
1. Introduction
2. Theory
3. Experiment
4. Conclusion
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
8. Theory
Chest radiography
Principle
ISBI2020 Workshop Presentation 8
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
𝐼𝐼 = 𝜇𝜇𝑏𝑏𝑙𝑙𝑏𝑏 + 𝜇𝜇𝑠𝑠𝑙𝑙𝑠𝑠 + 𝑤𝑤
𝜇𝜇𝑏𝑏 𝜇𝜇𝑠𝑠
X-ray
Source
X-ray
Film
• Obtained by attenuation coefficients 𝜇𝜇 and thickness 𝑙𝑙 of tissues
Chest radiography image, 𝐼𝐼
Bone Soft tissue
𝑙𝑙𝑏𝑏 𝑙𝑙𝑠𝑠
9. Theory
Chest radiography
Principle
ISBI2020 Workshop Presentation 9
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
𝐼𝐼 = 𝜇𝜇𝑏𝑏𝑙𝑙𝑏𝑏 + 𝜇𝜇𝑠𝑠𝑙𝑙𝑠𝑠 + 𝑤𝑤
X-ray
Source
X-ray
Film
• Obtained by attenuation coefficients 𝜇𝜇 and thickness 𝑙𝑙 of tissues
Chest radiography image, 𝐼𝐼
Bone Soft tissue
𝑙𝑙𝑏𝑏 𝑙𝑙𝑠𝑠
Different attenuation coefficient values
according to the X-ray energy levels
𝜇𝜇𝑏𝑏 𝜇𝜇𝑠𝑠
Kim, D-H., et al. Journal of Instrumentation, 2013
10. Theory
Multi-energy chest radiography
Images obtained by different X-ray energy levels
ISBI2020 Workshop Presentation 10
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
• Can be represented with a matrix form with 𝜇𝜇 and 𝑙𝑙
Multi-energy chest radiography image, 𝐈𝐈
𝐼𝐼1 𝐼𝐼2 𝐼𝐼3 𝐼𝐼𝑁𝑁
𝐈𝐈 =
𝐼𝐼1
𝐼𝐼2
⋮
𝐼𝐼𝑁𝑁
=
𝜇𝜇𝑏𝑏
1
𝜇𝜇𝑠𝑠
1
𝜇𝜇𝑏𝑏
2
𝜇𝜇𝑠𝑠
2
⋮ ⋮
𝜇𝜇𝑏𝑏
4
𝜇𝜇𝑠𝑠
4
𝑙𝑙𝑏𝑏
𝑙𝑙𝑠𝑠
+ 𝑤𝑤
= 𝝁𝝁𝑏𝑏𝑙𝑙𝑏𝑏 + 𝝁𝝁𝑠𝑠𝑙𝑙𝑠𝑠
= 𝐈𝐈𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏 + 𝐈𝐈𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠
11. Multi-energy chest radiography image, 𝐈𝐈
𝐼𝐼1 𝐼𝐼2 𝐼𝐼3 𝐼𝐼𝑁𝑁
𝐈𝐈 =
𝐼𝐼1
𝐼𝐼2
⋮
𝐼𝐼𝑁𝑁
=
𝜇𝜇𝑏𝑏
1
𝜇𝜇𝑠𝑠
1
𝜇𝜇𝑏𝑏
2
𝜇𝜇𝑠𝑠
2
⋮ ⋮
𝜇𝜇𝑏𝑏
4
𝜇𝜇𝑠𝑠
4
𝑙𝑙𝑏𝑏
𝑙𝑙𝑠𝑠
+ 𝑤𝑤
= 𝝁𝝁𝑏𝑏𝑙𝑙𝑏𝑏 + 𝝁𝝁𝑠𝑠𝑙𝑙𝑠𝑠 + w
= 𝐈𝐈𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏 + 𝐈𝐈𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 + w
Theory
Multi-energy chest radiography
Images obtained by different X-ray energy levels
ISBI2020 Workshop Presentation 11
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
If 𝐈𝐈𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏 is estimated,
𝐈𝐈𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 can be obtained.
• Can be represented with a matrix form with 𝜇𝜇 and 𝑙𝑙
12. Theory
Bone removal method
Estimation of bone & soft tissue images from multi-energy data
ISBI2020 Workshop Presentation 12
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
• Assumption
Multi-energy CR images are
all aligned.
• Two steps
1. Estimation of bone image
- use eigenvalue decomposition
2. Subtraction of bone to get
soft tissue images
- use correlation
Flow chart of our proposed method
13. Theory
Bone removal method
1st: Estimation of bone image
ISBI2020 Workshop Presentation 13
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
14. Theory
Bone removal method
1st: Estimation of bone image
ISBI2020 Workshop Presentation 14
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
1) Get a matrix 𝐘𝐘 with MxM kernel
𝐘𝐘 = 𝐀𝐀𝐀𝐀 + 𝑤𝑤
𝐀𝐀 : Attenuation coefficient matrix
𝐗𝐗 : path length matrix on the kernel
15. Theory
Bone removal method
1st: Estimation of bone image
ISBI2020 Workshop Presentation 15
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
1) Get a matrix 𝐘𝐘 with MxM kernel
2) Find 𝐀𝐀𝐀𝐀 by cost function
- With the optimal point 𝐗𝐗∗
= 𝐀𝐀T
𝐀𝐀
−𝟏𝟏
𝐀𝐀T
𝐘𝐘,
min 𝐘𝐘 − 𝐀𝐀𝐀𝐀 2
𝐘𝐘 = 𝐀𝐀𝐀𝐀 + 𝑤𝑤
𝐀𝐀 : Attenuation coefficient matrix
𝐗𝐗 : path length matrix on the kernel
𝐘𝐘 − 𝐀𝐀𝐗𝐗∗ 2
= 𝐘𝐘 2
− P𝐀𝐀𝐘𝐘 2
= 𝐀𝐀 𝐀𝐀T
𝐀𝐀
−𝟏𝟏
𝐀𝐀T
= 𝐮𝐮𝐮𝐮T
- Assume 𝐀𝐀T
𝐀𝐀 = 𝐼𝐼
- Consider only 𝐮𝐮
𝐮𝐮
𝐈𝐈 =
𝐼𝐼1
𝐼𝐼2
⋮
𝐼𝐼𝑁𝑁
=
𝜇𝜇𝑏𝑏
1
𝜇𝜇𝑠𝑠
1
𝜇𝜇𝑏𝑏
2
𝜇𝜇𝑠𝑠
2
⋮ ⋮
𝜇𝜇𝑏𝑏
4
𝜇𝜇𝑠𝑠
4
𝑙𝑙𝑏𝑏
𝑙𝑙𝑠𝑠
+ 𝑤𝑤
16. Theory
Bone removal method
1st: Estimation of bone image
ISBI2020 Workshop Presentation 16
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
1) Get a matrix 𝐘𝐘 with MxM kernel
2) Find 𝐀𝐀𝐀𝐀 by cost function
𝐘𝐘 = 𝐀𝐀𝐀𝐀 + 𝑤𝑤
𝐀𝐀 : Attenuation coefficient matrix
𝐗𝐗 : path length matrix on the kernel
min 𝐘𝐘 − 𝐀𝐀𝐀𝐀 2
= min 𝐘𝐘 2
− P𝐀𝐀𝐘𝐘 2
= max P𝐀𝐀𝐘𝐘 2
= max
𝐮𝐮
𝐮𝐮𝐮𝐮T
𝐘𝐘
2
= max
𝐮𝐮
𝐮𝐮T
𝐘𝐘𝐘𝐘T
𝐮𝐮
17. Theory
Bone removal method
1st: Estimation of bone image
ISBI2020 Workshop Presentation 17
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
1) Get a matrix 𝐘𝐘 with MxM kernel
2) Find 𝐀𝐀𝐀𝐀 by cost function
𝐘𝐘 = 𝐀𝐀𝐀𝐀 + 𝑤𝑤
𝐀𝐀 : Attenuation coefficient matrix
𝐗𝐗 : path length matrix on the kernel
min 𝐘𝐘 − 𝐀𝐀𝐀𝐀 2
= min 𝐘𝐘 2
− P𝐀𝐀𝐘𝐘 2
= max P𝐀𝐀𝐘𝐘 2
= max
𝐮𝐮
𝐮𝐮𝐮𝐮T
𝐘𝐘
2
= max
𝐮𝐮
𝐮𝐮T
𝐘𝐘𝐘𝐘T
𝐮𝐮
Covariance of 𝐘𝐘
18. Theory
Bone removal method
1st: Estimation of bone image
ISBI2020 Workshop Presentation 18
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
3) Perform eigen decomposition
𝐘𝐘𝐘𝐘T
= 𝐔𝐔𝐔𝐔𝐔𝐔T
- One of eigenvectors 𝐮𝐮𝐛𝐛
provides bone information.
19. Theory
Bone removal method
1st: Estimation of bone image
ISBI2020 Workshop Presentation 19
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
3) Perform eigen decomposition
𝐘𝐘𝐘𝐘T
= 𝐔𝐔𝐔𝐔𝐔𝐔T
- One of eigenvectors 𝐮𝐮𝐛𝐛
provides bone information.
4) Estimate bone image by 𝐮𝐮𝐛𝐛
𝐈𝐈𝒃𝒃𝒃𝒃𝒃𝒃𝒃𝒃 = 𝐮𝐮𝐛𝐛𝐮𝐮𝐛𝐛
T
𝐘𝐘
20. Theory
Bone removal method
2nd: Estimation of soft tissue image
ISBI2020 Workshop Presentation 20
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
𝐈𝐈𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 = 𝐈𝐈 − 𝐈𝐈𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏 − 𝑤𝑤
= 𝐈𝐈 − 𝛼𝛼𝐈𝐈𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏
Approximate noise 𝑤𝑤
min
𝛼𝛼
Corr < 𝐈𝐈𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏 , 𝐈𝐈𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 >
= min
𝛼𝛼
𝐈𝐈𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏
T
(𝐈𝐈 −𝛼𝛼𝐈𝐈𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏)
𝐈𝐈 −𝛼𝛼𝐈𝐈𝑏𝑏𝑏𝑏𝑏𝑏𝑏𝑏
• Subtract bone from CR data
• Compute 𝛼𝛼 by correlation
21. Contents
ISBI2020 Workshop Presentation 21
Boah Kim
1. Introduction
2. Method
3. Experiment
4. Conclusion
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
22. Experiment
Simulated chest radiography
ISBI2020 Workshop Presentation 22
Boah Kim
• Provided by DRTECH Corporation
• Data with multi-energy levels : [60, 120] kVp images with 10kVp interval
Dataset
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
Simulation CR data with multi-energy levels
60kVp 70kVp 80kVp 90kVp 100kVp 110kVp 120kVp
23. Experiment
Bone removal results from our proposed method
ISBI2020 Workshop Presentation 23
Boah Kim
Results
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
Result: Bone Result: Soft tissue
Data: 60kVp Data: 120kVp
24. Experiment
Application to deep learning
ISBI2020 Workshop Presentation 24
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
25. Experiment
Application to deep learning
GAN-based multi-energy image generation
ISBI2020 Workshop Presentation 25
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
• Synthesize the image of unknown energies by CollaGAN
• Proposed method could be applicable even with dual-energy data
Flow chart of CollaGAN
Lee, et al. CVPR, 2019
26. Experiment
Application to deep learning
GAN-based multi-energy image generation
ISBI2020 Workshop Presentation 26
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
• Synthesize the image of unknown energies by CollaGAN
• Proposed method could be applicable even with dual-energy data
Multi-energy image generation
Flow chart of CollaGAN
Lee, et al. CVPR, 2019
27. Experiment
Bone removal results with CollaGAN
ISBI2020 Workshop Presentation 27
Boah Kim
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
CollaGAN Result: 110kVp Result: Soft tissue
Data: 60kVp Data: 120kVp
Application to deep learning
28. Contents
ISBI2020 Workshop Presentation 28
Boah Kim
1. Introduction
2. Method
3. Experiment
4. Conclusion
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
29. Conclusion
Propose a novel bone subtraction method using multi-energy chest radiography
eigenvalue decomposition
ISBI2020 Workshop Presentation 29
Boah Kim
Demonstrate the effectiveness of our method with the simulated multi-energy chest
radiography data
Can be used in deep learning application of X-ray data in various ways
Multi-energy Bone Subtraction in Chest Radiography by Eigenvalue Decomposition
30. Acknowledgement
• This work was supported by Institute of Information & Communications Technology Planning & Evaluation
(IITP) grant funded by the Korea government(MSIT) [2016-0-00562(R0124-16-0002), Emotional
Intelligence Technology to Infer Human Emotion and Carry on Dialogue Accordingly]
• This work was also supported by the Industrial Strategic technology development program (10072064,
Development of Novel Artificial Intelligence Technologies To Assist Imaging Diagnosis of Pulmonary,
Hepatic, and Cardiac Diseases and Their Integration into Commercial Clinical PACS Platforms) funded by
the Ministry of Trade Industry and Energy (MI, Korea).
Medical Image Synthesis with Improved Cycle-GAN: CT from CECT
ISBI2020 Workshop Presentation 30
Boah Kim
31. Workshop: Deep Learning for Biomedical Image Reconstruction
Medical Image Synthesis with Improved Cycle-GAN: CT from CECT
ISBI2020 Workshop Presentation 31
Boah Kim
Thank you.