Attenuation Correction in Positron Emission Tomography Using Simultaneous Activity and Attenuation Reconstruction and Deep Neural Network = 동시영상재구성 기법과 심층 신경망을 이용한 양전자방출단층촬영의 감쇠보정

다국어 초록 (Multilingual Abstract)

Attenuation correction (AC) is essential for yielding quantitative and qualitative accurate functional information in positron emission tomography (PET). In clinical PET/CT, CT image is converted to attenuation map of 511 keV photon, and used in attenuation correction. However, CT-based PET AC has a fundamental limitation in its quantitative accuracy, because position mismatches between PET and CT scans lead to artifacts. AC in PET/MRI has difficulty in estimating accurate bone structures, and also suffers from position mismatches between PET and MR scans.
Simultaneous reconstruction of activity and attenuation reconstruction is a potential alternative method for PET AC. With the recent advancement of time-of-flight (TOF) technology, the simultaneous reconstruction algorithm allows estimating attenuation maps using only emission data. Because no anatomical images are necessary for the AC if the simultaneous reconstruction works properly, it is a potentially significant approach to overcoming the above-mentioned limitations of PET AC in PET/CT and PET/MRI. However, because of the limited timing resolution of current clinical PET scanners, the MLAA suffers from several problems, including the crosstalk artifacts (between activity and attenuation maps), slow convergence speed, and noisy attenuation maps. Another limitation of AC when using simultaneous reconstruction is the chicken-egg dilemma of the scatter estimation.
In this thesis, a deep learning-based PET attenuation map generation method that combines simultaneous reconstruction and convolution neural networks (CNNs) is proposed to improve the accuracy of PET AC and the feasibility of proposed method was validate. Combination of non-attenuation-corrected (NAC) PET and CNNs is also proposed to estimate with scatter distribution. To fully utilize emission data, an investigation of input combinations for CNNs was also conducted. In addition, this PET attenuation map generation method was applied to alleviating the respiratory motion, which causes degradation of PET image quality.
The proposed method was effectively mitigated noise and cross-talk artifacts in the attenuation map of simultaneous reconstruction in both brain and whole-body PET studies. The proposed method has the advantage that the output attenuation maps are free from the influence of PET and CT mismatches as the inputs of CNNs were derived from emission data and therefore CT scan is unnecessary.
NAC PET was useful for estimating the scatter distribution, which is required for applying simultaneous reconstruction. Also, AC performance with different combinations of CNN inputs was evaluated and compared.
With the proposed method, attenuation maps corresponding to each respiratory phase could be generated from gated PET. These phase-matched attenuation maps enabled motion estimation among respiratory phases as well as the phase-matched PET attenuation correction. Consequently, motion-free PET images could be generated, showing better lesion detectability and mitigated respiratory motion artifacts.
In conclusion, the PET attenuation map generation method developed in this thesis yielded high-quality attenuation maps, which enables accurate PET AC. This method has the potential to supplement and/or replace the conventional AC methods in PET/CT, PET/MRI, and even stand-alone PET.

국문 초록 (Abstract)

체내에 주입한 방사성의약품의 분포를 알기 위해 실시하는 양전자방출단층촬영 (PET)에서 정량적 및 정성적으로 정확한 정보를 얻기 위해서는 감쇠보정이 필수적이다. PET/CT에서는 CT영상을 511keV 광자의 감쇠영상으로 변환하여 감쇠보정을 실시한다. 그러나 CT영상 기반의 감쇠보정은 양전자방출단층촬영 영상과 CT영상 간의 위치 불일치 오류로 인한 아티팩트가 발생하여 정확도에 한계가 있다. PET/MRI에서는 MR 영상을 이용하여 감쇠보정이 가능하나, MR 영상에서 정확한 뼈의 구조를 얻기 어려우며, PET/CT에서와 마찬가지로 양전자방출단층촬영 영상과 MR 영상 간의 위치 불일치 오류 역시 한계로 지적된다.
감쇠보정을 위한 다른 방법은 동시영상재구성 기법을 이용하는 것이다. 비정 시차 정보를 이용하는 동시영상재구성 기법을 사용하면 방출 영상만 이용하여 감쇠영상을 추정 가능하다. 이 과정에서 해부학적 정보가 필요하지 않기 때문에, PET/CT와 PET/MRI에서 지적되었던 위치 불일치 오류를 극복 가능하다. 하지만, 현재 사용되는 양전자방출단층촬영 장비들의 비정 시차 성능이 충분하지 못해, 높은 잡음 레벨을 갖고 혼선 아티팩트가 나타나는 불완전한 감쇠영상이 산출된다.
본 논문에서는 양전자방출단층촬영 감쇠보정의 정확도를 향상시키고자 동시영상재구성 기법에 합성곱 신경망을 결합하여 딥러닝 기반의 감쇠보정을 제안하고, 이의 타당성을 검증하였다. 동시영상재구성 기법에서 한 가지 제한점은 감쇠영상과 산란 추정의 얽힘 현상이다. 이를 해결하고자 비감쇠보정 영상과 합성곱 신경망을 결합하여 산란 분포를 추정하는 것을 제안하였다. 또한, 합성곱 신경망에 가장 적합한 입력 조합에 대한 평가를 진행하였다. 추가적으로, 양전자방출단층촬영 영상의 품질을 저하시키는 호흡 운동의 영향을 완화하는데 제안된 감쇠보정을 적용해보았다.
제안된 감쇠보정은 뇌 및 전신 양전자방출단층촬영 연구 모두에서 동시영상재구성 기법의 감쇠영상의 노이즈와 혼선 아티팩트를 효과적으로 완화하였다. 방출 영상만이 합성곱 신경망의 입력으로 사용되기 때문에, 출력된 감쇠영상은 양전자방출단층촬영 영상과 CT 영상 간의 위치 불일치 오류의 영향을 받지 않는다는 장점을 갖는다.
비감쇠보정 영상은 동시영상재구성 기법에 필요한 산란 분포를 추정하는데 유용함이 검증되었다. 또한 합성곱 신경망 입력의 서로 다른 조합에 대한 감쇠 보정 성능을 평가 및 비교하였다.
해당 감쇠보정을 적용하여 각 호흡 단계에 대응되는 감쇠영상을 생성하였다. 이러한 위상 일치 감쇠영상은 호흡 단계간의 모션 추정 및 위상 일치 감쇠보정을 가능케했다. 결과적으로, 움직임이 없는 양전자방출단층촬영 영상을 획득할 수 있었으며, 병변 검출성능이 향상되고 및 호흡 운동으로 인한 아티팩트는 완화되었다.
결론적으로, 본 논문에서 개발한 방법은 정확한 감쇠보정을 가능하게 하는 고품질의 감쇠영상을 산출하였다. 이 감쇠보정은 PET/CT, PET/MRI는 물론, 독립형 PET에서 기존 감쇠보정을 보완 및 대체할 수 있을 것으로 기대된다.

목차 (Table of Contents)

• Chapter 1. Introduction 1
• 1.1. Attenuation Correction in PET 1
• 1.2. Simultaneous Reconstruction 2
• 1.2.1. MLAA algorithm 2
• 1.2.2. Limitations of MLAA 3
• 1.3. Contributions 4
• Chapter 2. Improving Accuracy of Simultaneously Reconstructed Activity and Attenuation Maps Using Deep Learning 5
• 2.1. Background 5
• 2.2. Materials and Methods 6
• 2.2.1. Dataset 6
• 2.2.2. Network Architecture 8
• 2.2.3. Data Augmentation and Training 10
• 2.2.4. Image Analysis 11
• 2.3. Results 12
• 2.4. Discussion 18
• Chapter 3. Generation of PET Attenuation Map for Whole-body Time-of-flight 18F-FDG PET/MRI Using a Deep Neural Network Trained with Simultaneously Reconstructed Activity and Attenuation Maps 21
• 3.1. Background 21
• 3.2. Materials and Methods 22
• 3.2.1. Subjects and Image Acquisition 22
• 3.2.2. Dataset 23
• 3.2.3. Network Architecture 25
• 3.2.4. Network Training and Loss Function 25
• 3.2.5. Image Analysis 26
• 3.3. Results 27
• 3.3.1. Attenuation Maps 27
• 3.3.2. Activity Maps 32
• 3.3.3. Comparison to 2D network 35
• 3.3.4. Computation Time 35
• 3.4. Discussion 36
• Chapter 4. Comparison of Deep Learning-based Emission-Only Attenuation Correction Methods for Positron Emission Tomography 39
• 4.1. Background 39
• 4.2. Materials and Methods 41
• 4.2.1. Dataset 41
• 4.2.2. Network Architectures 42
• 4.2.3. Scatter Estimate Comparison 44
• 4.2.4. Comparison of Attenuation and Activity Estimates 44
• 4.3. Results 46
• 4.3.1. Scatter Estimation using μ-CNNNAC 46
• 4.3.2. Attenuation Maps 50
• 4.3.3. Activity Images 54
• 4.4. Discussion 57
• Chapter 5. Data-driven Respiratory Phase-matched PET Attenuation Correction without CT 60
• 5.1. Background 60
• 5.2. Materials and Methods 63
• 5.2.1. Dataset 63
• 5.2.2. Simultaneous Image Reconstruction 64
• 5.2.3. Data-driven Respiratory Signal Measurement 64
• 5.2.4. Network Architecture and Training 66
• 5.2.5. Additional Activity Image Reconstruction 67
• 5.2.6. Estimating Motion Vector Field and Non-rigid Registration 67
• 5.2.7. Image Analysis 68
• 5.2.8. Comparison with Other Registration Schemes 68
• 5.3. Results 69
• 5.3.1. Phase-matched Attenuation Correction 69
• 5.3.2. Motion-free Image Generation 72
• 5.3.3. Comparison with Other Registration Scheme 75
• 5.4. Discussion 75
• Chapter 6. Summary and Conclusion 80
• Bibliography 81
• Abstract in Korean 90

참고문헌 (Reference)

1. `` Deep learning in medical image analysis, D. Shen , G. Wu , and H.-I . Suk, vol . 19 , pp . 221-248, , 2017

2. `` Machine learning in biomedical engineering, C. Park , C. C. Took , and J. K. Seong, vol . 8 , pp . 1-3, , 2018

3. Deep Learning in neural networks : An overview, J. Schmidhuber, vol . 61 , pp . 85-117, , 2015

4. `` Low-dose CT via convolutional neural network, H. Chen et al., vol . 8 , pp . 679-694, , 2017

5. `` Model-based scatter correction for fully 3D PET, J. M. Ollinger ,, vol . 41 , pp . 153-176, , 1996

6. `` Integrated whole body MR/PET : where are we ? ``, H. J. Yoo , J. S. Lee , and J. M. Lee, vol . 16 , pp . 32-49, , 2015

7. Representation learning : A review and new perspectives, Y. Bengio , A. Courville , and P. Vincent, vol . 35 , pp . 1798-1828, , 2013

8. `` Quantitation of respiratory motion during 4DPET/ CT acquisition, S. A. Nehmeh et al. ,, vol . 31 , pp . 1333-8, , 2004

9. `` PET image denoising using a deep neural network through fine tuning, K. Gong , J. Guan , C.-C. Liu , and J. Qi ,, vol . 3 , pp . 153-161, , 2018

10. `` Deep learning-based image segmentation on multimodal medical imaging, Z. Guo , X. Li , H. Huang , N. Guo , and Q. Li ,, vol . 3 , pp . 162-169, , 2019

11. `` Amyloid PET quantification via end-to-end training of a deep learning, J. Y. Kim et al., vol . 53 , pp . 340-348, , 2019

12. `` On predicting epileptic seizures from intracranial electroencephalography, Y. Yoo, vol . 7 , pp . 1-5, , 2017

13. `` Sources of attenuation-correction artefacts in cardiac PET/CT and SPECT/CT, S. J. McQuaid and B. F. Hutton ,, vol . 35 , pp . 1117-1123, , 2008

14. `` Adaptive template generation for amyloid PET using a deep learning approach, S. K. Kang et al., vol . 39 , pp . 3769-3778, , 2018

15. `` Analysis and comparison of two methods for motion correction in PET imaging ,, I. Polycarpou , C. Tsoumpas , and P. K. Marsden ,, vol . 39 , pp . 6474-6483 ,, , 2012

16. `` Simultaneous PET-MRI : a new approach for functional and morphological imaging, M. S. Judenhofer et al. ,, vol . 14 , pp . 459-465, , 2008

17. `` Time-of-flight PET data determine the attenuation sinogram up to a constant ,, M. Defrise , A. Rezaei , and J. Nuyts ,, vol . 57 , pp . 885-899, , 2012

18. `` Concept of an upright wearable positron emission tomography imager in humans ,, C. E. Bauer et al. ,, vol . 6 , pp . e00530 ,, , 2016

19. `` Attenuation correction in emission tomography using the emission data ? a review, Y. Berker and Y. Li ,, vol . 43 , pp . 807-832 ,, , 2016

20. `` Simultaneous reconstruction of activity and attenuation in time-of-flight PET ,, A. Rezaei et al. ,, vol . 31 , pp . 2224-2233, , 2012

21. `` Analysis of functional brain images using population-based probabilistic atlas ,, J. S. Lee and D. S. Lee ,, vol . 1 , pp . 81-87 ,, , 2005

22. `` MR-based synthetic CT generation using a deep convolutional neural network method, X. Han, vol . 44 , pp . 1408-1419, , 2017

23. `` First prototyping of a dedicated PET system with the hemisphere detector arrangement, H. Tashima et al. ,, vol . 64 , pp . 065004, , 2019

24. `` Markerless motion tracking and correction for PET , MRI , and simultaneous PET/MRI ,, J. M. Slipsager et al. ,, vol . 14 , pp . e0215524 ,, , 2019

25. `` Deep-learning-based automatic computer-aided diagnosis system for diabetic retinopathy, R. F. Mansour ,, vol . 8 , pp . 41-57, , 2018

26. `` Convolutional neural networks for medical image analysis : Full training or fine tuning ? ``, N. Tajbakhsh et al. ,, vol . 35 , pp . 1299-1312, , 2016

27. `` Synthetic CT generation from non-attenuation corrected PET images for whole-body PET imaging, X. Dong et al. ,, vol . 64 , pp . 215016, , 2019

28. `` The validation problem of joint emission/transmission reconstruction from TOF-PET projections ,, J. Nuyts , A. Rezaei , and M. Defrise ,, vol . 2 , pp . 273- 278, , 2018

29. `` X-ray-based attenuation correction for positron emission tomography/computed tomography scanners ,, P. E. Kinahan , B. H. Hasegawa , and T. Beyer ,, vol . 33 , pp . 166-79 ,, , 2003

30. `` 3D position estimation using an artificial neural network for a continuous scintillator PET detector, Y. Wang , W. Zhu , X. Cheng , and D. Li, vol . 58 , pp . 1375-1390 ,, , 2013

31. `` Deep learning in nuclear medicine and molecular imaging : current perspectives and future directions, H. Choi, vol . 52 , pp . 109-118, , 2018

32. Improving the accuracy of simultaneously reconstructed activity and attenuation maps using deep learning, D. Hwang et al., vol . 59 , pp . 1624-1629, , 2018

33. `` Proof-of-concept prototype time-of-flight PET system based on high-quantum-efficiency multianode PMTs, J. W. Son et al. ,, vol . 44 , pp . 5314-5324, , 2017

34. `` A deep convolutional neural network using directional wavelets for low ? dose X ? ray CT reconstruction, E. Kang , J. Min , and J. C. Ye ,, vol . 44 , pp . e360-e375, , 2017

35. `` Quantification of F-18 FDG PET images in temporal lobe epilepsy patients using probabilistic brain atlas, K. W. Kang et al., vol . 14 , pp . 1-6, , 2001

36. `` Novel adversarial semantic structure deep learning for MRI-guided attenuation correction in brain PET/MRI, H. Arabi , G. Zeng , G. Zheng , and H. Zaidi ,, vol . 46 , pp . 2746-2759, , 2019

37. `` Practical joint reconstruction of activity and attenuation with autonomous scaling for time-of-flight PET, Y. Li , S. Matej , and J. S. Karp ,, vol . 65 , pp . 235037 ,, , 2020

38. `` Multi class disorder detection of magnetic resonance brain images using composite features and neural network, V. V. Kale , S. T. Hamde , and R. S. Holambe ,, vol . 9 , pp . 221-231, , 2019

39. `` Neural network evaluation of PET scans of the liver : a potentially useful adjunct in clinical interpretation, O. Preis , M. A. Blake , and J . A. Scott ,, vol . 258 , pp . 714-21, , 2011

40. `` Artificial neural network classifier for the diagnosis of Parkinson 's disease using [ 99mTc ] TRODAT-1 and SPECT, P. D. Acton and A. Newberg ,, vol . 51 , pp . 3057-66, , 2006

41. `` Initial results of simultaneous PET/MRI experiments with an MRI-compatible silicon photomultiplier PET scanner ,, H. S. Yoon et al. ,, vol . 53 , pp . 608-614, , 2012

42. Deep-dose : a voxel dose estimation method using deep convolutional neural network for personalized internal dosimetry ,, M. S. Lee , D. Hwang , J. H. Kim , and J. S. Lee, vol . 9 , pp . 10308, , 2019

43. Simultaneous Multiparametric PET/MRI with Silicon Photomultiplier PET and Ultra-High-Field MRI for Small-Animal Imaging ,, G. B. Ko et al. ,, vol . 57 , pp . 1309-1315, , 2016

44. `` A novel loss function incorporating imaging acquisition physics for PET attenuation map generation using deep learning, L. Shi et al., pp . 723-731, , 2019

45. `` Toward implementing an MRI-based PET attenuationcorrection method for neurologic studies on the MR-PET brain prototype, C. Catana et al. ,, vol . 51 , pp . 1431-1438, , 2010

46. `` The effect of regularization in motion compensated PET image reconstruction : a realistic numerical 4D simulation study, C. Tsoumpas et al. ,, vol . 58 , pp . 1759, , 2013

47. `` Respiratory motion compensation for PET/CT with motion information derived from matched attenuation-corrected gated PET data, Y. Lu et al., vol . 59 , pp . 1480-1486, , 2018

48. `` Quantitative Evaluation of Atlas-based Attenuation Correction for Brain PET in an Integrated Time-of-Flight PET/MR Imaging System, J. Yang et al. ,, vol . 284 , pp . 169-179, , 2017

49. `` Clinical assessment of emission-and segmentation-based MR-guided attenuation correction in whole-body timeof- flight PET/MR imaging, A. Mehranian and H. Zaidi ,, vol . 56 , pp . 877-883, , 2015

50. `` Clinical evaluation of zero-echo-time attenuation correction for brain 18F-FDG PET/MRI : comparison with atlas attenuation correction ,, T. Sekine et al. ,, vol . 57 , pp . 1927-1932, , 2016

51. `` Joint correction of attenuation and scatter in image space using deep convolutional neural networks for dedicated brain ( 18 ) F-FDG PET, J. Yang , D. Park , G. T. Gullberg , and Y. Seo ,, vol . 64 , pp . 075019, , 2019

52. `` Obtaining PET/CT images from non-attenuation corrected PET images in a single PET system using Wasserstein generative adversarial networks, Z. Hu et al., vol . 65 , pp . 215010 ,, , 2020

53. Sensitivity increase through a neural network method for LOR recovery of ICS triple coincidences in high-resolution pixelateddetectors PET scanners, J . B. Michaud et al. ,, vol . 62 , pp . 82- 94, , 2015

54. `` Blind separation of cardiac components and extraction of input function from H ( 2 ) ( 15 ) O dynamic myocardial PET using independent component analysis, J. S. Lee et al., vol . 42 , pp . 938-943, , 2001

55. `` Comparison of segmentation-based attenuation correction methods for PET/MRI : evaluation of bone and liver standardized uptake value with oncologic PET/CT data, J. H. Kim , J. S. Lee , I.-C. Song , and D. S. Lee, vol . 53 , pp . 1878-1882, , 2012

56. `` Generation of PET attenuation map for whole-body time-of-flight 18F-FDG PET/MRI using a deep neural network trained with simultaneously reconstructed activity and attenuation maps, D. Hwang et al., vol . 60 , pp . 1183-1189, , 2019