Advances in Hardware and Software for Nuclear Medicine

Transcription

1 Advances in Hardware and Software for Nuclear Medicine Jae Sung Lee, PhD Nuclear Medicine & Biomedical Sciences Seoul National University College of Medicine Seoul National University FMIS Lab

2 Advances in SPECT/CT Deep learning for NM Seoul Seoul National University FMIS Lab

3 Advances in SPECT/CT Seoul Seoul National University FMIS Lab

4 SPECT: Tomographic Imaging Single Photon Emission Computed Tomography Dual-head SPECT/CT Triple-head A gamma camera collects SPECT data by means of rotating around the object and acquiring multiple planar images from different angular projections Seoul National University FMIS Lab

5 Dedicated Cardiac SPECT Systems Seoul National University FMIS Lab

6 Cardio-focused Collimator Parallel Conversing Courtesy of Siemens Healthcare Seoul National University FMIS Lab

7 Advances in Software Seoul Seoul National University FMIS Lab

8 Spatially Adaptive Noise Reduction Filter Narrow Moderate Wide Necessary kernel size of smoothing filter Seoul National University FMIS Lab

9 Flash 3D Filter (PIXON, Siemens) Standard time without filter Half time without filter Half time with filter Conventional PIXON Courtesy of Siemens Healthcare & U of Erlangen Conventional 20 min PIXON (onco-flash) 10 min Lee JS & Kim JH, Semin Musculoskelet Radiol 2014 Seoul National University FMIS Lab

10 Iterative Image Reconstruction Estimated projection Measured projection 1st 2nd 3rd 4th Current estimate Forward Projection Backward Projection (update) 1st 2nd 3rd 4th No Comparator Stop? Yes Final solution Error Seoul Seoul National University FMIS Lab

11 Incorporation of Detector Response Estimated projection Measured projection 1st 4th 3rd2nd Current estimate Forward Projection Backward Projection (update) 4th 3rd2nd Comparator Stop? Final Error Conventional: line integration solution Resolution recovery: system & position dependent Seoul Seoul blurring National University FMIS Lab 1st No Yes

12 Collimator-detector Response Modeling Point source Point source Collimator Detector Sharp image (closer point source) Less sharp image (more distant point source) Courtesy of GE Healthcare Seoul National University FMIS Lab

13 Collimator-detector Response Modeling Spatial resolution Parallel hole collimator Gamma camera with parallel hole collimator Distance (or depth) Depth-dependent resolution degradation Projection profile Space invariant response at same depth Seoul Seoul National University FMIS Lab

14 Collimator-detector Response Modeling Tc-99m 8.15 mci (12 sec/frame) 2D MLEM 3D OSEM ASTONISH (Philips) Seung J et al., SNM (2010) Seoul Seoul National University FMIS Lab

15 Collimator-detector Response Modeling Conventional FBP OSEM with CDR modeling 3D OSEM Stansfield et al., Radiology (2010) Seoul Seoul National University FMIS Lab

16 Collimator-detector Response Modeling 2D MLEM ASTONISH 23 sec/frame (Feb 2, 2009) Seung J et al., SNM (2010) 30% scan time reduction 16 sec/frame (Apr 22, 2010) Seoul National University FMIS Lab

17 Collimator-detector Response Modeling FBP Tc-99m MDP SPECT CDR modeling Stress Cardiac SPECT FBP (12 min) CDR (6 min) FBP (12 min) CDR (6 min) Courtesy of GE Healthcare (Evolution software) Seoul National University FMIS Lab

18 Advances in Hardware Seoul Seoul National University FMIS Lab

19 CdZnTe (CZT) SPECT Seoul National University FMIS Lab

20 CdZnTe (CZT) SPECT CZT is the direct convertor of gamma-ray to electric signal. Multi-pinhole collimator Stationary data acquisition without camera rotation CZT semiconductor detector Seoul National University Courtesy of GE Healthcare FMIS Lab

21 CdZnTe (CZT) SPECT Better energy resolution provides the improved performance in the rejection of scatter and crosstalk photons Multi-pinhole collimator Stationary data acquisition without camera rotation CZT semiconductor detector Seoul National University Courtesy of GE Healthcare FMIS Lab

22 CZT vs Conventional SPECT Conventional 14 min rest &12 min stress CZT+Multipinhole 4 min rest & 2 min stress Garcia et al. JNM (2011) Energy resolution Spatial resolution (mm) Central Tangentia l Sensitivity (kcps/mci) Radial Point Phantom Patients CZT 5.7% Ventri 9.4% Ratio x 2 x 2 x 6 Seoul National University FMIS Lab

23 Quantitative Dynamic SPECT Myocardial Perfusion Reserve (MPR) = K1_stress/K1_rest Seoul National Ben-Haim S et University al. JNM (2013) FMIS Lab

24 Whole-body CZT SPECT/CT GE NM/CT 870 CZT Spectrum Dynamics VERITON-CT Seoul National University FMIS Lab

25 SPECT/CT Cone-beam CT based MD CT based Seoul Seoul National University FMIS Lab

26 Advantages of SPECT/CT Simultaneous acquisition of functional and anatomic imaging dataset Anatomical localization and correlation with SPECT abnormality using CT More accurate reconstruction of SPECT data with quantitative capabilities using CTbased attenuation map Used in wide range of diseases including endocrinology, cardiology, and neurology Seoul Seoul National University FMIS Lab

27 More Confident Anatomical Localization A B C Anterior Posterior A 76-year-old female patient who experienced lower back pain underwent bone scan and SPECT/CT imaging after injection of 99m Tc-MDP to find the focus of active lesion. (a) On bone scan, a focal increased uptake was found in the left side of the L4 L5 intervertebral area. (b) On SPECT/CT, however, the focal uptake was revealed to be on the right facet joint of L4 L5. (c) In addition, a focal uptake was also found on the left facet joint of L3 L4 that was not noticed on the bone scan. Lee JS & Seoul Kim JH, Semin National Musculoskelet University Radiol 2014 FMIS Lab

28 Improved Quantitative Accuracy Seoul National University FMIS Lab

29 Evaluation of Accuracy of Quantitative SPECT Petri dish (PETRI) Small phantom (SMALL) Small uniform phantom (S.UNIFORM) Uniform phantom (UNIFORM) : 15 cm, h: 1.5 cm Dimension not measured (200 ml) : 12 cm, h: 15.5 cm (1573 ml) : 20 cm, h: 21 cm (6250 ml) Jaszczak phantom w/o sphere (JASZCZAK) Eliptical Jaszczak phantom w/o sphere (E. JASZCZAK) IEC body phantom (IEC BODY) : 21.6 cm, h: 18.6 cm (6900 ml) major axis : 30.5 cm, minor axis : 22.1 cm, h: 18.6cm (9847 ml) Sphere : 10, 13, 17, 22, 37 mm Cylindrical insert : 51 x h 180 mm Seoul (9700 National ml) University FMIS Lab

30 Evaluation of Accuracy of Quantitative SPECT Measured Activity (mci) Measured Activity (mci) Planar Sensitivity of 99m Tc ± 0.02 R 2 = Calculated Activity (mci) Volume Sensitivity of 99m Tc ± 0.02 R 2 = Calculated Activity (mci) Seoul National University FMIS Lab

31 State-of-art SPECT/CT Scanners Compact detector heads (mostly dual) Digital Anger logic Auto-contour Adaptive filters Resolution recovery reconstruction Combined with MDCT Iterative reconstruction for dose reduction Real-time CT dose modulation Seoul Seoul National University FMIS Lab

32 Summary SPECT is a tomographic imaging technique in which the different angular projections are acquired by rotating the gamma cameras. Recent technical advances in SPECT include Adaptive noise reduction filter, collimator-detector response modeling, and stationary cardiac imaging using CZT cameras SPECT/CT systems with better quantitative and diagnostic, and radiation dose management Seoul National University FMIS Lab

33 for Nuclear Medicine Imaging

34 Image Restoration and Inpanting Seoul National University FMISL

35 PET Image Denoising Low dose PET image High dose PET image EM+Filter CNN-Finetune CNN-W/O-Finetune Seoul National University FMISL Courtesy of Prof. Qi Jinyi (UC Davis) Gong K, Qi J et al. (2017 IEEE MIC)

36 Reconstruction using CNN Prior Cost Function Full dose (185 MBq) Low dose (1/10) OPOSEM OPOSEM OPOSEM + Gaussian CNN Prior Seoul National University FMISL Courtesy of Dr. Kim Kyungsang (MGH) Kim K, Li Q et al. (2017 IEEE MIC)

37 Joint Reconstruction with TOF + Deep Learning Activity MLAA with TOF Deep learning X-ray CT Emission PET Sinogram Attenuation map Seoul National University FMISL Hwang D, Kim KY et al. (J Nucl Med, 2018)

38 Results: Case 1 CNN1 CNN2 CT-based

39 Super Resolution & Pixel Restoration Seoul National University FMISL 30 AMAZING APPLICATIONS OF DEEP LEARNING (

40 Input Ground truth Sharp Filter Richardson Lucy deblurring Deep Learning Seoul National University FMISL Park J et al., Phys Med Biol 2018

41 Input Ground truth Sharp Filter Richardson Lucy deblurring Deep Learning Seoul National University FMISL Park J et al., Phys Med Biol 2018

42 Training Deep Neural Network Under-sampled Fully sampled Seoul National University FMISL

43 Which is Real? Seoul National University FMISL Kang SK et al. (unpublished data)

44 Which is Real? Seoul National University FMISL Kang SK et al. (unpublished data)

45 ½ Under-sampled Reconstructed images Seoul National University FMISL Kang SK et al. (unpublished data)

46 ¼ Under-sampled Seoul National University FMISL Kang SK et al. (unpublished data)

47 Renal Parenchyma Segmentation 3D Smoothing Preprocessing Image cropping (x, y, z) & Down sampling 3D Smoothing Target Data 3D Deep Learning Seoul National University FMISL GFR Measurement Park J et al., (unpublished data)

48 Manual vs. U-net (A) (B) SPECT X-ray CT Seoul National University FMISL Park J et al., (unpublished data)

49 Manual vs. U-net (A) (B) SPECT X-ray CT Seoul National University FMISL Park J et al., (unpublished data)

50 Manual vs. U-net (A) (B) SPECT X-ray CT Seoul National University FMISL Park J et al., (unpublished data)

51 (A) Manual vs. U-net (B) SPECT X-ray CT Seoul National University FMISL Park J et al., (unpublished data)

52 Manual vs. U-net (A) (B) SPECT X-ray CT Seoul National University FMISL Park J et al., (unpublished data)

53 Deep Learning Researches Artifact reduction Organ segmentation Super resolution 61 Scan time reduction