Basics of computational assessment for COPD: IWPFI 2017
3 likes1,402 views
1) The document discusses computational assessment methods for chronic obstructive pulmonary disease (COPD) using medical imaging data. It covers topics like lung segmentation, lobe segmentation, airway measurement, vessel quantification, and texture-based emphysema quantification. 2) Various algorithms are presented for tasks like robust lung and lobe segmentation, left/right lung splitting, airway skeletonization and labeling, and classification of pulmonary arteries and veins. 3) Quantitative image analysis methods are discussed for measuring airway wall thickness, quantifying emphysema heterogeneity, and classifying COPD patterns based on texture and shape features.
![Robust Airway Measurement
Inflammation -> Airway Remodeling -> Airway
Narrowing
quantify the extent of airway remodeling in vivo
using CT
Typical surrogate marker
MDCT
Provides bronchial tree geometry with sub-
millimeter resolution
Measure the airway wall thickness, luminal diameter,
wall area, lumen area, wall-lumen area ratio and
wall-lumen diameter ratio
Evaluate the regional airway physiology and
structure
– For the development of disease affecting the airways,
such as asthma and chronic obstructive pulmonary
disease (COPD),
Quantitative Imaging Biomarker
For the determination of bronchial tree dimensions
to assess the efficacy of new drug trial [1,2]
1. Weibel ER, et al. Design and structure of human lung. In: Pulmonary disease and disorders. New York: McGraw-Hill, 1988:11-60
2. Barnes PJ, et al. Lancet 2004;364:985-996
Airway Wall Measurement
With FHWM*](https://image.slidesharecdn.com/basicsofcomputationalassessmentforcopd-170325213933/85/Basics-of-computational-assessment-for-COPD-IWPFI-2017-9-320.jpg)
More Related Content
Similar to Basics of computational assessment for COPD: IWPFI 2017 (20)
Basics of computational assessment for COPD: IWPFI 2017
- 1. BASICS OF COMPUTATIONAL ASSESSMENT FOR COPD 1st PERFUSE Registry Workshop Namkug Kim, PhD Medical Imaging & Robotics Lab. University of Ulsan College of Medicine Asan Medical Center South Korea
- 2. Researches with Hyundai Heavy Industries Co. Ltd. LG Electronics Coreline Soft Inc. Osstem Implant CGBio VUNO Kakaobrain Conflict of Interests Stockholder Coreline Soft, Inc. AnyMedi Co-Founder Somansa Inc. Cybermed Inc. Clinical Imaging Solution, Inc AnyMedi, Inc. Selected Grants as PI National Research Foundation (한국연구재단), South Korea 7T용 4D 자기공명유속영상을이용한 심뇌혈관 질환의 in-vivo 유동 정량화 SW개발, 2016 4D flow MRI을 이용한 심혈관 질환의 in-vivo유동 연구, 2015-7 자기공명분광영상 및 MRI의 통합 분석 소프트웨어 개발 KEIT (산업부), South Korea 의료영상 인공지능 과제, 2016-20 3DP 척추 맞춤형 임플란트, 2016-20 3D 프린터 기반 무치악 및 두개악안면결손환자용 수복 보철물 제작, 재건 시스템 개발, 2015-9 근골격계 복구 수술 로봇 개발, 2012-7 영상중재시술 로봇시스템 개발, 2012-7 Spine및Neurosurgery 수술보조용항법 시스템 개발, 2001 의료용 3차원 모델 제작 S/W 기술 개발, 정통부, 2000 의료영상재구성에 의한 가상시술 소프트웨어 개발, 중소기업기술혁신개발,중기청, 2001 KHIDI (보건복지부), South Korea 영상 뇌졸중 예후 예측 및 치료방침 결정 시스템 개발, 2012-8 관동맥 관류 CT 의 자동 진단 프로그램을활용한 허혈성 질환의 진단과 치료, 2013-6 RP를 이용한 척추나사못 삽입술 계획 프로그램 개발, 2000 Commercial collaboration Hyundai Heavy Industry,Osstem Implant,S&G Biotech, Coreline soft, MidasIT, AnyMedi,Hitachi Medical, Japan,
- 3. Overview of Lung Analysis* 512x512x(512~3000) voxels Anatomical imaging, Qualitative Diagnosis Functional Imaging MR/CT perfusion/ventilation Classification On texture , shape, etc CADD (Computer Aided Differential Diagnosis) On diseases distribution, heterogeneity Surgery Support (replica model, surgery planning, Robot) CBIR (Content based Image Retrieval), CAD Image Repositories Image Queries Image Retrieval Virtual bronchoscopy Quantification On airway measurement, EI, etc Segmentation / Registration Lung/Lobe/Airway Seg. Full ins/ex lung registration *Seo JB, Lee SM @ Radiology AMC (2004.8~) On image based DB query and retrieval
- 4. Basic Lung Segmentation Lung Thoracic Trunk Rolling Ball AlgorithmRegion Growing AlgorithmAirway segmentation and left and right lung split Overall procedure of lung and airway segmentation Fast Seed Based Region Growing Rolling Ball Algorithm Valid Region? Input Seed Points Save Seed Points Retrieve a Seed Point Region segmentation using region growing Save Seed Point Queue Empty Segmentation Result
- 5. Hyper-inflated Lung Left/Right Split Lee MH, Kim N, et al, KSIIM 2011, Best Poster Award, J Digit Imag 2014 Surface fitting with iterative 3D morphological operator with Hessian matrix analysis For automatic lobe segmentation Robust left/right split algorithm for preserving lung volume with same threshold
- 6. Thoracic Cavity Segmentation COPD is systematic diseases Quantification of fat contents in the thoracic cavity. For robust segmentation Surface-fitting method with 5 surfaces including inner rib boundary and diaphragm A two-stage level set method using a shape prior. Additional heart and its surrounding tissue Volumetric overlap ratio (VOR) 98.17 ± 0.84%, 6/33Bae JP, Kim N, et al, Med Physics, 2014 Heart and its surrounding tissues Segmentation Results Surface fitting results
- 7. Robust Lobe Segmentation Park JH, Kim N, et al, RSNA 2011, IWPFI 2017 Robust lobe segmentation is not easy Anatomic & diseases variations including fake, incomplete fissures, COPD and ILD For robust segmentation Surface-fitting method with Hessian matrix and machine learning Comparison among machine learnings
- 8. Diaphragm Segmentation & Quantification Chang YJ, Kim N , IWPFI 2013, Med Phys 2016
- 9. Robust Airway Measurement Inflammation -> Airway Remodeling -> Airway Narrowing quantify the extent of airway remodeling in vivo using CT Typical surrogate marker MDCT Provides bronchial tree geometry with sub- millimeter resolution Measure the airway wall thickness, luminal diameter, wall area, lumen area, wall-lumen area ratio and wall-lumen diameter ratio Evaluate the regional airway physiology and structure – For the development of disease affecting the airways, such as asthma and chronic obstructive pulmonary disease (COPD), Quantitative Imaging Biomarker For the determination of bronchial tree dimensions to assess the efficacy of new drug trial [1,2] 1. Weibel ER, et al. Design and structure of human lung. In: Pulmonary disease and disorders. New York: McGraw-Hill, 1988:11-60 2. Barnes PJ, et al. Lancet 2004;364:985-996 Airway Wall Measurement With FHWM*
- 10. Airway Segmentation / Skeletonization / Labelling Airway Tree Segmentation Lung Segmentation/Split Input: Output: 512x512x(300~2000) voxels Anatomical imaging, Qualitative Diagnosis Skeletonization Labelling Lung 2008, KJ Radiol 2008, KJ Radiol 2008
- 11. Airway Phantom Measurement FWHM method B: the physical phantom filled with poly-urethan form A: eleven artificial tubes of the physical phantom without filled poly-urethane foam C: axial slice of phantom at no tilt. D: axial slice of phantom tilted at 45’ to the scan plane. * Kim N, Seo JB et al, Part I, II, Korean J Radiol 2008 Number of tube Inner Radius Mean ± SD Outer Radius Mean ± SD Wall thickness Mean ± SD 1 0.66 1.56 ± 0.01 0.90 ± 0.01 2 0.63 1.08 ± 0.10 0.45 ± 0.10 3 2.13 5.21 ± 0.02 3.08 ± 0.02 4 1.66 4.12 ± 0.01 2.46 ± 0.01 5 1.8 3.01 ± 0.01 1.21 ± 0.01 6 1.63 2.59 ± 0.00 0.96 ± 0.00 7 1.5 2.06 ± 0.12 0.56 ± 0.12 8 3.23 6.04 ± 0.02 2.81 ± 0.02 9 2.34 4.07 ± 0.01 1.73 ± 0.01 10 4.23 6.01 ± 0.01 1.78 ± 0.01 11 3.51 5.09 ± 0.02 1.58 ± 0.02 Table 1. physical dimensions of artificial airways Overall flow of airway measurement Phantom Study Airway measurement (Green – lumen, Blue – normal wall Cyan – mean of nl wall Pink – outside of 2SD of nl wall Red – mean of nl wall Full Width at Half Maximum* Lung 2008, KJ Radiol 2008, KJ Radiol 2008
- 12. Airway Phantom Measurement : Band Integral Method B: the physical phantom filled with poly-urethan form A: eleven artificial tubes of the physical phantom without filled poly-urethane foam C: axial slice of phantom at no tilt. D: axial slice of phantom tilted at 45’ to the scan plane. Number of tube Inner Radius Mean ± SD Outer Radius Mean ± SD Wall thickness Mean ± SD 1 0.66 1.56 ± 0.01 0.90 ± 0.01 2 0.63 1.08 ± 0.10 0.45 ± 0.10 3 2.13 5.21 ± 0.02 3.08 ± 0.02 4 1.66 4.12 ± 0.01 2.46 ± 0.01 5 1.8 3.01 ± 0.01 1.21 ± 0.01 6 1.63 2.59 ± 0.00 0.96 ± 0.00 7 1.5 2.06 ± 0.12 0.56 ± 0.12 8 3.23 6.04 ± 0.02 2.81 ± 0.02 9 2.34 4.07 ± 0.01 1.73 ± 0.01 10 4.23 6.01 ± 0.01 1.78 ± 0.01 11 3.51 5.09 ± 0.02 1.58 ± 0.02 Table 1. physical dimensions of artificial airways Overall flow of airway measurement Phantom Study Band based avg Density Profile J Comput Assist Tomogr. 2015
- 13. Classification of Pulmonary Artery and Vein For COPD and Pulmonary HT Subtree extraction Weighted minimal spanning tree by cutting branches with lower labels Park SY, Kim N, Seo JB, et al, MIRL, AMCMed Phys. 2013
- 14. Vessel Quantification 14/33 • 10 control with non-contrast volumetric chest CT scans • The radius error :1.57±0.51 mm • The direction error : 8.77±17.20%. Bae JP, Kim N, et al, RSNA 2014
- 15. COPD Quantification S/W Quantification S/W of Emphysema index on HRCT LAA (Low-Attenuation Area), Emphysema Index: Area (volume) % below threshold (-950HU), Mean Lung Density, Lung Volume Lee YK, Kim N, Seo JB, et al, Lung 2008
- 16. Size based Emphysema Analysis Flow Hwang JE, Kim N, et al, IJ COPD 2016, Under Review : IJ COPD 2017
- 17. Quantitative Assessment of Regional Heterogeneity of Emphysema Functional silence of upper lung Automatic quantification of heterogeneity Central to Peripheral Anterior to Posterior Upper to lower Correlation with PFT FEV1 = 24.9 FEV1 = 22.5 • The severity of emphysema in lower lung affects values of PFT more significantly than the severity of emphysema in upper lung. EJ Choi, N Kim, JB Seo, et al, AJ Radiol 2010
- 18. What is Texture? Texton : fundamental element Texture : statistical distribution of texton * P. Brodatz: Textures, A photographic album for artists and designers, Dover Publications, New York, 1966. Examples * Texton Statistical distribution
- 19. Classification of COPD Parenchyma PLE or severe CLE Mild CLE Bronchiolitis obliterans (BO) Normal *Lee YJ, Kim N, Seo JB et al, provisionally accepted at CMPBInv Radiol 2008, Kim N, Seo JB, et al, J Digit Imag 2009
- 20. Shape Features Cluster analysis Preprocessing: segmentation (threshold: - 960HU) and filtering Cluster features Number of Cluster Size (Mean, SD) Circularity (Mean, SD) Aspect ratio: LR/SR (Mean, SD) Top-hat Transformation Extract contrasted component according to the size or shape Suppress the effect of breathhold variation Features from Top-hat White top-hat: mean, SD Black top-hat: mead, SD Original Black Top-hat White Top-hat Comput Meth Prog Bio 2009
- 21. Sensitivity/Specificity & Improvement * Statistically significant difference (p<0.05) Texture Shape Texture+S hape Normal 92.6 72.5 93.8 BO 76.9 68.1 83.9 Mild CLE 78.5 82.2 92.8 PLE/severe CLE 95.9 87.3 99 Overall 85.8 77.2 92.2 0 10 20 30 40 50 60 70 80 90 100 Normal BO Mild CLE PLE/severe CLE Overall Class Sensitivity(%) Texture Shape Texture+Shape 0 10 20 30 40 50 60 70 80 90 100 Normal BO Mild CLE PLE/severe CLE Class Specificity(%) Texture Shape Texture+Shape Texture Shape Texture+S hape Normal 96 90.3 97.3 BO 96.5 90.4 97.4 Mild CLE 91.6 93.2 96.5 PLE/severe CLE 98 96.1 98.6 0 2 4 6 8 10 12 14 16 Normal BO Mild CLE PLE/severe CLE Improvementofsensitivity(%) Improvement of sensitivity (%) after adding shape features Normal 1.2 BO 7 Mild CLE 14.3 PLE/severe CLE 3.1 Kim N, Seo JB, et al, JDI 2011 2nd best paperComput Meth Prog Bio 2009
- 22. Texture-based Quantification TEI : texture-based emphysema index of HRCT = 0.3 x ME% + SE% DEI : density-based emphysema index DEIVol of volumetric CT DEIHR of HRCT SE : Severe Emphysema ME : Mild Emphysema BO : Bronchiolitis Obliterans NL : Normal Lung Park YS, Kim N, Seo JB, et al, Inv Radiol 2008
- 23. Texture-based Quantification 24 Mean area fraction of texture-based quantification • Severe Emphysema : 12.5 ± 16.3 % • Mild Emphysema : 24.0 ± 10.1 % • Bronchiolitis Obliterans : 16.0 ± 10.1 % • Normal Lung : 47.4 ± 26.3 % Severe Emphysema (SE) Mild Emphysema (ME) Bronchiolitis Obliterans (BO) Normal Lung (NL) ** Texture-based Quantification Image Color SE ME BO NL Park YS, Kim N, Seo JB, et al, IR 2008Park YS, Kim N, Seo JB, et al, Inv Radiol 2008
- 24. Texture analysis over cross-vendors Study design Training set Test set Bayesian SVM p-value GE GE GE 82.35 ± 2.82 92.34 ± 2.26 <0.001 Siemens Siemens Siemens 86.26 ± 3.16 91.53 ± 2.07 <0.001 IntegratedSet GE+Siemens GE+Siemens 76.65 ± 2.57 91.18 ± 1.91 <0.001 GE->Siemens GE Siemens 71.98 ± 2.94 82.33 ± 5.75 <0.001 Siemens->GE Siemens GE 71.71 ± 3.6 79.07 ± 3.27 <0.001 Park YJ, Kim N, Med Physics 2013 6 classes (a) Normal lung parenchyma, (b) ground-glass opacity, (c) consolidation, (d) reticular opacity, (e) emphysema, and (f) honeycombing. (normal; green, ground-glass opacity, yellow; reticular opacity, cyan, honeycombing, blue; emphysema, red; and consolidation, pink). (a) GE CT images (b) GE training on GE CT images, (c) the Siemens training on GE CT images, (d) integrated training data on GE CT images. Flow chart
- 25. Fissure Integrity Twenty patients with severe COPD for endobronchial valve volume reduction procedure Fissure Integrity Evaluation Process Lung left and right split lobe segmentation Histogram analysis with maximum likelihood threshold method. Gold standards Two thoracic radiologists (rad1, rad2) Results Completeness (CAD) : 0.982 Accuracy between computer and radiologists : 85% Cohen’s kappa values : rad1 vs rad2, 0.694 / CAD vs rad1, 0.681 / CAD vs rad2, 0.588 / CAD vs radc, 0.700). 26/33Lee MH, Kim N, et al, IWPFI 2017 Read CT data Airway segmentation Lung segmentation Pulmonary vessel segmentation Lobe segmentation Fissure detection Complete fissure Find FIR Maximum density projection Thresholding Incomplete fissure Subtraction Segmentation Fissure detection Find FIR Visual Scoring by radiologists Validation Find Maximum density value at z- axis line 3D fissure surface Binary fissure mask
- 26. Lung Registration for Air Trapping Full inspiration CT Registered full expiration CT Full Expiration CT + Inspiration + Expiration Subtraction color map Deformation map** (with respect to COI) B-spline+ levelset registration Non-rigid Registration Result **Color means deformable distance ( Near : R-G-B : Far) Initial Rigid Registration
- 27. Air Trapping by Using Automatic Registration in COPD Inspiration Expiration Registration Lee HJ,Seo JB, unpublished dataLung 2008, KJ Radiol 2015, Eur Radiol 2016, Eur Radiol 2016
- 28. DECT Xenon Ventilation :AMC First clinical study in the world Research contract with Siemens Medical, CT (advisory board) Pilot study Normal volunteers Patients: COPD, BE, BO Protocol DECT 30% Xenon inhalation Xe ventilator, monitoring devices Dynamic and static whole lung scan Detect in vivo inert Xenon gas Radiology 2008, Radiology 2010, Inv Radiol 2010, Inv Radiol 2016, Eur Radiol 2016 CT images K map in wash-in K map in wash-out AUC map in wash-in Kinetics of Xe ventilation according to Kety’s model Ct = A (1 – e-Kt)
- 29. V/Q Mismatch) Iodine Perfusion(Q) mapXenon Ventilation(V) map V/Q mismatch map unpublishedRSNA 2014
- 30. Quantitative Imaging : Phantom calibration waterarc Outer air1 Inner air Bed 1 Bed 2 Outer air2 CT Phantom Manufact urer Scanner kVp Tube cur rent (mA) Average e ffective tu be current (mA) Slice thick ness Pitch Gantry r otation ti me Reconstruc tion Filter Siemens Sensation 16 140 200 100 0.7 1.000 0.5 B30 Siemens Sensation 64 140 270 99.9 0.7 1.000 0.37 B30 GE LightSpeed 16 140 190 101.2 0.625 0.938 0.5 Standard GE LightSpeed VCT 64 140 250 101.6 0.625 0.984 0.4 Standard Philips Brilliance 16 140 142 100.2 0.8 1.063 0.75 B Philips Brilliance 64 140 135 99.9 0.625 1.014 0.75 B Variation in Emphysema indexes (%) from four different CT scanners. Before DC After DC 3.4 3.6 3.8 4 4.2 4.4 4.6 4.8 5 Time Point(s) EmphysemaIndex(%) Density Correction (outair) ; Standard ; -950 HU Thresholding Siemens 16 Philips 16 (2) Philips 40 Toshiba 64 Outsider air volume density correction based on water and air in four different CT scanners (Water was assumed as 0 HU and air as -1000 HU). FEV1 FEV1/FVC Emphysema index (Base) -0.318 0.002 -0.510 <0.001 Emphysema index (Inner air correction) -0.597 <0.001 -0.612 <0.001 Emphysema index (Outer air correction) -0.394 <0.001 -0.497 <0.001 Mean lung density (Base) 0.259 0.011 0.460 <0.001 Mean lung density (Inner air correction) 0.487 <0.001 0.528 <0.001 Mean lung density (Outer air correction) 0.383 <0.001 0.499 <0.001. Partial correlation analysis adjusted by age and sex between CT and PFT parameters in Philips and Toshiba (n=98).
- 31. Quantification of EI DECT Xe Ventilation DECT Perfusion MR PFT O2 Ventilation DCE-MRI perfusion Wall thickness Lung Evaluation Metrics Lobe Segmentation VQ Mismatch Map Size based EI Disease Classification using Texture Vessel analysis Airway analysis MRICT Diaphragm Thoracic Cavity 32 *Radiology, Med Phys 2013, Eur Radiol, JDI, Med Phys 2016, …
- 32. COPD Characterization 33 Perfusion MR Perfusion Ventilation Emphysema Structure Micro Structure Machine Learning Deep Learning Air flowHemodynamics Hwang HJ, Investigative Radiology (2016), E Beek, Clinics in chest medicine (2015), J thoracic imaging 29(2):80- 91, Hwang JE, IJ COPD (2016), J Applied Physiology (2007) DECT Perfusion MR Ventilation DECT Ventilation VQ mismatch VQ 3He MRI Diffusion For Emphysema hol Segmentation / (B0, B1 Correction) / Registration Single Voxel : Multi- dimensional Data DWI/DTI Perfusion Ventilation Texture Analysis Modeling Structure VQ mismatch Diffusion
- 33. Semi-Automatic Lobe Segmentation Efficient semi-automatic segmentation even in severe COPD patients Whole Working Time < 5 min
- 34. Size based Emphysema Index Efficient semi-automatic segmentation even in severe COPD patients Whole Working Time < 5 min
- 36. Airway Lumen/Wall Analysis Analyze whole airway tree automatically.
- 37. Deep Learning : Semantic Segmentation Detection : RCNN, Fast RCNN, Faster RCNN, YOLO, … Segmentation : FCN, Deconvnet, DCN with CRF, … Super-resolution : SRCNN, DRCN
- 38. MR Segmentation Supine (dimension reduction : 5 times) About 160 slices 3Hr -> 10 sec
- 39. 2.5D CNN Airway Segmentation 80 COPD Patients’ Inspiration CT 69 CT volumes are included in training 11 CT volumes are NOT included in training GS : Manual segmentation 41 Axial 3 slices, Sagittal 3 slices, Coronal 3 slices 32 x 32 x 3 x 3 Weights are shared 2-class classification CT Volume Probabilit y Volume CNN For each voxels inside lungs Segment ed Airway Hard thresholding (0.51) and Select the connected component
- 40. CNN Airway Segmentation Initial Manual 1~2Hr -> 2 min
- 41. Collaborators • MIRL • Clinical Collaborators@Asan Medical Center, SNU BH – Radiology : Chest, Cardiac, Abdomen, Neuro/Brain – Neurology :Dongwha Kang, Chongsik Lee, Jaehong Lee, Sangbeom Jun, Misun Kwon, Beomjun Kim – Cardiology ; Jaekwan Song, Jongmin Song, Younghak Kim – Internal Medicine : Jeongsik Byeon – Pathology : Hyunhee Go – Surgery : Bumsuk Go, JongHun Jeong, Songchuk Kim