Auto-Segmentation Using Deformable Image Registration Lei Dong, Ph.D. Dept. of Radiation Physics University of Texas MD Anderson Cancer Center, Houston, Texas AAPM Therapy Educational Course Aug. 4th 2011 Disclosure Sponsored research and software licensing agreement with Varian Medical Systems Research grants from NIH and the State of Texas on adaptive radiotherapy Objectives Describe deformable image registration methods for auto-segmentation Understand achievable accuracies and validation methods for anatomysegmentation Illustrate applications of atlas-based autosegmentation in radiation therapy 1
Definition: Registration Finding the correct geometrical transformation that brings one image in precise spatial correspondence with another image Deformable Transformation y y Transform Fixed Image x Moving Image x Deformable image registration is a voxel mapping process Definition: Segmentation Resolving the boundaries of an object 2
Atlas Based Segmentation Labeled Objects Contours Ref Image Target Image New Contours Transform Metric Optimizer Deformable Registration Contour Mapping There are many approaches for deformable image registration Sparse (feature) model-based Dense (image intensity) image-based Hybrid methods Validation of Deformable image registration There is no unique solution (degeneracy) Voxels of similar intensities can be grouped differently based on different rules. It is difficult to obtain the ground-truth in patient images Surrogates of truth are needed 3
Validation of Auto-Segmentation Boundary of a structure is usually visible Comparison of expert contours vs. computer-generated contours Impact of Intra- and inter-observer variations Quantitative Validation of Segmentation DICE Similarity Coefficient: Slice-wise 2D Hausdorff distance Evaluation of Atlas-based Segmentation Precision Quality of segmentation on atlas (reference) image set Inter- and intra- observer variations Quality of deformable image registration Accuracy Comparison with independent segmentation method Effectiveness Speed Robustness Practicality 4
Clinical Application of Atlasbased Segmentation Intra-Object Auto-Segmentation 4D CT Adaptive Radiotherapy 4D CT Contour Propagation Transverse Sagittal Coronal Dong et al. 5
Adaptive RT Reference Planning CT Bone Alignment (Daily CT) Adaptive RT Reference Planning CT Adapt to Anatomy (Daily CT) Prostate Radiotherapy Planning contours mapped to 24 in-room CTs Wang H. et al. IJROBP 72 (1):210-219, 2008. 6
Inter-Object Auto-Segmentation Auto-segmentation on a new patient Normal structures Clinical target volumes Contouring from scratch vs. computer assistant (BOT) Contouring from scratch Contouring using deformed template Chao KSC, et al. Int J Radiat Oncol Biol Phys 68 (5):1512-1521, 2007. Contouring from scratch vs. computer assistant (NPX) Contouring from scratch Contouring using deformed template 7
Modified contours vs. unmodified (deformed) contours ROIs (cc) VOI (Modified Contours vs. Computed Generated Contours) VOI (min) VOI (max) Distance 1SD (mm) Agreement (mm) CTV1 93% 88% 96% 1.0 0.3 CTV2 95% 93% 97% 0.9 0.3 CTV3 91% 87% 94% 0.9 0.6 L parotid 89% 84% 96% 0.7 0.3 R parotid 91% 86% 97% 0.6 0.4 Spinal cord 93% 76% 100% 0.3 0.3 Brainstem 88% 77% 100% 1.0 0.6 Larynx 82% 57% 96% 1.1 0.9 Average= 90.2% 81.1% 97.0% 0.8 0.5 VOI Volume Overlap Index (VOI): Vmodified Vdeformed VOI V V 2 modified deformed 3D Surface Distance (BOT Case) Time Savings (minutes) BOT Case NPX Case Physicians Time to contour from scratch (min) Time to modify from deformed contours (min) Ratio Time Saving (min) 1 24 29 1.21-5 2 43 24 0.56 19 3 38 31 0.82 7 4 65 16 0.25 49 5 50 37 0.74 13 6 45 29 0.64 16 7 69 20 0.29 49 8 60 35 0.58 25 Average 49 28 0.64 22 Time Time to contour from Time to modify from Saving Physicians scratch (min) deformed contours (min) Ratio (min) 1 38 33 0.87 5 2 45 35 0.78 10 3 33 35 1.06-2 4 76 39 0.51 37 5 75 38 0.51 37 6 75 40 0.53 35 7 79 30 0.38 49 8 65 45 0.69 20 Average 61 37 0.67 24 Benefits for Experienced and Inexperienced Physicians Experienced H&N IMRT Physicians Time Saving (min) Time to contour from scratch (min) Time to modify from deformed contours (min) Ratio 1 24 29 1.21-5 3 38 31 0.82 7 6 45 29 0.64 16 7 69 20 0.29 49 Average 44 27 0.74 17 Inexperienced H&N IMRT Physicians Time Saving (min) Time to contour from scratch (min) Time to modify from deformed contours (min) Ratio 2 43 24 0.56 19 4 65 16 0.25 49 5 50 37 0.74 13 8 60 35 0.58 25 Average 55 28 0.53 27 BOT Case 8
Benefits for Experienced and Inexperienced Physicians Experienced H&N IMRT Physicians Time Saving (min) Time to contour from scratch (min) Time to modify from deformed contours (min) Ratio 1 38 33 0.87 5 3 33 35 1.06-2 6 75 40 0.53 35 7 79 30 0.38 49 Average 56 35 0.71 22 Inexperienced H&N IMRT Physicians Time Saving (min) Time to contour from scratch (min) Time to modify from deformed contours (min) Ratio 2 45 35 0.78 10 4 76 39 0.51 37 5 75 38 0.51 37 8 65 45 0.69 20 Average 65 39 0.62 26 NPX Case Landmark-based CTV Delineation University of Marburg, Marburg, Germany Atlas based software using 14 landmark points Semi-automatic Mean time: 2.7 min Strassmann et al. IJROBP 2010 Landmark-based CTV Delineation Strassmann et al. IJROBP 2010 9
Atlas-based Lymph Node Segmentation Five physicians contoured on five patients The Simultaneous Truth and Performance Level Estimation (STAPLE) algorithm was used to estimate the truth. Auto-segmentation is more consistent than manual contours, and is closer to the truth Stapleford et al., IJROBP v77 No. 3, 2010 Stapleford et al., IJROBP v77 No. 3, 2010 Stapleford et al., IJROBP v77 No. 3, 2010 10
Auto-Segmentation of Parotid Intra- and inter- observer variations in contouring the atlas case Various image presentation of parotid for different patients J Yang et al. (MDACC), MICCAI, 2010 Multi-Atlas Approach STAPLE J Yang et al. (MDACC), MICCAI, 2010 Auto-Segmentation of Parotid Use STAPLE algorithm to reduce inter-observer variations and deformable registration errors J Yang et al. (MDACC), MICCAI, 2010 11
Auto-plan: automatic segmentation without contour modification Manual-plan: manually drawn contours Auto-plan has low target coverage and higher dose to spinal cord Not clinically acceptable Auto-segmentation may be acceptable for normal structure segmentation, but not robust for CTV/GTV. Tsuji et al., IJROBP v77 No. 3, 2010 Summary of Atlas-based Segmentation Highly practical and relevant to radiation therapy Especially for adaptive RT Accuracy depends on deformable image registration and inter- and intra-observer variations in defining reference atlas Auto-segmentation of normal structure contours may not be perfect but perhaps acceptable for treatment planning CTV segmentation requires physician s validation Improvements are found in efficiency and consistency Acknowledgement Computational Scientists Yongbin Zhang, M.S. Jinzhong Yang, Ph.D. Lifei Zhang, Ph.D. MD Anderson Colleagues Adam Garden, David Schwartz, K. K. Ang, David Rosenthal, Ritsuko Komaki, Zhongxing Liao etc. Catherine Wang, Song Gao etc. 12