Combination of Markerless Surrogates for Motion Estimation in Radiation Therapy

Transcription

1 Combination of Markerless Surrogates for Motion Estimation in Radiation Therapy CARS 2016 T. Geimer, M. Unberath, O. Taubmann, C. Bert, A. Maier June 24, 2016 Pattern Recognition Lab (CS 5) FAU Erlangen-Nu rnberg

2 Motivation External Beam Radiation Therapy Planning CT Optimized dose distribution Respiratory motion Target volume displacement! Potential survival of malignant cells c BrainLAB AG June 24, 2016 T. Geimer Pattern Recognition Lab (CS 5) Combination of Markerless Surrogates for Motion Estimation 2

3 Motion Estimation Clinical State of the Art Marker Tracking Control points based Only sparse deformation kv tube IR cam kv tube June 24, 2016 T. Geimer Pattern Recognition Lab (CS 5) Combination of Markerless Surrogates for Motion Estimation 3

4 Motion Estimation Clinical State of the Art Marker Tracking Control points based Only sparse deformation Novel Approaches Range Imaging Dense surface information Non-intrusive kv tube RI cam kv tube Fluoroscopy Image-based signal Internal motion models Dense internal deformation June 24, 2016 T. Geimer Pattern Recognition Lab (CS 5) Combination of Markerless Surrogates for Motion Estimation 3

5 Outline Motion Estimation Motion Models Correlation & Prediction Multi Surrogate Approach Results & Discussion Outlook June 24, 2016 T. Geimer Pattern Recognition Lab (CS 5) Combination of Markerless Surrogates for Motion Estimation 4

6 Motion Models Reference Surface extracted from CT 3D + t CT n+1volumes Internal Motion Non-rigid 3D/3D registration w.r.t. reference phase tref 3D/3D Registration Cropping to internal ROI Interpolation of vector fields n internal deformation fields Surface Motion n surface deformation fields PCA PCA Surface extracted from 3-D data set Deformation fields d i interpolated at mesh-vertices Principle Components of Internal Motion Model 2 n feature vectors 1. dim 2. dim 3. dim Internal ASM June 24, 2016 T. Geimer n feature vectors 1. dim 2. dim 3. dim Principle Components of Surface Motion Model Surface ASM Pattern Recognition Lab (CS 5) Combination of Markerless Surrogates for Motion Estimation 5

7 Framework Overview Internal Motion Surface Surrogate Fluoro Surrogate Range Imaging n surface deformation fields X-ray Fluoroscopy n images Active Shape Model Active Shape Model 3D + t CT n+1volumes 3D/3D Registration Individual n internal deformation fields June 24, 2016 T. Geimer Active Shape Model Correlation Multi-linear regression between model weights Pattern Recognition Lab (CS 5) Combination of Markerless Surrogates for Motion Estimation 6

8 Framework Overview Internal Motion Surface Surrogate Fluoro Surrogate Range Imaging n surface deformation fields X-ray Fluoroscopy n images Active Shape Model Active Shape Model 3D + t CT n+1volumes 3D/3D Registration Individual n internal deformation fields June 24, 2016 T. Geimer Active Shape Model Correlation Multi-linear regression between model weights Pattern Recognition Lab (CS 5) Combination of Markerless Surrogates for Motion Estimation 6

9 Framework Overview Internal Motion Surface Surrogate Fluoro Surrogate Range Imaging n surface deformation fields X-ray Fluoroscopy n images Active Shape Model Active Shape Model 3D + t CT n+1volumes 3D/3D Registration Combination n internal deformation fields June 24, 2016 T. Geimer Active Shape Model Correlation Multi-linear regression between model weights Pattern Recognition Lab (CS 5) Combination of Markerless Surrogates for Motion Estimation 6

10 Principle Components of Fluoro Motion Model 1. dim 2. dim 3. dim Fluoroscopy Surrogate Model Preprocessing DRRs from volumetric CT data According to Vero kv beam geometry Breathing Signal Extraction Linearized images 3D + t CT n volumes (no ref) forward projection n DRRs PCA of the change in intensity First few principle component correlate highly with breathing phase PCA n feature vectors June 24, 2016 T. Geimer Pattern Recognition Lab (CS 5) Combination of Markerless Surrogates for Motion Estimation 7

11 Multi-surrogate Approach Motivation Combined information from range imaging and fluoroscopy Improve the estimation or compensate for one surrogate failing Combination Concatenated [ ] feature vector σrii σ CB = R (f RI+f FL ) 1 i σ FL i Regression matrix R CB R l (f RI+f FL ) Example: R CB R 3 (2+2) σ φ RI1 1 φ 2 = R CB σ RI2 σ φ FL1 3 σ FL2 June 24, 2016 T. Geimer Pattern Recognition Lab (CS 5) Combination of Markerless Surrogates for Motion Estimation 8

12 Evaluation Data Nine 4-D CT patient data sets Ten volumes each (voxel size: mm 3 ) DRRs: detector with pixels and 0.39 mm pixel spacing Experiments Correlation study on feature weights Leave-one-out-assessment of the estimation error w.r.t. ground-truth June 24, 2016 T. Geimer Pattern Recognition Lab (CS 5) Combination of Markerless Surrogates for Motion Estimation 9

13 Results: Correlation Pat3 Pat9 1. PC 2. PC 3. PC June 24, 2016 T. Geimer Pattern Recognition Lab (CS 5) Combination of Markerless Surrogates for Motion Estimation 10

14 Results: Estimation Estimation error / mm RI FL CB (1/1) CB (2/2) CB (3/3) CB (4/4) 0 l=1 l=2 l=3 l=4 Internal model dimensionality reference mean: 2.31 ± 0.70 mm June 24, 2016 T. Geimer Pattern Recognition Lab (CS 5) Combination of Markerless Surrogates for Motion Estimation 11

15 Discussion Single Surrogate Fluoro outperforms surface surrogate: 0.67 ± 0.33 mm for l = 2 No improvement with higher internal model dimension Combination Approach No consistent improvement f RI, f FL = 1: 1-D respiratory phase Linearly dependent Best overall for l = 2 from (2+2) or higher combined features: 0.62 ± 0.28 mm Surrogate combination useful under certain circumstances June 24, 2016 T. Geimer Pattern Recognition Lab (CS 5) Combination of Markerless Surrogates for Motion Estimation 12

16 Outlook Surrogate Extraction DR eliminates non-redundant information Sophisticated ways to extract mutually exclusive information Data Acquisition & Training Training currently only done on 6 of 9 phases (necessity of leave-one-out approach) ASMs do not describe an entire cycle Training on entire cycle Prediction on another of the same patient June 24, 2016 T. Geimer Pattern Recognition Lab (CS 5) Combination of Markerless Surrogates for Motion Estimation 13

17 Thank you for your attention. Questions? June 24, 2016 T. Geimer Pattern Recognition Lab (CS 5) Combination of Markerless Surrogates for Motion Estimation 14