Seite - 51 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

Sim.Metric #1 #2 #3 Mean Std. dmin (mm) 2.43 2.32 2.95 2.57 0.34 dortho (mm) 2.36 2.05 3.36 2.59 0.68 DOV (%) 98.01 92.87 82.11 91.00 8.11 VOD (%) 74.72 80.13 68.12 74.32 6.01 κ (%) 87.49 90.41 79.75 85.88 5.51 Table2. EvaluationofLVshaperecovery fromthree in-vivoangiograms. 5. DiscussionandConclusion In thiswork,anewmethodforrecoveringtheLVfromcontrast-enhancedbi-planarcine-angiographic x-ray images has been proposed. The novelty of our approach is that a-priori information about the LV anatomy is learned from high-resolution CT images, modeled as a SSM and utilized for recon- struction. A2-D/3-D registration technique is applied to fit theSSMtoangiographicprojections. When only two (noisy) projections are available, the reconstruction problem usually becomes under- determinedandambiguous. Insuchcases, theincorporationofa-priori informationplaysanimportant role, since this can limit the space of possible solutions and improve the ability to deal with noisy data. In contrast to [7], anatomical a-priori information is derived from data of in-vivo instead of post-mortem subjects; other approaches often do not utilize this kind of information at all. Although onlyonebi-planaracquisition isused for reconstruction,ourapproach isgenerallynot limitedby the number of projections. However, since additional acquisitions increase the amount of radiation and bolus, thisnumber is usuallykept toaminimum. Using a SSM for reconstruction allows to generate statistically plausible and patient specific shapes. Unlike other 3-D LV SSMs often found in literature, anatomical areas like the apex, the atrial con- cavity and the aortic valve region are preserved in our model. This is necessary to generate complete contour and densitometric information; otherwise, additional errors are introduced in the reconstruc- tion process. Further note that these areas typically overlap with the ventricular cavity in projection imagesandare thereforehard to recover withoutprior knowledge. Evaluation with both simulated data and real patient data shows promising results. The LV volume is recovered at high accuracy. This is important for assessing volumetric diagnosis parameters, like EF.Concerningshapeconformity, theoverlapbetweenoriginalandrecoveredvolumeishigh, though there is still place for minor improvements. Future work will focus on improving the model fitting processandon evaluating ourapproachwith more in-vivoangiograms. References [1] S. Benameur, M. Mignotte, S. Parent, H. Labelle, W. Skalli, and J. de Guise. 3D/2D Registra- tion and Segmentation of Scoliotic Vertebrae using Statistical Models. Computerized Medical ImagingandGraphics, 27:321–337,2003. [2] T. F. Cootes, C. J. Taylor, D. H. Cooper, and J. Graham. Active Shape Models - Their Training andApplication. ComputerVision and ImageUnderstanding, 61:38–59,1995. 51