Registration of Functional Data: Aligning Inner Carotid Artery
Transcript
Registration of Functional Data: Aligning Inner Carotid Artery
Registration of Functional Data: Aligning Inner Carotid Artery Centerlines (⋆) Registrazione di dati funzionali: allineamento delle linee assiali di arterie carotidi interne Laura Maria Sangalli, Simone Vantini MOX, Dipartimento di Matematica - Politecnico di Milano e-mail: [email protected] Keywords: functional data analysis, curve registration, amplitude and phase variability 1. Introduction This work is a product of the AneuRisk Project, a scientific program that aims at evaluating the role of vascular geometry and hemodynamics in the pathogenesis of cerebral aneurysms. By means of functional data analysis, we explore the AneuRisk dataset to highlight the relations between the geometric features of the internal carotid artery (ICA), expressed by its radius profile and centerline curvature, and the aneurysm location (Sangalli et al. 2007a, Sangalli et al. 2007b). The dataset is based on 3Dangiographies of 65 patients hospitalized at the Neuroradiology Department of Niguarda Ca’ Granda Hospital, Milano, from September 2002 to October 2005: the i-th sample unit is a regular curve in R3 describing, as function of an abscissa s, the three spatial coordinates (xi (s), yi (s), zi (s)) of the ICA centerline of the i-th patient. 2. Methods and results We focus here on the centerlines registration procedure conceived and performed within the former functional data analysis. This procedure is the key to statistically meaningful comparisons among the geometrical features, radius and curvature, observed in different patients. Indeed, it separates two types of variability affecting functional data: the phase variability, related to the abscissa s and strongly dependent on the dimensions and proportions of patients skulls, and the amplitude variability, the residuals variability related to the geometrical features (Ramsay and Silverman, 2005). In detail, the phase variability is removed by finding 65 warping functions hi of the abscissa s, leading from the original centerlines (xi (s), yi (s), zi (s)) to new registered centerlines −1 −1 (x̃i (s), ỹi (s), z̃i (s)), where x̃i = xi ◦ h−1 i , ỹi = yi ◦ hi , and z̃i = zi ◦ hi . We look for the optimal warping functions in the class W of increasing affine transformations, maximizing the following similarity index between the i-th centerline (x̃i (s), ỹi (s), z̃i (s)) and a reference centerline (x0 (s), y0 (s), z0 (s)): (⋆) This research has been carried out within AneuRisk Project, a joint research program involving MOX Laboratory for Modeling and Scientific Computing (Dip. di Matematica, Politecnico di Milano), Laboratory of Biological Structures (Dip. di Ingegneria Strutturale, Politecnico di Milano), Istituto Mario Negri (Ranica), Ospedale Niguarda Ca’ Granda (Milano), and Ospedale Maggiore Policlinico (Milano). The Project is supported by Fondazione Politecnico di Milano and Siemens-Medical Solutions, Italia. R R ′ ′ ′ ′ ′ ′ ds x̃ x ds ds ỹ y z̃ z 1 Si i 0 Si i 0 Si i 0 qR qR qR qR + qR + qR 3 ′2 ′2 ′2 ′2 ′2 ′2 ds ds ds ds ds ds x̃ x ỹ y z̃ z Si i Si 0 Si i Si 0 Si i Si 0 R (1) where Si is the support of the i-th centerline. The index (1) is a suitable 3D-generalization in L2 (R; R3 ) of the cosine of the angle between the first derivatives of two functions in L2 (R; R). The class W and the similarity index (1), jointly provide useful properties in terms of formal consistency (e.g., the similarity index between two curves does not change if the two curves are warped along the same function of class W ), pertinence to the real problem and algorithm implementation. A Procrustes fitting criterion is used to jointly estimate both the 65 warping functions and the reference centerline. Alternate expectation and maximization steps are performed: at each expectation step, the reference curve is estimated using all the curves obtained at the previous iteration; at each maximization step, all curves are warped in order to maximize theirs similarity with the estimated reference curve. The optimal warping functions hi are then obtained by composition of the maximizing warping functions found at each iteration step. The algorithm converges in few steps. 1.0 x~i ′(s ) 0.0 0.5 −1.0 −1.0 x i ′(s ) 0.0 0.5 1.0 Figure 1: First derivatives of the 65 centerlines before (left) and after (right) registration, together with first derivatives of the reference centerline (solid black line) −100 −80 −60 −40 −20 0 −100 −80 −60 −40 −20 0 −40 −20 0 −40 −20 0 0.5 −1.0 −1.0 y~i ′(s ) 0.0 y i ′(s ) 0.0 0.5 1.0 s 1.0 s −100 −80 −60 −40 −20 0 −100 −80 −60 z~i ′(s ) 0.0 0.5 −1.0 −1.0 z i ′(s ) 0.0 0.5 1.0 s 1.0 s −100 −80 −60 −40 s −20 0 −100 −80 −60 s References Ramsay J.O., Silverman B.W. (2005) Functional Data Analysis, II ed., Springer, New York. Sangalli L.M., Secchi P., Vantini S., Veneziani A. (2007a) Functional data analysis. A case study: geometrical features of the internal carotid artery, Tech. Rep. 24/2007, MOX, Dipartimento di Matematica, Politecnico di Milano, available at http://mox.polimi.it/it/progetti/pubblicazioni. Sangalli L.M., Secchi P., Vantini S., Veneziani A. (2007b) Smoothing and dimension reduction of 3D centerlines of inner carotid arteries by free-knots regression splines, Tech. Rep. 23/2007, MOX, Dipartimento di Matematica, Politecnico di Milano, available at http://mox.polimi.it/it/progetti/pubblicazioni.