Automated intracranial aneurysm isolation and quantification Ignacio Larrabide 1,2 , Maria Cruz Villa-Urio 2,1 , Rub´ en Cardenes 2,1 , Jose Maria Pozo 2,1 D. Rod Hose 3 and Alejandro F. Frangi 2,1,4 Abstract— Until today, geometrical descriptors of intracra- nial aneurysms are largely used for diagnosis and treatment selection. Nevertheless, relatively little work has been devoted to automatize these measurements. In this work we propose a methodology for the automated isolation and quantification from vascular segmentation. The proposed methodology is based on skeleton topology analysis, geometrical analysis and deformable models to isolate, automatically, the aneurysm dome as well as its geometrical characteristics. The accuracy of this methodology when compared to manually isolated aneurysms is evaluated in 10 patient-specific vascular geometries. Good correspondence is observed between manual and automated results. I. INTRODUCTION Nowadays, simple geometrical measures of intracranial aneurysms like depth, ostium diameter, volume and aspect ratio (AR, computed as dome maximum depth/ostium diam- eter) are largely used for diagnosis, prognosis and treatment selection in the clinical practice. These measurements are manually obtained by radiologists or clinicians with the subsequent subjectivity and variability upon different ob- servers. Depending on the imaging modality, the detection and quantification of intracranial aneurysms by the “naked eye” can be unprecise and observer dependant [1]. Ford et al. [2] proposed a methodology for aneurysm removal which is based on the reconstruction of the original vessel lumen (i.e., before the aneurysm exists). In the work by Lauric et al. [3], is presented a method for aneurysm isolation based on geometrical characteristics and topology of the vasculature. Still, none of this methodologies has been compared or assessed against a ground truth, or at least obtained manually from aneurysms. In order to isolate and automatically quan- tify intracranial aneurysms, the detection of the aneurysm ostium is mandatory. With this in mind, we propose a methodology that uses deformable models, topology and geometry analysis for automatically obtaining the aneurysm ostium(s), the corresponding centroid(s), aneurysm depth, This work was partially supported within the CENIT-CDTEAM and CENIT-cvREMOD projects funded by the Spanish CDTI and partly within the framework of the @neurIST Project (IST-2005-027703), which is co- financed by the European Commission within the IST Program of the Sixth Framework Programm. I. Larrabide (ignacio.larrabide@upf.edu), M. C. Villa-Uriol, R. Cardenes, J. M. Pozo and A. Frangi are with 1 Networking Research Center on Bioengi- neering, Biomaterials and Nanomedicine (CIBER-BBN), 2 Computational Imaging and Simulation Technologies in Biomedicine (CISTIB), Universitat Pompeu Fabra and 4 Instituci´ o Catalana de Recerca i Estudis Avanc ¸ats (ICREA), Barcelona, Spain D. R. Hose is with 3 Academic Unit of Medical Physics, Faculty of Medicine and Biomedical Sciences, University of Sheffield, Sheffield, UK. ostium(s) maximum and minimum diameters, volume and AR. II. METHODS In the following section we describe a methodology for automated analysis of intracranial aneurysm geometry. This sequence of algorithmic steps starts from the image segmen- tation and finishes in the aneurysm morphology quantifica- tion. The process is based on the vascular geometry on a region of interest (ROI) around the aneurysm. This analysis is performed following four steps: i) image segmentation, ii) skeleton processing, iii) aneurysm identification and iv) quantification, which are described below (see Figure 2(a)). A. Algorithm description In Figure 1 the algorithm description of the proposed method. We define the input medical image as I is presented. The output of the algorithm are the quantitative indicators of aneurysm shape and size, namely: aneurysm depth d , ostium maximum diameter o max , ostium minimum diameter o min , ostium mean diameter o av , aneurysm volume a vol and aspect ratio AR= d o av . The aneurysm volume and surface are also computed. It is assumed that the aneurysm presents only one ostium and that the aneurysm ostium is located at the limit with the parent vessel. B. Segmentation Medical image segmentation is among the most difficult problems of medical image analysis and, in many aspects, it remains an open research topic. In this work, the methods proposed by Hernandez and Frangi [4] and that of Bogunovic et al. [5], are used. This method is based Geodesic Active Regions (GAR) and is a comprehensive approach to the problem of neurovascular segmentation. C. Skeleton processing To properly describe the skeleton topology process, let us first introduce some definitions. We define the surface of the vascular geometry as S and the skeleton topology of such geometry as l s . Then, l s will be recovered from S, providing relevant information of the vascular district regarding its topology and shape. Skeleton computation: For the extraction of l s from S,a flux driven homotopic thinning algorithm is used [6]. From the skeleton computation, a tree like structure is obtained where each branch is a 3D curve l i ∈ l s . For each point x ∈ l s , an estimation of the vessel diameter is computed as the distance to the closest point on S.