Despite improving survival rates, patients diagnosed with aortic coarctation (CoA) suffer from late complications post CoA repair and regular screening via cardiac imaging is warranted [1]. Cardiovascular magnetic resonance (CMR) imaging being a cost-effective and accurate tool to provide detailed 3D anatomical shape information, imaging databases are growing and novel approaches to structure such data are required. We hypothesised that combining statistical shape modelling (SSM) with hierarchical clustering techniques from the field of data mining may aid in discovering previously unknown shape patterns in a population of aortic arches.A retrospective cohort of 45 CoA patients post repair (aged 15-38 years; BSA 1.47–2.2m2, Fig. 1) and 20 healthy Control subjects (aged 10-19 years; BSA 1.3-2.3m²) who underwent routine CMR imaging (1.5T Avanto scanner, Siemens Medical Solutions, Erlangen, Germany; 3D balanced steady-state free precession sequence, SSFP) were included in the study. Aortas were segmented semi-automatically [2] from the CMR data and reconstructed [3] to obtain 3D surface models of the arch. The arches were cut at the level of the sub-annular plane and the level of the diaphragm; head and neck vessels and coronary arteries were cut as close as possible to the arch. The meshed 3D surface models constituted the input for the SSM framework Deformetrica [4], which essentially computes the mean anatomical shape (atlas) and its deformations towards each subject shape within a population of surface shapes without requiring any landmarking. Our approach was two-fold: First we computed the atlas of the Control population as an ideal reference shape (Fig. 2). All CoA arch models were then rigidly registered [5] to the Control atlas and the surface deformations of the atlas towards each CoA subject were calculated in Deformetrica (Fig. 3). The set of all subject-specific deformation vectors constituted the input for a hierarchical clustering algorithm (MATLAB, Natick, MA) that does not require specifying an expected number of clusters prior to calculation. Results were visualised as a dendrogram where objects are grouped together based on their arch shape similarity. Cutting the dendrogram horizontally at a particular height partitions the data into shape sub-groups [6]. The obtained arch clusters were compared with regard to traditional morphometric parameters (surface to volume ratio SVol; centreline tortuosity CLT; minimum to maximum diameter ratio Dmin/Dmax; ascending aortic diameter to proximal descending diameter ratio Dasc/Ddesc) and functional parameters derived from CMR (left ventricular ejection fraction LVEF; indexed end-diastolic volume iLVEDV; left ventricular mass LVmass) using Kruskal-Wallis tests (IBM SPSS Statistics, Chicago, IL).The clustering yielded two main aortic arch shape clusters, with one splitting into two large subgroups when cut at the same height of the dendrogram (Fig. 4). Distributions of the measured morphometric parameters SVol (p < .001), Dmin/Dmax (p = .006) and Dasc/Ddesc (p = .005) were significantly different between the three main shape clusters. Interestingly, distributions of iLVEDV were significantly different (p = .032) between the three groups, with larger values occurring in group three, which clustered subjects with a dilated aortic root together.We present the first application of data mining techniques and 3D statistical shape modelling to a population of aortic arch shapes post CoA repair. Visually, similar arch shapes were clustered together correctly, which was corroborated by traditional shape parameters being different between the groups. Found differences between clusters in iLVEDV suggest that clinical outcomes can be associated with arch shape features detected via the proposed deformation-based shape clustering. Studies in a larger cohort and including more clinically relevant data are necessary to support our findings. Yet, particularly the high degree of automation makes our method an attractive tool for the retrieval of similarly shaped subjects from medical imaging databases for clinical decision support [7].The combination of SSFP CMR data acquisition and advanced 3D shape modelling with data mining techniques provides a sensitive analytical platform, which could improve and facilitate risk assessment, follow-up strategies and treatment planning in complex cardiac lesions.