The radiotherapic treatment of cerebral arterious malformations (AVM) requires an accurate estimation of the AVM shape. This estimation is classically obtained from the delineation of the AVM in several 2D angiographic views. In this paper, a clinical study of the inter-observer variability in the AVM detection is first performed. It proves that the estimated volume varies a lot between observers. For these reasons, we propose a framework for AVM delineation which makes use of 2D and 3D angiographic images: the initial estimate obtained with 2D angiographic images is then refined within the 3D volume using deformable models. Results are presented demonstrating shape delineation on various AVMs.