Our medical objective is to match multimodal 3D medical images into a coherent patient model, from which diagnosis can be assessed and therapeutics guided. 3D image segmentation is absolutely necessary to reach this objective. We investigate two complementary approaches for segmenting 3D medical images. We firstly present some definitions, basic properties and recent theoretical results about formal neural networks, and show that these results can be applied to brain tumor segmentation. A variational approach (called the "snake spline" method) is then detailed. We finally show how segmented images can be used for multimodal images matching, and conclude on the industrial challenges to take up.