In this work we analyse the performance of Convolutional Neural Networks (CNN) on medical data by benchmarking the capabilities of different network architectures to solve tasks such as segmentation and anatomy localisation, under clinically realistic constraints. We propose several CNN architectures with varying data abstraction capabilities and complexity, and we train them with different amounts of annotated image patches in order to analyse their performance when training sets of different size are used. Our implementation is capable of performing convolutions on volumetric data using 3D kernels, which enables further performance analysis when 2D, 2.5D and 3D data are employed during training and testing. Our novel segmentation approach does not only make use of voxel-wise classification outcomes produced by the network, but also considers high level features obtained from the second-last fully connected layer of the CNNs in a voting-based strategy. Through Hough voting and patch-wise back-projection from a multi-atlas, anatomy localisation and segmentation are performed and the region of interest is delineated in 3D by a smooth and regular contour. Our method has been tested on volumetric medical images acquired in two modalities for segmentation of deep-brain structures, such as the midbrain in ultrasound and a set of 26 regions in MRI.