The document analysis of music score images is a key step in the development of successful Optical Music Recognition systems. The current state of the art considers the use of deep neural networks trained to classify every pixel of the image according to the image layer it belongs to. This process, however, involves a high computational cost that prevents its use in interactive machine learning scenarios. In this paper, we propose the use of a set of deep selectional auto-encoders, implemented as fully-convolutional networks, to perform image-to-image categorizations. This strategy retains the advantages of using deep neural networks, which have demonstrated their ability to perform this task, while dramatically increasing the efficiency by processing a large number of pixels in a single step. The results of an experiment performed with a set of high-resolution images taken from Medieval manuscripts successfully validate this approach, with a similar accuracy to that of the state of the art but with a computational time orders of magnitude smaller, making this approach appropriate for being used in interactive applications.