In medical imaging, there are clinically relevant segmentation tasks where\nthe output mask is a projection to a subset of input image dimensions. In this\nwork, we propose a novel convolutional neural network architecture that can\neffectively learn to produce a lower-dimensional segmentation mask than the\ninput image. The network restores encoded representation only in a subset of\ninput spatial dimensions and keeps the representation unchanged in the others.\nThe newly proposed projective skip-connections allow linking the encoder and\ndecoder in a UNet-like structure. We evaluated the proposed method on two\nclinically relevant tasks in retinal Optical Coherence Tomography (OCT):\ngeographic atrophy and retinal blood vessel segmentation. The proposed method\noutperformed the current state-of-the-art approaches on all the OCT datasets\nused, consisting of 3D volumes and corresponding 2D en-face masks. The proposed\narchitecture fills the methodological gap between image classification and ND\nimage segmentation.\n