Novelty detection is a process of identifying if new data does not belong to the previously observed patterns; this is important in several applications such as medical diagnosis and autonomous driving, where wrongly identifying a novel pattern could be catastrophic. We focus on the problem of novelty detection in images based on deep learning (DL) techniques. DL based models are trained through an optimization process in which a loss function is minimized. For visual novelty detection, this loss function must be able to capture in a scalar value the visual properties of the normal samples, usually the mean square error (MSE). While satisfactory performance has been obtained with this formulation, it is often complicated for a single loss function to capture all of the relevant information to discriminate novelty from normality. In this paper we propose novel reconstruction loss that combines pixel-wise and perceptual information. Its performance is experimentally evaluated using generative adversarial networks (GANs), and compared against autoencoders based on MSE, showing superior performance. Additionally, we present an application to novelty detection in X-Ray images, which is relevant for current work con COVID diagnosis.