Every segmentation task is uncertain due to image resolution, artefacts, annotation protocol etc. Propagating those uncertainties in a segmentation pipeline can improve the segmentation. This article aims to assess if segmentation can benefit from uncertainty of an auxiliary unsupervised task - the reconstruction of the input image. This auxillary task could help the network focus on rare examples that are otherwise poorly segmented. The method was applied to segmentation of myocardial infarction areas on cardiac magnetic resonance images.